model the migration of Grevy`s zebras - endeleo

Transcription

Faculteit Bio-ingenieurswetenschappen
Academiejaar 2008-2009
MODELLING THE MIGRATION OF GREVY’S
ZEBRA IN FUNCTION OF HABITAT TYPE USING
REMOTE SENSING
Eline HOSTENS
Promotor: Prof. Dr. ir. Robert R. D E W ULF
Masterproef voorgedragen tot het behalen van de graad van
B IO - INGENIEUR IN HET BOS - EN NATUURBEHEER
Faculteit Bio-ingenieurswetenschappen
Academiejaar 2008-2009
MODELLING THE MIGRATION OF GREVY’S
ZEBRA IN FUNCTION OF HABITAT TYPE USING
REMOTE SENSING
Eline HOSTENS
Promotor: Prof. Dr. ir. Robert R. D E W ULF
Masterproef voorgedragen tot het behalen van de graad van
B IO - INGENIEUR IN HET BOS - EN NATUURBEHEER
De auteur en de promotor geven de toelating deze masterproef voor consultatie beschikbaar te stellen
en delen ervan te kopiëren voor persoonlijk gebruik. Elk ander gebruik valt onder de beperkingen van
het auteursrecht, in het bijzonder met betrekking tot de verplichting uitdrukkelijk de bron te vermelden
bij het aanhalen van resultaten uit deze scriptie.
The author and promotor give the permission to use this thesis for consultation and to copy parts of it
for personal use. Every other use is subjected to the copyright laws, more specifically the source must
be exensively specified when using results from this thesis.
The promotor:
Prof. dr. ir. R. De Wulf
The author:
Eline Hostens
Foreword
The making of a thesis is quite a challenge. I would never have been able to do this without the help
of a lot of people. Here I would like to take the opportunity to thank all the people who contributed to
the success of this work.
First let me express my sincere thanks to my supervisor prof. dr. ir. Robert R. de Wulf who gave me
the opportunity to make this thesis about a passion of mine, i.e. animals. I would also like to thank
Toon Westra for the support during the year. I could always go to him for advice about practical work
or for any other questions.
I am grateful to Northern Rangelands Trust for the collection of ground truth data and the delivery
of GPS tracking data and especially to Juliet King fot the coordination. I’d also like to thank Else
Swinnen of VITO for preparing the SPOT-Vegetation ten-day composites.
I would like to show my appreciation to Kenny Devos and Els Verdonck who read and improved my
thesis. I would like to thank my father Ivan Hostens, who has always helped me where possible during
my studies, for reading this work and for a lot of other problems and jobs he has taken for his account.
A word of gratitude goes to all the people who made my student days one of the best periods of my
life so far. All the new friends I made in Gent, all the people of my year and especially my collegueroomers to whom I could always go to have a good chat and for support. I am really going to miss
them.
My parents as well deserve appreciation as they made a great effort to give me the opportunity to
study and explore my possibilities. Therefore I will always be grateful to them. I would also like to
thank them for supporting me during the more difficult times and for their trust in me.
Handzame, mei 2009
Eline Hostens
List of Abbreviations
2-D
:
3-D
:
AVHRR :
CITES :
EVI
:
GPS
:
LAI
:
LCCS :
LiDAR :
MCP
:
MIR
:
NASA :
NDVI :
NIR
:
NN
:
NOAA :
NRT
:
PAs
:
PC
:
PCA
:
PDOP :
PTT
:
SA
:
SAR
:
SSC
:
TLU
:
UNEP :
VHF
:
two-dimensional
three-dimensional
Advanced Very High Resolution Radiometer
Convention on International Trade in Endangered Species
Enhanced Vegetation Index
Global Positioning System
Leaf Area Index
Land Cover Classification System
airborne lasers
Minimum Convex Polygon
Mid Infra Red
National Aeronautics and Space Administration
Normalized Difference Vegetation Index
Near Infra Red
Artificial Neural Networks
National Oceanic and Atmospheric Administration
Northern Rangelands Trust
Protected Areas
Principal Component
Principal Component Analysis
Positional Dilution Of Precision
Platform Transmitter Terminals
Selective Availability
Synthetic Aperture Radar
Species Survival Commision
Tropical Livestock Unit
United Nations Environment Program
Very High Frequency
Contents
1
Introduction
1
2
Grevy’s Zebra (Equus grevyi)
3
1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3
2
Social structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4
3
Habitat and diet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5
4
Breeding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7
5
Predators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8
6
Threats and conservation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8
3
Study area
11
1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11
2
Climate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
13
3
Livestock . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
13
4
Vegetation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
14
5
Conservancies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
16
5.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
16
5.2
Conservation of Grevy’s zebras . . . . . . . . . . . . . . . . . . . . . . . .
17
II
Contents
4
5
6
Wildlife telemetry
18
1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
18
2
Very-High-Frequency (VHF) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
19
3
Satellite tracking: Argos system . . . . . . . . . . . . . . . . . . . . . . . . . . . .
20
4
Global Positioning System (GPS) tracking . . . . . . . . . . . . . . . . . . . . . . .
22
4.1
Operation of the system . . . . . . . . . . . . . . . . . . . . . . . . . . . .
22
4.2
Accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
23
4.3
Obstructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
24
4.4
Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
27
4.5
Examples of studies using GPS telemetry . . . . . . . . . . . . . . . . . . .
28
Wildlife tracking and remote sensing
29
1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
29
2
Habitat maps and habitat suitability mapping . . . . . . . . . . . . . . . . . . . . .
29
2.1
Habitat maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
29
2.2
Habitat suitability maps . . . . . . . . . . . . . . . . . . . . . . . . . . . .
30
3
Spatial heterogeneity assessment based on primary productivity . . . . . . . . . . .
31
4
Temporal heterogeneity assessment . . . . . . . . . . . . . . . . . . . . . . . . . .
32
5
Heterogeneity assessment based on landscape structural properties . . . . . . . . . .
33
6
Heterogeneity assessment based on plant chemical constituents . . . . . . . . . . . .
34
Data and methods
35
1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
35
2
Satellite images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
35
2.1
35
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
III
Contents
7
2.2
Landsat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
36
2.3
MODIS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
38
2.4
SPOT-Vegetation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
39
3
Tracking data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
40
4
Vector data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
42
5
Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
43
5.1
Ground truth data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
43
5.2
Artificial Neural Networks (NN) . . . . . . . . . . . . . . . . . . . . . . . .
44
5.3
Classification methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
46
5.4
Accuracy assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
47
6
Analysis of Grevy’s zebra tracking data . . . . . . . . . . . . . . . . . . . . . . . .
48
7
Analysis of Grevy’s zebras’ migration . . . . . . . . . . . . . . . . . . . . . . . . .
48
7.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
48
7.2
Correlation of the zebras’ migration with biomass . . . . . . . . . . . . . . .
49
7.2.1
Linking NDVI and tracking datasets . . . . . . . . . . . . . . . .
49
7.2.2
Statistical analysis . . . . . . . . . . . . . . . . . . . . . . . . . .
50
7.3
Correlation between zebra presence and water . . . . . . . . . . . . . . . . .
50
7.4
Correlation between zebra presence and livestock . . . . . . . . . . . . . . .
51
7.5
Correlation between zebra presence and towns . . . . . . . . . . . . . . . .
51
7.6
Habitat preference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
51
7.7
Integration of all factors influencing the migration . . . . . . . . . . . . . .
53
Results and discussion
54
1
Habitat classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
54
1.1
54
Landsat-based habitat classification . . . . . . . . . . . . . . . . . . . . . .
IV
Contents
1.2
MODIS-based habitat classification . . . . . . . . . . . . . . . . . . . . . .
56
1.3
Analysis of the result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
59
1.4
Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
63
2
Analysis of tracking data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
64
3
Correlation between tracking data and biomass . . . . . . . . . . . . . . . . . . . .
69
4
Correlation between tracking data and water . . . . . . . . . . . . . . . . . . . . . .
74
5
Correlation between tracking data and livestock . . . . . . . . . . . . . . . . . . . .
75
6
Correlation between tracking data and towns . . . . . . . . . . . . . . . . . . . . . .
77
7
Habitat preference
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
79
7.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
79
7.2
Habitat preference tested on the MODIS classification . . . . . . . . . . . .
80
7.2.1
First level comparison: testing for non-random use . . . . . . . . .
80
7.2.2
First level comparison: ranking of the habitat types in order of preference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
81
Second level comparison: testing for non-random use . . . . . . .
83
Habitat preference tested on Africover . . . . . . . . . . . . . . . . . . . . .
84
7.3.1
First level comparison: testing for non-random use . . . . . . . . .
85
7.3.2
First level comparison: ranking of the habitat types in order of preference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
85
7.3.3
First level comparison: integration over all sixteen zebras . . . . .
88
7.3.4
Second level comparison: testing for non-random use . . . . . . .
88
7.3.5
Second level comparison: integration over all sixteen zebras . . . .
88
Integration of all factors influencing the occurrence . . . . . . . . . . . . . . . . . .
89
7.2.3
7.3
8
8
Conclusion
94
V
Contents
9
Nederlandse samenvatting
97
1
Inleiding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
97
2
Literatuurstudie . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
97
2.1
Grevy’s zebra (Equus grevyi) . . . . . . . . . . . . . . . . . . . . . . . . . .
97
2.2
Studiegebied . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
98
2.3
Wildlife telemetrie . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
99
2.4
Tracking van wild en teledetectie . . . . . . . . . . . . . . . . . . . . . . . 101
3
4
Data en methoden . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
3.1
Satellietbeelden . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
3.2
Tracking data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
3.3
Classificatie . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
3.4
Analyse van de Grevy’s zebra’s tracking data en migratie . . . . . . . . . . . 103
Resultaten en discussie . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.1
Classificatie . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.2
Analyse van de Grevy’s zebras tracking data en migratie . . . . . . . . . . . 105
4.2.1
Correlatie tussen tracking data en biomassa . . . . . . . . . . . . . 105
4.2.2
Correlatie tussen tracking data en aanwezigheid van water, vee en
dorpen . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
4.2.3
Habitatpreferentie . . . . . . . . . . . . . . . . . . . . . . . . . . 106
4.2.4
Integratie van alle factoren . . . . . . . . . . . . . . . . . . . . . 107
Reference
108
A Ground truth collection form
115
B Classes of the Africover classification of the study area
117
VI
Contents
C Boxplots for the different seasons
119
D Habitatpreference based on made classification
121
E Habitatpreference based on the Africover reclass classification
123
F Histograms for the different seasons
125
VII
Chapter 1
Introduction
The Grevy’s zebra (Equus grevyi) is listed on the IUCN red list as an endangered species that can only
be found in the eastern part of Ethiopia and the northern part of Kenya. There has been a fast decline of
the remaining population in the past decades. The major threat for this species is introduced livestock
that compete for grazing. As cattle is mostly kept nearby water, Grevy’s zebras are sometimes forced
to drink at night, when they are more vulnerable to predation. As the zebras can travel large distances,
much of their home range is located outside protected areas. They are mostly found in the arid and
semi-arid rangelands.
In this thesis migration of Grevy’s zebras is modelled in function of habitat type and plant biomass
using remote sensing. As the Grevy’s zebra is a threatened species, it is very important to monitor
their movement and to increase the knowledge about their behaviour. The more is known about the
use of resources and migration, the more efforts can be made to preserve them. There are two major
objectives in this thesis. The first is to perform a habitat classification of the study area with the aim
of determining the habitat use of the Grevy’s zebras. The second objective is the modelling of the
migration of the Grevy’s zebras.
The habitat classification of the study area will be based on Landsat and MODIS satellite images.
There will be searched for the best method of classifying the study area. Habitat classes will be
derived from ground truth data and the classification will be conducted with the Maximum Likelihood
classifier and with Neural Networks. An Africover map, a rough habitat map of Africa is already
available for the study area, but there will be tried to make a more detailed map. Additionally, attempts
will be made to make a ranking of the habitat preference of Grevy’s zebras based on the habitat
classification and the Africover classification.
The objective of the modelling of the Grevy’s zebras’ migration will be divided into some subobjectives: the correlation of the tracking data with biomass, with available water, with livestock
presence and with the presence of towns. Data about the migration and location of the Grevy’s zebras
1
CHAPTER 1. Introduction
is obtained by applying GPS-collars to sixteen zebras. There are several factors influencing the animals’ movement and these will be investigated separately. The most important influence is probably
the availability of food sources. As Grevy’s zebras are herbivores, biomass can be used as an indicator
for the available amount of food. This will be modelled using the Normalised Difference Vegetation
Index (NDVI) as a proxy. SPOT-Vegetation NDVI images will be applied to derive time series of
NDVI for the study area. To model the influence of water availability, a map of the distance to the
nearest water source is being used. The impact of livestock will also be examined. The influence of
towns will also be examined by calculating the distance to the nearest town and finding the relationship between this distance and zebra occurence. Finally all these factors will be merged together to
make a prediction of the areas within the study area that are best suitable for the Grevy’s zebras.
First, there is a brief overview of the literature. The species Grevy’s zebras will be discussed as well as
the study area. To understand the tracking technique, GPS tracking is handled. To make a comparison
with the other possibilities, VHF and satellite tracking will also be discussed. Last, a link will be
made between animal movement and remote sensing.
2
Chapter 2
Grevy’s Zebra (Equus grevyi)
1
Introduction
Zebras are still numerous and widespread in Africa. There are three species: The plains zebra (Equus
burchelli Gray), the Grevy’s zebra (Equus grevyi Oustalet) and the mountain zebra (Equus zebra L.).
The Grevy’s zebra is listed on the IUCN red list as an endangered species that can only be found in the
eastern part of Ethiopia and the northern part of Kenya. The Grevy’s zebra is the biggest species of the
wild equids. It can easily be distinguished from other zebra species by its larger size, big rounded ears,
narrow, evenly divided stripes, a white unmarked belly, and a brown spot on the nose (Rubenstein,
2004). They are about 250–275 cm long and have a shoulder height of about 140–160 cm. Females
weigh about 350–400 kg, males 380–450 kg (ARKive, Images of life on Earth, read 07/2008).
Figure 2.1: A Grevy’s zebra (Gardner, read 08/2008)
3
CHAPTER 2. Grevy’s Zebra (Equus grevyi)
Table 2.1: Taxonomic classification of the Grevy’s zebra.
Kingdom
Phylum
Class
Order
Family
Genus
Species
2
Animalia
Chordata
Mammalia
Perissodactyla
Equidae
Equus
Equus grevyi
Social structure
Social structures of all species, like group size, spatial dispersion, and mating systems, are shaped
by the environment. The major force leading to sociality for zebras was probably the need to protect
against predation. Of all predatory attacks on zebras by lions, 35% are successful when zebras are
solitary, whereas only 22% when zebras live in moderate-sized groups. Social relationships may also
be influenced by the fulfilling of other needs, such as acquiring food, water, and mates (Rubenstein,
1986).
The social structure of the Grevy’s zebra is different from the other zebra species as it is a much opener
society (ARKive, Images of life on Earth, read 07/2008). They are loosely social animals, which can
be found in the most distinct groupings. There are groups of mostly young stallions without territory,
who live in bachelor groups; there are groups of mares with or without foals; and there are also mixed
groups of stallions and mares. The herd composition varies constantly as the members do not have an
individual connection with each other. The formation of mixed big flocks is connected to the seasonal
migration. The most stable relationships are those of a stallion to his territory and of a mare to her
foal (Grzimek, 1972).
The female’s movements and association choices are primarily thought to be dependent on water and
forage distribution. There’s a difference in dietary needs, both quantitative and qualitative, and susceptibility to predation between lactating and non-lactating females (Sundaresan et al., 2007b). Having
a good condition is important for survival, embryo development, and the raising of young to independence. Non-lactating females put efforts in acquiring large quantities of high-quality vegetation,
while lactating females both want to acquire food and access to predator-free sources of water. If high
quality food and safe water coincide, then the different reproductive classes can be found together.
Otherwise, they are distributed in different areas (Rubenstein, 1986). Grevy’s zebras live in arid areas
with scarce water. Only lactating females need to drink every day. When a foal is killed, the mothers
go to sites more distant from water and with more plentiful vegetation (Rubenstein, 2004). These
4
different needs prevent them to form stable bonds among each other. Grevy’s zebra’s females range
between 10 to 15 kilometres per day (Rubenstein, 1986). Competition for vegetation is rare among
females. Regardless of abundance, they avoid interfering with each other as they try to consume as
much food as quickly as possible by adjusting their spacing (Rubenstein, 2004).
About 10% of a population’s mature stallions (ARKive, Images of life on Earth, read 07/2008) have a
territory of 2.5–10.5 km2 , which is huge in comparison to other herbivores’ territories. The strongest
males claim prime watering and grazing areas. These factors attract other zebras to the territory. The
territory stallion tolerates other stallions if they don’t approach receptive mares; the resident male has
the exclusivity to mate in his territory. If they do approach the mares, they will be attacked and chased
away from the mare about 30–100m. They are rarely chased away from the territory. The attacked
stallion admits to the dominance of the territory owner and will not defend himself. Stallions without a
territory will fight each other to mate with a mare. The territories are located along recognition points
in the landscape. The main marking of the territory is by the presence of the owner. The sound and
smell signals, which indicate the borders, are presumably of subordinate role. These piles of manure
seemingly help the animal orientate in its terrain. The piles are several square meters in size and about
40cm high (Grzimek, 1972).
3
Habitat and diet
Today, Grevy’s zebra can only be found in the northern parts of Kenya and in the south of Ethiopia.
This is due to a rapid decline in their population. They used to roam in semi arid shrublands and plains
of Somalia, Ethiopia, Eritrea, Djibouti, and Kenya (African Wildlife Foundation, read 07/2008). They
are presumed extinct in Somalia since its last sighting in 1973 (figure2.2) (ARKive, Images of life on
Earth, read 07/2008).
Grevy’s zebras live in a more arid semi-desert habitat compared to the other zebra species (Youth,
Howard, 2004). These habitats include arid grasslands and dusty acacia savannas. The bushed grassland habitats have woody vegetation that is dominated by Acacia species. The grasses are primarily
of the genera Themeda, Cynodon and Pennisetum (Sundaresan et al., 2007a).
5
Figure 2.2: Historical and current range of Grevy’s zebras (Grevy’s zebra Trust, read 08/2008)
Grevy’s zebras nearly always coexist with people. Therefore they have a trade-off between locations with good quality vegetation and proximity to human activities. According to Sundaresan et al.
(2007a) forage quantity and quality, and habitat openness are vegetation features important to zebras.
The ability to detect predators is affected by the visibility, which in turn is affected by the bush density. Grevy’s zebras may avoid areas close to humans and their livestock, due to direct disturbance or
because of indirect competition with domestic ungulates for forage (Sundaresan et al., 2007a).
Eating is a major occupancy for the Grevy’s zebras. They spend nearly two-thirds of their day on it
(Saint Louis Zoo, read 07/2008). They are predominately grazers. Forbs, shrubs, and trees are alternative foods if grasses are scarce. Leaves can constitute up to 30% of their diet (Smithsonian National
Zoological Park, read 07/2008). They can digest many types and parts of plants that cattle cannot
(African Wildlife Foundation, read 07/2008). They are also beneficial to other wild grazers because
they clear off the tops of coarse grasses that are difficult for other herbivores to digest. Zebras also eat
coarse grasses that grow on marginal lands where cattle do not dwell (Seaworld Adventure Parks, read
07/2008). Zebras ferment vegetation after digestion in the stomach. Therefore food processing is not
slowed down as in ruminant grazers (Rubenstein, 2004). The contact with the absorptive surfaces of
the intestine is limited. To survive they must therefore consume large quantities of vegetation, which
can be of low quality. Zebra foraging is consequently only limited by the time they can devote to
feeding (Rubenstein, 1994).
6
Grevy’s zebras are mostly observed in areas of short, green grass. It can be expected that they seek
out areas with high-quality forage. However, lactating females and bachelors use areas with greener
but shorter grass, seeking higher-quality forage at the cost of reduced quantity. The specific nutrient
demands of lactation may drive the choice for the females with foals. For the bachelors there can be
several possible explanations. The presence of lactating females may attract them, as these females
come into predictable oestrus. They may require particular micronutrients, more abundant in growing
grass, because many are still growing. Or bachelors may be avoiding territorial males who can harass
them. Lactating females are also more often seen in dense woody vegetation, which is strange as these
areas are thought to be unsafe as they provide cover for lions and given the fact that foals are very
vulnerable to predation. The use of denser bush by lactating females suggests a trade-off between the
risk of predation and other benefits of these areas, such as proximity to water or high-quality forage.
The bachelors’ greater use of medium bush area can be due to their avoidance of territorial males.
Non-lactating females and territorial males may pursuit a strategy of gaining weight by using areas
with lower-quality, higher bulk forage (Sundaresan et al., 2007a).
Water is also indispensable and a key to Grevy’s zebras’ survival and reproductive success. The
animals must always be within fairly easy reach of water holes or rivers. If water is available they
will drink daily, but as an adaptation to living in semi-desert, they can go without water for 2–5 days.
The rain is the primary source of these water sources, and it also transforms the land around them.
After the rain, the dusty plains are transformed into fertile pastures, peaking the zebras access to water
and their breeding. As lactating females are forced to drink every day, they stay close to permanent
water sources and the groups of mothers and foals often travel together. Zebras prefer drinking during
the day, when they can easier see danger coming. During the daytime, some of the water sources are
shielded off because cattle is grazing. Then, the zebras are forced to drink at night, after herders and
their livestock left. This implies a greater risk of being caught by predators (Youth, Howard, 2004).
4
Breeding
The Grevy’s zebra females wander through the territories of up to four males in one day. They mate
with several males with which they associate; they are polyandrous. The females with newborn foals,
remain near permanent sources of water in one male’s territory and mate exclusively with this male;
they are monandrous. The males copulate twice as frequently with polyandrous females then when
consorting with relatively sedentary monandrous females (Ginsberg & Rubenstein, 1990).
Grevy’s zebras mate year round, with a gestation period of 13 months. A mare gives birth to only one
foal. The peak birth and mating periods are from July through August and October through November.
The breeding starts at an age of three for the females and six for the males and usually follows a two
year interval (African Wildlife Foundation, read 07/2008). When there’s a shortage of food or water,
the interval can become once every three years. In longer dry periods, the breeding ceases because
7
the females go out of oestrus as their bodies adjust more to a state of survival than one of readiness to
mate (Youth, Howard, 2004).
The newborn foals have a long hair crest down the back and belly, and their stripes are more brownish
(African Wildlife Foundation, read 07/2008). They are able to stand a mere six minutes after birth, and
run after 45 minutes, (ARKive, Images of life on Earth, read 07/2008) which is necessary because the
foals are specifically vulnerable to predators. They start with the learning of the mother’s individual
stripe pattern and smell before the mother lets any other zebra get close. The foal follows the mother
every step and they spend time together playing, nuzzling, and nursing (African Wildlife Foundation,
read 07/2008). The foals remain dependent on their mother’s milk until six to eight months of age
and the young zebra stays about 2–3 years with its mother (ARKive, Images of life on Earth, read
07/2008).
5
Predators
The main predator of all zebra species is the lion (Panthera leo L.). The zebras are most attacked
during the night at waterholes (Grzimek, 1972). Lion activity peaks at night. The darkness provides
them adequate concealment to hunt in open field, thereby shifting their habitat use from woodland
to grassland. It is therefore more dangerous to be in open areas for zebras at night time because
then lions are more likely to be present. Zebras can minimize the risk of an attack by reducing the
number of encounters with lions, for instance by looking up more bushy habitats. However, their
digestion system of hindgut fermentation forces a zebra to graze frequently throughout the day and
night. Grazing occupies about 60% of their time. They prefer grassland, so moving to a safe place
and waiting out the lions is not always an option (Fischhoff et al., 2007).
By associating with other ungulates, the Grevy’s zebras obtain an advantage to predators. Wildebeest
(Connochaetes taurinus Burchell), beisa oryx (Oryx gazelle beisa L.), eland (Taurotragus oryx Pallas),
and plains zebras are some of the species with which they sometimes graze and travel (Youth, Howard,
2004).
6
Threats and conservation
The Grevy’s zebra is classified as an endangered species on the IUCN Red List 2008 (IUCN/SSC, read
07/2008). They are also listed on Appendix I of the Convention on International Trade in Endangered
Species (CITES), which effectively bans international trade in the species (ARKive, Images of life
on Earth, read 07/2008). In the late 1970s, the remaining population was estimated at about 15000.
Recent estimates are 2000 remaining wild individuals in Kenya and about 120–250 in three isolated
8
Ethiopian populations. The species is considered extinct in Somalia (Saint Louis Zoo, read 07/2008).
The species occurs in several protected reserves throughout much of its current range (ARKive, Images of life on Earth, read 07/2008).
The first major threat to the Grevy’s zebra is introduced livestock that compete for grazing. Cows
chew the plants to the ground which results in a considerable environmental degradation due to the
highly erosive soil and fragile vegetation (Youth, Howard, 2004). The large cattle are mostly kept in
grasslands nearby water, thereby making the water unreachable for the zebras during daytime, and
forcing them to drink during the night, when they are more vulnerable to predation. This is one of the
reasons why the population in Kenya declined with about 70% between 1977 and 1988 (IUCN/SSC,
read 07/2008). The traditional water sources are sometimes dried up due to the intensive irrigation in
farm areas upstream. Some herders block the zebra’s access to water by fencing it with thorn covered
acacia limbs. These are all reasons why there is a constant decline in the water reserves for the Grevy’s
zebras (Saint Louis Zoo, read 07/2008). Non-lactating females avoid livestock more than any other
age or sex class. As livestock exploit the best grazing sites and females, in need of replenishing their
body reserves after giving birth, try to avoid these areas, this could lengthen the inter-birth interval,
and slow down the population growth (Rubenstein, 2004).
Another threat is the hunting by poachers for zebra skins. High prices were paid for the zebra fur. The
hunt is the reason why the species is threatened in Ethiopia. The species is extinct in Somalia because
of hunting and wars. Thanks to CITES, this trade is now banned (African Wildlife Foundation, read
07/2008). The species was declared protected by the Kenyan and Ethiopian governments about 20
years ago. Despite the laws, the animals are still hunted for food and are used in traditional medicine
(Youth, Howard, 2004).
In some Kenyan reserves, Grevy’s zebras can drink and feed in cattle and gun free refuges. Unfortunately, these protected areas can only meet their needs for part of the year. The zebras keep spending
much of their time on unprotected lands. Less than 0.5% of Grevy’s zebras’ range falls within protected areas, according to the IUCN Species Survival Commision’s (SSC) action plan. Even in these
areas the animals encounter stress, caused by tour vehicles ignoring the rules and driving off-road,
disturbing the zebras and other wildlife, causing erosion, and destroying fragile vegetation. Zebras
sometimes stay away from waterholes during the day because tourists use them for swimming pools
(Youth, Howard, 2004).
Another serious problem can be due to the plains zebras. Whenever they outnumber the Grevy’s zebras, they significantly lower the feeding rate of the Grevy’s zebras. The replenishment of previously
poached wildlife populations within National Parks is one of the goals of the Kenyan government.
This can be achieved by translocations from densely to sparsely populated areas. The removal of
plains zebras from areas where Grevy’s zebras are abundant, but where their numbers are not increasing, may help reduce competition and increase Grevy’s zebra birth rates (Rubenstein, 2004).
9
The fact of habitat loss is the most serious threat today in the already restricted area where the Grevy’s
zebra lives. Grasslands are still cleared to make way for agriculture (African Wildlife Foundation,
read 07/2008).
However, there are also positive initiatives. The Grevy’s zebra is kept in zoos and breeding programmes have been started to preserve the species. Scientist and local communities in Africa are also
working together to stop the decline and try to multiply the number of Grevy’s left (Saint Louis Zoo,
read 07/2008). Zoos play vital roles as they educate people about the Grevy’s zebra’s situation and
provide opportunities to observe the animals (Youth, Howard, 2004).
10
Chapter 3
Study area
1
Introduction
The Republic of Kenya is situated on the east coast of Africa, on the equator. Kenya has several
physical features, from low lying arid and semi-arid lands to a coastal belt, plateaus, highlands and
the lake basin around Lake Victoria. The Great Rift Valley, which extends for 8700km from the Dead
Sea in Jordan to Beira in Mozambique, bisects the country. Mount Kenya, rising to a height of 5199m,
is the second highest snow capped mountain in Africa after Mount Kilimanjaro.
Kenya has a diverse population of an estimated 34 million people of 42 ethnic groups. The capital city
is Nairobi (Government of Kenya, read 11/2008). Kenya contains 8 provinces (figure 3.1(a)), namely
Central, Coast, Eastern, Nairobi, North Eastern, Nyanza, Rift Valley and Western (Kenya-space, read
11/2008). The study area is the area where all tracking data of the zebras was collected. It is located in
the centre of Kenya (figure 3.1), between latitudes 0.3◦ and 2◦ North and longitudes 36.99◦ and 38.1◦
East. It is located in parts of 6 districts: Laikipia District, Isiolo District, Samburu District, Marsabit
District, Meru District and Nyambene District.
Kenya mainly consists of savanna and grassland ecosystems (39%), as well as bushland and woodland
ecosystems (36%). Agroecosystems cover 19% of the country, forests make up 1.7% and urban land
takes only 0.2%. There is a small percentage of areas that are naturally devoid of vegetation, bare
grounds (World Resources Institute et al., 2007).
11
CHAPTER 3. Study area
(a) Provinces of Kenya
(b) Districts within the study area
Figure 3.1: Location of the study area (International Livestock Research Instistute, read 2009)
Kenya has a very rich biodiversity. The country is home to 6500 plant species, more than 260 of
which are found nowhere else in the world. Kenya has second place among African countries in
species richness for birds (1000 species) and mammals (350 species). As Kenya is on the boundary
between Africa’s northern and southern savannas, more species of large mammals are concentrated
in its rangelands than in almost any other African country (World Resources Institute et al., 2007).
Rangelands are all the habitats suitable for grazing livestock or wildlife (Dictionary, read 04/2009).
The rangelands sustain livestock and wildlife. The wildlife species are an important income for the
country through tourism. The wildlife and livestock census of 1994-96 showed that rangelands were
dominated by livestock, with about 84% of all grazing animals in that area consisting of cattle, sheep
or goat. There was a decline of 61% of all large grazing wildlife species in the rangelands between
1977-78 and 1994-96. The main reasons for this decline were the competition for land and water with
humans and their livestock, as well as illegal hunting. It has been shown that livestock near water
points push wildlife away from water. In almost all the rangeland districts, water demand is greater by
livestock than by wildlife. Only in a few areas near or within protected areas, the water consumption
by wildlife is larger than the water consumption by livestock. It has been shown that the density of
human settlements has an impact on wildlife densities as well. The lower the human densities are, the
higher the wildlife densities (World Resources Institute et al., 2007).
12
2
Climate
In Kenya there is a tropical climate with moderate temperatures averaging about 22°C throughout the
year. The coast is hot and humid, the inland is temperate and the north and northeast parts of the
country are dry (Government of Kenya, read 11/2008). The mean annual temperatures in Laikipia
District range between 16°C and 26°C. The mean temperature in Samburu District is 29°C, the one in
Isiolo District is about 27°C. So the study area is on average in the hotter parts of the country (Ministry
of state for the Development of Northern Kenya and other arid lands, read 11/2008).
For a country on the equator, the annual rainfall in Kenya is low with an annual average of about
630mm per year. This is very unevenly distributed over the land and varies greatly between the years.
In each decade over the past 30 years, there have been regularly major droughts and floods. Distinct
seasonal patterns can also be discriminated. There are two rainy seasons: the short rains are from
October to December and the long rains from March to June, with the hottest period being January
to March. This high variability of rainfall throughout the seasons, between years, and across space
has influenced the distribution of plants, animals and humans within the country (World Resources
Institute et al., 2007).
The northern and eastern parts of the country get about 200–400mm, while the western part bordering
lake Victoria, and central Kenya close to the high mountain ranges receive more than 1600mm. In
Laikipia District, there is an annual rainfall between 400–750mm, in Samburu District between 250
and 1250mm in the higher parts, Isiolo district has an average annual rainfall of about 580mm, and
Marsabit District between 200–1000mm. So the study area is located in the drier parts of the country
with only higher rainfall averages on the more elevated parts (Ministry of state for the Development
of Northern Kenya and other arid lands, read 11/2008).
Kenya consists of more than 80% arid and semi-arid land. Only about 15% of the country receives
enough rain to grow non-drought resistant crops, 13% has marginal rainfall sufficient to grow selected
drought resistant crops, while the other 72% has no agronomic useful growing season. On these latter
grounds, no agriculture is possible without irrigation. When no irrigation is applied, the land consists
of a mixture of grasses, shrubs and trees, with water availability and soil type determining the exact
spatial patterns of the plant communities (World Resources Institute et al., 2007).
3
Livestock
There has been an increased grazing pressure on the semi-arid rangelands of northern Kenya during
the last decades (Cornelius & Schultka, 1997). The semi-arid savannas in the Isiolo and Samburu
Districts used to be pastures for cattle during the rainy season. In the dry season, the herds moved to
wetter grazing refugia on the Laikipia plateau and on the Wamba mountains. During the past decades,
a successive change of the grazing schemes was observed to a year round grazing of cattle. Grasses are
13
the main component of cattle’s diet, even during the dry season. The rangelands of northern Kenya,
have very limited biomass of valuable standing hay, and there is a quick deterioration of the forage
quality of the herb layer after the rains. As the rainfall is extremely unreliable, the forage supply varies
greatly between the years. As this region has so much limitations and uncertainties, year round cattle
grazing is not a suitable land use here (Schultka & Cornelius, 1997). The consequence of this yearround grazing is a deterioration of the natural pastures. The overgrazing is often accompanied by an
decrease of perennials in favour of annuals. The vegetational degradation also causes a replacement
of indigenous flora by invaders (Cornelius & Schultka, 1997).
However, the rangelands possess a huge potential for food production. Besides grasses and forbs, there
is the available biomass of dwarf shrubs, shrubs and trees. These plant forms can provide forage for
smaller livestock (sheep, goats, . . . ), donkeys and camels. When a mixture of grazers, browsers and
intermediate types of feeders is used, the rangelands are best utilized and risks of climatic uncertainties
and prolonged droughts are less severe.
Livestock can have an influence on vegetation patterns. One example is the encroachment of Acacia
species which results in thickets. This encroachment into thickets is a widespread problem in African
savanna that is mainly attributed to overstocking. There are different origins for thickets, some occur
on soils eroded by heavy trampling like Acacia reficiens and Acacia horrida thickets, others are limited
to eutrophicated sites like juveniles of Acacia tortilis. As soil erosion is irreversible, the first thickets
are very hard to restore (Schultka & Cornelius, 1997). Grasses and herbs are suppressed by these
impenetrable thickets. Overgrazing is believed to be the cause for woody plant encroachment due
to changed grass-tree competitive interactions. Another reason can be the loss of fuel leading to
a disrupted fire regime (Wiegand et al., 2006). This increase in woody plant abundance over the
past century in rangeland results in a decline in the suitability of rangeland for cattle production.
Native ungulates, especially elephants, can play an important role in reducing and even reversing
shrub encroachment (Augustine & McNaughton, 2004).
4
Vegetation
In northern Kenya the savannas receive low and erratic rainfall that is coupled with a high evaporative
demand. Between the rainy seasons long dry spells occur, with plant opportunities limited by a short
rainy season, normally lasting about 60 days. Plants that establish and quickly mature have a good
chance of surviving to the next generation (Keya, 1997). The study site mostly consists of savannas,
communities composed of more or less continuous herbaceous layers and of a discontinuous shrubarborescent layer. Water is collected from different pedological horizons by grasses and trees. Grasses
use the shallow water rather than the deeper water and this allows the coexistence between the trees
and grasses (Akpo, 1997). Savanna grasses’ growth is largely confined to the wet season. They have a
rapid growth response to the first rains as that is the moment where nutrients become available through
decomposition of litter accumulated during the dry season. Woody plant species grow throughout the
14
year with a top growth during the wet season. Woody trees and shrubs, contrary to herbs, produce new
leaves before the first rains, possibly triggered by photoperiodicity, temperature and moisture (de Bie
et al., 1998).
A lot of trees within the study area have a multi-stem growth form. Some contrasting growth forms
of trees that occur regularly are the dense umbrella-shaped canopy of Acacia tortilis and the open,
irregular canopy of Commiphora Africana. There is also a common occurrence of the evergreen tree
Boscia coriacea. Some nearly closed woody vegetation along rivers and channels contain trees as the
most conspicuous life form, but are dominated by shrub life forms. The most characteristic species
of these gallery woods are Grewia bicolor and Cordia crenata. Acacia xanthophloea is a tall growing
tree confined to the banks of some permanent rivers. This tree overgrows A. tortilis by about 4m.
Some characteristic shrub species are Acacia mellifera, Grewia villosa, Caucanthus albidus, Cadaba
farinosa, Grewia tenax and Cordia sinensis. Some patches are more composed of thickets, which can
be formed by the shrubs Acacia horrida or Acacia reficiens.
Some locations contain a well developed understorey of dwarf shrubs with some dominating species
being Lippia carviodora, Vernonia cinerascens and Sericocomopsis pallida. All these (dwarf) shrub
species are indigenous to the semi-arid lowlands of northern Kenya. A much occurring dwarf shrub
is Indigofera spinosa, a species of the semi-desert grassland and drier Acacia-Commiphora bushland.
Some others are Hibiscus micranthus, Indigofera volkensii and Pavonia patens which are characteristic of dry savannas (Schultka & Cornelius, 1997).
Sometimes there are vegetation patches around shrubs. These originate from a positive response of
plants to an increased infiltration, a reduced soil moisture evaporation and the protection from herbivores created by these shrubs. So within these patches, there is a concentration of cycling resources,
with a limited movement of water, nutrients and propagules outward into the inter-shrub areas (King,
2008).
The herbaceous layer consists of grasses and forbs. Some species that occur with high frequencies as
widespread weeds on arable fields are Ipomoea plebeia, Oxygonum sinuatum, Ocimum americanum
and Pupalia lappacea. Some annual grasses that are species typical of disturbed habitats like heavy
grazed pastures, arable fields and roadsides are Tragus berteronianus and Setaria verticillata. There
are also some forbs typical of disturbed habitats, they are occurring in arable fields or in the vicinity
of settlements like Digera muricata, Cyathula orthacantha, Tribulus cistoides, Achyranthes aspera,
Leucas urticifolia, Commelina benghalensis and Erucastrum arabicum. Sporobolus nervosus is a savanna grass species, Chrysophogon plumulosus and Oropetium minimum are perennials with the latter
also being adapted to more arid conditions. Some annual grasses are Cyperus blysmoides, Tetrapogon
cenchriformis, a typical plant of semi-deserts and the pioneer Setaria acromelaena. Some indigenous
savanna pioneer forbs are Blepharis linnarifolia, Ipomoea cordofana and Farsetia stenoptera. Some
species live in the bed of dried channels and rivers like the annual forb Mollugo cervinia, the annual
grass Eragrostis aspera and the perennal grass Cynodon plectostachyus that experiences seasonal
15
flooding (Cornelius & Schultka, 1997).
5
5.1
Conservancies
Introduction
A large part of the study area consists of conservancies, community-led conservation initiatives. Conservancies contribute to the protection of specific biodiversity, they provide green corridors for the
movement of game, or they can be protected habitats where rare and endangered species occur. The
registration of a conservancy does not involve a change in land use, there are for instance many farms
that are part of conservancies. The only requirement is that the land is managed by good environmental practices (conservancies.co.za, read 11/2008).
The conservancy areas in the study area are managed by the local communities with the support of
a local institution, the Northern Rangelands Trust. The membership of community conservancies is
expanding, the area is already about 600000 hectares, and home to approximately 60000 pastoralists
of different ethnic origin. The goal of the Northern Rangelands Trust is to solve the local problems by
creating long-lasting local solutions, and by this, leading the community to development and preserving the resident wildlife. The growing recognition of the value of wildlife as an alternative income
strategy and contributor to development for the community at large, is one of the main reasons for
conservancy establishment. The wildlife value is clear in the demarcation of core conservation areas
within conservancies, which are livestock free and focused on the development of wildlife and tourism
(NRT, read 11/2008). The conservancies are: Il Ngwesi, Kalama, Lekurruki, Meibae, Melako, Namunyak, Sera, West Gate, Ruko, Naibunga, Ltungai, and Newlew.
Figure 3.2: Conservancies (NRT, read 11/2008)
16
5.2
Conservation of Grevy’s zebras
The community-owned rangelands of northern Kenya are one of the few places left in Africa where
wildlife can move freely across a vast area without fences, that is protected by communities (NRT,
read 11/2008). Large areas of land are secured, allowing a continued migration of wildlife through
their natural range, with complementary protection, monitoring and management of wildlife and
its rangeland. The communities have already undertaken several actions to protect the endangered
Grevy’s zebras.
An anthrax outbreak in Wamba area in December 2005, disproportionately affected Grevy’s zebras.
After a test period on a small group of animals, a broader vaccination was successfully conducted
on approximately 620 Grevy’s zebras. This vaccination was led by the Kenya Wildlife Service in
association with the Lewa Wildlife Conservancy and Northern Rangelands Trust. They also had the
assistance from County Councils and communities. The main targets were the populations most at
risk from the disease. To reduce the level of environmental contamination, a mass vaccination of over
50000 head of livestock was performed.
Since May 2003, there is a Grevy’s Zebra Scout Programme, which employs women and men of
the communities to collect data on the distribution and abundance of Grevy’s zebras. The Northern
Rangelands Trust coordinate the programme. It receives funding support from Saint Louis Zoo and
technical support from Dr. D. Rubenstein of Princeton University. The collected information provides
a better understanding of the ecological pressure on this species in areas of high livestock density
and guides the community conservation programs so that community members themselves have the
opportunity to make recommendations on ways to reduce competition between Grevy’s zebra and
livestock.
Wildlife and telemetry experts have been able to develop advanced tracking systems for Grevy’s zebras. Collars for Grevy’s zebra were developed and deployed by the Northern Rangelands Trust, Lewa
Wildlife Conservancy and Save The Elephants in June 2006. The collars measure GPS position every hour. This information is critical in helping the communities manage their conservation areas to
benefit Grevy’s zebra (NRT, read 11/2008). In this work, the data collected by these collars will be
used.
17
Chapter 4
Wildlife telemetry
1
Introduction
According to the dictionary, the definition of telemetry is: ”The science and technology of automatic
measurement and transmission of data from remote sources by wire, radio, or other means to receiving
stations for recording and analysis” (Dictionary, read 04/2009). Telemetry will be discussed here as
it is used to monitor the migration of the Grevy’s zebras.
The term wildlife telemetry is generally associated with the study of animal movements with the use
of radio tags. Radio-tracking is like wildlife telemetry but without the transmittance of data on the
status of the animal. When using wildlife telemetry, the disturbance of the normal behaviour of the
animal being studied should be avoided. The basic idea of wildlife telemetry is to study living animals
in their natural environment.
There are several ways to collect measurements by remote means. There is always the need for
interception of energy radiated by the animal or reflected from the animal. Wildlife telemetry has to
use a transmission mode not detectable by the animal, to avoid the disturbance of their communication
with one another or their sensing of the environment (Priede & Swift, 1992).
In the past, data were often obtained from field surveys using direct observation of the animals. Transect surveys were conducted where animals in the vicinity of a set of sampling lines or points were
recorded. The problem with these methods was that it yielded relatively few sightings, particularly
for rare species living in inaccessible environments. By using the advances in communication and
information technology, radio- and satellite-telemetry became available and increased the amount of
data on animal movement, with a focus on the individual animal (Aerts et al., 2008). Nowadays, there
is also the possibility of GPS tracking. Other basic information like survival, mortality, migration
periods, home range, and territoriality can also be achieved using telemetry methods. In addition, this
information can be related to other individuals: which animals share their home range, which ones
18
CHAPTER 4. Wildlife telemetry
avoid each other, which are the likely partners,. . . Telemetry can also be helpful to locate the animals
for direct observation. The data obtained from telemetry studies are usually coupled with remote sensing data and GIS technology. More about the link between tracking data and remote sensing will be
handled in chapter 5. In the following sections all three of the telemetry technologies, VHF-tracking,
satellite tracking and GPS tracking, will be discussed (Priede & Swift, 1992). GPS-tracking was the
method used in this thesis to collect data about the migration of Grevy’s zebras.
2
Very-High-Frequency (VHF)
Ground based very-high-frequency (VHF) radio tracking became possible in the 1960s, and it allows
the scientist to monitor species movements and home ranges over 50 to 300 km2 . VHF tracking can
record species location to within a couple of meters and can be undertaken in areas with dense canopy
cover, which is an advantage over satellite tracking (Gillespie, 2001). This classical radio tracking
technique uses very-high-frequencies; this is the wavelengths between 1m and 10m. The animals
have a radio transmitter in a collar or a tag attached, and with the use of a hand-held directional
antenna, a receiver and headphones, a field researcher is able to track them. An animal’s location is
determined from a series of bearings, which is determined by listening for peaks or nulls in the signal
level, and it is usually confirmed by visual sighting (Priede & Swift, 1992).
The choice for VHF band has several reasons. VHF is the highest frequency at which simple crystal
oscillators can be used to generate the carrier frequency directly. The transmitters can therefore be
made with less than ten components and can weigh less than 1g. There are several transmitters varying
in size, complexity and performance. The antenna dimensions are also advantageous. As the antenna
size is directly proportional to wavelength, the VHF frequencies give a reasonable practical portable
directional antenna. To achieve optimal performance, it is important that the transmitter is carefully
matched to the application. The transmitters typically emit a 20 ms long pulse every 0.5–1 seconds to
minimize power consumption. To distinguish between different individuals, different pulse rates and
frequencies are used, but the combinations are limited in the narrow bands allocated in most countries
for biotelemetry.
There is one major problem with VHF tracking, which is unfortunately often ignored, and that is
the poor location precision of the technique. A visual confirmation of the animal wipes out this
problem. The simple, relatively cheap equipment and the own manufacture of transmitters are the
major advantages. However it is still a very labour intensive method and the use of it in routine
investigations can not always be justified as it has a huge demand for resources. Another disadvantage
is that the information is only gained when researchers are actively receiving signals, although the
radio is constantly transmitting. The result is small sample sizes with only a few locations per day.
With this conventional system it is difficult for a person to collect more than three locations for 20
animals per day. To collect more locations, more people are needed or fewer animals can be tracked.
19
More people will increase the errors due to different operators. Instead, it is common to take bearings
intensively over a short period of time. The loss of dispersing individuals during non-telemetry times
can obstruct further data collection. Typically only one person takes all bearings which results in a
lower accuracy as animals move between the bearings. If the operator is too close, he can cause a
disturbance of the animals (Priede & Swift, 1992).
VHF telemetry is not as commonly used anymore as there are easier and more efficient methods available. Four recent studies that use VHF telemetry are given as an example. In Belgium, a study was
conducted on red deer (Cervus elaphus L.) to investigate their habitat use. The location of the tagged
animals was recorded once a week (Licoppe, 2006). In Norway, a study was conducted on ringed
seals (Pusa hispida Schreber) to examine their haul-out behaviour. In addition to visual counts, some
seals were VHF-tagged and their haul-out behaviour was monitored via an automatic recording station
(Carlens et al., 2006). In Utah and Idaho (USA), pumas (Puma concolor L.) were VHF tracked to
estimate their preying behaviour (Laundré, 2008). In Northern Chile, Humboldt penguins (Spheniscus
humboldti Meyen) were VHF tracked to determine their at-sea behaviour (Culik et al., 1998).
The use of radio telemetry is generally restricted to a limited area or to species with a limited range
of movement. Unless observers are able to stay within several kilometres of the animals, it is rather
difficult to apply it to study long-distance migrants. The receiving of signals and following of animals
often becomes constraining in hilly terrains or dense vegetation, challenging the use of this technology
(Javed et al., 2003).
3
Satellite tracking: Argos system
A major challenge in satellite tracking is the requirement of a radio signal, coming from a small
transmitter package on the animal, which is powerful enough to be received on board a spacecraft.
The use of high altitude geostationary orbits was therefore excluded and receivers were located on
polar-orbiting remote-sensing satellites. There is currently only one operational system namely the
US/French Argos system which began service in 1978. The program is the result of a far-reaching cooperation between the Centre National d’Etudes Spatiales (CNES, France), the National Aeronautics
and Space Administration (NASA, USA) and the National Oceanic and Atmospheric Administration
(NOAA, USA). The receivers are carried on board the NOAA series of satellites, which are spacecrafts in circular, polar orbits at 850km altitude. These satellites are scheduled to provide a complete
global coverage to the Argos system, so that it can collect locations of fixed and moving platforms
worldwide. The transmittance at 401.650 MHz by the Argos platform transmitter terminals (PTTs),
makes conveniently small antennas and very high transmission rates possible (Priede & Swift, 1992).
Service Argos estimates the PTT’s location from Doppler shifts in frequency. The Doppler effect is
the change in frequency of the electromagnetic wave caused by the motion of the transmitter and the
receiver relative to each other. The frequency of the signal measured by the satellite receiver is higher
20
than the actual transmitted frequency when the satellite approaches the transmitter, and lower when
the satellite moves away. These location measurements are then relayed to ground stations in USA
and France from where users can directly retrieve data from Service Argos’ website or via electronic
mail. Argos provides a range of location accuracies. The most accurate location, class 3 (LC3), has an
estimated range of radius of 150m. LC2 has a radius range of 350m, LC1 of 1000m, and LC0 of more
than 1000m. Argos also provides a signal quality index with each location. After PTT purchase, the
PTTs need to be registered with Service Argos and an agreement has to be signed (Javed et al., 2003).
On this agreement form there is information about the programme, the objectives, the requirements of
the system, the duration of the program, . . .
A spacecraft’s pass over a given position lasts 10–12 minutes on average and the Argos PTTs transmit
messages every 90 seconds. Data collection is possible from a single message, but the location of an
animal is determined using two messages. For tracking very mobile species, there is the possibility to
request a shorter repetition period of 60 seconds between messages.
Several thousand PTTs are in operation in the world today. Researchers and manufacturers in satellitebased tracking face major problems with the development of PTT technology, including environmental capability, matching the PTT to the animal, the PTT power supply and sensor data. The production
of hardware that is suitable for the animal and withstands its natural environment is a significant part
of the effort. PTTs must be resistant to corrosion and sea water, total leak-tight, resistant to impact
and resistant to pressure. The suitability of a PTT for an animal is dependent on several aspects like
weight, shape and size, and attachment method (Priede & Swift, 1992). Currently the PTTs manufactured can weigh less than 50g (Telonics, read 10/2008). The PTTs include solar panels or lithium
batteries. A PTT must not disturb an animal’s aerodynamics or hydrodynamics and must not modify
its temperature regulation. There are several attachment methods available including subdermal anchoring, boding inside fur, and PTTs inside collars, harnesses, and harpoons attached to float. The
chance that the animal can be freed at the end of the programme should be maximized as a result of
the design of the attachment method. As long-term animal tracking programmes are now possible
with the use of satellite telemetry, the production of reliable equipment and the use of long-life power
supplies is encouraged. PTTs can also be used for other purposes besides animal tracking. They can
provide data on a range of behavioural and physiological characteristics, for example the monitoring
of animal activity over short and long periods, number, duration and depth of dives in marine animals,
water temperature, air temperature and barometric pressure . . .
There are two ways to collect location data namely continuous monitoring, where each movement of
an animal is noted, or interval sampling. To collect behavioural data, or to track animals in terrains that
are difficult to reach is only practical by using continuous monitoring. To analyse the data, a software
package is used which is able to calculate some summary statistics for each monitoring session with
a particular animal and some proportions of fixes in particular categories (Priede & Swift, 1992).
The locations recorded by the receivers in the NOAA satellites are in the form of latitude and lon21
gitude. With the use of habitat information gathered via remote sensing and the tracking data, it is
possible to develop a more complete picture of the animal’s long-distance movements, an aspect of its
ecology especially important for conservation of species and their habitats (Javed et al., 2003).
The satellite tracking method has its own advantages. Once a transmitter is attached, the researcher
does not need to undertake extensive field triangulation. It is also easier to study wide-ranging species
that regularly cross international boundaries (Gillespie, 2001). In relation to VHF radio- tracking, the
Argos system is highly affordable and competitive if full programme costs are taken into account.
These costs include satellite-based wildlife tracking and Argos data distribution, journey and staff
costs and other travel expenses, the capital equipment needed for fieldwork and the associated logistics
burden, and the purchase of hardware such as receivers (Priede & Swift, 1992).
There are a lot of studies where satellite tracking is used. Only a small number of them will be
given as an example. In Mexico, humpback whales (Megaptera novaeangliae Borowski) were satellite tagged to follow their migratory movements and surfacing rates (Lagerquist et al., 2008). In
Mongolia, Mongolian gazelles (Procapra gutturosa Pallas) were satellite tracked to compare their
migration to seasonal ranges with biomass via NDVI (Ito et al., 2006). In West Greenland, satellite
tracking of caribou (Rangifer tarandus L.) was a valuable tool to identify critical caribou areas in
summer. That makes it possible to change tourism and mineral exploitation to have a minimal impact
on caribou population (Tamstorf et al., 2005). A study using satellite tracking of leatherback turtles
(Dermochelys coriacea Vandelli) across the North Atlantic ocean, showed that some turtles are not
foraging at particular hotspots but have a pattern of near continuous travelling (Hays et al., 2006). In
Sweden, satellite tracking of common buzzard (Buteo buteo L.) revealed their short-distance migration
pattern (Strandberg et al., 2009).
4
4.1
Global Positioning System (GPS) tracking
Operation of the system
The relatively new tool, Global Positioning System (GPS) collars, can solve a lot of problems associated with traditional radio-telemetric techniques (Johnson et al., 2002). This tool became widely
available to wildlife biologists in the mid 1990s (Blake et al., 2001). The determination of the location by GPS is based on a measurement of the distance between satellite and receiver. As the position
of the satellites is known, the location can be derived from the time the radio waves take to get from
the satellite to the receiver. GPS has the advantage of monitoring the most precise locations with the
fewest biases and has the possibility of relocating animals frequently, up to once per second. GPS
also works 24 hours, so data is even collected during the night, and during any weather condition
(Johnson et al., 2002). Like traditional radio-collars, GPS collars require the capture of the animal
and the fitting of the collar. The collars are normally pre-programmed to collect data according to
22
specified schedules (Coelho et al., 2007). The collar can be located due to a traditional VHF receiver
and with a UHF modem link it is possible to communicate between the collar and a remote laptop
computer. This link allows data download, RAM memory clearing, and reprogramming of the collar
(Blake et al., 2001).
The collected data consists of the wearer’s identity, time of day, date and coordinates, de PDOP
value (Positional Dilution Of Precision) and if the acquired signal is two-dimensional (2-D) or threedimensional (3-D) (Coelho et al., 2007). 2-D fixes are recorded when the GPS unit simultaneously
contacts three satellites, and 3-D fixes are those recorded when the GPS contacts four or more satellites
(Lewis et al., 2007). So, researchers obtain substantially larger spatial location datasets with greater
precision and significant cost savings per location. This brings also the challenge of managing and
analyzing huge volumes of data constantly updated. A standard, complete and cost-effective software
system to fully support management, modelling and analysis of GPS collar data is not yet available to
the scientific community (Cagnacci & Urbano, 2008). This technology should if possible be combined
with field research. The cost of GPS collars is more the price of conventional VHF radio-collars, but
GPS tracking makes it possible to simultaneously collect spatial data on many individuals (Coelho
et al., 2007).
Using GPS data, several attributes of free-ranging animals can be calculated. Location can be estimated from a single GPS fix, speed can be calculated from a minimum of two points and by using
three points, turning angle calculations become possible. These estimates are based on straight line
distances between location points. In the intervening period between a long fix interval, there is uncertainty about an animal’s activity. These long fix intervals underestimate the actual distance travelled
by the animal. Only the minimum speed and minimum distance travelled can be calculated between
two consecutive pairs of location fixes. In reality, animals take less direct routes to the second point
and therefore could travel faster and this higher speed enables them to access a wider variety of resources. With increasing time between GPS fixes, there is an increasing prediction error (Swain et al.,
2008).
GPS gives the possibility of obtaining worldwide locations at 1-second intervals throughout the 24
hour cycle, regardless of weather and roughness of terrain. The major advantages of GPS for wildlife
biologists is its simple use, cheapness, lightweight and it can give instantaneous locational information
in real time.
4.2
Accuracy
Until May 2000, the accuracy of GPS locations was downgraded by the process called Selective
Availability (SA) intentionally imposed by the US department of Defence. Before this date, only
uncorrected or post-processed differential GPS data could be used. In the future it will still be possible
23
that SA is reactivated. Uncorrected GPS data had a location error between 20m and 80m and postprocessed differential GPS data had an error between 4m and 8m (Hulbert & French, 2001). In
the differential method, post-treatment corrects distances to individual satellites in the GPS collar
with data obtained simultaneously by a GPS reference base station (Adrados et al., 2002). Both the
receiver and the reference station record errors in time and hence distance between the GPS receiver
and from all visible satellites. After the post-treatment, locations are accurately determined. With the
disablement of SA, the accuracy of GPS locations is considered to be less than 1m. This accuracy was
previously not possible without the use of complex or expensive equipment.
Data accuracy can also be improved using more satellites to calculate the location. A fourfold improvement in accuracy can be achieved between locations obtained from four satellites and those
calculated using data from five or more satellites. Some errors can also be caused by the receiver
itself, topographically induced errors or errors due to ionospheric and tropospheric delays. The objectives of the study and data requirements of the hypothesis being tested will determine whether further
complex tasks should be performed on the data to remove these errors and obtain an even better accuracy. It is important that the accuracy of the GPS-derived locations is better than the resolution of
the map used for tracking. Before each study, it is important to test the GPS device for instrument
errors and to test for errors specific to each site before deploying this technique on animal collars or
for mapping.
As a consequence of suboptimal satellite geometry, the accuracy of many locations can degrade beyond that quoted by the manufacturer. Purchasers of GPS collars have the option to employ differential correction, which can increase precision (Hulbert & French, 2001). Differential correction can
remove other sources of error besides SA, namely satellite clock errors, ionospheric and tropospheric
obstruction (Adrados et al., 2002). However, differential correction can have many unforeseen drawbacks that can add to project costs, or reduce immediate usefulness of the data. With the disablement
of SA, the use of it is substantially reduced. So, differential correction is not necessary for all projects
as it requires a large amount of computing time, more memory per location, more frequent retrieval
of data, greater power demands, and therefore results in a reduction of the collar’s field life (Johnson
et al., 2002). The choice to use differential correction will be determined by the scale needed in the
study. It opens new approaches to wildlife research as it allows the study of animal-habitat relationships at a very fine spatio-temporal scale, never achieved before with other techniques (Adrados et al.,
2002).
4.3
Obstructions
The GPS collars have to be small enough for an animal not to be hindered (Sprague et al., 2004). The
recommended weight of a collar has to be lower or equal to a limit of 5% of the body weight (Coelho
et al., 2007). In using GPS collars, there is always the trade-off between the weight and functions
24
of the collar. Functions that require electricity and large battery must be reduced to acquire a lighter
GPS collar. These collars may limit the satellite search time, record fewer positions, and need to
be recovered for data download after detaching automatically. As batteries, antennas, and electrical
efficiency keep improving, it will be possible to get better acquisition rates and increasing number of
positions recorded. It may even become available to have data download and detachment on radio
command for smaller GPS collars.
Satellite signals can be blocked by canopy enabling the GPS device to calculate a location. Although
even in forests, sufficient opportunities exist for the GPS device to receive satellite signals to calculate
the position. Animals sometimes are in clearings or under deciduous trees without leaves, where
there is a very good reception of the satellites. Some animals are able to climb in trees, where at that
height they have a good reception (Sprague et al., 2004). If the receiver is not capable of obtaining
signals from a minimum of three satellites, it cannot calculate a location. Before using this technology,
researchers should perform an assessment of the GPS device’s performance across the habitat types
animals are expected to use. Generally, the reception of satellite signals will be degraded by large
diameter, dense and tall vegetation, and steep topography. Consequently there can be a large variation
in location acquisition rates within and among study areas (Johnson et al., 2002). GPS receivers can
be confused by multi-path effects, where multiple signals are bounced off of tree trunks or moving
canopy in windy conditions.
The topography plays an important role in acquiring contact with satellites as well, hills for instance
can block the sky (Sprague et al., 2004). The orientation of a hill on which an animal is present can
also play a role in location determination. It is however safe for the researcher to assume that large
biases into GPS radio-telemetry data will not be imposed by orientation alone (D’Eon & Delparte,
2005).
Animal behaviour can also play a significant role. In contrast to stationary collars, movement of the
collars on free-ranging animals may result in much lower GPS location performance. The position
of the GPS antenna attached to the collar has also an effect on the collar performance. In an open
flat terrain, there is no difference in performance between a vertical or horizontal antenna position,
possibly because there are more than enough satellites available to calculate the location in the open
area. However, in under closed canopy, horizontal positions, for instance on laying animals, exhibit
increased location time and location error, decreased fix rate, number of satellites available, and 3-D
proportion. So there is a negative effect of the horizontal antenna position in forests on the location
performance, which suggests that the vegetation makes it difficult to collect information from satellites
with a horizontal antenna. If the number of satellites available reduces, it becomes impossible for the
GPS to choose an ideal constellation distribution in order to calculate a more accurate location or
even impossible to acquire a location at all. An antenna on a collar worn by an animal can easily
shift to horizontal position when the animal is feeding, sleeping or resting resulting in a decrease in
fix rate (Jiang et al., 2007). Attention has to be paid for biases introduced by this effect of antenna
25
position. For example, the activity of an animal digging while foraging may potentially translate into
significantly lower fix rates than at other times, and would result in proportionally fewer recorded
locations for this habitat type. The researcher could therefore draw the conclusion that these locations
are used infrequently or even be avoided, when in fact the opposite is true (D’Eon & Delparte, 2005).
The location time is affected by all habitat features except for available sky. This suggests that battery
longevity will be shortened by all of the obstacles that limit the number of satellites available or that
interfere with the connection between the satellite and the GPS collar. Poor dispersion of satellites
will result in a poor triangulation and thus in low-quality location data (Jiang et al., 2007).
In some studies, the collars can be subjected to extreme variations in temperature over short periods of
time, rapid changes in humidity, and complete immersion in water. Some collars, operating under extreme conditions, can face premature failure. This failure can cause extra costs by other factors. First,
the malfunction is sometimes not immediately diagnosed due to the infrequent animal relocations.
If a collar failure is detected, there is some additional time needed to arrange a recapture operation,
which then can be delayed due to poor weather or unsuitable terrain. Once recovered, collars have
to be diagnosed, repaired and returned by the manufacturer. So collar malfunctions, together with
organisation, logistics and weather delays, contribute to a significant loss of potential data. To ensure
the collection of enough data, there should always be a minimum number of collars in the field. This
can be assured by keeping a pre-determined number of collars in reserve to replace the ones who fail.
Three manufacturers of GPS collars (Lotek Engineering, Newmarket, Ontario, Canada; Televilt International AB, Lindesber, Sweden; Telonics, Inc., Mesa, Arizona, USA) make the remote retrieval
of data and diagnostics capable. The remotely programming with new location and communication
schedules, for instance where sampling strategies need to be adjusted in accordance with unpredictable
animal behaviour, is possible with the GPS collars from Lotek Engineering. Depending on the species
and study duration, these features can be very useful. The user-collar communication is not necessary
when animal capture can be performed year-round and is inexpensive, or when information about animal movement is required only for short periods. If study lengths exceed collar memory and animals
are difficult to capture or where animals periodically move large distances and are difficult to relocate,
remote data retrieval is recommended (Johnson et al., 2002).
The ability to record fixes at a high frequency, even though GPS data is extremely accurate, results
in increased noise/signal ratios when there is little movement or when speeds are low. It is best to
analyse GPS data where movements are at least twice the minimum resolution per sampling interval
to obtain a respectable signal/noise ratio (Ryan et al., 2004).
26
4.4
Errors
The detailed information acquired with GPS can be used to evaluate wildlife movement, space use and
resource selection with a high degree of precision and accuracy. Nevertheless, there are two types of
errors that can bias analysed data on GPS locations, namely missed location fixes and location error,
the difference between an objects actual or true position and that estimated by a GPS fix.
First there is the unsuccessful fix acquisition, which leads to missing location data. Stationary GPS
collars have fix rates ranging from 68–100%, with most collars above 85%, but sometimes rates as
low as 13% (Lewis et al., 2007). Missing locations equate to a loss of information, which implicates a
reduced efficiency and potential biases. As failed location attempts do not occur randomly but systematically, bias is likely to occur in GPS telemetry studies. The conditions that can affect GPS location
acquisition are canopy type, percentage canopy cover, tree density, tree height, and tree basal area,
which all can be strengthened by the interaction with a mountainous study area (Frair et al., 2004).
GPS collar data may therefore be biased towards acquiring satellite fixes in more open habitats with
favourable topography. Another major factor affecting GPS collar fix rates is animal behaviour. Collars that attempt location fixes at shorter trials have also higher fix rates, so collar location acquisition
schedules also influences fix rates. Development and application of correction factors can be applied
to GPS location data sets to counter biases associated with missed locations.
The second error type is inherent in all telemetry systems: location error. Dependent on the magnitude of location error and the degree of landscape heterogeneity, this can lead to misclassification of
habitats used by the animal. Habitat components, like canopy cover and terrain obstructions, largely
influence location error. Atmospheric conditions can also contribute to this. The 3-D fixes are generally more accurate. With increasing canopy cover, 2-D fixes increase as a result of satellite signal
obstruction.
For each location, there is a PDOP value recorded. This is a measurement of satellite geometry. Lower
PDOP values indicate a wider satellite spacing able to minimize triangulation error and thus a more
accurate location estimate. Screening of location data is sometimes used to reduce location error.
There are several screening mechanisms, for instance, screening out of 2-D fixes, removing data with
high PDOP values. . . However, this screening can also lead to significant reduction of location data
and introduce additional biases into analyses of animal locations. The seemingly most effective data
approach could be the screening out of 2-D fixes at a specific PDOP cut-off; this can be a suitable
compromise between reducing large location errors and minimizing data reduction. By evaluating
the proportion of locations with relatively large errors, an appropriate PDOP threshold value for data
screening can be chosen. Screening should be used with precaution not to potentially introduce biases that affect estimates of habitat and space use. These extra biases can be caused by eliminating
locations associated with habitats that induce greater errors. Sites with high terrain obstruction, high
canopy cover or a combination of both have most missing fixes. In addition, these sites could have
greater location error because collars receive location signals from fewer satellites that exhibit poorer
27
satellite configuration. The PDOP values and number of 2-D fixes will therefore increase (Lewis et al.,
2007).
4.5
Examples of studies using GPS telemetry
There is an increasing amount of studies using the GPS technique to follow animal movement. In
Brazil, wild maned wolves (Chrysocyon brachyurus Illiger) were GPS tracked during full and new
moon nights to discover the effect of the full moon on the activity of the predator (Sàbato et al.,
2006). In Kenya, African elephants (Loxodonta Africana Blumenbach) were GPS collared to investigate their movements and use of corridors in relation to protected areas (Douglas-Hamilton et al.,
2005). In Greece, loggerhead sea turtles (Caretta caretta L.) were GPS tracked to examine their use
of microhabitat as the reason of their reproductive success at a margin of their range (Schofield et al.,
2009). In Wales (UK), Manx shearwaters (Puffinus puffinus Brünnich) were GPS tracked to monitor
their foraging movements with a GPS device weighing 17g (Guilford et al., 2008). On Europa Island, in the Mozambique Channel, red-footed boobies’ (Sula sula L.) flight pattern and the way they
forage over tropical waters was examined using GPS packages (Weimerskirch et al., 2005). In Sweden, moose (Alces alces L.) and roe deer (Capreolus capreolus L.) were GPS tracked to examine the
effectiveness of a highway overpass (Olsson et al., 2008).
28
Chapter 5
Wildlife tracking and remote sensing
1
Introduction
The most important methodology for plant diversity assessment is the direct mapping of species and
associations, based on characteristic spectral reflectance features of plant species or plant communities. Faunal species, which are mobile, complicate the assessment of species occurrence and richness.
This is especially the case for migrants, which can move long distances occupying a wide range of
habitats. For these non-sessile animals approaches are based upon proxies and surrogates (Leyequien
et al., 2007). Animals’ basic needs, forage, water, and shelter, mostly vary spatially and temporally,
which makes their movements not randomly, but distributed in relation to the variation of their needs.
Remote sensing techniques can provide the opportunity to map these basic needs. The physiognomic
landscape features, shelter and shade, can be derived directly from spectral information on various
imagery bands. For forage distribution, several indices can be applied, like the widely used Normalised Difference Vegetation Index (NDVI) (van Bommel et al., 2006). In the following sections
some methodologies are discussed to map the distribution of animals.
2
2.1
Habitat maps and habitat suitability mapping
Habitat maps
Land cover is the observed physical description of the earth’s surface, and is the attribute most commonly mapped with remote sensing methods. This first data layer is then combined with additional
information to derive more useful spatial products. These land cover maps are usually not enough to
reveal underlying mechanisms and the dynamics of complex natural landscapes or to improve predictions of species distributions. The more useful habitat maps can be derived indirectly from land cover
maps or they can be modelled through integration with other environmental factors.
29
CHAPTER 5. Wildlife tracking and remote sensing
Habitat mapping can be conducted on various spatial scales. The labour-intensive manual interpretation of aerial photographs is limited in range, and frequently used as a complement to field surveys.
It is used in studies of species with limited ranges and in the analysis of relatively small areas. The
skilled interpreters, who perform these analyses, generate detailed, high-quality information.
Digital processing of high spatial resolution imagery, like Landsat-7 (30m), medium spatial resolution imagery, like MODIS (250m), and low spatial resolution imagery, like SPOT-Vegetation (1km)
is possible for larger study areas (McDermid et al., 2005). The use of time series of satellite data can
even improve this habitat mapping. The changes over time of vegetation structure, productivity, and
phenology are as important for some species perception of habitat quality as temperature and precipitation (Bradley & Fleishman, 2008).
A habitat classification of a larger area is based on training data collected during field surveys. This
training data are being digitized and the spectral properties of the different habitat classes are determined. The result is a classification of the area of interest into discrete habitat types. A major
drawback of this technique is the assumption that the conditions at the training locations may be extrapolated over a larger area. To make sure that no biased result is obtained, these assumptions need
to be tested carefully.
2.2
Habitat suitability maps
Habitat suitability is a widely used remotely sensed proxy for species distribution and richness. It
relies on the fact that each animal has its own environment in which it lives and grows. A habitat map
is created with the use of airborne or satellite data, biophysical, geophysical, and meteorological data
in combination with the knowledge of habitat preferences and requirements of the species of interest.
Data on species distribution, habitat use or characteristics, can be collected by field surveys or by
analysing the movements of collared individuals. These findings can be extrapolated to cover a large
region of interest and to estimate habitat suitability.
This discrete classification approach is not always sufficient for ecological purposes. A lot of species
require the micro-heterogeneity of areas, and many herbivore species use more than one distinct vegetation type (Leyequien et al., 2007). The quality of the habitat is also very important, for instance,
the structural complexity of vegetation and the relative proportion of cover in the understory, shrub
layer, and canopy (Bradley & Fleishman, 2008). Some non-herbivore species on the other hand show
little direct association with a habitat or vegetation type, and many species, regardless of the degree of
habitat specificity, do not occupy the full extent of their preferred habitat type. To make a successful
correlation of animal occurrence and remotely sensed habitat data, the animals need to be well studied
and their habitat preference well documented. Other species have a changing habitat preference with
geographical position, which restricts the predictive value of the animal-habitat relationship. Some
predicted distributions are wrong due to the socio-biology of the species, for example an interspecific
competition or anthropogenic influences can force them to use other less suitable vegetation classes.
30
The limited accuracy in some studies using this approach can be caused by applying proxies at inappropriate spatial, spectral, and temporal resolutions. So this technique should always be applied with
precaution.
3
Spatial heterogeneity assessment based on primary productivity
Spatial heterogeneity is a key component in the explanation of species richness. Environmental heterogene ecosystems have more different niches in comparison to simple ecosystems and can thus
support more species. It is determined by factors like temporal and spatial variation in the biological,
physical, and chemical features of the environment. The species distribution and local abundance
of individuals is thought to be influenced by the spatial and temporal varying plant productivity and
biomass of ecosystems (Leyequien et al., 2007).
There are several vegetation indices used in remote sensing to represent the presence and condition
of green vegetation. These vegetation indices are mathematical combinations of the red (R) and
Near-Infrared (NIR) bands of several sensors. The most commonly used vegetation index is the
Normalised Difference Vegetation Index (NDVI): NDVI=(NIR-R)/(NIR+R) (Lillesand et al., 2004). It
is a proxy for photosynthetic activity as it is based on the strong absorption of the incident radiation
by chlorophyll in the red, and the contrasting high reflectance by plant cells in the NIR spectral region
(Mutanga & Skidmore, 2004b). As NDVI seems to be a suitable indicator for vegetation parameters
like biomass and aboveground primary productivity, it is often correlated to faunal species occurrence
and diversity (Leyequien et al., 2007).
High NDVI values indicate vegetated areas, as these have a relatively high NIR reflectance and low red
reflectance. Clouds, water, and snow have negative values as these areas have larger red reflectance
than NIR reflectance. Rock and bare soil areas show similar reflectances in both NIR and red and thus
have NDVI values near zero. An advantage of NDVI is that it helps compensate for extraneous factors
like changing illumination conditions, surface slope, aspect, . . . (Lillesand et al., 2004). Caution has to
be taken with the use of NDVI in semi-arid areas, because of soil interference and darkening effects.
If the study area consists of a single soil type with only some sparse parts of other material, the NDVI
can be used without big negative effects (Verlinden & Masogo, 1997). There are also some factors
influencing NDVI observations that are unrelated to vegetation conditions, these are for instance the
variability in incident solar radiation, radiometric response characteristics of the sensor, atmospheric
effects and off-nadir viewing effects (Lillesand et al., 2004).
NDVI is correlated to vegetation biomass and dynamics in various ecosystems worldwide. It has
been used to monitor vegetation, estimate primary production and detect environmental change. It is
determined by the composition of species within the plant community, the vegetation form, growth
and structure, the vertical and horizontal vegetation density, and by the reflection, absorption and
31
transmission within and on the surface of the vegetation or ground, and by the atmosphere, clouds and
atmospheric contaminants (Pettorelli et al., 2006).
In this approach, animal occurrence and diversity is related to terrestrial features by means of an ecological, trophic link. This means that herbivore animals can be related to the vegetation that they
consume. If additional environmental variables, like landscape diversity, evapotranspiration, land surface temperature, rainfall, altitude,. . . are included, together with primary productivity, considerable
variation in species richness can be explained. However, scale or resolution is the main factor influencing the accuracy of predictions of species richness using primary productivity indicators (NDVI).
It is more difficult to correlate NDVI with the distribution of less abundant species as these might not
occupy all suitable habitats. This biomass-based approach is only successful for herbivorous species
that are sensitive to differences in vegetation characteristics across an area (Leyequien et al., 2007).
The use of NDVI always has to be used with elaborate ground thruthing. For instance, high NDVI
values in heavily grazed areas probably indicate high bush cover rather than green grass (Verlinden &
Masogo, 1997).
The main weakness of NDVI is its asymptotical approach to a saturation level above a certain biomass
density and leaf area index (LAI) (Gao et al., 2000). The technique has therefore a limited value in
assessing biomass during, for example, the peak of seasons. This problem can be overcome by using
narrow band vegetation indices in areas with dense vegetation. These indices are calculated using
narrow bands in the whole electromagnetic spectrum (350-2500nm) (Mutanga & Skidmore, 2004b).
More recent remote sensing products and techniques, like the Enhanced Vegetation Index (EVI) from
the MODIS product suite, can overcome this saturation problem. The goal of the EVI is to optimize the vegetation signal with improved sensitivity in high biomass regions and improved vegetation
monitoring through a decoupling of the canopy background signal and a reduction in atmosphere influences. The EVI has a higher range in values in high biomass regions, making it possible to detect
more variation in these areas. However, the NDVI has a higher range in values over semi-arid regions,
making it possible to detect more variation in biomass in these areas. For intermediate regions, both
indices show an identical range in values (Huete et al., 2002).
4
Temporal heterogeneity assessment
Seasonal climatic variations cause differences in plant species growth and establishment, leading to
changes in species composition and distributions. As a result there are changes in the spatial distribution of plant phenology and growth. When looking at the land cover data of multiple years, a vision
can be made of the influence of climate variability on ecosystems. Ecosystems are the most important
feature in biodiversity assessment and multitemporal satellite data can have the potential to describe
interactions among seasonal, annual and long-term climate variability to understand species diversity. As many animal species are very mobile over time, multi-temporal data can also provide a more
32
complete view of their occurrence and distribution unlike single-date studies that do not cover their
complete range of habitats. With the establishment of the Advanced Very High Resolution Radiometer (AVHRR) meteorological satellite series in 1980, continuous data to study ecoclimatic dynamics
became available (Leyequien et al., 2007).
5
Heterogeneity assessment based on landscape structural properties
Many species select their habitat based on structural properties of the habitat instead of species assemblage. It is stated that, in general, the more vertically diverse a forest is, the more diverse its biota
is. It is possible to estimate the structural properties and assess their heterogeneity with the use of
remote sensing (Nelson et al., 2005). Most common remote sensing techniques for relating landscape
structural properties to animal diversity, are Synthetic Aperture Radar (SAR) and the height measuring technologie of airborne lasers (i.e. airborne LiDAR). Both are active remote sensing systems
(Lillesand et al., 2004). These tools are used to map vegetation height and its variability, percent
canopy cover, field boundary height, fractional vegetation cover, and aboveground biomass (Nelson
et al., 2005; Hinsley et al., 2002).
Radar uses microwave energy while LiDAR sensors use pulses of laser light. Radar measures the
strength and origin of echoes or reflections received from objects within the system’s field of view.
LiDAR measures the time of pulse return, which is then processed to calculate the variable distances
between the sensor and the surfaces present on the ground. LiDAR has not only the possibility to
discriminate features as forest canopy and bare ground but also surfaces in between. An advantage of
LiDAR is that the data is georeferenced which makes it compatible with GIS applications (Lillesand
et al., 2004). There is extreme potential for high resolution mapping of wildlife habitats by combining
these techniques, that measure vegetation structural types, and information obtained from other remote
sensing techniques, like multispectral satellite images (Hinsley et al., 2002; Imhoff et al., 1997).
These techniques are mostly used in forest ecosystems. Several examples illustrate the use of these
techniques. In Delaware, LiDAR was used to identify and locate forested sites potentially supportive
of populations of the Delmarva fox squirrel (Sciurus niger cinereus L.) (Nelson et al., 2005). In
Australia, vegetation heterogeneity mapped with the use of SAR and aerial photography were related
to field studies of bird abundances (Imhoff et al., 1997). In England, the quality of woodland for
Great Tits (Parus major L.) and Blue Tits (Parus caeruleus L.) was estimated with the use of airborne
LiDAR (Hinsley et al., 2002).
33
6
Heterogeneity assessment based on plant chemical constituents
This last approach uses plant chemical constituents to define habitat heterogeneity and eventually assess and predict species richness. Animal species are attracted to a habitat by the spatial and structural
composition, but also by the forage quality that an animal perceives in that habitat. For example, the
spatial distribution of many wildlife species in the African savannas is influenced by the variation in
grass quality (Leyequien et al., 2007). High productivity areas can sometimes be limited for herbivores in plant chemical composition. There is a decrease in forage quality as grass matures by the
accumulation of structural tissues and their fibre content decreases as well, reducing their digestibility. It may therefore be important for broad-scale satellite-based habitat models for wild ungulates
to consider the forage quality-quantity trade-offs (Mueller et al., 2008). A canopy quality estimation
on a large scale appears thus relevant to understand wildlife diversity. Broadband satellites such as
Landsat TM or SPOT are not spectrally detailed enough to detect or estimate the concentration of
chemical constituents. Imaging spectrometers, on the contrary, can detect and quantify canopy biochemical components by measuring canopy reflectance in narrow and contiguous spectral bands in
a wide wavelength range. The many subtle absorption features of the spectrometer data allows the
identification of a wide range of plant compounds and their concentration. The relationship between
spectral properties and foliar chemicals have been examined from dried and fresh leaves, to entire
canopies. The estimation of biochemicals of entire canopies brings along complicating factors, like
the masking effect of leaf water absorption, the complexity of the canopy architecture, variation in
leaf internal structure and directional, atmospheric and background effects. Several methods, including band ratios and difference indices, have been developed to maximize sensitivity to the vegetation
characteristics, while minimizing confounding factors.
A number of studies have shown the potential of this technique in understanding the movement and
distribution of wildlife, particularly in areas where herbivorous wildlife is known to be limited by
nutrients. In Australia, chemical constituents of leaves of four Eucalyptus species were investigated
to predict herbivory by greater gliders (Petauroides volans Kerr) and common ringtail possums (Pseudocheirus peregrinus Boddaert) (McIlwee et al., 2001). In South-Africa, the different levels of nitrogen concentration in grass was mapped with the use of imaging spectroscopy and neural networks
(Mutanga & Skidmore, 2004a). In the future, monitoring of seasonal changes in foliar nutrient concentration as well as extending the method to predict other macro nutrients and secondary compounds
in both grass and tree canopies may be possible. The major constraint is the little contribution of
foliar chemicals to the canopy optical properties. Currently, It is a prerequisite to further investigate
the spectral features of attractants and repellents of forage and their influence on faunal species distributions to successfully upscale these findings to large areas for monitoring and conserving faunal
species (Leyequien et al., 2007).
34
Chapter 6
Data and methods
1
Introduction
First the different satellite images will be discussed, followed by the tracking data and ground truth
data. Later in this chapter, the classification methods and statistical analysis are mentioned. Most of
the tasks conducted on satellite images were performed using the program Idrisi Andes and in some
cases Arview 3.1. The statistical analysis is conducted with the help of S-Plus 8.
2
2.1
Satellite images
Introduction
As different satellites have different properties, three types of satellite images will be used. Landsat7 and MODIS images are used to create a classification of the study area. The Landsat-7 images
were chosen for their high spatial resolution to select the training data. As the MODIS images have
a coarser spatial resolution, they contain more mixed pixels, which makes it more difficult to select
appropriate training data. However, MODIS images were chosen for their high temporal resolution
(every 2 days a global coverage), which makes it possible to obtain time series. These time series
make it possible to monitor the vegetation phenology over the year, so that the different vegetation
classes can easier be distinguished. Landsat and MODIS data are distributed by the Land Processes
Distributed Active Archive Center (LP DAAC), located at the U.S. Geological Survey (USGS) Earth
Resources Observation and Science (EROS) Center (lpdaac.usgs.gov). SPOT-Vegetation images were
used to find a relationship between zebras and biomass, with the Normalised Difference Vegetation
Index (NDVI) as the biomass indicator.
35
CHAPTER 6. Data and methods
2.2
Landsat
Two landsat-7 images were downloaded from USGS Global Visualisation Viewer (GloVis) (USGS,
read 12/2008). These images were acquired on February 21th, 2000, with the Enhanced Thematic
Mapper Plus (ETM+) sensor on board Landsat-7. The first image corresponds with path 168 and row
59 in World Reference System (WRS), the second image with path 168 and row 60. The WRS is a
notation system for Landsat data, which divides the world in a global grid of 233 paths by 248 rows.
It enables a user to choose a scene by specifying the path and row number (figure6.1).
Figure 6.1: WRS path/row numbering scheme (NASA, read 2009)
The reference system of the Landsat images is UTM-37N in meters. The reference datum and reference ellipsoid are WGS84. The ETM+ collects 15m resolution panchromatic data and six bands
of data in the visible, near-Infrared (NIR) and mid-Infrared (MIR) spectral regions at a resolution of
30m. The seventh, thermal band has a resolution of 60m (table 6.1).
In the first image, bands 1–5 have 8713 columns and 7573 rows, in the second image 8741 columns
and 7599 rows. For all the different bands, the two images were mosaicked together and an area was
extracted on which the classification was performed. This extracted area is smaller than the study area
as first different classification methods are tested. The image of the extracted area has 3266 columns
and 4330 rows.(figure 6.2)
36
Table 6.1: Bandwidth and resolution of the different ETM+ bands
Bandwidth
Name
Resolution
(1) 0.45 to 0.52
(2) 0.52 to 0.60
(3) 0.63 to 0.69
(4) 0.76 to 0.90
(5) 1.55 to 1.75
(6) 10.4 to 12.5
(7) 2.08 to 2.35
PAN 0.50 to 0.90
Blue-Green
Green
red
NIR
MIR
Thermal-IR
MIR
Panchromatic
30
30
30
30
30
60
30
15
Figure 6.2: Two Landsat images and the extracted area with the coordinates in UTM-37N at each corner.
37
2.3
MODIS
The MODIS images were downloaded from the NASA Warehouse Inventory Search Tool (WIST)
(NASA, read 01/2009). In the WIST tool MODIS Terra, Vegetation Indices 16-Day L3 Global 250m
SIN grid (short name: MOD13Q1) was selected. These are 16 day composites. Eighteen images were
downloaded from the year 2008, with the images from half March until the end of May missing (table
6.2).
Table 6.2: Start and end date of the 16 day periods of the different images
Image
1
2
3
4
5
6
7
8
9
Start date
End date
19 Dec 2007
17 Jan 2008
02 Feb 2008
18 Feb 2008
24 May 2008
09 Jun 2008
25 Jun 2008
11 Jul 2008
27 Jul 2008
03 Jan 2008
01 Feb 2008
17 Feb 2008
04 Mar 2008
08 Jun 2008
24 Jun 2008
10 Jul 2008
26 Jul 2008
11 Aug 2008
Image
10
11
12
13
14
15
16
17
18
Start date
End date
12 Aug 2008
28 Aug 2008
13 Sep 2008
29 Sep 2008
15 Oct 2008
31 Oct 2008
16 Nov 2008
02 Dec 2008
18 Dec 2008
27 Aug 2008
12 Sep 2008
28 Sep 2008
14 Oct 2008
30 Oct 2008
15 Nov 2008
01 Dec 2008
17 Dec 2008
02 Jan 2009
Each MOD13Q product contains 6 bands. There are four composited surface reflectance bands: red
(band 1) , NIR (band 2), blue (band 3), and MIR (band 7) (table 6.3). The other two bands are vegetation indices: the Normalized Difference Vegetation Index (NDVI) and the Enhanced Vegetation Index
(EVI). These vegetation indices give an indication of the biomass present on the ground. The EVI
minimizes canopy background variations and optimizes sensitivity in dense vegetation conditions.
It also includes the blue band to reduce atmosphere influences caused by smoke and sub-pixel thin
clouds (Huete et al., 2002).
Table 6.3: Bandwidth and spatial resolution of the different surface reflectance bands downloaded
Band
Bandwidth
Resolution
(1) red
(2) NIR
(3) blue
(7) MIR
620-670 nm
841-876 nm
459-479 nm
2105-2155 nm
250
250
500
500
The images were downloaded in HDF-EOS format with a sinusoidal projection. The coordinate system was converted to UTM-37N with reference datum WGS84, and the study area was extracted. The
38
images of the study area were 527x821 pixels. (figure6.3)
Figure 6.3: MODIS image of the study area with the coordinates in UTM-37N of the corners
2.4
SPOT-Vegetation
The SPOT-Vegetation NDVI images were delivered by VITO (Vlaamse Instelling voor Technologisch
Onderzoek, Flemish institution for technological research). There are 36 images for the years 2006
and 2007, and 34 images for the year 2008. There are 3 images per month, this is for days 1–10, days
11–20, and day 21 till the end of the month. It consists of synthesis products over 10 day periods.
These images are obtained from the compilation of daily atmospherically corrected images of ten
consecutive days taken by the SPOT-Vegetation sensor on board SPOT-5. The resulting value for each
pixel corresponds to the value of the date with maximum NDVI for that pixel, so the synthesis is thus
composed of pixels with values from different dates (SPOT-Vegetation, read 03/2009). The NDVI
values of these SPOT-Vegetation images are rescaled between 0 and 250 using a linear model with
intersect -0.08 and slope 0.004. Some additional values are assigned to the missing pixels, namely
251 to a missing pixel, 252 to a cloud pixel, 253 to a snow pixel, 254 to a sea pixel, and 255 to a
back pixel. The reference system is UTM-37S with Arc1960 the reference datum. These images were
converted to the latitude/longitude reference system and the WGS84 reference datum. The spatial
resolution is 1km and each image has 205 x 232 pixels. (figure 6.4)
39
Figure 6.4: SPOT-Vegetation image with the coordinates in Latitude/Longitude (Degrees) of the corners
3
Tracking data
The tracking data of the Grevy’s zebras were collected using GPS collars. This was part of the ’Save
the Elephants Animal Tracking Project’. The data were delivered by the Northern Rangelands Trust
(NRT). Data are available from the period June 2006 till August 2008, but the period of data collection
and the amount of data are different for each animal (figure 6.5). The reason why a collar stops
measuring locations is mostly due to an equipment failure or sometimes due to the dead of the animal.
In total, sixteen Grevy’s zebras have been collared.
40
Figure 6.5: Period of data collection for each zebra
For each zebra there is a database file containing the information collected with the GPS collar (table
6.4). From these database files, point vector files were created with the use of ArcView 3.1, and then
imported into Idrisi Andes. In figure 6.6, the vector files of four zebras are shown, with a colour palette
indicating the movement: Change in colour over time. For the background image, a SPOT-Vegetation
NDVI image was used with a greenscale as colour palette.
Table 6.4: The information contained in the database file for each fixed location
Information
Description
objectID
collarID
Fix time
Download time
Location coordinates
Height above sea level
Temperature
Number of the location fix
Number of the collar
Date when the location measurement was obtained
Date when the location measurement was downloaded from the collar
in latitude and longitude
in metres
in degrees Celsius
41
Figure 6.6: Tracking data shapefiles for 4 zebras, with a change in colour indicative of the change in location
over time
4
Vector data
Vector data on the district boundaries, major rivers, roads and towns, protected areas, water bodies
and waterpoints in Kenya were downloaded from the World Resources Institute (International Livestock Research Instistute, read 2009). Vector data indicating the livestock density from 1990 was
collected as well. NRT provided vector files containing the conservancies. An Africover land cover
map of Kenya was used for comparison with the land cover classification map produced in this study.
Africover is an initiative of the Food and Agriculture Organization of the United Nations (FAO). This
land cover map has classes based on the FAO/UNEP (United Nations Environment Program) international standard Land Cover Classification System (LCCS).
42
5
5.1
Classification
Ground truth data
Classification of satellite images requires ground truth data to assist in the interpretation of the different land cover classes in the image and for the selection of training data. Ground truth data was
provided by NRT. For each sampled vegetation point, the GPS location was measured, a vegetation
description was performed by filling in a form, and a photograph was acquired. The form is shown in
Appendix A. In figure 6.7 some examples of photographs taken by NRT are shown. Vegetation description consisted of estimating the percent cover in the herbaceous, shrub and tree layer. In the form,
the percent of trees, shrubs and herbaceous was indicated as closed (C: 70%–100% cover, crowns
overlapping, touching, or very slightly separated), open (O: 20%–70% cover, crowns not touching,
distance between crowns up to twice the average crown diameter), sparse (S: 2%–20% cover, distance
between crowns more than twice the average crown diameter), or absent (A). For shrubs, it was indicated whether the average height was more or less then half a meter. For herbaceous, the composition
was indicated as forbs (F: >75% cover of forbs), grasses (G: >75% cover of grasses), or mixed (M:
forbs cover less then 75% and grasses cover less then 75%). In total, 65 GPS locations were measured.
(a) Shrubland class
(b) Woodland more than 70% trees class
(c) Herbaceous class
Figure 6.7: Examples of photographs taken from different vegetation classes
43
5.2
Artificial Neural Networks (NN)
Artificial Neural Networks (NN) will be discussed in this section as it will be used to perform classifications. NN are based on biological nervous systems’ information processing. They are composed
of a large number of interconnected artificial neurons with several inputs and a single output. The
network consists of one input layer, some hidden layers and an output layer. It is able to process information and analyze patterns in data that are too difficult to distinguish for humans and other computer
techniques.
Figure 6.8: Artificial neural network with the three layers: input, hidden and output
When the neuron is in training mode, it is trained to associate outputs with input patterns. In the using
mode, the neuron fires, this means is activated, when it recognizes the input pattern. If not, a firing
rule is used to determine whether it should fire or not. These rules account for the high flexibility of
NN. So the network tries to identify the input pattern and match the associated output pattern with it.
When an input is not known, it is given an output of an input pattern that is least different from it.
In more sophisticated neurons, the connections between the neurons have weights so that every input
has a different effect on the output. The network is trained for a specific application by a learning
process which involves adjustments to the connections between the neurons, to the weights. The
weights are adjusted so that the error between the desired and actual output is reduced. To control this
process, the network calculates how the error changes as each weight is increased or decreased.
44
Figure 6.9: Neuron with several weighted inputs and a single output (Stergiou & Siganos, read 04/2009)
There are two learning methods. There is supervised learning where the network is given the input
and matched output, the network learns to infer the relationship between the two. If the network is
then properly trained, it has learned to model the function that relates the input variables to the output
variables. It can then be used to make predictions for inputs with unknown outputs. Unsupervised
learning is based on local information. The network organizes the data presented by itself and detects
emergent collective properties. When the network is given an input of which there is no matching
output, the network assigns the output of the input that is most closely related.
There are three categories of transfer functions. In linear units, the output activity is proportional to
the total weighted output. In threshold units, the input is multiplied with the weight, this gives the
weighted input, and if the sum of these exceeds a pre-set threshold value, the neuron is activated. In
sigmoid units, the output varies continuously but not linearly as the input changes. When the neuron
fires, the activation signal is passed through an activation function to produce the output of the neuron.
In a feed-forward NN, signals can only travel one way, from input to output. In feedback networks,
signals can travel both ways by introducing feedback loops in the network. These are very powerful
networks but they can get extremely complicated. A great advantage of NN is that users don’t need
to understand the internal mechanism of the task and they are very well suited for real time systems
because of their fast response and computational times which are due to their parallel architecture.
There are several parameters in a NN. The learning rate determines by how much the weights are
changed at each step. In the used algorithm, the learning rate is 0.01. The momentum is 0.1 and
allows the change to the weights to persist for a number of adjustment cycles. The number of cells in
the hidden layer is 10 and the maximum number of cycles the network is run is 1000. After every 5
cycles, the error on the test set is calculated. The activation function of the used network is the tangent
hyperbolic (tanh). The training fraction is 0.5, meaning that half of the data is used as training set and
half is used as test set (Statsoft, read 04/2009; Stergiou & Siganos, read 04/2009).
45
5.3
Classification methods
Several methods (Maximum Likelihood and NN with different inputs) were used to perform a classification of the study area. First training data were digitized on the Landsat images using the ground
truth data. On the ground truth form the direction in which the photographs were taken is mentioned,
so the training pixels were also selected on that side of the GPS location point. Initially, water was
also included in the trainingdata, but as the Landsat image is from the dry season, hardly no water is
visible, so water was excluded from the classification. In total more than 2000 pixels per class were
selected. On these pixels several classification methods were tested.
A supervised Maximum Likelihood classification, was performed using the signatures of the training
pixels. The signature statistics are extracted over all available bands (bands 1–5) for each informational class. This classification method is based on the probability density function associated with a
particular training site signature. Each pixel is assigned to the most likely class based on a comparison
of the probabilities that the pixel belongs to each class (Idrisi Andes Help). As no knowledge exists
about the prior probabilities with which each class can occur, equal prior probabilities are used.
As it is better to use independent validation or testpixels for the calculation of the accuracy of the
classification, classifications using half of the training pixels were performed. Half of the training
pixels were selected to calculate the signatures, the other half is being used as test pixels. The training
pixels were split using the program randompixelselection (Frieke Van Coillie).
Secondly, NN was used to make a classification. The NN program used is called pixelclass (Toon
Westra). The program needs some network parameters contained in a parameter file, the trainingpixels
and a file containing all information that can be used to base the classification on. The output is a
classification in raster file. The program selects a trainingset of pixels and a testset of pixels. The
trainingset is used to train the network and to make the classification while the testset is used to test
the classification’s accuracy. As input, all the available bands from the Landsat image were used.
The temporal information contained in the MODIS time series can contribute to a better differentiation between the different classes. The trainingpixels from the Landsat images were reused on the
MODIS images, but as the spatial resolution of the two satellites is different, the trainingpixels from
the Landsat images needed to be enlarged on the MODIS images. In total about 100 pixels per class
were selected. The maximum likelihood classification was used again, in the same manner as above
mentioned. In the NN method, several combinations of input images were evaluated. The following
combinations of input images were evaluated:
• Spectral bands from all images: to make a classification based on the difference in spectral
reflectances over the year between the vegetation classes.
46
• All NDVI images: to make a classification based on the differences in biomass over the year
between the different habitat classes.
• First three components from the Principal Component Analysis (PCA) of the NDVI images
with first three components form the PCA of the EVI images: When there are a lot of data, part
of the information is surplus as it is correlated with other variables. PCA is used to transform
data in such a way that new variables are created that are not correlated. The eigenvalues
and eigenvectors of the original covariance matrix are being calculated. Every eigenvalue and
associated eigenvector describes a principal component, with the eigenvector being the direction
of the new component and the eigenvalue a measure of the amount of information contained in
that principle component (Lillesand et al., 2004). Here, only the first three components are
used, as these already contain most of the variation in the information. By using PCA, the
variation between different pixels is maximized in the components and this may make it easier
to distinguish between the different habitat classes.
• All spectral bands from all images, together with the first three components from the PCA of
the NDVI images and the first three components from the PCA of the EVI images: combination
between the differences in spectral reflectances over the year and the maximized variations in
biomass created by PCA of NDVI and EVI.
• All spectral bands from all images, together with all NDVI images and all EVI images: all the
spectral reflectances over the year and all the biomass changes over the year described by the
vegetation indices NDVI and EVI.
5.4
Accuracy assessment
An error matrix can be calculated based on the training data, from which several accuracies can be
derived. The Kappa values already give a first indication of the accuracy of the classification result.
They indicate the amount of ’true’ agreement of the percentage of correct values in the error matrix
by taken out the percentage of correct values due to a ’chance’ agreement. It can be calculated as
follows:
kappa =
N
Pr
P
− ri=1 (xi+ ∗ x+i )
i=1 xii
P
N 2 − ri=1 (xi+ ∗ x+i )
where r=number of rows in the error
xii =number of observations in row i and column i (on the main diagonal)
xi+ =total of observations in row i
x+i =total of observations in column i
N=total number of observations included in matrix
47
The nondiagonal elements in the columns represent errors of omission, those in the rows errors of
commission. The overall accuracy is determined by the quotient of the total number of correctly
classified pixels and the total number of reference pixels. The producer’s accuracies are those resulting
from the quotient of the number of correctly classified pixels for each category and the number of
pixels of that category in the ground truth data. This measurement indicates how well the training
pixels of the given habitat type are classified. The user’s accuracies are those resulting from the
quotient of the number of correctly classified pixels in each category and the total number of pixels
that were classified in that category. This indicates the probability that a pixel classified into a given
category actually represents that category on the ground (Lillesand et al., 2004). The classification
with the highest accuracy was also compared with Africover, by means of a comparison matrix.
6
Analysis of Grevy’s zebra tracking data
First some general information was extracted from the tracking data to get a view of the animals
followed. For each animal, the home range, distance moved and number of fixes within protected
areas (PAs) was calculated. The home range was calculated using the Minimum Convex Polygon
(MCP) method. This simply draws a polygon around all the fixes and thus tends to exaggerate the
total home range area. However, MCP is still widely in use. As the variation in number of fixes and
time period of data collection between the different animals is great, this will affect the home range
size and therefore the MCPs cannot strictly be compared.
As the Grevy’s zebra is a threatened species, it is important to know how much time a zebra spends
within PAs, where they are better protected. The percentage of fixes for each zebra falling within PAs
was determined. PAs include community conservancies, National Reserves and National Parks.
In Arcview the total distance moved by each animal and the mean distance moved per day can be
calculated. As the period of data collection plays an important role in the total distance moved, this
is only calculated for interest. Contrarely, the mean distance moved per day is comparable between
animals and is indicative of how mobile each animal was.
7
7.1
Analysis of Grevy’s zebras’ migration
Introduction
The main objective of this thesis is to model the migration of Grevy’s zebras. There are many factors
influencing the movement of the animals. First vegetation biomass will be investigated. Grevy’s
zebras are herbivores so the vegetation distribution and biomass will probably play an important role
in their migration. As animals cannot survive without water, this source will also be investigated as an
influence on their behaviour. The presence of livestock will be taken into account, as livestock is an
48
important competitor for resources. Based on the obtained land cover classification map and on the
Africover map, the habitat preference of the Grevy’s zebras will be investigated.
7.2
Correlation of the zebras’ migration with biomass
The aim of this part is to seek for a relationship between the migration and the available biomass,
using NDVI as the indicator.
7.2.1
Linking NDVI and tracking datasets
The objective was to obtain a dataset containing the NDVI values for the regions where zebra presence
has been tracked. NDVI values of locations with no zebras present will also be determined as this is
necessary to make a comparison between the NDVI values of the preferred and other areas.
First the vector file containing the tracking data was split into subsets in Arcview, using a script (Toon
Westra). For each ten day period within the NDVI time series, a vector file per zebra was created with
the location points recorded during that ten day period. So for every zebra, three vectorfiles per month
were obtained: the file for day 1–10, the file for day 11–20 and the file from day 21 till the end of
the month. All of these vectorfiles were then converted to rasterfiles in Idrisi Andes, with the same
pixelsize as the SPOT-Vegetation NDVI images. Every pixels has the value of the amount of GPS
location points that it contains. This is done to make the images compatible with the program for the
extraction of the NDVI values and amount of zebra location points.
Secondly for every zebra a mask was created: per zebra all the pixels were selected where the zebra
occurred at least once during the study period. This mask will be used as the zebra’s range. In everey
ten day period, the NDVI values of all the pixels of the range are extracted. As a comparison needs to
be made between the NDVI values of the pixels where the zebra is present and pixels where the zebra
is absent at that time, both values need to be known. When all the NDVI values of the pixels from
the range are extracted for every ten day period, many NDVI values are known from pixels that are
not being used by zebras at that time. However, all the pixels within the range are accessible to the
zebras, so no NDVI values are obtained from pixels that are inaccessible and thus impossible to use
by the zebras.
For every ten day period, there is a rasterfile containing the location points for the zebra in that period
and a SPOT-Vegetation image with the NDVI values. Next, for every 10-day period, the SPOTVegetation NDVI value and the corresponding amount of zebra location points is determined for all
pixels within the zebras’ home range. As the mask contains all the pixels where the zebra is present at
least once in the entire period, a lot of pixels do not have any zebra location points in a ten day period.
This data extraction is done with the program zebra-extract (Toon Westra).
As a result, an excel file is obtained per zebra containing the period (year-month-period, for example
49
200662: year 2006, June, from day 11–20), the NDVI and the amount of zebras present per pixel. This
file will be reduced for the statistical analysis. All sixteen zebras will be merged together, per period.
For every period all the pixels with zebras present will be selected. The amount of zebra location
points for each NDVI value will be determined, and this NDVI value will then be listed that amount
of times. This is done so that for each NDVI value the preference of use will be accounted. After this
is done for all the NDVI values in that ten day period, the total amount of records is determined. Then
an equal amount of non-zebra NDVI values will be selected at random. It is better for the statistical
analysis that there is an equal amount of data in the two groups that need to be compared. At the end,
a dataset is obtained with three columns: the period, the NDVI and a last column indicating if it is a
zebra present (1) or zebra absent (0) record.
7.2.2
Statistical analysis
The statistical analysis was performed in S-Plus 8. The test variable is always the NDVI and zebra
present/absent is the grouping variable. The tests were done for several combinations of periods. First
all periods together, this is all the data of the entire study period together. To get some better idea
of the preferred NDVI values, the rainy and dry seasons were tested separately. Many statistical tests
require that the data are normal. The data are tested for normality with the Kolmogorov-Smirnov test.
In this test the null hypothesis is that the distribution is normal. Attention should be paid to the central
limit theorem. This says that a sample of more than 30 observations has an average that approaches
quite good the asymptotical normal distribution. Practically this means that a p-value of a parametric
test close to the nominal significance level should be handled with caution, in other cases, the small
deviation of normality does not affect the result. As the equity of variances is important too, the next
test consists of the Levene test. This test is a homogeneity-of-variances test that is not dependent on
the assumption that the data need to be from a normal distribution. As the data contain more than 30
observations it is allowed to test parametrically. To compare the averages of the two groups, zebra
present and zebra absent, a Student’s t-test was performed. As a control the non-parametric test was
also done, namely the Wilcoxon rank test. Almost all the t-tests were done one-sided, there was tested
whether the average NDVI of the pixels with zebras present was higher than the average NDVI of the
pixels with zebras absent. Only for the first and second rainy seasons other tests were performed. For
the first rainy season, there was tested whether the average NDVI of the pixels with zebras present was
lower than the average NDVI of the pixels with zebras absent. The test for the second rainy season
was done two-sided.
7.3
Correlation between zebra presence and water
To search for a correlation between the tracking of the zebras and the availability of water, the distance
to water is used. Two shapefiles are used as the sources for water. The shapefile of waterbodies
50
contains the lakes and permanent rivers. As water is only limited in the dry season no temporal rivers
are included in the analysis, as these are mostly non-existent during the dry season. The other shapefile
contains the waterpoints in Northern Kenya. The shapefiles with waterbodies and waterpoints in
Northern Kenya are merged together and a raster file is created giving a continuous scale of the
distance from water in every pixel. Next, the amount of zebra location points at every distance from
water is determined. This output is being redistributed in intervals of half a kilometre and put in a
graph together with the area of each distance class.
7.4
Correlation between zebra presence and livestock
To search for a correlation between zebra tracking and livestock, the shapefile containing data about
livestock density in 1990 is used. The amount of zebra location points in each livestock density class
is determined and put out graphically.
7.5
Correlation between zebra presence and towns
To search for a correlation between the tracking of the zebras and the presence of towns, the distance
to the nearest town is used as indicator. A shapefile containing all the towns in the study area was
used to create a map indicating the distance to the nearest town in kilometres. Then the amount of
zebra location points at every kilometre from the nearest town was determined. This output was put
in a graph together with the area of each distance class within the study area.
7.6
Habitat preference
To assess the habitat preference of Grevy’s zebras, the tracking data of all 16 zebras is used together
with the habitat classification of the study area and the Africover classification. The method used is
based on the article of Aebischer et al. (1993). The comparison of utilized and available habitat is
performed on two levels: home range composition versus total study area, and proportional habitat
use based on GPS locations versus home range composition. The habitat use of an animal is the
proportion of the animal’s path contained within each habitat. The tracking data approximates this
path, so the proportion of GPS locations in each habitat estimates the use of each habitat. The home
range of an animal is the area in which its path is located during a given period. The area within the
home range occupied by each habitat type can be expressed as a proportion of the total home range
area. Based on its widespread use, the home range is estimated using the Minimum Convex Polygon
(MCP) method. In Arcview the extension ’animal movement’ is used to do this. ’Extract’ in Idrisi
Andes was used to calculate the habitat composition of the total study area and of each animal’s home
range. It was also used to determine the number of GPS locations from each animal within each
51
habitat type. The percentage of each habitat type in the total study area and the MCP’s is calculated,
as is the percentage of GPS locations from each animal in each habitat type.
In the ideal case, all habitat types are available and all are used by each animal. In practice, not all
the habitats may be utilized by the animals according to the tracking data. If the habitat is not present
in the MCP, or no GPS data falls within the habitat, a percentage of zero usage for this habitat is
obtained. The zero percentage of utilized habitat implies that the use is so low that it was not detected.
As a zero numerator or denominator in log-ratio transformation is invalid, a small positive value will
be substituted, here 0.01%.
First it is checked whether the habitat selection is random or not. The available (total study area) and
utilized (MCP home range) habitat compositions are transformed to log-ratios yA and yU using the
proportion of woodland (<70% trees) as the denominator. According to the article Aebischer et al.
(1993), for any component xj of a composition, the log-ratio transformation y=ln(xi /xj ) renders
the yi linearly independent. If there is a random use of the habitat types, yU equals yA or the pair
wise differences d=yU -yA between matching log-ratios for utilized and available habitat follows a
multivariate normal distribution such that d=0. So after the log-ratio transformation, the difference
d= yA -yU is calculated. A residual matrix R2 is created, this is the matrix of raw sums of squares and
cross-products calculated from d. R1 a matrix of mean-corrected sums of squares and cross-products
is also calculated from d. This is used to calculate Λ=|R2 |/|R1 | and the quantity -N*ln(Λ) is then χ2
distributed. This gives an idea whether the habitat use is random or non-random.
When habitat use is non-random, the second step is to rank the habitat types in order of preference.
A preferred habitat type is one that is used more than expected from its availability. The concept of
preference allows the ranking of habitat types from least preferred to most preferred. This ranking
can be achieved by comparisons based on the pair wise differences d. When di >0, habitat i is used
more than expected relatively to habitat j, or habitat j is used less than expected relatively to habitat
i. When di >0 for all i, habitat j is used less than expected relatively to all other habitat types, it is
the relatively least used habitat type. So the habitat types are ranked by calculating the matrix (d1 ,
. . . ,dD ) as illustrated in table 6.5, for each zebra. The matrix columns are indexed by the habitat type
used as denominator in the log-ratio, and the rows by the numerator. This is an antisymmetric matrix,
and because of this and the independence property of log-ratios, each element is independent of the
others in the same row or column. The number of positive elements in each row ranks the habitats in
order of increasing relative use, with 0 the worst and D-1 the best. To combine all 16 zebras, the mean
and standard error of the elements at each position is calculated. The ratio mean/standard error gives
a t-value. As the non-random use was already checked, the significance level stays 5% rather than for
instance Bonferroni levels. It is important to know that the ranking of the sample of the population is
subject to error, and the pattern of t-values can be used to asses which ranks give a reliable order and
which ones are interchangeable.
52
Table 6.5: Matrix used to establish habitat rankings. The number of positive values ranks the habitats in increasing order of preference
Habitat types
(numerator)
1
2
.
.
.
D
7.7
Habitat
1
ln(xU2/xU1)-ln(xA2/xA1)
.
.
.
ln(xUD/xU1)-ln(xAD/xA1)
types
...
(denominator)
D
Positive values
(total)
...
...
.
.
.
...
ln(xU1/xUD)-ln(xA1/xAD)
r1
r2
.
.
.
rn
ln(xU2/xUD)-ln(xA2/xAD)
.
.
.
.
Integration of all factors influencing the migration
Until now several factors having an influence on the migration of Grevy’s zebra were treated as distinct features. In reality a complex interaction between all these factors and others determines the
migration pattern. The aim of this part is to search whether it is possible to predict which areas in
the study area are best suitable for Grevy’s zebras. The different factors influencing their movement
and occurence will first be investigated separately. These results will then be combined to produce a
general suitability map for Grevy’s Zebras for the entire study area.
53
Chapter 7
Results and discussion
1
Habitat classification
Several habitat classifications of the study area were performed in order to investigate the relationship
between the habitat and the zebra tracking data: which habitats do they prefer, which habitats are
being avoided. First a Landsat image from the dry season of 2000 was used as input for the habitat
classification. Next, a time series of eighteen 16-day composite MODIS images from the year 2008
were applied, as these might reveal more distinction between the different classes based on the differences in plant behaviour throughout the year. Two classification techniques were tested: the Maximum
Likelihood Classifier and Neural Networks (NN). There were six habitat classes distinguished:
• Herbaceous: cover of the herbaceous layer is more than 50% with a shrub and tree cover lower
than 50%
• Low vegetation cover: vegetation cover lower than 20%
• Shrubland: cover of shrubs more than 50%
• Woodland (<70% trees): cover of trees between 50–70%
• Woodland (>70% trees): cover of trees more than 70%
• Forests: closed tree cover, could easily be distinguished based on their spectral properties
1.1
Landsat-based habitat classification
Nine classifications were based on the Landsat image from february 2000, using the spectral bands
1–5. Six of them were performed with the Maximum Likelihood Classifier (table 7.1). The class
54
CHAPTER 7. Results and discussion
crops was excluded from the classification as there was no ground truth of cropland present lying
within the study area. The intention to include cropland in the classification was based upon the presence of cropland within the study area on the Africover classification. Classification 2–5 were done
with the class water included. Water was excluded as the image was from the dry season and not
enough trainingdata was available for this class. Only one permanent river could be distinguished.
The first classification result revealed that most of the area was classified as woodland. To give a
better idea of habitat variability, the woodland class was therefore split into two groups based on the
vegetation description in the ground truth data form. This subdivision was based on percentage of
tree cover. Woodland 1 indicates the class woodland with more than 70% tree cover (closed woodland) and woodland 2 indicates the class woodland with less than 70% tree cover (open woodland).
Classifications 7, 8 and 9 were performed using NN. For classification 7, the same training pixels
were used as in classification 6. The training pixels were adjusted between classification 7 and 8 to
try to obtain a better classification result. This was done by selecting a bigger region at every ground
truth point. In the final trainingset, 2000 pixels per class were selected. Classification 9 was based on
the same trainingpixels as classification 8, but only half of the training pixels were used to make the
classification. The other half was used as testset.
Table 7.1: Classifications made on the Landsat image
number
classification method
classes present
1
6
7
Maximum Likelihood
Maximum Likelihood
NN
8
9
NN
NN
herbaceous, low veg. cover, shrubland, woodland, forest
idem 5 minus class water and with more training pixels
herbaceous, low veg. cover, shrubland, woodland1,
forest, woodland2
idem 7
idem 7
Kappa value
39.93%
54.02%
70.51%
70.91%
63.36%
As only one image was available from the dry season, no good result was obtained. The result with the
best Kappa value, classification 8 can be seen in figure 7.1. The only habitat class that could easily be
distinguished from the others using Maximum Likelihood was the forest class. The other classes, all
subclasses of savanna, gave no good result. In Stuart et al. (2006) it is also stated that a classification
based on Landsat data using conventional Maximum Likelihood Classification is only suitable for
extracting the overall boundaries of savannas with associated vegetation types (like forests), but that
it is not able to make a reliable map of the distribution of vegetation formations within savanna areas.
NN gives better classification results, but there are still quite some misclassifications. The class best
mapped is again the forest class.
55
Figure 7.1: The Landsat classification with the best Kappa value
1.2
MODIS-based habitat classification
Habitat classifications were also performed based on MODIS time series using Maximum Likelihood
and NN classification techniques. The training set derived from the Landsat image was first used,
but was then adjusted and enlarged. This training set was used again as the Landsat image showed
a lot more details than the MODIS images and so it was easier to indicate the training sites on the
Landsat image. These training sites had to be enlarged on the MODIS images as MODIS images have
a coarser spatial resolution and a lot of the training sites from the Landsat image didn’t even cover
one MODIS pixel. There were several combinations of input images used for the classification.
1. All spectral bands from all 18 MODIS images
2. All 18 NDVI images
3. First three components from the PCA of NDVI and the PCA of EVI
4. All spectral bands of all images and the first three components from the two PCAs
5. All spectral bands from all images with all NDVI images and all EVI images
The Principal Components of the NDVI and EVI of these MODIS images were used to reduce the
amount of images for classification. As the first three components of the PCA contain most information and explain the greatest variation between areas, these were used (for loadings see table 7.2).
In the PCA of the EVI images, the first component contained 75.93% of the variation, the second
56
component 6.90% and the third component 5%. By using the first three components of the PCA of
the EVI, 87.83% of all the variation contained in the EVI images is used for classification. In the PCA
of the NDVI images, the first component contained 84.69% of the variation, the second component
5% and the third component 3.02%. By using the first three components of the PCA of the NDVI,
92.71% of all the variation contained in the NDVI images is used for classification. From table 7.2
it can be observed that all images contribute highly positive to the first Principal Component (PC1).
Each factor has a lot less contribution to PC2 and PC3 and some have positive while others have a
negative contribution.
Table 7.2: Loadings from the PCA of the EVI and NDVI images
EVI
Image
M1EVI
M2EVI
M3EVI
M4EVI
M5EVI
M6EVI
M7EVI
M8EVI
M9EVI
M10EVI
M11EVI
M12EVI
M13EVI
M14EVI
M15EVI
M16EVI
M17EVI
M18EVI
NDVI
PC1
0.883
0.870
0.881
0.862
0.903
0.922
0.928
0.877
0.837
0.816
0.852
0.813
0.848
0.843
0.859
0.845
0.897
0.922
PC2
-0.168
-0.212
-0.195
-0.038
0.052
0.137
0.168
0.348
0.478
0.513
0.461
0.495
0.341
-0.176
-0.186
-0.317
-0.126
-0.022
PC3
-0.229
-0.336
-0.339
-0.370
-0.096
-0.136
-0.133
-0.022
0.084
0.085
0.039
0.125
0.102
0.217
0.316
0.295
0.177
0.002
Image
M1NDVI
M2NDVI
M3NDVI
M4NDVI
M6NDVI
M5NDVI
M7NDVI
M8NDVI
M9NDVI
M10NDVI
M11NDVI
M12NDVI
M13NDVI
M14NDVI
M15NDVI
M16NDVI
M17NDVI
M18NDVI
PC1
0.921
0.925
0.915
0.940
0.950
0.941
0.954
0.928
0.894
0.894
0.914
0.894
0.921
0.868
0.913
0.915
0.935
0.959
PC2
-0.152
-0.188
-0.196
-0.015
0.071
-0.015
0.141
0.225
0.403
0.403
0.365
0.392
0.273
-0.238
-0.211
-0.215
-0.064
0.010
PC3
-0.193
-0.254
-0.290
-0.246
-0.125
-0.109
-0.118
-0.017
0.086
0.067
0.014
0.073
0.047
0.205
0.221
0.224
0.148
0.023
The classifications performed on the entire study area, based on MODIS images are listed in table 7.3,
as well as the Kappa values obtained with the training set used as test set. From all the classifications
conducted on the MODIS images, the first 4 classifications were performed using the original Landsat
image training sites. As already mentioned, these contained too little training pixels on the MODIS
images to obtain good results. There were high Kappa values obtained for these classifications, but
this can be explained by the fact that only a small amount of pixels from the training set were used
to test the accuracy. The training pixels were enlarged several times to obtain better classification
57
results. This was done by selecting the entire pixel instead of only a small part. The final training set
contained 200 pixels per class. From classification 24 onward the entire study area was used instead
of a smaller sample to obtain a good classification.
Classification 30 was completely the same as classification 28. The NN was ran a second time and
a different output was obtained. In the last three classifications (28, 29 and 30) an independent test
set too was used to calculate the Kappa value. This resulted in kappa values of 84.41%, 83.45% and
82.34% respectively. Therefore, the best classification result for the study area was classification 28,
obtained using NN and all spectral bands from all images, all NDVI images and all EVI images (figure
7.2). In further discussion this classification will be referred to as the ’MODIS classification’. The
error matrix using the training data as test data is given in table 7.4. The accuracies are given in table
7.5.
The classification based on the PCA did not give a better result compared to when all NDVI and all
EVI images were used, due to the fact that NN were able to process all the available information. It
was not necessary to reduce the amount of information to speed up the processing as the amount of
time needed to make a classification was limited.
Table 7.3: Classifications made on the MODIS images
Number
24
25
26
27
28
29
30
Classification method
Used images
NN
NN
NN
NN
NN
NN
NN
All spectral
All spectral + PCA
All spectral, NDVI and EVI
Trainingset used
Kappa value
S1
S1
S1
S2
S3
S4
S3
87.01%
86.59%
87.18%
87.13%
90.39%
90.61%
88.79%
Table 7.4: Error matrix of the MODIS classification with all trainingdata used as testdata
1
2
3
4
5
6 Total
ErrorC
1
2
3
4
5
6
213
11
10
6
1
12
15
130
6
1
0
2
11
9
115
7
0
1
8
4
4
403
3
4
0
0
0
1
1141
0
4
5
7
14
2
108
251
159
142
432
1147
127
Total
ErrorO
253
0.1581
154
0.1558
143
0.1958
426
0.0540
1142
0.0009
140
0.2286
2258
0.1514
0.1824
0.1901
0.0671
0.0052
0.1496
0.0655
58
Table 7.5: Accuracies obtained from the error matrix for the MODIS classification
Kappa value
Overall accuracy
producer’s accuracy
user’s accuracy
1.3
Class
Accuracy
herbaceous
low veg. cover
shrubland
woodland1
woodland2
forest
herbaceous
low veg. cover
shrubland
woodland1
woodland2
forest
84.41%
93.45%
84.19%
84.41%
80.42%
94.60%
77.14%
99.91%
84.86%
81.76%
80.99%
93.29%
85.04%
99.48%
Analysis of the result
As already mentioned, the overall Kappa of the MODIS classification obtained after calculation of the
Error Matrix using all the training data as input, is 90.39%. The Kappa obtained using an independent
test set is 84.41%. However, these Kappa values are not an ultimate indicator of a good classification
result. This high value means that the classification strategy employed works well in the training
areas. The accuracies based on training data are a bit too optimistic, especially when derived from
limited data sets (Lillesand et al., 2004). As the Kappa obtained with the independent test set is also
quite good, the classification result may be a good indicator of reality. However, the test set was rather
small because the total amount of ground truth data was small. This means that the Kappa value only
gives an indication of the classification on a small part of the study area. So an absolute decision
whether a good classification result was obtained or not is rather difficult to make as the Kappa values
have only a limited value to make a decision of accuracy.
The area of the different classes within the study area, extracted from the MODIS classification are
listed in table 7.6. The herbaceous class is the largest, followed by woodland (more than 70% trees).
Forest covers the smallest area within the study area.
59
Figure 7.2: MODIS classification: classification result with the highest accuracy
60
Table 7.6: Area of the different classes within the study area
class
area (km2 )
herbaceous
low vegetation cover
shrubland
woodland1
forest
woodland2
6863.38
3387.01
3257.84
5461.98
1586.12
2717.07
When the classification is further investigated it is clear that the result probably shows some differences with reality. This further investigation can be done by comparing the classification with the
Africover classification. Africover is only a rough classification of Africa, so there are misclassifications on Africover as well. But it can be used as an indicator to compare with the MODIS classification. The study area was extracted from Africover and reclassed into bigger groups resembling the
selected habitat types. Table 7.7 gives the reclassification scheme. The class names and class numbers
of the Africover classification can be found in Appendix B.
Table 7.7: reclassification scheme to compare Africover with the made classification
class in the made classification
class numbers of Africover
classes not able to match the MODIS classification (0)
herbaceous (1)
low vegetation cover (2)
shrubland (3)
woodland1 (4)
forest (5)
woodland2 (6)
1, 2, 20, 231 and 232
125, 126, 131, 132, 133, 162 and 163
10, 127 and 134
121, 122 and 124
114, 115, 116 and 145
112 and 113
117 and 118
In the comparison matrix (table 7.8), made with Africover and the MODIS classification, it can be
seen that a lot of pixels are classified differently. Only the elements on the major diagonal of the error
matrix are those that are classified into the same land cover categories. The calculated accuracies
can be found in table 7.9. The Africover herbaceous class is a very large class, as it includes all
classes with the main vegetation type herbeaceous. So this is probably an overestimation of the class
herbaceous, which can explain the huge amount of Africover herbaceous pixels that are classified into
other habitat groups in the MODIS classification. However, there are still a lot of pixels classified
as herbaceous that do not fall within the herbaceous class of Africover. The shrubland class on the
MODIS classification covers only a small amount of the shrubland pixels on Africover, only the low
61
vegetation cover class has less Africover shrubland pixels. Only the forest class of Africover falls
relatively well within the forest class of the MODIS classification. But here again a lot of forest pixels
on the MODIS classification are non-forest on Africover. So in general it can be stated that there is
not a good resemblance between the MODIS classification and Africover.
Table 7.8: Error matrix of the MODIS classification and Africover. In the columns are the Africover pixels and
on the rows the MODIS classification pixels.
1
2
3
4
5
6
Total
ErrorC
1
2
3
4
5
6
119422
60540
54723
86753
5535
38530
1253
64
1044
314
214
749
2388
456
1405
5167
8374
6230
2385
1145
1500
4857
4358
2158
146
32
47
469
9483
55
944
115
401
1178
429
949
126538
62352
59120
98738
28393
48671
0.0562
0.9990
0.9762
0.9508
0.6660
0.9805
Total
ErrorO
365503
0.6733
3638
0.9824
24020
0.9415
16403
0.7039
10232
0.0732
4016
0.7637
423812
0.6787
Table 7.9: Accuracies obtained from the error matrix
Kappa value
Overall accuracy
producer’s accuracy
user’s accuracy
Class
Accuracy
herbaceous
low veg. cover
shrubland
woodland1
woodland2
forest
herbaceous
low veg. cover
shrubland
woodland1
woodland2
forest
5.95%
32.13%
32.67%
1.76%
5.85%
29.61%
23.63%
92.68%
94.38%
0.10%
2.38%
4.92%
1.95%
33.40%
62
1.4
Discussion
There is great uncertainty about the accuracy of the classification and whether a good classification
result was obtained or not. There was only a small amount of ground truth data and already a great
uncertainty existed about the definition of the classes. The MODIS classification was presented to the
Northern Rangelands Trust to check with the reality and no comments were received.
Savanna ecosystems are very difficult to classify into subtypes. Even the distinction of these subtypes
on the ground can be challenging, and they have quite similar reflectance spectra (Stuart et al., 2006).
In savannas there are a lot of subtypes where the vegetation consists of various plant forms, for instance combinations of shrubs and grasses, woodland with a understorey of grasses and forbs, and
even a combination of the three main vegetation types: trees, shrubs and herbs. In this study, only six
habitat classes were distinguished so these are certainly classes composed of combinations of plant
forms. It should be better to make more distinction between all of these combination types, based
on different cover percentages, but then a lot more ground truth data should be collected. As there
were only 65 data points, there could only be a limited amount of classes, covering distinct vegetation
types. There will always be a certain amount of mixture of herbs and shrubs, but with more ground
truth data, more distinctions could be made and a better classification result could be obtained. The
method of collection is also very important to obtain accurate ground truth data. Stuart et al. (2006)
mention that it is also important to locate homogeneuos areas that are larger than the spatial resolution of the satellite images used. Then accurate ground truth data is obtained and complete pixels
can be selected as training data. For instance, when Landsat images are used, homogeneous areas of
about 30m diameter should be selected. As the MODIS images have a spatial resolution of 250 m,
the number of homogeneous pixels reduces considerably. Many pixels will contain several classes
(mixed pixels), which make the classification process more difficult. It might be possible to obtain
a more accurate classification using Landsat images when more ground truth data are collected in
homogeneous areas.
The low classification accuracy might also be partially explained by errors during ground data collection. The ground data collection included estimation of ground cover for the herbaceous, shrub
and tree layer. Human misjudgements in estimation of ground cover could have induced classification mistakes. If the ground data is collected by several persons, vegetation cover might be estimated
differently by each person. The photographs acquired for each sampled point were sometimes misleading, as some were taken in bird perspective, only showing a small piece of the area. There was
also only one photograph per GPS location, showing the vegetation in only one direction. It could
have been possible that some photos were taken on the edge of vegetation classes inducing location
points to be classified as one class while they were on the edge of different classes. On the Landsat
image the training pixels were selected at the side of the location point in which the photograph was
taken. The Landsat training pixels were enlarged on the MODIS images to cover complete pixels.
As the MODIS pixels are already mixed pixels the mistakes of taken a photograph on the edge of a
63
vegetation type and classifying the entire pixel as that class is probably neglible. In general, it would
have been better that a photograph was taken in each direction with a horizontal angle.
It could be possible to obtain better classification results by using other classification techniques.
For instance, finer resolution imagery like IKONOS (1m spatial resolution) imagery might be used.
However, the cost of acquiring the IKONOS data covering large study areas as in this thesis will be
too high. Another possibility is the combination of optical and radar satellite images. By combining
these two data types, vegetation classes could be distinguished based on their spectral differences
and texture differences (measured by radar). Haack & Bechdol (2000) investigated the use of Shuttle
Imaging Radar and optical Landsat Thematic Mapper (TM) satellite images for mapping savanna and
woodland vegetation in eastern Africa. The results indicated that there is a high potential in combining
optical and radar data for mapping the basic land cover patterns. The radar data by itself had good
classification accuracies, but the combinations of radar and optical data improved the classification
result.
As a conclusion it should be mentioned that it is extremely difficult to make a classification based on
ground truth data collected by others without the own knowledge of the study area. To obtain a good
classification of the study area more data should be collected and other classification techniques could
be applied: combination radar and optical imagery, more Landsat images of different dates . . .
2
Analysis of tracking data
In this section, some general characteristics of the movement of the Grevy’s zebras are extracted from
the tracking data. The proportion of tracking data within protected areas (PAs) is also investigated.
First the location of the tracking data within the study area was analysed (Figure 7.3 and 7.4). There
are two major hotspots for the tracked zebras within the study area, one in the Nort-Eastern part around
Laisamis and the other in the South-West from Wamba over Barsalinga till the South at Archers Post
and near Isiolo. In between these two hotspots no zebra location data points were recorded.
In figure 7.3, the distribution is shown of the zebras: Hiroya, Kobosa, Dableya, Martha, Johnna, Njeri,
Belinda, Lepere, Liz and Silurian2. In figure 7.4, the tracking data are shown of the zebras: Rose,
Petra, Jeff, Samburu, Loijuk and Samburu2. The zebras Hiroya, Kobosa and Dableya are located in
the North-Eastern part of the study area, near the town Laisamis. Liz, Petra and Lepere have smaller
home ranges located in the Western part of the study area, west of the town Wamba and north of the
town Barsalinga. North of Barsalinga part of the tracking data of Loijuk is located as well, but she also
ranges more south-east, passing East from Barsalinga till the western part of Archers Post. Belinda
and Johnna range from Archers Post till Barsalinga. The home ranges of Martha, Jeff and Rose are
situated near Archers Post. Njeri has some location point East from Barsalinga but also a smaller
amount of location points are located at the West side of the town. Silurian2’s home range is located
in the Southern part of the study area, West from Isiolo. Most location data points of Samburu are
64
located in the surrounding of Archers Post, with some data extending to the West in the direction of
Barsalinga. Samburu2 has location data in the South-Western part and in the North-Eastern part of the
study area. It seems that this zebra has two home ranges. It is very unlikely for a zebra to be located
on such a large home range without any location data inbetween the two hotspots. It seems that the
data of two zebras were accidently merged together into one dataset. For the further investigations
Samburu2 will be handled as one zebra with all the given location points. As all location points from
all zebras are always merged together for most analysis, this will not have any effect on the result.
Only for the analysis of speed and distance Samburu2 is left out.
Figure 7.3: The location within the study area of the home ranges of Belinda, Dableya, Hiroya, Johnna, Kobosa,
Lepere, Liz, Martha, Njeri and Silurian2.
65
Figure 7.4: The location within the study area of the home ranges of Jeff, Loijuk, Petra, Rose, Samburu and
Samburu2.
Secondly, some general characteristics were determined in Arcview: total distance moved, mean
distance moved between fixes, minimum speed per day, the maximum speed per day, the mean daily
speed and the Minimum Convex Polygon (MCP) area. In table 7.10 the results are shown for all
sixteen zebras. Njeri showed the highest mean movement between fixes (930m). Jeff and Silurian2
show also high mean movement rates between fixes, especially in comparison to their MCP area. This
means that these zebras do not undertake large scale movements, but move very extensively within
their home range. The average over all zebras of the mean distance between fixes is 500m. Samburu
has a negative minimum speed, which is probabely due to the fact that data of some days are missing.
Samburu disregarded, the minimum speed ranges from 5.13 m/day for Belinda till 23.09 m/day for
Njeri. This low value of 5.13 m/day can be explained by the fact that sometimes measurements of
location are limited to two observations per day. If these are recorded on a relatively short interval,
the distance travelled that day is very low. The value for Njeri of 92 km/day as maximum speed
is completely unrealistic. This is probably caused by some missing values. The maximum speed
otherwise ranges from 0.85 km/day for Silurian2 till 13.41 km/day for Samburu. The average mean
daily speed for all zebras is about 10 km/day, ranging from 15.22 km/day for Dableya till 7.37 km/day
for Loijuk and hiroya. These are realistic values as in literature the average is set between 10 and 15
km/day (Rubenstein, 1986).
66
Table 7.10: Analysis of tracking data: the number of bearings per zebra, the distances travelled, the speed
analysis and MCP areas
ZEBRA
Total distance
(km)
Mean distance
(m)
Min speed
(m/day)
Max speed
(km/day)
Mean daily
speed (km/day)
MCP Area
(km2 )
belinda
dableya
hiroya
jeff
johnna
kobosa
lepere
liz
loijuk
martha
njeri
petra
rose
samburu
silurian2
5072.45
3881.43
700.00
577.67
1963.87
827.53
1631.03
1483.11
4512.06
3029.20
757.66
1153.35
76.50
2087.19
95.81
475.35
599.54
568.64
651.26
464.71
586.07
360.77
436.46
411.35
309.10
929.64
363.95
382.50
551.87
573.68
5.13
5.76
11.00
8.06
5.96
5.50
7.38
7.02
6.17
5.58
23.09
6.30
21.90
-0.74
8.38
12.67
5.58
3.42
9.26
6.21
7.71
3.19
3.62
6.63
4.80
92.23
5.51
1.05
13.41
0.85
7.47
15.22
7.37
11.54
11.76
15.06
9.27
9.96
7.37
8.16
9.13
9.38
10.93
8.96
13.69
1508.24
2815.41
1256.38
114.34
1340.22
935.24
201.80
319.74
1607.91
297.51
1003.09
159.60
36.13
1370.43
45.53
In table 7.11 the analysis of tracking data within PAs is given. PAs include National Reserves, Forest
Reserves and community conservancies. The National Reserves located within the study area are
Shaba, Samburu, Losai and Buffalo Springs. The Forest Reserves in the study area are Matthews
Range, Ngaia, Ndotos Range and Mukogodo. The community conservancies within the study area
are Melako, Sera, Namunyak, Kalama, West Gate, Meibae, Naibunga, Lekurruki, Il Ngwesi and a
small part of Lewa (figure 7.5).
From table 7.11 it is clear that still half of the time zebras move outside of PAs. The conservancies
play an important role in the conservation, as some animals do stay within these protected areas all of
the time (Lepere, Liz and Petra). They account for 54% of the total amount of zebra location points
within protected areas. Only 4.67% of the location points of all zebras is located within National
Reserves or Forest Reserves. Five out of the sixteen collared zebras spent more than 90% of their time
within PAs; five spent between 50 and 90% of their time within PAs. The remaining six animals spent
less than 30% of their time within PAs.
67
Table 7.11: Percentage of bearings in PAs
ZEBRA
% in reserves
% in conservancies
% in PAs
belinda
dableya
hiroya
jeff
johnna
kobosa
lepere
liz
loijuk
martha
njeri
petra
rose
samburu
samburu2
silurian2
Total
0.45
3.10
4.63
36.94
5.39
26.82
0
0
6.36
0
0
0
68.66
30.66
4.12
0
4.67
21.58
0.54
2.60
52.48
0.12
2.90
100
100
92.70
82.63
95.45
100
0
37.19
52.36
0
54.01
22.03
3.61
5.28
89.41
5.51
27.25
100
100
99.06
82.63
95.45
100
68.66
67.86
56.37
0
58.57
68
Figure 7.5: The location of the protected areas within the study area.
3
Correlation between tracking data and biomass
The most important factor influencing zebra migration is probably the biomass distribution. As zebras
are herbivores, biomass is directly linked with their food resources. The proxy for biomass used is
the NDVI, which is determined from SPOT-Vegetation NDVI ten-day composites. The NDVI values
are determined for each pixel within each zebra’s range for every ten day period. The range is the
pixels at least once used by the zebra during the study period. All pixels where zebras are absent
during a certain period can be reached and used by the zebras as they do this at other times. So there
are values obtained for pixels where zebras are present and pixels where zebras are absent during that
period. The goal is to determine which areas are being used by zebras on specific times based on
NDVI values. The principal idea is to compare the NDVI values of pixels where zebras are present
and pixels where zebras are absent. The names of the ten day periods are always year/month/ten-day
period of that month, for example 200662 is the second period of June in the year 2006.
To form a general idea of the difference in NDVI values between zebra present and zebra absent pixels,
the averages for every ten day period for the zebra absent and zebra present data were calculated and
put in figure 7.6. The figure clearly shows the difference between the rainy seasons (peaks) and the
69
dry seasons. In almost all cases the average NDVI value of the pixels with zebras present is higher
than the average NDVI value of the pixels with zebras absent. Only in the first rainy season, from
November 2006 till February 2007, the average of zebra present pixels is lower. This rainy season
was a very wet one, and the values lie well above the other rainy season values.
Figure 7.6: Graph showing the average NDVI value for every image (ten day period) for the data with zebra
present and the data with zebra absent
The tests were done on several testsets: one global testset, over the entire study period and one testset
for every season. For all the testsets, normality was never present, but as the dataset is always much
bigger than 30 measurements, it is allowed to use the limit theorem, and the tests can be conducted
parametric. To compare the averages between the NDVI values in pixels with zebras present and the
NDVI values in pixels with zebras absent, two sample t-tests were conducted. When the variances are
equal, this was marked in the t-test. As a control the non-parametric test, the Wilcoxon rank test, was
also done but as could be expected this always gave the same result. S-Plus did have some difficulties
calculating the exact p-values, probably because of the size of the dataset. Almost all tests gave a
p-value of zero. To check whether this was no mistake, some tests were done in SPSS and R as well,
but these programs gave a p-value of zero too. So it can be assumed that the output of a p-value of
zero in S-Plus means an extremely small p-value and a rejection of the null hypothesis.
70
The first statistical test was performed on the entire dataset, including all the rain seasons and all
the dry seasons. The t-test was conducted one-sided, in other words it is tested that the average of
the pixels with zebra present is larger than the average of the pixels with no zebras. The result was
significant, so zebras use areas with on average larger NDVI values within their home range. The
average NDVI over the entire study period for pixels with zebras is 0.296; the average NDVI for
pixels without zebras is 0.272. The boxplot (figure 7.7) shows that there is a lot of overlap between
the two groups and that there are a lot of outliers, especially in the larger NDVI values. The fact
that the result is significant although there is only a small difference between the two averages can be
explained by the huge amount of data available. The dataset converges to infinite. The characteristics
of the two groups are given in the table 7.12. The distribution of the NDVI values used by zebras can
be seen in the histogram shown in figure 7.8. Here it can be seen that the zebras select NDVI values
between 39 (0.076) and 195 (0.7), with the core amount of date between 52 (0.128) and 143 (0.492).
The data do not follow a normal distribution but rather a right-skewed normal distribution, where the
right tail is longer and heavier than the left one.
Figure 7.7: Boxplot showing the distribution of the NDVI values for pixels with zebras (1) and for pixels
without zebras (0)
71
Table 7.12: Summary statistics for the entire dataset
Zebras absent
Zebras present
Min: 20
1st Qu.: 66
Mean: 88.38
Median: 80
3rd Qu.: 102
Max: 227
Total N: 117347
Variance: 973
Std Dev.: 31.2
SE Mean: 9.10e-002
Min: 26
1st Qu.: 73
Mean: 93.82
Median: 86
3rd Qu.: 114
Max: 220
Total N: 117444
Variance: 720
Std Dev.: 26.83
SE Mean: 7.83e-002
Figure 7.8: Histogram indicating the amount of NDVI values within each interval, for all the pixels used by
zebras over the entire study period.
To get some better idea of the use of areas with specific NDVI values, the dry and wet seasons were
tested separately. To split the dataset in subsets indicative of the seasons, the figure 7.6 was used to
have an idea of when the NDVI values increased or decreased. The seasons chosen here probably do
72
not coincide with the actual rainy and dry seasons as the reaction of biomass on the rain or drought can
sometimes be a bit delayed. It is however best to use the NDVI as the indicator to choose the seasons
rather than the actual rain pattern, as food is delivered by biomass and thus this is the indicator for the
zebra migration.
Table 7.13 shows the derived periods for the different seasons together with the NDVI averages for
the zebra present and zebra absent data. The values between brackets are the rescaled NDVI values.
For the dry seasons, all t-tests were performed one-sided, in other words it was tested that the average
NDVI of the pixels with zebras present is larger than the average NDVI of the pixels with zebras
absent. The first rainy season was tested vice versa, it was tested that the pixels with zebras present
have a smaller NDVI average than those with zebras absent. The second rainy season was tested twosided as the figure gave no clear idea about whether larger or smaller NDVI values were preferred. For
the other rainy seasons the same test was performed as for the dry seasons. Except for the second rainy
season (p-value = 0.8757), all the tests were significant. So in general, zebras choose larger NDVI
values than in the surroundings. A possible explanation for the selection of smaller NDVI values in
the first rainy season can be that the higher NDVI values after this very wet season are in regions with
more woody biomass. As zebras prefer forbs and grasses they still choose these habitats and not the
woody vegetation with the higher biomass and NDVI values. In the boxplots (Appendix C) can be
seen that the range of values is big and that there is quite some overlap between the NDVI values of
areas where zebras are present and NDVI values of areas where zebras are absent. This makes it rather
difficult to make a selection of the NDVI values chosen by zebras.
Table 7.13: Periods and results for the different seasons
Period
Starting period
Ending period
zebra absent
average
zebra present
average
dry 1
wet 1
dry 2
wet 2
dry 3
wet 3
dry 4
wet 4
dry 5
200662
2006111
200723
200743
200762
2007112
200823
200841
200861
2006103
200722
200742
200761
2007111
200822
200833
200853
200881
0.188 (67)
0.416 (124)
0.232 (78)
0.32 (100)
0.204 (71)
0.26 (85)
0.196 (69)
0.28 (90)
0.188 (67)
0.208 (72)
0.412 (123)
0.236 (79)
0.32 (100)
0.244 (81)
0.336 (104)
0.232 (78)
0.364 (111)
0.212 (73)
Derived from these results it’s difficult to predict zebra presence based on NDVI. The range has too
much overlap to select the preference NDVI of Grevy’s zebras. There should also be some knowledge
about the habitat type corresponding to the NDVI values as shown for the first rainy season where
lower NDVI values were selected by the zebras.
73
4
Correlation between tracking data and water
A distance to water map was created to test the relationship between zebra movement and availability
of water. As the distance to water is an indicator for the time they have to spent to get to water, zebras
will always have to be in areas where they can reach water in time to drink. First a map (figure 7.9)
was created indicating the distance in kilometres to the nearest water body. The available water is in
waterpoints, permanent rivers or lakes.
Figure 7.9: Map showing the distance to water for the study area
After the extraction of the distance from water for all the zebra location points the different distances
were aggregated into classes of 0.5km. This was done to obtain a more continuous graph instead
of location points every meter (figure 7.10). The area present in the different distance classes was
also calculated, so that a comparison is possible between the distributaion of the distance to water
classes that occur in the study area and the distribution of the distance to water classes preferred by
the Grevy’s zebras.
74
Figure 7.10: Graph showing the amount of zebra location points in relation to the distance to water and the
area covered by each distance class
The graph has a peak in the range 0–7km. The amount of zebras increases from the distance 0-3.5km,
and then decreases rapidly. At a distance of about 10–18km the amount of zebra location points is
almost zero. If the distance is too large, zebras will not occur as they need water to survive. In this
study, all zebras are in relative close proximity to water as they can go without water for about 2–5
days and can travel about 10–15km per day. Very close to water, the amount of zebras is lower than
in the 2.5–4.5 km distance range, probably because of the high chance of predation near waterpoints
or the interference of livestock. In comparison with the available area, the zebras show less usage of
the areas closer to water and a faster decline in usage after the peak. The peak shows more usage of
these distance classes in comparison to the available amount.
5
Correlation between tracking data and livestock
Based on the map of the livestock density in the study area (figure 7.11), the number of zebras present
in the different livestock density areas is extracted. The extracted values are aggregated in classes of
5 units per square kilometres. These values are then put in a graph (figure 7.12) using the middle
of the classes as x-value. On the map, the livestock density is expressed as Tropical Livestock Units
(TLU). This is a common unit used in the tropics, in which different kinds of livestock (cattle, small
ruminants etc) can be compared. One TLU is equal to an animal weight of 250kg. For instance one
cow equals 0.7 TLU, one camel accounts for 1.8 TLUs, and 14 goats or sheep are needed to make up
75
one TLU. Even wildlife species can be expressed as TLU. One elephant for example is equivalent to
7 TLUs, one buffalo to 2.5 TLUs and one wildebeest to 0.9 TLU (World Resources Institute et al.,
2007).
Figure 7.11: Map of the livestock density in the study area in units livestock per square kilometre
76
Figure 7.12: Graph of the amount of zebra GPS data per livestock density
In figure 7.12 it is easy to see that the amount of zebras present decreases with an increasing amount
of livestock present. This is logical as livestock is a direct competitor of food and water. Sometimes
the areas where a lot of livestock are present are sealed off from the surroundings so that zebras and
other wildlife cannot enter these areas, so they cannot be present there. An exponential curve is fit
through the data but is not a very good indicator for the smaller livestock values, where the amount of
zebra location points increases more rapidly.
6
Correlation between tracking data and towns
A map was created indicating the distance to the nearest town (figure 7.13). This was used to test
the effect of towns on the presence of Grevy’s zebras. The amount of zebra location points per
kilometre was extracted and the area covered by the different distance classes within the study area
was determined. Both the amount of zebra location points and the area were put in a graph (figure
7.14). In the graph there is a first peak at about 3km from the nearest town, then a second peak
at about 8km of the nearest town. After the second peak, the amount of zebras declines to become
approximately zero at about 33km from the nearest town. When compared to the amount of area
available in the distance classes, the zebra graph shows an earlier peak and a faster decline. Grevy’s
zebras do not occur in very close proximity to towns, but have a peak from 3–13km from the nearest
77
town. They may stay within relative close proximity of humans as these might be located on the
best grazing grounds. A lot of people are dependent upon livestock so they might live near the best
pastures. As zebras occupy the same habitat, they can be found in relative close proximity of towns.
Towns are also mostly nearby water, which is a possible explanation for the shape of the graph as well.
It seems as that other factors have much more influence on the occurence and migration of Grevy’s
zebras than the distance to towns.
Figure 7.13: Map showing the distance to the nearest town within the study area
78
Figure 7.14: Graph showing the amount of zebra location points in relation to the distance to the nearest town
and the area covered by each distance class
7
7.1
Habitat preference
Introduction
In this section the habitat preference of the zebras will be examined. This will be based on two
classifications: the MODIS classification and a reclass of Africover. The calculation of the habitat
preference is divided in different steps. First it is tested whether there is a non-random use of the
available habitats. If this is not the case, zebras use the habitats in the same amount as could be
expected from the availability of the habitats. A ranking of preferred habitats can only be made when
the habitat use is non-random. Secondly, a comparison will be made between the available habitat
and the used habitat. This can be performed on two levels, the first level is the comparison between
the amount of each habitat in the study area and the amount of each habitat within each animals’
home range. The second level comparison is that of the amount of each habitat in each animals’ home
range and the number of GPS locations recorded within each habitat. Preference ranking is performed
for each zebra separately, so for each zebra a different ranking is made. To have a general idea of the
preference of habitats for Grevy’s zebras, results from all sixteen zebras were integrated by calculating
the mean and standard error of all log-ratio differences between the available and utilised habitat.
79
7.2
Habitat preference tested on the MODIS classification
First the habitat compositions in the total study area and in each animal’s MCP were calculated, and
the percentage of GPS locations from each zebra in each habitat was determined using extract in Idrisi
Andes. The table showing these values can be found in appendix D. The missing habitat types were
treated by changing a 0% use of available habitat in a 0.01% use of that habitat. These proportions
were then transformed into log-ratios, using the proportion of woodland (<70% trees) as denominator.
The choice of the denominator is arbitrarely because it is only used to determine whether there is a
non-random use or not.
7.2.1
First level comparison: testing for non-random use
The first level comparison between the utilized and available habitat is that of home range composition
versus total study area. The difference matrix d, the difference between log-ratios of available habitat
and log-ratios of utilized habitat, was calculated. R1 , the matrix of mean-corrected sums of squares
and cross-products, and R2 , the matrix of raw sums of squares and cross-products, were extracted
from d and used to calculate Λ = |R1 |/|R2 |.
48.728 67.379 35.005 47.168 23.351
67.379 103.496 46.710 66.627 34.049
35.005 46.710 30.208 30.191 19.555
47.168 66.627 30.191 76.713 19.864
23.351 34.049 19.555 19.864 61.633
Λ = 56.247
76.809 40.680
76.809 115.322 53.827
40.680
53.827 34.491
49.507
69.561 31.957
−14.568 −13.507 −9.063
49.507 −14.568
69.561 −13.507
31.957 −9.063 77.440
8.066 8.066 252.869 So -N*ln(Λ) = -16*ln(0.1655) = 28.78 this yields a p-value of 0.00003 < 0.05 when compared to a
chi-squared distribution with 5 degrees of freedom. There can be concluded that there is a significant
non-random use of the available habitat types.
80
7.2.2
First level comparison: ranking of the habitat types in order of preference
The second part is the ranking of the habitat types in order of use, or preference. Per zebra, a matrix
was set up like the one in chapter Materials and methods, table 6.5. In table 7.14 the preference
ranking is given for the zebra Belinda as an example. This is done in the same way for the other
zebras as well. In table 7.15 habitat preference ranking for all zebras is summarized. The habitats
are ranked from 0–5, where the habitat with index 0 is the least preferred and the one with index 5
the most preferred. There are some differences in preference amongst the different zebras. This can
be explained by the fact that not every zebra is present in the same area of the study area. As this
differs, the composition of the habitats can also differ so their preference for other habitats can be
due to the fact that other habitats occur more. The habitat forest is always least preferred. The low
vegetation cover habitat is most preferred for 7 zebras, the others prefer woodland 2. In figure 7.15
the proportion of each habitat type in each zebra’s MCP is represented graphically. Herbaceous is
almost always present for about 20% of the home range. The habitat with low vegetation cover can
reach up to 40% of some zebra’s home ranges. For the zebras living in the Northern part of the study
area, woodland 2 is absent from their home ranges, in the others it can make up as much as 20%.
Table 7.14: Preference ranking of habitat types for Belinda
Belinda
Herbaceous
sparse
shrubland
woodland1
forest
woodland2
herbaceous
-0.339
0.162
0.063
-1.910
0.318
low veg. cover
shrubland
woodland1
forest
woodland2
rank
0.339
-0.162
-0.501
-0.063
-0.401
0.100
1.910
1.571
2.072
1.972
-0.318
-0.656
-0.155
-0.255
-2.227
2
1
4
3
0
5
0.501
0.401
-1.571
0.656
-0.100
-2.072
0.155
-1.972
0.255
2.227
81
Table 7.15: Preference ranking of habitat types per zebra
zebra
herbaceous
low veg. cover
shrubland
woodland1
forest
woodland2
Belinda
Dableya
Hiroya
Jeff
Johnna
Kobosa
Lepere
Liz
Loijuk
Martha
Njeri
Petra
Rose
Samburu
Samburu2
Silurian2
2
4
4
2
2
4
4
2
1
5
2
4
4
3
3
3
1
5
5
1
1
5
5
4
4
1
4
5
5
1
5
2
4
2
2
4
3
3
2
1
2
4
3
2
2
4
2
5
3
3
3
3
4
2
3
3
3
3
5
3
1
2
4
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
5
1
1
5
5
1
1
5
5
2
1
1
3
5
1
4
Figure 7.15: Percentage of habitat use based on MCP for each zebra
82
To make a preference ranking of the Grevy’s zebras as a species, all sixteen zebras were integrated.
At each position in the matrix, the mean and standard error over all 16 zebras was calculated. The
significance of the ratio was evaluated with t-values compared to t-distributions with 15 degrees of
freedom. From these t-tests the interchangeability in preference of the habitats could be determined.
Only the forest habitat was significantly less preferred in comparison to the others. For the five other
habitat types, the ranking was not significant.
7.2.3
Second level comparison: testing for non-random use
The second level comparison is that of the habitat use based on GPS locations versus home range
composition. This time, the habitat forest is left out, as this is practically absent in all zebra location
data points and very low in area in the MCPs. So the further analysis is done for the five remaining
habitats. The difference matrix d, difference between log-ratios of home range and log-ratios of
tracking data was calculated. R1 and R2 were extracted again from d and used to calculate Λ.
27.714
29.714
16.050
23.045
Λ = 37.373
37.327
23.235
30.563
29.714 16.050 23.045
47.398 19.530 23.038
19.530 15.900 20.025
23.038 20.025 35.046
37.327 23.235 30.563
53.400 25.194 28.964
25.194 21.246 25.618
28.964 25.618 40.897
The calculation of -N*ln(Λ)= -16*ln(0.6317)= 7.35 resulted in a p-value of 0.1186 when compared to
a chi-squared distribution with 4 degrees of freedom. As this is larger than 0.05 there is a significant
random use of the available habitat types, meaning that the zebras use the habitat in the same amount
as would be expected from the habitat availability. The reason for this random use of habitats can
be that the classification does not really resemble reality or it can be that Grevy’s zebras show no
significant preference for the available habitat types. As there is a random use of habitats, it has no
point to rank the habitat types in order of preference. Only the summary of the percentage of location
points per zebra in each habitat type is represented graphically in figure 7.16. For each zebra, except
Jeff, about 20% of their location point falls within herbaceous (for Njeri up to more than 60%). The
habitat class with low vegetation cover can contain more than 60% of some zebras’ location points.
Only Jeff has no location points in this class. This could be explained by the fact that Jeff is the only
male animal with and can have a territory. Male Grevy’s zebras choose territories wich are attractive
83
to females, so territories with higher amount of vegetation.
Figure 7.16: Percentage of habitat use based on tracking data for each zebra
It is not possible to draw a conclusion from this part. It is not possible to make a preference ranking
of the available habitats on the MODIS classification due to the fact that the MODIS classification
does not show a good resemblance to reality or due to the fact that the Grevy’s zebras do not show
any habitat preference. The habitat preference of the Grevy’s zebras should be further investigated on
other classifications, for instance on the Africover classification (See next subsection).
7.3
Habitat preference tested on Africover
As no habitat preference was concluded from the MODIS classification, the test was also conducted
on the Africover classification. Africover was first reclassed into larger groups so that the amount of
classes reduced. The reclassification scheme is given in the table 7.16 and the result in figure 7.17.
From this classification, the amount of location points per class and per zebra was determined as was
the area of each habitat in the study area and the different MCPs (appendix E). As there were no
location points in the classes forest, closed shrubs and crops, these classes were left out. The zero
percentages in the remaining classes were changed to 0.01%.
84
Table 7.16: Reclassification scheme for the Africover classification
class number
1
2
3
4
5
6
7
8
9
10
11
7.3.1
class name
classes from Africover
settlements
bare
water
forest
open woody
very open woody
closed shrubs
open-sparse shrubs
herbaceous and shrubs
herbaceous
crops
class 1 and 2
class 10
class20
class 112 and 113
class 114,115,116,145
class 117 and 118
class 121 and 122
class 124,125,126, 127
class 131,132,162
class 133, 134, 163
class 231, 232
First level comparison: testing for non-random use
The proportions were transformed into log-ratios, using the proportion of class herbaceous as denominator. To compare the utilized (home range) with the available (study area) habitat, the difference
matrix d was calculated and R1 and R2 were extracted. The Λ was calculated in the same way as
above and -N*ln(Λ) equaled to 36.15 with a p-value of 0.00001 when compared to a chi squared distribution with 7 degrees of freedom. So there is a significant non-random use of the available habitat
types.
7.3.2
First level comparison: ranking of the habitat types in order of preference
For each zebra a matrix like in chapter Materials and methods, table 6.5, was made and the habitat
types were ranked in order of preference. The result for all the zebras can be found in table 7.17. There
is again a difference between the different zebras. The least preferred habitat is very open woody (6),
as this habitat has 6 rank zero values and 6 rank one values. The most preferred habitat is herbaceous
and shrubs (9) with 6 rank seven values and 5 rank six values. In figure 7.18 the percentage of each
habitat in the MCP is given per zebra. The habitat classes open-sparse shrubs (8), herbaceous and
shrubs (9) and herbacous (10) are most abundant. Especially class 9, which can be found in almost
100% of the MCP of some zebras (Lepere, Petra).
85
Figure 7.17: Result from the reclass of the Africover classification into larger groups.
86
Table 7.17: Preference ranking of habitat types per zebra
zebra
1
2
3
5
6
8
9
10
Belinda
dableya
hiroya
jeff
johnna
kobosa
lepere
liz
loijuk
martha
njeri
petra
rose
samburu
samburu2
silurian2
4
4
4
5
7
4
6
6
4
5
3
6
7
7
6
6
7
7
2
2
1
2
2
2
2
3
1
5
3
3
7
4
3
1
1
1
0
0
5
4
5
2
2
4
2
0
5
3
1
3
3
6
5
3
3
3
3
4
6
2
0
1
1
1
0
0
0
0
4
1
1
1
0
1
7
1
1
4
0
2
6
2
5
3
3
5
4
5
6
7
4
3
4
5
3
5
5
5
6
7
6
6
7
7
7
6
5
7
5
6
4
7
2
6
7
4
2
7
0
0
1
0
0
0
6
2
2
0
Figure 7.18: Percentage of habitat use based on MCP for each zebra
87
7.3.3
First level comparison: integration over all sixteen zebras
At each position in the matrix, the mean and standard error over all 16 zebras was calculated. The
significance of the ratio was evaluated with t-values compared to t-distributions with 15 degrees of
freedom. Classes 1 and 9 are the most preferred. The preference between these classes is not significantly different. However, both classes have a significantly higher preference compared to all other
classes. The third preferred habitat type is class 8, which is significantly less preferred than 1 and
9 and significantly more preferred than the others. The outcome of the ranking for the other habitat
types is 5–2–10–3–6 with habitat type 5 being most preferred. These last habitat types however are
interchangeable. The relationships that are not significant according to the t-tests are: 2 versus 3, 2
versus 5, 2 versus 10, 3 versus 6, 3 versus 10, 5 versus 10, and 6 versus 10.
7.3.4
Second level comparison: testing for non-random use
Next, a comparison can be made between the GPS locations and the home range composition. Again
a difference matrix d was calculated, being the difference between the log-ratios of home range and
log-ratios of tracking data. R1 and R2 were extracted and used to calculate Λ= -16*ln|0.2246| =
23.89. This results in a p-value of 0.0012 when compared to a chi squared distribution with 7 degrees
of freedom. So there is a significant non-random use within the home range of the different habitat
types. It was however not possible to make a ranking of the habitat types per zebra as a lot of habitat
types showed an equal proportion in the MCP and in the GPS data, which resulted in a difference
value of zero.
7.3.5
Second level comparison: integration over all sixteen zebras
However, when the mean and standard error of the log-ratio differences is calculated over all 16
zebras, there were no zero values and a ranking could be made. The ranking made was (from most
preferred to least preferred): 1–10–9–6–2–3–8–5. However only a small number of relationships are
significant, namely 1 versus 2, 1 versus 5, 1 versus 6, 1 versus 8, 5 versus 9, 5 versus 10, 6 versus 8, 8
versus 9, and 8 versus 10. So it is rather difficult to make a significant ranking of the preferred habitat
types. When only the habitats 5 (open woody), 8 (open-sparse shrubs), 9 (herbaceous and shrubs) and
10 (herbaceous) are taken into account, it is possible to make a ranking. These habitats are chosen as
they compose most of the areas in the study area and in the MCPs. Habitat 5 and 8 are significantly
less preferred than habitat types 9 and 10. Habitat type 5 and 8, and habitat type 9 and 10 are not
significantly more or less preferred from each other. In figure 7.19 the percentage of GPS data in each
habitat type are shown per zebra. Habitat class 9 is the main habitat type where zebra location points
occur for almost all zebras. Only Rose has a dominant use of the class herbaceous (10). Dableya has
an almost equal amount of location points in classes 9 and 10.
88
Figure 7.19: Percentage of habitat use based on tracking data for each zebra
As a conclusion it can be stated that Grevy’s zebras prefer habitat types with herbaceous as main cover
type. This can be in mixture with shrubs as well. This outcome could be expected from literature
where the Grevy’s zebras diet is said to consist mainly of grasses and forbs, the primary components
of herbaceous habitat.
8
Integration of all factors influencing the occurrence
In this section, all the factors influencing the Grevy’s zebras’ migration, are being integrated to determine the parts within the study area that are most suitable for Grevy’s zebras. All the areas that are
not being used by the Grevy’s zebras were extracted, based upon the obtained results. Then the other
areas are divided into several preference classes based upon their distance to the nearest water point
and their NDVI value.
As could be seen in the section about habitat preference (section 7), Grevy’s zebras avoid forest
habitat. So the forest habitat areas are extracted from the MODIS classification and considered as
non-suitable Grevy’s zebra area.
Based on figure 7.10, the map of the distance to the nearest water point was divided into four classes.
A distance more than 20km was indicated as non-suitable area. The edge of 20km is rather low, as
zebras can be much further away from water, but in this study, the amount of zebra GPS points drops
to zero at a location 18km of the nearest water point. Next, a value of one was assigned to the areas
with a distance from water of 11–20km, a value of two was assigned to the distance classes 0–2km and
89
6–11km. The peak of zebra values, the areas between 2–6km of the nearest water point were given
a value of 3. So the higher the value, the more suitable for Grevy’s zebras. Water is only a limited
factor during the dry seasons, thus the effect of water is only of importance during these seasons. In
the rainy seasons, a lot more water is being available and the distribution of water has no longer an
influence on the distribution of Grevy’s zebras. However, the dry seasons are the most limited for
survival of the zebras, making it therefore important to base the indication of best suitable areas upon
these periods.
From figure 7.12 can be observed that the amount of zebras reaches an extremely low value at a
livestock density of 20TLU/km2 . This livestock density is chosen to select the areas not suitable for
the Grevy’s zebras: the areas with a livestock density above 20TLU per square kilometre.
For every season (seasons are again defined as the ones in table 7.13), the histogram is made for all
NDVI values of the locations where zebras were present during that season. These histograms can be
found in appendix F. For each season a lower and upper boundary was selected. These boundaries
were not chosen at the absolute edges as some very high or very low NDVI values whith hardly no
observations were left out. The cut-off value was different for every season, as it was dependent
on the total amount of observations. For all the dry seasons together and for all the rainy seasons
together, the average was calculated of the upper and lower boundaries and a range was obtained for
the dry and rainy seasons within which almost all observations were found. The boundaries for every
season and the overall range can be found in table 7.18. An average SPOT-Vegetation NDVI image
was created for the dry seasons. The average NDVI value for each pixel over all the ten day periods
within the five dry seasons was therefore calculated. The same was calculated for the rainy seasons
with an average SPOT-Vegetation NDVI image for the rainy seasons as a result. On these averaged
SPOT-Vegetation NDVI images, the areas are extracted that did not fall within the determined NDVI
ranges. The area that is indicated as non-used on both SPOT-Vegetation images was then taken into
account as non-suitable for Grevy’s zebras.
Table 7.18: The selected upper- and lower NDVI boundaries for each season and the extracted averages as
NDVI ranges for the dry and rainy seasons
Dry seasons
dry 1
dry 2
dry 3
dry 4
dry 5
Average
Lower boundary
Upper boundary
54
54
43
64
62
55
94
108
133
84
90
102
Rainy seasons
Lower boundary
Upper boundary
wet 1
wet 2
wet 3
wet 4
75
61
59
91
174
145
158
136
Average
72
153
It was also determined in which areas zebras occurred most based on the histograms of the NDVI
90
values in all five dry seasons and all four rainy seasons. These histograms can be found in figure 7.20.
The NDVI ranges extracted for the dry seasons was 61–88, this is the range where the intervals have
more than 10000 observations and for the rainy seasons 91–143, where the intervals have more than
7000 observations. This difference in treshold value is due to the fact that a different distribution is
observed between the dry and rainy seasons. These core areas obtained a value of 2, while the other
areas were given a value of 1. The two images, of the dry and of the rainy seasons, were multiplied.
An image was obtained which had areas with values 1, 2 and 4. This image was reclassified by replacing the value 4 with a value of 3.
(a) Histogram of all five dry seasons
(b) Histogram of all four rainy seasons
Figure 7.20: Histograms of both the dry and rainy seasons, indicating the distribution of the NDVI values of
the zebra present pixels
91
Then in a final step, all images were merged together. All areas not suitable for the Grevy’s zebras
were assembled. These areas were indicated on the image of the distance classes and the image of the
NDVI classes as being an area with value zero. These last two images were then summed up. The
final result was an image indicating areas with values between 0–5 (figure 7.21). The areas with value
5 are supposed to be most used by the Grevy’s zebras, while the areas with value zero are supposed
to be avoided. To test this result, the amount of zebra location points within each class was extracted.
The area of the different classes was also calculated. To get an idea of the usage of the areas by the
zebras, the percentage of zebra point and the percentage of the study area was calculated for each
class and the ratio determined. If the ratio is more than 1, the zebras use this class more than expected
from the availability of the class. A ratio below 1 means that the class is less used than expected from
the availability. The results can be seen in table 7.19. It can be seen that class 5, being the expected
best class is used about 2.4 times more than would be expected from its area. So this class is definitely
preferred by the zebras. The other classes are all used less than would be expected from their area.
Figure 7.21: Areas suitable for the Grevy’s zebras
92
Table 7.19: Results of the analysis of the integration map
class
zebra
area
%zebra
%area
ratio
0
1
2
3
4
5
Total
7000
4088
4692
7860
23642
71225
118507
5650
1046
1632
3385
5851
5860
23424
5.91
3.45
3.96
6.63
19.95
60.10
24.12
4.47
6.97
14.45
24.98
25.02
0.24
0.77
0.57
0.46
0.80
2.40
The integration of all these factors does not give an exclusive idea of where the Grevy’s zebras would
occur. There are besides the factors examined here also other factors influencing the occurence of
Grevy’s zebras. For instance predators have a high influence on their prey. When lions are present,
Grevy’s zebras will try to avoid these areas, sometimes by departing to other less suitable areas (Fischhoff et al., 2007). Another factor that has a high influence on zebra occurrence is the reproductive
state of the females. Lactating females have other nutritive needs than non-lactating females. They
also have to be in closer proximity to water, as they have to drink every day (Rubenstein, 1986). Competition with other ungulates can also affect Grevy’s zebras area use. For instance, plains zebras can
outnumber the Grevy’s zebras in good grazing areas, forcing the Grevy’s zebras to use less appropriate areas (Rubenstein, 2004). To integrate all the factors influencing Grevy’s zebras occurence and
migration, a lot more data should be obtained, not only about the Grevy’s zebra, but also about other
ungulates and predator species.
93
Chapter 8
Conclusion
As the Grevy’s zebra is a threatened species, it is important to know as much as possible about their
habitat use and migration pattern. This thesis had two main objectives: the creation of a habitat
classification and the analysis of the Grevy’s zebras migration. The habitat classification was based
on Landsat and MODIS images. Both Maximum Likelihood and Neural Networks were used to
conduct the classification. To analyse the migration, data obtained from the GPS-tracking of sixteen
Grevy’s zebras was used. Several factors with a possible influence on the migration were examined:
distribution of biomass, water, livestock and towns. The final step was to make an integration of all
these factors to predict the areas within the study area that are most suitable for Grevy’s zebras.
The first objective of this thesis was to make a habitat classification of the study area. The use of
Landsat satellite images was abandoned as no good result was obtained using these images. Instead
time series of MODIS images were used which enhanced the distinction between different classes
providing information on the plant phenology. The Maximum Likelihood classification method only
made a good separation of the forest class from the other habitat classes. Using the Neural Networks
classification technique, a better distinction between the different savanna sub-classes was obtained.
The best classification result was obtained with NN using all MODIS spectral images, all NDVI
images and all EVI images as input. However, there might still be some distinctions between the
classification result and reality. The reason for this is the small amount of ground truth data points and
the collection method.
The second objective was to model the migration of the Grevy’s zebras. The most important factor
influencing the migration of the Grevy’s zebras was the available biomass as food source. NDVI was
used as a proxy for available biomass. The Grevy’s zebras almost always used areas with significantly
higher NDVI values than in the surroundings. Only during the first rainy season they preferred areas
with significantly lower NDVI values and in the second rainy season there was no significant difference between the NDVI values in pixels where zebras were absent or present. The fact that in the
first rainy season areas with lower NDVI values were chosen can be explained by the very wet rainy
94
CHAPTER 8. Conclusion
season.
The other factors influencing Grevy’s zebra migration are proximity to water and livestock density.
The zebras mostly prefer areas between 0–15km of water. They are most present within the range of
2.5–4.5km from the nearest water point. Areas very close to water are less preferred as there is more
competition in these areas with other wildlife and livestock. In this study, all zebras were always in
relatively close proximity to water, as they can go without water for 2–5 days and can travel between
10–15km per day.
When comparing the tracking data and livestock density it was found that Grevy’s zebras avoid areas
with high livestock density. This can be explained by the direct competition between zebras and
livestock for water and food. The relationship of the Grevy’s zebras and the distance to the nearest
town resembles the relationship between the zebras and the distance to the nearest water point. Their
migration and occurrence is probably not very affected by the towns in the study area.
Based on the MODIS and Africover classification, a habitat preference ranking for the Grevy’s zebras
was performed. First it was tested whether there was a random use of habitat or not. In the case of
a random use, the zebras use the available habitat in proportion to the area of each habitat type. In
case of a non-random use of habitats, a ranking was made per zebra of which habitat they preferred.
Finally, the result of all sixteen zebras was integrated to obtain an overall habitat preference ranking
for all Grevy’s zebras tracked in the study area. From the preference ranking based on the MODIS
classification, it could only be concluded that Grevy’s zebras avoid forest habitat. Between the other
habitat types no significant distinction in preference could be made. A possible explanation is that the
classification does not correspond with reality very well.
From the preference ranking based on the Africover classification could be concluded that in the first
level comparison, between the composition of the study area and that of the home ranges of each
animal, there is a significant preference of the habitats settlements and shrubs & herbaceous, followed
by a preference for open-sparse shrubs. All other habitat types could not be ranked in a significant
order. For the second level comparison, this is between the home range compositions and the GPS
data, there is a significant preference of the habitat types herbaceous & shrubs, and herbaceous. The
next habitat types in the preference ranking are open woody and open-sparse shrubs. The other habitat
types could be left out as most of the MCPs were composed of these four habitat types.
Finally an integration of all the factors influencing the migration was made based on the obtained
results. The areas not suitable for Grevy’s zebras were determined. For the other areas the influence
of the distance to the nearest water point and of the NDVI was taken into account to divide these areas
into different preference classes. The result showed an 2.4 times more usage of the most suitable areas
by the Grevy’s zebras than would be expected from the area of this class. However, there are a lot
more factors influencing the occurrence and migration of the Grevy’s zebras. For instance, there is
an influence of predators, other ungulates and reproductive state of the Grevy’s zebra females. Data
about all these influences and maybe even more should be collected and taken into account to get a
95
CHAPTER 8. Conclusion
better idea of the areas preferred and used by Grevy’s zebras.
96
Chapter 9
Nederlandse samenvatting
1
Inleiding
Deze masterproef handelt over de migratie van Grevy’s zebra’s (Equus grevyi) in functie van habitat
type en vegetatie biomassa, gebruik makend van teledetectie. Aangezien de Grevy’s zebra een uiterst
bedreigde diersoort is, is het belangrijk om hun bewegingen te kennen en om zoveel mogelijk te weten
over hun gedrag. Hoe meer geweten is over hun gebruik van voedsel, water, beschutting . . . , hoe meer
inspanning kan geleverd worden om de soort te behouden. Deze masterproef heeft dan ook twee
objectieven. Ten eerste zal getracht worden een habitatclassificatie op te stellen van het studiegebied,
zodat het habitatgebruik van Grevy’s zebra’s kan onderzocht worden. Het tweede objectief is de
modellering van de migratie van Grevy’s zebra’s. Dit laatste wordt onderverdeeld in sub-objectieven.
Er worden verschillende factoren onderzocht die mogelijks een invloed hebben op de migratie zoals
biomassa, water, vee en de aanwezigheid van dorpen.
2
2.1
Literatuurstudie
Grevy’s zebra (Equus grevyi)
De Grevy’s zebra is een uiterst bedreigde diersoort die enkel nog voorkomt in het noorden van Kenia
en het oosten van Ethiopië. Het is de grootste zebra soort en kan gemakkelijk onderscheiden worden
van de andere soorten door de grote ronde oren, nauwe gelijk verdeelde strepen, een witte buik en een
bruine vlek op de neus.
De sociale structuur van de Grevy’s zebra is eveneens verschillend van de andere zebra soorten. Ze
leven in een veel opener gemeenschap, waarbij zo’n 10% van de mannetjes territoria hebben. Hun
leefgebied ligt gelokaliseerd in ariede gebieden met schaars water. Alleen lacterende vrouwtjes dienen
97
CHAPTER 9. Nederlandse samenvatting
iedere dag te drinken, de anderen kunnen 2–5 dagen zonder water. Hun verplaatsing bedraagt gemiddeld 10–15km per dag.
De kwaliteit en kwantiteit van het voedsel en de openheid van de vegetatie zijn belangrijke kenmerken
voor Grevy’s zebra’s. Ze brengen ongeveer twee derden van hun tijd door al etend. Het zijn grazers,
die ook wel eens kruiden, struiken en bomen consumeren wanneer gras schaars is. Bladeren kunnen
tot 30% van hun dieet uitmaken. Ze mijden meestal erg gesloten vegetatie, omdat de kans op een confrontatie met predatoren zoals bijvoorbeeld leeuwen er groter is. Zebra’s verkiezen ook om overdag
te drinken, omdat dan eveneens de kans lager is op een confrontatie. Er zijn echter waterplassen die
overdag afgeschermd worden voor het wild, zodat het vee er ongestoord kan grazen. Dan worden de
zebra’s gedwongen om ’s nachts te drinken wanneer het predatierisico veel groter is.
De overblijvende Grevy’s zebrapopulatie werd in 1970 geschat op 15000 individuen, recente schattingen zijn 2000 resterende individuen in Kenia en ongeveer 120–250 in Ethiopië. De eerste grote
bedreiging vormt het vee die voor competitie zorgt voor voedsel en water. Koeien kunnen onder andere
zorgen voor een degradatie van het milieu door toegenomen erosie en een fragielere vegetatie. Een
andere reden van de afname van de soort zijn stropers, maar dankzij CITES is de handel in Grevy’s
zebra producten nu verboden. In reservaten kunnen zebra’s drinken en eten in vee- en wapenvrije
zones, maar deze gebieden bedekken slechts 0.5% van hun home ranges volgens het IUCN/SSC actie
plan. De steppezebra kan ook voor competitie zorgen. Het ernstigste probleem is het habitatverlies
van de reeds gelimiteerde oppervlakte waar de Grevy’s zebra voorkomt. Er zijn gelukkig ook positieve zaken, er zijn reeds kweekprogramma’s opgestart en wetenschappers en locale gemeenschappen
werken samen om de achteruitgang van de soort te stoppen en het aantal terug op te krikken.
2.2
Studiegebied
De Republiek Kenia is gesitueerd aan de oostkust van Afrika. Kenia bestaat hoofdzakelijk uit savanne
en grasland ecosystemen (39%) en bushland en woodland ecosystemen (36%). Landbouw bedekt
19% van het land, bossen 1.7% en stedelijk gebied slechts 0.2%. Het studiegebied ligt centraal in het
land tussen 0.3◦ and 2◦ Noord en 36.99◦ en 38.1◦ Oost. Het is gelegen in 6 verschillende districten:
Laikipia, Isiolo, Samburu, Marsabit, Meru en Nyambene.
Kenia heeft een tropisch klimaat met gemiddelde jaartemperaturen rond de 22°C. De kust is warm
en vochtig, het binnenland is gematigd en het noorden en noordoosten van het land is droog. De
gemiddelde neerslag is erg laag voor een land op de evenaar, slechts een gemiddelde van 630mm per
jaar. Dit is zeer onevenredig verdeeld over het land en varieert sterk tussen de jaren. Er kunnen ook
twee regenseizoenen onderscheiden worden: de korte regens van oktober tot december en de lange
regens van maart tot juni. Kenia bestaat voor meer dan 80% uit ariede en semi-ariede gebieden.
Het studiegebied bestaat grotendeels uit savanne ecosystemen opgebouwd uit een min of meer continue kruidlaag en een discontinue struiken boomlaag. De meest voorkomende soorten in de struik98
en boomlaag zijn Acacia soorten. De afgelopen jaren is er in de semi-ariede rangelands een toegenomen
graasdruk waargenomen. Het gevolg van deze overbegrazing is een achteruitgang van de natuurlijke
graslanden. Er is een overgang vastgesteld van overblijvende planten naar eenjarigen en een vervanging van de inheemse flora door exoten. Vee kan ook een effect hebben op het vegetatiepatroon, bijvoorbeeld de verstruiking naar struwelen met hoofdzakelijk Acacia soorten. Dit is een veelvoorkomend
probleem in alle Afrikaanse savannes.
Een groot deel van het studiegebied bestaat uit conservancies, gemeenschapsgeleide initiatieven. Ze
kunnen overal voorkomen waar het land beheerd wordt volgens goede milieupraktijken. Ze dragen
bij tot de bescherming van specifieke biodiversiteit, ze zorgen voor groene corridors voor de beweging van wild of ze kunnen beschermde gebieden zijn waarin zeldzame en bedreigde diersoorten
voorkomen. De conservancies in het studiegebied worden gesteund door een lokale organisatie,
de Northern Rangelands Trust. Er wordt gezocht naar oplossingen voor lokale problemen met een
langdurige lokale oplossing. Dit leidt tot de ontwikkeling en bescherming van het aanwezige wild.
De gemeenschappen hebben reeds enkele acties ondernomen om de Grevy’s zebra’s te beschermen.
De Grevy’s zebra’s werden gevaccineerd tijdens een anthrax uitbraak, er is een Grevy’s zebra scout
programma opgestart waarin lokale mensen data verzamelen over de distributie en aantallen van de
Grevy’s zebra’s en er werd een tracking project opgezet met GPS halsbanden om de Grevy’s zebra’s
te volgen. De data hiervan werd ook voor deze masterproef aangewend.
2.3
Wildlife telemetrie
Telemetrie is de wetenschap en technologie om automatisch metingen uit te voeren en de data van op
een afstand te verzenden met behulp van draad, radio of nog andere manieren, naar ontvangststations
voor opslag en analyse. Er zijn drie belangrijke telemetrie methodes: VHF-tracking, satelliet tracking
en GPS tracking.
De VHF-tracking techniek gebruikt heel hoge frequenties, dit zijn de golflengtes tussen 1 en 10m. De
dieren dragen een zender in een halsband en met behulp van een draagbare antenne, een ontvanger en
koptelefoon is een onderzoeker in staat het dier te volgen. Uit het signaal kunnen pieken en nullen
afgeleid worden en uit deze serie kan de locatie bepaald worden. Dit wordt dan meestal bevestigd
door een visuele waarneming, omdat de locatie precisie anders erg laag is. Een ander nadeel is dat
een onderzoeker actief moet bezig zijn met het ontvangen van signalen terwijl de zender constant
signalen uitzendt. Het resultaat hiervan is een kleine steekproef met slechts een paar locaties per dag.
Het gebruik van VHF is meestal gelimiteerd tot soorten met een beperkt oppervlaktegebruik of een
beperkte beweging.
Bij de satelliet tracking techniek is er momenteel slechts 1 operationeel systeem, namelijk het VS/Frans
Argos systeem. De ontvangers bevinden zich aan boord de NOAA series van satellieten. Dit zijn
ruimtetuigen in een circulair, polaire orbit op 850km hoogte. De locatie wordt berekend aan de hand
99
van een Doppler shift in frequentie. Er kan ook extra informatie geleverd worden naast de locatie
van het dier, namelijk een hele reeks van gedrag en fysiologische karakteristieken, bijvoorbeeld de
activiteit over korte of langere periodes; aantal, duur en diepte van een duik bij mariene dieren, water
temperatuur, luchttemperatuur en barometrische druk, . . . Met deze methode is het makkelijker om
dieren te bestuderen die over een grote oppervlakte bewegen en regelmatig internationale grenzen
kruisen.
Het laatste systeem, de GPS tracking werd toegepast in dit onderzoek om de Grevy’s zebra’s te volgen. De locatie wordt bepaald door het meten van de afstand tussen satelliet en ontvanger. The positie
van de satelliet is hierbij gekend en vanuit de tijd die de radiogolven nodig hadden om tot de ontvanger te komen kan de locatie bepaald worden. GPS berekent de meest precieze locatie. Het zou
een theoretische precisie hebben van minder dan een meter. Het grote voordeel van GPS is dat het
overal kan gebruikt worden, dat er locatie metingen kunnen gebeuren tot een keer per seconde en het
werkt 24u per dag. De data bevat informatie over de eigenaar, tijd van de dag, coördinaten, de PDOP
waarde en of het signaal 2D (GPS heeft contact met 3 satellieten) of 3D (GPS heeft contact met 4 of
meer satellieten) is.
Tot mei 2000 werd de accuraatheid van GPS locaties gedegradeerd door het proces van selectieve
beschikbaarheid opzettelijk opgelegd door het Amerikaanse Ministerie van Defensie. Voor deze datum konden alleen ongecorrigeerde of nabehandelde differentiële GPS data gebruikt worden. Ongecorrigeerde GPS data hebben een locatie fout van 20–80m, nabehandelde differentiële GPS data een fout
van 4–8m. Deze nabehandeling houdt een correctie in gebaseerd op de simultane locatie meting van
de ontvanger en een referentie grondstation. Aangezien beiden dezelfde fouten registreren en de locatie van het grondstation gekend is kan de fout worden verbeterd.
Obstructies, zoals gesloten kroonlaag kunnen ervoor zorgen dat het GPS toestel niet in staat is een
locatie te berekenen. Dit kan zijn omdat er niet genoeg satellieten binnen het bereik liggen. De topografie van het terrein speelt hierin ook een belangrijke rol, heuvels kunnen bijvoorbeeld het signaal
blokkeren. Het gedrag van het dier zelf kan ook een invloed uitoefenen. Wanneer de dieren bewegen zal een lagere precisie gehaald worden dan wanneer ze stil staan. De antenna kan ook door de
stand van het dier een horizontale positie aannemen met een hogere locatie fout als gevolg in gesloten
vegetatie.
Er zijn twee soorten fouten die kunnen optreden. Er zijn ten eerste de gemiste metingen, die leiden tot
ontbrekende data. Stationaire halsbanden hebben een fix rate van 68–100% met de meeste boven de
85%. Deze gemiste locaties gebeuren echter niet random, waardoor bias hoogstwaarschijnlijk is. De
condities die dit beı̈nvloeden zijn kroonlaag type, kroonlaag bedekking, boomdensiteit, boomhoogte
en basale oppervlakte. Een heuvelachtig studiegebied kan dit alles nog eens versterken. Dus de data
kan gebiased zijn naar meer open habitat. Het tweede type fout is de locatie fout. De PDOP-waarde
is een meting van de satelliet geometrie, waarbij lagere PDOP waarden bredere satelliet spatiëring
voorstellen die de triangulatie fout kunnen minimaliseren en betere resultaten opleveren. De data
kunnen gescreend worden om de ergste fouten te verwijderen alvorens verdere berekeningen worden
100
uitgevoerd.
2.4
Tracking van wild en teledetectie
Plantendiversiteit gebaseerd op de spectrale karakteristieken van de verschillende plantensoorten of
gemeenschappen kan rechtstreeks in kaart worden gebracht. Diersoorten, die meestal mobiel zijn,
maken de zaak wat ingewikkelder. Hun diversiteit en verdeling dient meestal in kaart gebracht te
worden door gebruik te maken van benaderingen.
Landbedekking is de geobserveerde fysische beschrijving van het aardoppervlak en is het attribuut die
meestal gekarteerd wordt met behulp van teledetectie. Deze laag wordt dan meestal gecombineerd met
additionele informatie zodat habitatkaarten kunnen ontstaan. Habitatgeschiktheid is een veelgebruikte
benadering voor de modellering van soortendiversiteit en rijkdom. Dit kan bekomen worden door
satellietbeelden of luchtfoto’s, biofysische, geofysische en meteorologische data te combineren met
de kennis van habitatpreferentie en eisen van een bepaalde diersoort. Data over de verspreiding van de
soort, hun habitatgebruik of karakteristieken kunnen verzameld worden door veldonderzoek of door
het analyseren van de bewegingen van individuen die gevolgd worden via wildlife tracking. Dit kan
dan geëxtrapoleerd worden naar grotere gebieden.
Ruimtelijke heterogeniteit is een sleutelcomponent in het verklaren van soortenrijkdom. Hoe heterogener ecosystemen zijn, hoe meer niches ze bevatten en hoe meer soorten ze dus kunnen onderhouden.
De distributie van soorten wordt beı̈nvloed door ruimtelijke en temporele variatie in plantproductiviteit
en biomassa van ecosystemen. Er worden verschillende vegetatie indices gebruikt in de teledetectie
om de aanwezigheid en toestand van vegetatie te meten. De meest gebruikte is de Normalised Difference Vegetation Index (NDVI). Hoge NDVI waarden duiden op plantrijke gebieden. Wolken, water en
sneeuw hebben negatieve waarden terwijl stenen en naakte grond waarden hebben rond de nul. NDVI
wordt gebruikt om vegetatie te modelleren, primaire productie te schatten en milieuveranderingen te
detecteren. Bij deze benadering wordt het voorkomen van bepaalde diersoorten gerelateerd aan terrestrische features door middel van een ecologische, trofische link. Herbivoren worden gerelateerd
aan het voedsel dat ze consumeren.
Seizoensgebonden klimaatsveranderingen kunnen verschillen veroorzaken in platensoorten, hun groei
en vestiging. Dit leidt tot veranderingen in soortensamenstelling en distributie. Wanneer de landgebruikdata van meerdere jaren wordt geı̈ntegreerd, dan kan een visie gevormd worden over de invloed
van klimaat op de variabiliteit binnen ecosystemen. Ook doordat veel soorten mobiel zijn in de tijd,
kunnen multitemporele data een completer beeld geven van hun voorkomen en distributie.
Er zijn ook veel soorten die hun habitat selecteren op basis van structurele kenmerken in plaats van
soortensamenstelling. Structurele kenmerken kunnen ingeschat worden met gebruik van teledetectie.
Hiervoor worden actieve sensors gebruikt, namelijk LiDAR en radar. Radar gebruikt microgolf
101
energie terwijl LiDAR pulsen van laser licht gebruikt.
Habitatheterogeniteit kan tenslotte ook beschreven worden aan de hand van de chemische bestanddelen van de plant. Voedselkwaliteit is een belangrijke factor bij het aantrekken van bepaalde soorten.
Beeldvormende spectrometers kunnen biochemische componenten detecteren en kwantificeren door
het meten van de plantreflectie in de nauwe en aaneengesloten spectrale banden van een breed golflengten bereik.
3
3.1
Data en methoden
Satellietbeelden
Er werden drie soorten satellietbeelden gebruikt voor dit onderzoek. Landsat en MODIS beelden
werden gebruikt om een habitatclassificatie te maken, terwijl SPOT-Vegetation NDVI beelden gebruikt werden om de migratie van de zebra’s in functie van biomassa te analyseren.
Twee Landsat-7 beelden gemaakt met de ETM+ sensor op 21 februari 2000 werden gedownload van
de USGS Global Visualisation Viewer (GloVis). Deze beelden hebben een ruimtelijke resolutie van
30m. Voor de classificatie werden ze hoofdzakelijk gebruikt om de trainingdata op aan te duiden.
Achttien MODIS beelden van het jaar 2008 werden gedownload van de NASA Warehouse Inventory
Search Tool (WIST). Dit zijn 16 dagen composieten met een resolutie van 250m . Naast de spectrale banden rood, NIR, blauw en MIR, zijn ook twee vegetatie indices beschikbaar, namelijk NDVI
beelden en EVI beelden.
De SPOT-Vegetation NDVI beelden werden bekomen via VITO (Vlaamse Instelling voor Technologisch Onderzoek). Er zijn 36 beelden beschikbaar voor het jaar 2006 en 2007 en 34 beelden voor het
jaar 2008. Dit zijn tien-dagen composieten die bekomen werden door het compileren van dagelijks
atmosferisch gecorrigeerde beelden van tien opeenvolgende dagen. De resulterende waarde per pixel
is de maximum NDVI voor die pixel gedurende die tien dagen. De NDVI waarden werden lineair
getransformeerd naar waarden tussen 0 en 250.
3.2
Tracking data
Zestien Grevy’s zebra’s werden gevolgd via GPS-tracking. De data werd geleverd door de Northern
Rangelands Trust in Kenia. Data is beschikbaar van de periode juni 2006 tot augustus 2008, met
duidelijke verschillen in hoeveelheid data en periode van verzameling tussen de verschillende dieren.
De reden waarom een halsband stopt met data verzameling kan een apparatuurbreuk of de dood van
het dier zijn.
102
3.3
Classificatie
Het Northern Rangelands Trust zorgde eveneens voor ground truth data voor de classificatie, bestaande
uit een formulier met specificaties en een foto. Gebaseerd op deze data werden zes klassen onderscheiden: herbaceous, lage vegetatiebedekking, shrubland, woodland met meer en minder dan 70%
boombedekking en bos.
Artificiële neurale netwerken kunnen gebruikt worden om een classificatie uit te voeren. Het netwerk
wordt eerst getraind. Tijdens dit trainen leert het bepaalde input patronen te combineren met de
overeenkomstige output. Wanneer dan onbekende informatie aan het netwerk wordt voorgeschoteld,
wordt aan de hand van dezelfde regels een output gecreëerd. Aan de inputs kunnen verschillende
gewichten toegekend worden, zodat bepaalde factoren een grotere invloed uitoefenen op het uiteindelijke resultaat dan anderen.
Voor de classificatie werd gestart met Landsat beelden, waarop de training sites werden aangeduid.
Omdat dit Landsat beeld geen goed resultaat gaf, werd overgeschakeld op MODIS beelden. Door
de hogere temporele resolutie, werd getracht het onderscheid tussen de verschillende vegetatievormen te maken op hun verschillende fenologie. Er werden classificaties uitgevoerd met de Maximum
Likelihood classifier en met neurale netwerken.
3.4
Analyse van de Grevy’s zebra’s tracking data en migratie
Eerst werd gekeken naar de locatie van de verschillende zebra’s binnen het studiegebied evenals naar
hun gemiddelde snelheden en de oppervlakte van hun home range. Er werd ook gekeken naar de
hoeveelheid locaties die binnen beschermde gebieden zoals reservaten of conservancies vielen.
Aangezien er verschillende factoren zijn die de migratie van Grevy’s zebra’s beı̈nvloeden, werd
gekeken naar de afzonderlijke invloed van deze factoren en getracht deze ook gezamenlijk te integreren zodat een uitspraak kon gedaan worden over de geschikte gebieden. Een eerste belangrijke
factor is plantbiomassa. Aangezien zebra’s herbivoren zijn is er een directe link tussen biomassa en
voedsel. De NDVI werd hierbij gebruikt als indicator voor biomassa. Per zebra werd een range afgebakend als zijnde elke pixel waarin de zebra minstens eenmaal voorkomt tijdens de studieperiode.
Voor elke tien dagen periode werd voor elke pixel binnen deze range de NDVI waarde bepaald en
hoeveel zebra locatie punten er in die periode voorkwamen. Er werden dus heel veel NDVI waarden
bekomen waar op dat moment geen zebra’s voorkwamen. Dit is noodzakelijk om een vergelijking te
maken tussen de NDVI waarden van de verkozen gebieden en de andere NDVI waarden. Aan de hand
van t-testen werd gecontroleerd of er een significant verschil was tussen de beide groepen. Deze testen
werden uitgevoerd op de volledige dataset die alle regen- en alle droge seizoenen omvat en eveneens
op alle seizoenen afzonderlijk.
103
Een tweede factor die een belangrijke invloed uitoefent op de zebra migratie is de aanwezigheid van
water. Hierbij werd de afstand tot het dichtstbijzijnde waterpunt gebruikt als indicator. Een derde
factor is de aanwezigheid van vee, aangezien deze een rechtstreekse concurrent is voor voedsel en
water. Er werd ook nagezien of de aanwezigheid van dorpen een invloed heeft op de zebra’s, dit werd
uitgevoerd door de afstand tot het dichtstbijzijnde dorp te berekenen.
De habitatpreferentie van de Grevy’s zebra’s werd ook bepaald. Eerst werd getest worden of hun
habitatgebruik random is of niet. Indien hun habitatgebruik random is, gebruiken ze elke habitat
in proportie van de oppervlakte. Bij een non-random gebruik kan een preferentie rangschikking
opgesteld worden. Dit werd eerst gedaan voor elke zebra afzonderlijk. Daarna werd geı̈ntegreerd
over alle 16 zebra’s. Aan de hand van t-testen werd dan bepaald welke rangschikking significant is
of welke habitats verwisseld konden worden. Deze habitat preferentie test werd uitgevoerd op de
gemaakte classificatie en op een reclass van Africover.
Als allerlaatste werd getracht de verschillende factoren die een invloed hebben op de migratie te integreren. Voor de verschillende factoren werd gekeken welke gebieden geschikt waren voor de zebra’s
en welke niet. Al de ongeschikte gebieden werden samengebracht en voor de overgebleven gebieden
werd een indeling gemaakt op basis van de afstand tot water en de NDVI waarden. Het resultaat werd
gecontroleerd door de hoeveelheid zebra GPS-punten te bepalen in elke geschiktheidsklasse.
4
Resultaten en discussie
4.1
Classificatie
Het doel van deze habitatclassificatie was een link te onderzoeken tussen habitat en zebra-voorkomen.
Eerst werden classificaties uitgevoerd op een Landsat beeld uit het droge seizoen van 2000. Op het
Landsat beeld werden de training-data gedigitaliseerd. De klasse water werd uitgesloten omdat het
beeld van het droge seizoen was en er niet genoeg training pixels konden aangeduid worden. Er
werden classificaties uitgevoerd gebruik makende van de Maximum Likelihood classifier en met
Neurale Netwerken. Het Landsat beeld alleen gaf echter geen goed resultaat. Enkel de klasse bos
kon gemakkelijk onderscheiden worden van de rest.
Er werd overgeschakeld op het gebruik van achttien MODIS 16-dagen composiet beelden uit het jaar
2008. Het gebruik van een tijdserie maakt het mogelijk verschillende habitats te onderscheiden op
basis van hun fenologie. De meeste classificaties werden uitgevoerd met Neurale Netwerken, omdat
dit betere resultaten opleverde dan Maximum Likelihood. Er werd gebruik gemaakt van verschillende
combinaties van input beelden:
1. Alle spectrale banden van alle 18 MODIS beelden
104
2. Alle 18 NDVI beelden
3. Eerste drie componenten van de Principale componenten analyse van de NDVI en van de EVI
4. Alle spectrale banden van alle beelden en de eerste drie componenten van de twee PCAs
5. Alle spectrale banden van alle beelden met alle NDVI en alle EVI beelden
Het beste resultaat werd bekomen met NN en als input alle spectrale banden van alle beelden met alle
NDVI en alle EVI beelden. De kappa-waarde van dit resultaat bedroeg 90.39% wanneer de volledige
trainingset als testset werd gebruikt en 84.41% bij gebruik van een onafhankelijke testset. Aan de hand
van deze waarden kan geen eenduidige conclusie getrokken worden omtrent het resultaat. Door het
beperkt aantal referentiepunten geeft de kappa-waarde slechts een indicatie van het classificatieresultaat over een kleine oppervlakte van het studiegebied. Het bekomen resultaat, de MODIS classificatie
genaamd, werd ook vergeleken met Africover. Hieruit blijkt dat er heel wat verschillen zijn tussen
beide. Africover is echter slechts een grove classificatie, gemaakt op het niveau van Afrika, zodat hier
waarschijnlijk ook misclassificaties aanwezig zijn.
Het is dus heel moeilijk een uitspraak te doen over de kwaliteit van het resultaat. Een betere classificatie
zou eventueel bekomen kunnen worden door het gebruik van meer referentiedata. Eigen terreinkennis
zou hierbij zeker een pluspunt zijn. Fouten kunnen ook zijn opgetreden doordat de data hier door
verschillende personen werd verzameld. De inschatting van de kruid-, struiken boombedekking
kan verschillend zijn voor verschillende personen. Zo kan het gebeuren dat gebieden met eenzelfde
bedekking toch als verschillende habitats geclassificeerd werden.
4.2
4.2.1
Analyse van de Grevy’s zebras tracking data en migratie
Correlatie tussen tracking data en biomassa
Eerst werd de relatie onderzocht tussen de Grevy’s zebra tracking en de aanwezige biomassa aan
de hand van SPOT-Vegetation NDVI beelden. Er werd een dataset opgesteld met per datum NDVI
waarden voor alle punten waar zebra’s aanwezig zijn op dat moment en een gelijk aantal ad random
bepaalde NDVI waarden uit de overvloed aan waarden vanuit de range waar op dat moment geen
zebra GPS punt gelokaliseerd was. Er was een dataset bestaande uit alle data, dus voor alle regen- en
alle droge seizoenen en er was een dataset per seizoen. Op deze datasets werden t-testen uitgevoerd.
Er werd telkens, behalve voor de dataset van het eerste en tweede regenseizoen, getest of de gemiddelde NDVI van pixels met zebra’s aanwezig hoger was dan de gemiddelde NDVI van pixels zonder
zebra’s. Voor het eerste regenseizoen werd net het omgekeerde getest, namelijk of de gemiddelde
NDVI van pixels met zebra’s aanwezig lager was dan de gemiddelde NDVI van pixels zonder zebra’s.
Voor het tweede regenseizoen werd tweezijdig getest. De manier van testen en de afbakening van
105
de seizoenen werd bepaald uit de grafiek waarop alle gemiddeldes per tien dagen periode staan voor
alle pixels met zebra’s aanwezig en voor alle pixels zonder zebra’s. Alle testen, behalve deze voor
het tweede regenseizoen, waren significant. Dus algemeen gesteld verkiezen Grevy’s zebra’s hogere
NDVI waarden. Wanneer de boxplots bekeken werden, werd vastgesteld dat er een grote overlap is in
waarden tussen beide groepen. Dat de testen toch significant zijn kan verklaard worden door het feit
dat de dataset convergeert naar oneindig. Het is dus heel moeilijk om te beslissen welke waarden de
Grevy’s zebra’s nu juist zullen gebruiken. Het feit dat het eerste regenseizoen omgekeerd significant
is kan verklaard worden door het erg natte regenseizoen. Hierdoor komen de hogere NDVI waarden
waarschijnlijk overeen met houtige gewassen die minder verkozen worden als voedselbron.
4.2.2
Correlatie tussen tracking data en aanwezigheid van water, vee en dorpen
Wanneer de afstand tot water werd vergeleken met de aanwezigheid van de Grevy’s zebra’s, kon
besloten worden dat de zebra’s zich hoofdzakelijk bevinden tussen 0–10km afstand van het dichtstbijzijnde waterpunt. Vanaf een afstand van 18km valt het aantal aanwezige zebra’s bijna op nul. Het
aantal zebra’s neemt toe tussen 0 en 3.5km om daarna snel af te nemen. In deze studie bevonden de
Grevy’s zebra’s zich relatief dicht bij water aangezien ze gemakkelijk 2–5 dagen zonder water kunnen
en gemiddeld 10–15km per dag kunnen afleggen.
Bij een toename van de vee dichtheid neemt de hoeveelheid zebra’s sterk af. Dit kan verklaard worden
door het feit dat vee rechtstreeks in competitie treedt met de zebra’s voor voedsel en water.
De relatie tussen de aanwezigheid van Grevy’s zebra’s en dorpen was gelijkaardig aan de relatie
met water. Er is dus geen uitgesproken effect van de dorpen op de zebra’s, andere factoren zullen
waarschijnlijk belangrijker zijn in het bepalen van de migratie.
4.2.3
Habitatpreferentie
Er werd ook getest of de Grevy’s zebra’s een uitgesproken habitatpreferentie vertonen. Dit werd getest
op de MODIS classificatie en op Africover. Er werd een preferentie volgorde opgesteld van de verschillende habitats per zebra wanneer het habitatgebruik non-random was. Er werd ook geı̈ntegreerd
over de verschillende zebra’s zodat een algemeen besluit kon getrokken worden voor alle Grevy’s
zebra’s in het studie gebied. Indien er een random habitatgebruik is, gebruiken de zebra’s de habitats
in proportie tot hun oppervlakte. De habitatpreferentie werd getest op twee verschillende niveaus. De
vergelijking op het eerste niveau gebeurde tussen de samenstelling van het studiegebied en de samenstelling van de verschillende home ranges. De vergelijking op het tweede niveau was dan tussen
de samenstelling van de home ranges en de verdeling van de GPS metingen over de verschillende
habitats.
Uit de MODIS classificatie kon geen significante habitat preferentie besloten worden. Er werd alleen
aangetoond dat de Grevy’s zebra’s boshabitat significant minder gebruiken dan de andere habitat106
vormen. Uit de Africover classificatie kon na integratie over alle zebra’s op het eerste niveau besloten
worden dat de klassen dorpen en kruiden met struiken significant meest geprefereerd werden in de
home ranges. Daarna werd de klasse open-schaarse struiken verkozen. Tussen de andere habitats kon
geen significante volgorde opgesteld worden. Wanneer naar het tweede niveau werd gekeken bleven
er vier habitats over, degene die het grootste deel van de home ranges uitmaakten. Er kon besloten
worden dat de Grevy’s zebra’s de klassen kruiden met struiken en kruiden meest prefereerden, boven
de klassen open houtig en open-schaarse struiken. Onderling zijn deze twee klasses telkens uitwisselbaar.
4.2.4
Integratie van alle factoren
Door de verschillende factoren te combineren werd een kaartje gecreëerd waarop alle gebieden aangeduid staan die volgens de bekomen resultaten minder geschikt zijn voor de Grevy’s zebra’s en welke
gebieden juist heel geschikt zijn. Wanneer dit resultaat werd vergeleken met de locatie van de GPS
punten bleek dat de beste klasse 2.4 keer meer data punten bevatte dan van de oppervlakte zou
verwacht worden. Dit gebied wordt dus wel degelijk geprefereerd. De andere gebieden werden allemaal minder gebruikt dan van de oppervlakte zou verwacht worden.
Om echter een volledige uitspraak te kunnen doen over de geschikte gebieden voor de Grevy’s zebra’s
dienen veel meer factoren gekend te zijn. De aanwezigheid van predatoren heeft eveneens een invloed
op het voorkomen van de zebra’s. Andere factoren die mogelijks een invloed hebben zijn competitie
met andere grote grazers zoals bijvoorbeeld de steppezebra en ook de voorplantingstoestand van de
vrouwelijke Grevy’s zebra’s speelt een belangrijke rol. Lacterende vrouwtjes hebben andere voedselbehoeftes dan niet-lacterende wijfjes en ze gebruiken dan ook andere gebieden. Er dient dus nog heel
wat onderzoek te gebeuren om een echte voorspelling te maken van het voorkomen en de migratie van
de Grevy’s zebra’s.
107
Reference
Adrados, C., Girard, I., Gendner, J.-P. & Janeau, G. (2002). Global positioning system (GPS) location accuracy improvement due to Selective Availability removal. Comptes Rendus Biologies,
325(2):165–170.
Aebischer, N. J., Robertson, P. A. & Kenward, R. E. (1993). Compositional analysis of habitat use
from animal radio-tracking data. Ecology, 74(5):1313–1325.
Aerts, G., MacKenzie, M., McConnel, B., Fedak, M. & Matthiopoulos, J. (2008). Estimating spaceuse and habitat preference from wildlife telemetry data. Ecography, 31(1):140–160.
African Wildlife Foundation (read 07/2008). Grevy’s zebra. http://www.awf.org/content/
wildlife/detail/grevyszebra.
Akpo, L.-E. (1997). Phenological interactions between tree and understory herbaceous vegetation of
a sahelian semi-arid savanna. Plant Ecology, 131(2):241–248.
ARKive, Images of life on Earth (read 07/2008). Grevy’s zebra (Equus grevyi). http://www.
arkive.org/species/GES/mammals/Equus_grevyi/more_info.html.
Augustine, D. J. & McNaughton, S. J. (2004). Regulation of shrub dynamics by native browsing
ungulates on East African rangeland. Journal of Applied Ecology, 41(1):45–58.
Blake, S., Douglas-Hamilton, I. & Karesh, W. (2001). GPS telemetry of forest elephants in central
africa: results of a preliminary study. African Journal of Ecology, 39(2):178–186.
Bradley, B. A. & Fleishman, E. (2008). Can remote sensing of land cover improve species distribution
modelling. Journal of Biogeography, 35(7):1158–1159.
Cagnacci, F. & Urbano, F. (2008). Managing wildlife: a spatial information system for GPS collars
data. Environmental Modelling & Software, 23(7):957–959.
Carlens, H., Lydersen, C., Kraft, B. A. & Kovacs, K. M. (2006). Spring haul-out behavior of ringed
seals (Pusa hispida) in Kongsfjorden, Svalbard. Marine Mammal Science, 22(2):379–393.
108
CHAPTER 9. Reference
Coelho, C., Bandeira de Melo, L., Sàbato, M., Rizel, D. & Young, R. (2007). A note on the use of
GPS collars to monitor wild maned wolves (Chrysocyon brachyurus). Applied Animal behaviour
Science, 105(1–3):259–264.
conservancies.co.za (read 11/2008). What is a conservancy? http://www.conservancies.
co.za/.
Cornelius, R. & Schultka, W. (1997). Vegetation structure of a heavily grazed range in northern
Kenya: ground vegetation. Journal of Arid Environments, 36(3):459–474.
Culik, B., Luna-Jorquera, G., Oyarzo, H. & Correa, H. (1998). Humboldt penguins monitored via
VHF telemetry. Marine Ecology-Progress Series, 162:279–286.
de Bie, S., Ketner, P., Paasse, M. & Geerling, C. (1998). Woody plant phenology in the West Africa
savanna. Journal of Biogeography, 25(5):883–900.
D’Eon, R. G. & Delparte, D. (2005). Effects of radio-collar position and orientation on GPS radiocollar performance, and the implications of PDOP in data screening. Journal of Applied Ecology,
42(2):383–388.
Dictionary (read 04/2009). Dictionary. http://dictionary.reference.com/.
Douglas-Hamilton, I., Krink, T. & Vollrath, F. (2005). Movements and corridors of African elephants
in relation to protected areas. Naturwissenschaften, 92(4):158–163.
Fischhoff, I. R., Sundaresan, S. R., Cordingley, J. & Rubenstein, D. (2007). Habitat use and movements of plains zebra Equus burchelli in response to predation danger from lions. Behavioral
Ecology, 18(4):725–729.
Frair, J. L., Nielsen, S. E., Merrill, E. H., Lele, S. R., Boyce, M. S., Munro, R. H., Stenhouse, G. B.
& Beyer, Hawthorne, L. (2004). Removing GPS collar bias in habitat selection studies. Journal of
Applied Ecology, 41(2):201–212.
Gao, X., Huete, A. R., Ni, W. & Miura, T. (2000). Optical-biophysical relationships of vegetation
spectra without background contamination. Remote Sensing of Environment, 74(3):609–620.
Gardner, D. R. (read 08/2008). Photo grevy’s zebra. www.pbase.com/rgardner/image/
59846454.
Gillespie, T. W. (2001). Remote sensing of animals. Progress in Physical Geography, 25(3):355–362.
Ginsberg, J. & Rubenstein, D. (1990). Sperm competition and variation in zebra mating behavior.
Behavioral Ecology and Sociobiology, 26(6):427–434.
Government of Kenya (read 11/2008). Kenya. http://www.kenya.go.ke.
109
Grevy’s zebra Trust (read 08/2008). Status. http://grevyszebratrust.org/status.
html.
Grzimek, B. (1972). HET LEVEN DER DIEREN. deel Zoogdieren 3. Spectrum.
Guilford, T., Meade, J., Freeman, R., Biro, D., Evans, T., Bonadonna, F., Boyle, D., S., R. & Perrins, C. (2008). GPS tracking of the foraging movements of Manx Shearwaters Puffinus puffinus
breeding on Skomer Island, Wales. Ibis, 150(3):462–473.
Haack, B. & Bechdol, M. (2000). Integrating multisensor data and RADAR texture measures for land
cover mapping. Computers & Geosciences, 26(4):411–421.
Hays, G. C., Hobson, V. J., Metcalfe, J. D., Righton, D. & Sims, D. W. (2006). Flexible foraging
movements of Leatherback Turtles across the North Atlantic Ocean. Ecology, 87(10):2647–2656.
Hinsley, S., Hill, R., Gaveau, D. & Bellamy, P. (2002). Quantifying woodland structure and habitat
quality for birds using airborne laser scanning. Functional Ecology, 16(6):851–857.
Huete, A., Didan, K., Miura, T., Rodrigues, E., Gao, X. & Ferreira, L. (2002). Overview of the
radiometric and biophysical performance of the modis vegetation indices. Remote Sensing of Environment, 83(1–2):195–213.
Hulbert, I. & French, J. (2001). The accuracy of GPS for wildlife telemetry and habitat mapping.
Journal of Applied Ecology, 38(4):869–878.
Imhoff, M. L., Sisk, Thomas, D., Milne, A., Morgan, G. & Orr, T. (1997). Remotely sensed indicators
of habitat heterogeneity: use of synthetic aperture radar in mapping vegetation structure and bird
habitat. Remote Sensing of Environment, 60(3):217–227.
International Livestock Research Instistute (read 2009). Gis database. http://www.ilri.org/
gis/.
Ito, T., Miura, N., Lhagvasuren, B., Enkhbileg, D., Takatsuki, S., Tsunekawa, A. & Jiang, Z. (2006).
Satellite tracking of Mongolian gazelles (Procapra gutterosa) and habitat shifts in their seasonal
ranges. Journal of Zoology, 269(3):291–298.
IUCN/SSC (read 07/2008).
The IUCN red list of Threatened species).
iucnredlist.org/search/details.php/7950/all.
http://www.
Javed, S., Higuchi, H., Nagendran, M. & Takekawa, J. Y. (2003). Satellite telemetry and wildlife
studies in India: advantages, options and challenges. Current Science, 85(10):1439–1443.
Jiang, Z., Sugita, M., Kitahara, M., Takatsuki, S., Goto, T. & Yoshida, Y. (2007). Effects of habitat
feature, antenna position, movement, and fix interval on GPS radio collar performance in Mount
Fuji, central Japan. Ecological Research, 23(3):581–588.
110
Johnson, C. J., Heard, D. C. & Parker, K. L. (2002). Expectations and realities of GPS animal location
collars: results of three years in the field. Wildlife Biology, 8(2):153–159.
Kenya-space (read 11/2008).
kenyaprovinces.htm.
Kenya provinces.
http://www.kenyaspace.com/
Keya, G. A. (1997). Environmental triggers of germination and phenological events in an arid savannah region of northern Kenya. Journal of Arid Environments, 37(1):91–106.
King, E. (2008). Facilitative effects of Aloe secundiflora shrubs in degraded semi-arid rangelands in
Kenya. Journal of Arid Environments, 72(4):358–369.
Lagerquist, B. A., Mate, B. R., Ortega-Ortiz, J. G., Winsor, M. & Urban-Ramires, J. (2008). Migratory
movements and surfacing rates of humpback whales (Megaptera novaeangliae) satellite tagged at
Socorro Island, Mexico. Marine Mammal Science, 24(4):815–830.
Laundré, J. (2008). Summer predation rates on ungulate prey by a large keystone predator: how many
ungulates does a large predator kill? Journal of Zoology, 275(4):341–348.
Lewis, J. S., Rachlow, J. L., Garton, E. O. & Vierling, L. A. (2007). Effects of habitat on GPS
collar performance: using data screening to reduce location error. Journal of Applied Ecology,
44(3):663–671.
Leyequien, E., Verrelst, J., Slot, M., Schaepman-Stub, G., Heitkönig, I. M. & Skidmore, A. (2007).
Capturing the fugitive: applying remote sensing to terrestrial animal distribution and diversity.
International Journal of Applied Earth Observation and Geoinformation, 9(1):1–20.
Licoppe, A. M. (2006). The diurnal habitat used by red deer (Cervus elaphus l.) in the Haute Ardenne.
European Journal of Wildlife Research, 52(3):164–170.
Lillesand, T. M., Kiefer, R. W. & Chipman, J. W. (2004). Remote sensing and image interpretation.
John Wiley and Sons, Inc.
McDermid, G. J., Franklin, S. E. & LeDrew, E. F. (2005). Remote sensing for large-area habitat
mapping. Progress in Physical Geography, 29(4):449–474.
McIlwee, A. M., Lawler, I. R. & Cork, S. J. (2001). Coping with chemical complexity in mammalplant interactions: near-infrared spectroscopy as a predictor of Eucalyptus foliar nutrients and of
the feeding rates of folivorous marsupials. Oecologia, 128(4):539–548.
Ministry of state for the Development of Northern Kenya and other arid lands (read 11/2008). Arid
Lands Resource Management Project II: Project Districts. http://www.aridland.go.ke.
Mueller, T., Olson, K. A., Fuller, T. K., Schaller, G. B., Murray, M. G. & Leimgruber, P. (2008). In
search of forage: predicting dynamic habitats of Mongolian gazelles using satellite-based estimates
of vegetation productivity. Journal of Applied Ecology, 45(2):649–658.
111
Mutanga, O. & Skidmore, A. (2004a). Integrating imaging spectroscopy and neural networks to map
grass quality in the Kruger National Park, South Africa. Remote Sensing of Environment, 90:104–
115.
Mutanga, O. & Skidmore, A. (2004b). Narrow band vegetation indices overcome the saturation problem in biomass estimation. International Journal of Remote Sensing, 25(19):3999–4014.
NASA (read 01/2009).
imswelcome/.
WIST.
https://wist.echo.nasa.gov/˜wist/api/
NASA (read 2009). Landsat 7 science data users handbook. http://landsathandbook.
gsfc.nasa.gov/handbook/handbook_htmls/chapter5/chapter5.html.
Nelson, R., Keller, C. & Ratnaswamy, M. (2005). Locating and estimating the extent of Delmarva fox
squirrel habitat using an airborne LiDAR profiler. Remote Sensing of Environment, 96(3–4):292–
301.
NRT (read 11/2008). http://www.nrt-kenya.org/.
Olsson, M. P., Widén, P. & Larkin, J. L. (2008). Effectiveness of a highway overpass to promote
landscape connectivity and movement of moose and roe deer in Sweden. Landscape and Urban
Planning, 85(2):133–139.
Pettorelli, N., Gaillard, J.-M., Mysterud, A., Duncan, P., Stenseth, N. C., Delorme, D., Van Laere, G.,
Toı̈go, C. & Klein, F. (2006). Using a proxy of plant productivity (NDVI) to find key periods for
animal performance: the case of roe deer. OIKOS, 112(3):565–572.
Priede, I. G. & Swift, S. M. (1992). Wildlife Telemetry: remote monitoring and tracking of animals.
Ellis Horwood.
Rubenstein, D. I. (1986). Ecological Aspects of Social Evolution. Princeton University Press, Princeton, NJ.
Rubenstein, D. I. (1994). Animal Societies: Individuals, Interactions, and Organization. Kyoto
University Press.
Rubenstein, D. I. (2004). Zebra sociality: different stripes for different types. Balliol College Record,
pp. 18–23.
Ryan, P., Petersen, S., Peters, G. & Grémillet, D. (2004). GPS tracking a marine predator: the effects
of precision, resolution and sampling rate on foraging tracks of African penguins. Marine Biology,
145(2):215–223.
Saint Louis Zoo (read 07/2008). Grevy’s zebra. http://www.stlzoo.org/animals/
abouttheanimals/mammals/hoofedmammals/grevyszebra.htm.
112
Sàbato, M. A. L., de Melo, L. F. B., Vaz Magni, E. M., Young, R. J. & Coelho, C. M. (2006). A
note on the effect of the full moon on the activity of wild maned wolves, Chrysocyon brachyurus.
Behavioural Processes, 73(2):228–230.
Schofield, G., Bishop, C. M., Katselidis, K. A., Dimopoulos, P., Pantis, J. D. & Hays, G. C. (2009).
Microhabitat selection by sea turtles in a dynamic thermal marine environment. Journal of Animal
Ecology, 78(1):14–21.
Schultka, W. & Cornelius, R. (1997). Vegetation structure of a heavily grazed range in northern
Kenya: tree and shrub canopy. Journal of Arid Environments, 36(2):291–306.
Seaworld Adventure Parks (read 07/2008).
Grevy’s zebra.
http://www.seaworld.
org/Animal-info/animal-bytes/animalia/eumetazoa/coelomates/
deuterostomes/chordata/craniata/mammalia/perissodactyla/
grevys-zebra.htm.
Smithsonian National Zoological Park (read 07/2008). African Savanna Grevy’s zebra. http://
nationalzoo.si.edu/Animals/AfricanSavanna/fact-gzebra.cfm.
SPOT-Vegetation (read 03/2009).
spot-vegetation.com/.
SPOT 5 Vegetation Programme.
http://www.
Sprague, D. S., Kabaya, H. & Hagihara, K. (2004). Field testing a global positioning system (GPS)
collar on a Japanese monkey: reliability of automatic GPS positioning in a Japanese forest. Primates, 45(2):151–154.
Statsoft (read 04/2009).
NN.
http://www.statsoft.com/textbook/stneunet.
html;http://www.doc.ic.ac.uk/˜nd/surprise_96/journal/vol4/cs11/
report.html.
Stergiou, C. & Siganos, D. (read 04/2009). Neural Networks. http://www.doc.ic.ac.uk/
˜nd/surprise_96/journal/vol4/cs11/report.html.
Strandberg, R., Alerstam, T., Hake, M. & Kjellen, N. (2009). Short-distance migration of the Common
Buzzard Buteo buteo recorded by satellite tracking. Ibis, 151(1):200–206.
Stuart, N., Barratt, T. & Place, C. (2006). Classifying the neotropical savannas of belize using remote
sensing and ground survey. Journal of Biogeography, 33(3):476–490.
Sundaresan, S. R., Fischhoff, I. R., Hartung, H. M., Akilong, P. & Rubenstein, D. I. (2007a). Habitat
choice of Grevy’s zebras in Laikipia, Kenya. African Journal of Ecology, 46(3):359–364.
Sundaresan, S. R., Fischhoff, I. R. & Rubenstein, D. (2007b). Male harassment influences female
movements and associations in Grevy’s zebra (Equus grevyi). Behavioral Ecology, 18:860–865.
113
CHAPTER . Reference
Swain, D., Wark, T. & Bishop-Hurley, G. (2008). Using high fix rate GPS data to determine the
relationships between fix rate, prediction errors and patch selection. Ecological Modelling, 212(3–
4):273–279.
Tamstorf, M. P., Aastrup, P. & Cuyler, L. C. (2005). Modelling critical caribou summer ranges in
West Greenland. Polar biology, 28(9):714–724.
Telonics (read 10/2008). ST-18 and ST-19 Argos PTTs for avian species. http://www.
telonics.com/products/legacy/st18avian/st18avianintro.html.
USGS (read 12/2008). USGS Global Visualization Viewer. http://glovis.usgs.gov/.
van Bommel, F. P., Heitkönig, I. M., Epema, G. F., Ringrose, S., Bonyongo, C. & Veenendaal, E. M.
(2006). Remotely sensed habitat indicators for predicting distribution of impala (Aepyceros melampus) in the Okavango Delta, Botswana. Journal of Tropical Ecology, 22:101–110.
Verlinden, A. & Masogo, R. (1997). Satellite remote sensing of habitat suitability for ungulates and
ostrich in the Kalahari of Botswana. Journal of Arid Environments, 35(3):563–574.
Weimerskirch, H., Le Corre, M., Ropert-Coudert, Y., Kato, A. & Marsac, F. (2005). The threedimensional flight of red-footed boobies: adaptations to foraging in a tropical environment? Proceedings of the Royal Society of London series B-biological sciences, 272(1558):53–61.
Wiegand, K., Saltz, D. & Ward, D. (2006). A patch-dynamics approach to savanna dynamics and
woody plant encroachment-insights from an arid savanna. Perspectives in Plant Ecology, Evolution
and Systematics, 7(4):229–242.
World Resources Institute, Department of Resource Surveys and Remote Sensing, Ministry of Environment and Natural Resources, Kenya, Central Bureau of Statistics, Ministry of Planning and
National Development, Kenya & International Livestock Research Institute (2007). Nature’s Benefits in Kenya, An Atlas of Ecosystems and Human Well-Being. World Resources Institute.
Youth, Howard (2004). Thin stripes on a thin line. Zoogoer, 33(6).
114
Appendix A
Ground truth collection form
115
CHAPTER A. Ground truth collection form
GPS
point #
Direction
in which
picture is
taken
Date:
Vegetation description (circle the estimated
cover/ height/ composition using guidelines
below)
% cover
% cover of
% cover
of TREES SHURBS +
HERBACEOUS
average height + composition
C O
C
A
C
S A
>0.5 m <0.5 m
F
C O
C
A
C
S A
>0.5 m <0.5 m
F
C O
C
S A
C O
S A
C O
S A
O
O
S
A
C
>0.5 m <0.5 m
F
C
O
S
A
C
>0.5 m <0.5 m
F
C
O
S
O
S
A
C
>0.5 m <0.5 m
F
Guidelines
% Cover
C = Closed (70% - 100% cover, crowns overlapping,
touching, or very slightly separated)
O = Open (20% - 70% cover, crowns not touching,
distance between crowns up to twice the average crown
diameter)
S = Sparse (2 % - 20 % cover distance between crowns
more than twice the average crown diameter)
A = Absent
S
O
S
G
O
G
G
G
A
M
S
G
A
M
S
O
A
M
S
O
A
M
S
O
Specify cover if no
natural vegetation is
present (for example
settlement, rock,
bare soil, …)
A
M
Herbaceous composition
F = Forbs (> 75 % cover of forbs)
G = Grasses (> 75 % cover of grasses)
M = Mixed (forbs cover less than 75% and grasses cover
less than 75 %)
116
Appendix B
Classes of the Africover classification of
the study area
117
built up
refugee/rural settlement
bare
water bodies
closed woody+trees
closed woody + shrubs
open woody + shrubs
open woody + herbaceaous
open trees + herbaceaous + shrubs
very open trees + shrubs
very open trees + shrubs + herbaceous
closed shrubs + trees
closed shrubs
1
2
10
20
112
113
114
115
116
117
118
121
122
124
125
126
127
131
132
133
134
145
162
163
231
232
class number
class name
open shrubs + herbaceous
very open shrubs + herbaceous + sparse trees
very open shrubs + herbaceous
sparse shrubs + herbaceous
herbaceous + trees + shrubs
herbaceous + shrubs
closed to open herbaceous
sparse herbaceous
open woody - flooded
herbaceous + shrubs - flooded
herbaceous - flooded
herbaceous crops - RF
maize - RF
Table B.1: Africover classification classes
class name
class number
CHAPTER B. Classes of the Africover classification of the study area
118
Appendix C
Boxplots for the different seasons
(a) Boxplot of first dry season
(b) Boxplot of first wet season
(c) Boxplot of second dry season
119
CHAPTER C. Boxplots for the different seasons
(d) Boxplot of second wet season
(e) Boxplot of third dry season
(f) Boxplot of third wet season
(g) Boxplot of fourth dry season
(h) Boxplot of fourth wet season
(i) Boxplot of fifth dry season
120
Appendix D
Habitatpreference based on made
classification
Table D.1: Percentage of each habitat type in the MCP of each zebra
zebra
% MCP
herbaceous
sparse veg
shrubland
woodland1
forest
woodland2
belinda
dableya
hiroya
jeff
johnna
kobosa
lepere
liz
loijuk
martha
njeri
petra
rose
samburu
samburu2
silurian2
29.98
25.22
23.74
19.88
21.04
24.13
34.77
20.71
24.07
48.54
26.50
35.19
46.52
27.05
26.86
24.61
10.55
52.70
50.57
0.94
6.32
51.65
46.33
21.67
19.19
5.05
16.59
42.12
32.30
9.40
19.04
7.02
16.74
5.43
8.29
20.58
11.89
9.28
3.79
8.77
12.49
17.09
14.41
4.21
6.67
16.18
12.71
52.68
25.40
16.07
16.87
26.36
36.75
14.53
14.84
18.86
23.26
23.57
32.75
16.36
2.96
21.38
26.65
0.12
1.03
0.18
0.01
0.01
0.09
0.01
0.01
0.02
0.20
0.25
0.93
0.01
0.15
0.60
5.19
0.36
16.30
0.39
0.53
32.24
23.89
0.40
0.27
29.97
20.79
5.49
8.82
2.12
11.41
25.38
9.54
15.22
121
CHAPTER D. Habitatpreference based on made classification
Table D.2: Percentage of tracking data in each habitat type per zebra
zebra
% tracking data
herbaceous
sparse veg
shrubland
woodland1
forest
woodland2
belinda
dableya
hiroya
jeff
johnna
kobosa
lepere
liz
loijuk
martha
njeri
petra
rose
samburu
samburu2
silurian2
33.75
23.83
33.77
14.53
13.98
25.83
28.57
27.48
33.97
19.12
69.41
29.43
25.87
45.02
27.15
19.05
5.71
64.43
42.53
0.11
26.07
52.30
50.02
36.48
26.16
0.17
13.76
42.43
52.74
15.60
25.08
5.36
17.40
4.37
3.41
19.48
6.20
6.65
7.83
6.94
7.71
12.75
4.91
8.86
10.45
6.56
12.06
41.67
40.25
7.12
20.29
37.27
48.36
15.07
13.40
18.36
22.61
58.57
9.71
18.17
2.99
9.41
30.16
0.00
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.03
0.11
0.01
2.90
0.25
0.01
28.60
5.39
0.14
0.18
10.74
9.55
9.39
2.21
1.10
7.96
23.39
5.44
33.93
122
Appendix E
Habitatpreference based on the Africover
reclass classification
Table E.1: Percentage of each habitat type in the MCP of each zebra
Belinda
dableya
hiroya
jeff
johnna
kobosa
lepere
liz
loijuk
martha
njeri
petra
rose
samburu
samburu2
silurian2
1
2
3
5
6
8
9
10
0.01
0.01
0.01
0.01
0.30
0.01
0.01
0.01
0.01
0.01
0.01
0.01
10.67
0.31
0.03
0.01
0.29
1.05
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.20
0.30
0.01
0.18
0.01
0.01
0.01
0.01
0.01
0.11
0.22
0.48
0.01
0.28
0.03
0.01
0.05
1.08
0.01
0.43
1.73
0.79
3.52
5.89
0.78
0.21
0.74
0.91
1.28
13.32
0.10
0.01
1.14
2.83
0.01
0.01
0.01
0.01
0.01
1.44
0.01
0.01
0.01
0.01
0.01
3.46
0.01
0.01
0.96
0.45
0.01
55.34
19.24
23.72
10.90
31.93
42.01
2.96
14.74
31.73
60.63
24.45
1.35
9.48
47.60
46.55
8.92
41.77
42.12
47.09
81.81
56.49
40.62
96.72
84.30
66.70
38.09
58.50
98.52
34.37
43.77
36.38
91.08
1.63
35.86
28.41
3.76
3.95
16.58
0.01
0.01
0.17
0.01
0.01
0.01
45.48
5.78
9.94
0.01
123
CHAPTER E. Habitatpreference based on the Africover reclass classification
Table E.2: Percentage of tracking data in each habitat type per zebra
Belinda
dableya
hiroya
jeff
johnna
kobosa
lepere
liz
loijuk
martha
njeri
petra
rose
samburu
samburu2
silurian2
1
0.01
0.01
0.01
0.01
1.89
0.01
0.01
0.01
0.01
0.01
0.01
0.01
9.95
2.25
0.16
0.01
2
0.01
1.39
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.14
0.01
3
0.01
0.01
0.01
0.01
0.01
0.07
0.04
0.09
0.05
0.01
0.25
0.09
0.01
0.08
0.02
0.01
5
0.21
0.08
0.06
1.35
11.59
0.01
0.13
0.18
0.55
0.57
2.83
0.09
0.01
0.45
1.70
0.01
6
0.01
0.01
0.01
0.01
0.43
0.01
0.01
0.01
0.01
0.01
0.74
0.01
0.01
0.24
0.05
0.01
8
41.63
15.95
10.31
8.90
11.95
42.96
0.29
1.03
25.68
7.07
26.29
0.28
4.48
9.70
21.94
3.57
9
47.72
37.10
61.91
87.61
61.84
35.03
99.54
98.71
73.72
92.36
69.90
99.53
1.49
69.05
67.42
96.43
10
10.43
45.48
27.72
2.14
12.30
21.94
0.01
0.01
0.01
0.01
0.01
0.01
84.08
18.24
8.58
0.01
124
Appendix F
Histograms for the different seasons
(j) Histogram of first dry season
(k) Histogram of first wet season
(l) Histogram of second dry season
125
CHAPTER F. Histograms for the different seasons
(m) Histogram of second wet season
(n) Histogram of third dry season
(o) Histogram of third wet season
(p) Histogram of fourth dry season
(q) Histogram of fourth wet season
(r) Histogram of fifth dry season
126

model the migration of Grevy`s zebras - endeleo

Transcription

Similar documents

Brochure - Habitat for Zoomanity A Life Science Webquest

The truth about… The truth about… habitats

Zebra Mussels Press Kit Color

WILDLIFE WARRIORS - Grevy`s Zebra Trust

Artists - Sound the Trumpet

grand rapids, mi • van andel arena • sunday, october 16, 2016 • 6:00

College 4 (Crok) - Universiteit Leiden

Untitled - Spier Poetry

2011/01

Engels CSE KB - De Digitale School