2015 Macroscale - UAV classification

Transcription

2015 Macroscale - UAV classification
Macroscale Biology Meeting – UAV classification
Three centimeter classification – How cool is that?
Michael Palace, Christina Herrick
1Earth
System Research Center
Institute for the Study of Earth, Oceans, and Space
University of New Hampshire
Daniel Finnell2, Anthony John Garnello2, Carmondy
McCalley1, Ruth Varner1
2Virginia
Commonwealth University,
3University of Arizona
Talk Organization
UAV data collection
Field data collection
Stitching the images
Georeferencing and orthrectification
Textural analysis
Neural network
Final Classification
Balloons and birds
• Balloons are rare and local
– First remote sensing images
were from bird, balloon, and kite
First known photograph,
taken in 1827
1839 aerial photograph of a
supposedly empty street in Paris,
note enlargement
http://www.oneonta.edu/faculty/baumanpr/geosat2/RS%20History%20I/RS-History-Part-1.htm
One of Lawrence's 1906
photographs of San
Francisco
http://www.oneonta.edu/faculty/baumanpr/geosat2/RS%20History%20I/RS-History-Part-1.htm
Early satellites
• Corona
• Argon
• GRACE- Gravity Recovery
and Climate Experiment
Field data collection
•
Vegetation composition collected through the
use of a one-square meter quadrat divided into
64 sub-plots
•
An ADC-Lite (Figure 1) and GoPro camera were
used to collect images via pole based mount
•
A Teflon chip was used to normalize entropy
and evenness data for light conditions
•
Post-processing included a variety of texture
analysis including: Entropy, Lacunarity, Angular
Second Momentum (ASM), and Normalized
Difference Vegetation Index (NDVI)
Orthorectification and Georeferencing
Abisko
approx.
>1km x <.5km
The Process
1. Orthorectify WV-2 images
– Removes distortions due to terrain displacement
– Displacement due to off-nadir sensor angle
– Used 15m Aster GDEM
> 200m shift
The Process
1. Orthorectify WV-2 images
–
–
–
Removes distortions due to terrain displacement
Displacement due to off-nadir sensor angle
Used 15m Aster GDEM
2. Correct GPS data
GPS Data Correction
Typical
Error
“The closer, the longer, the better”
-- close to base station (less than 50km)
-- longer point collection (> 1hr)
Autonomous Differential
GPS
GPS
Satellite
Clocks
1.5
0
Orbit Errors
2.5
0
Ionosphere
5
0.4
Troposphere
0.5
0.2
Receiver
Noise
0.3
0.3
Multipath
0.6
0.6
Total (m)
10.4
1.5
GPS Reference Datum
GPS unit (rover) collects data as raw observations
Base station position & rover data in ECEF coordinates
ECEF coordinates converted to LLA of desired datum
Datum = model of the earth’s surface
The Process
1. Orthorectify WV-2 images
–
–
–
Removes distortions due to terrain displacement
Displacement due to off-nadir sensor angle
Used 15m Aster GDEM
2. Correct GPS data
– Obtain from Lantmäteriet ECEF antenna position
& 1-sec base data
– Convert antenna position to LLA
– Use Trimble’s Pathfinder Office to correct rover
data
3. Georectify UAV mosaic
Georectification
• GPS points (low HDOP & low RMS error)
• Links are added
• Mosaic is transformed to new coordinate
system
Georectification
The Process
1. Orthorectify WV-2 images
–
–
–
Removes distortions due to terrain displacement
Displacement due to off-nadir sensor angle
Used 15m Aster GDEM
2. Correct GPS data
–
–
–
Obtain from Lantmäteriet ECEF antenna position & 1-sec base data
Convert antenna position to LLA
Use Trimble’s Pathfinder Office to correct rover data
3. Georectify UAV mosaic
– 64 GPS points + WV-2 image
– Accuracy between 15cm (around boardwalk) & 50cm
(around edges)
4. Create training samples
Training Samples
• 200 random plots in random order
• 8 classes: TS, HM, SW, WT, TG, H2O, RK, OT
• What’s in a 50cm2 plot? (17x17 pixel window)
Training Samples
Color RGB
Texture analysis on green band – entropy, angular second momentum, and evenness
wet
rock
water
other
Semi-wet
hummock
Tall gram
Tall shrub
Mire classification
Legend
Wet
Tall Shrub
Tall Gram
Semi-wet
Rock
Other
Hummock
Water
Classification
Classification
Legend
Wet
Tall Shrub
Tall Gram
Semi-wet
Rock
Other
Hummock
Water
Classification
Change
Class
Water
Hummock
Other
Rock
Semiwet
Tall Gram
Tall Shurb
Wet
If these classes are near then
Semiwet Tall Gram Wet
Tall Gram Wet
Tall Gram
for x in range(20,(x1-20)):
print x, x1,(float(x)/float(x1))*100,"% done of model calculation "
for y in range(20,(y1-20)):
pp = b1[x,y]
for x3 in range(-10,11):
for y3 in range(-10,11):
ap=b1[x+x3,y+x3]
if pp==2 and ap==5 or ap==6 or ap==8:
change[x,y]=5
if pp==5 and ap==6 or ap==8:
change[x,y]=8
if pp==8 and ap==6:
change[x,y]=6
Change to
No change
Semiwet
No change
No change
Wet
Tall Gram
No change
Tall Gram