Data fusion for ecological studies
Transcription
Data fusion for ecological studies
Data fusion for ecological studies Jaime Collazo, Beth Gardner, Dorit Hammerling, Andrea Kostura, David Miller, Krishna Pacifici, Brian Reich, Susheela Singh, Glenn Stauffer SAMSI working group Data fusion for ecological studies 1 Motivation I Our group focused on developing methods to combine multiple data sources to estimate species distribution maps. I These maps are fundamental to ecology, e.g., to study effects of land use and climate change. I We apply the methods to jointly model 1. eBirds and SE US breeding-bird survey (BBS) data. 2. eBirds and PA Bird Atlas data. 3. ship and areal surveys of seabirds. SAMSI working group Data fusion for ecological studies 2 eBird effort (c) eBirds sqrt effort 40 40 36 30 20 10 32 −90 SAMSI working group −80 Data fusion for ecological studies 3 eBird number of observations (d) ebirds sqrt sample rates 40 3 36 2 1 0 32 −90 SAMSI working group −80 Data fusion for ecological studies 4 BBS effort (a) BBS effort 40 150 125 36 100 75 50 32 −90 SAMSI working group −80 Data fusion for ecological studies 5 BBS sample proportion (b) BBS sample proportions 40 36 0.2 0.1 0.0 32 −90 SAMSI working group −80 Data fusion for ecological studies 6 Models We considered four models: 1. BBS-only 2. Using EB as a covariate to predict BBS 3. Joint model for EB and BBS with shared occupancy 4. Joint model for EB and BBS with multivariate random effects SAMSI working group Data fusion for ecological studies 7 Spatial occupancy model for the BBS data I Let Ni and Yi be the number of sampling occasions and sightings, respectively, in grid cell i. I Yi ∼ Binomial(Ni , pi Zi ), where I I Zi = 1 indicates that the species occupies cell i I Zi = 0 indicates that the species doesn’t occupy cell i. I pi is the detection probability. Our objective to estimate Zi in all grid cells. SAMSI working group Data fusion for ecological studies 8 Spatial occupancy model for the BBS data I We use a bivariate spatial model for occupancy and detection. I Let θ i = (θ0i , θ1i )T be a random effect for site i, with Zi = I (θ0i > 0) and pi = Φ(θ1i ). I To model spatial variation in occupancy and detection, and their relationship use a multivariate CAR model I Given the random effects at all other sites, θ i ∼ Normal(ρθ̄ i , I I I 1 Σ) mi θ̄ i is the mean of θ j over the mi neighboring sites ρ controls spatial dependence Σ is the 2 × 2 covariance between occupancy and detection. SAMSI working group Data fusion for ecological studies 9 Estimated occupancy (posterior mean of Zi ) (a) Single 40 1.00 0.75 36 0.50 0.25 0.00 32 −90 SAMSI working group −80 Data fusion for ecological studies 10 Estimated detection (posterior mean of pi ) (a) Single 40 0.20 36 0.15 0.10 0.05 32 −90 SAMSI working group −80 Data fusion for ecological studies 11 Using EB as a covariate for BBS I The simplest data fusion method is to use BBS as a covariate in the prior mean for θ i I That is, let Xi be an initial estimate of EB abundance or occupancy at site i. I E(θji ) = Xi β j . I We included six constructed covariates. SAMSI working group Data fusion for ecological studies 12 Estimated EB abundance (Xi ) eBirds Abundance 40 −0.5 −1.0 36 −1.5 −2.0 32 −90 SAMSI working group −80 Data fusion for ecological studies 13 Estimated occupancy (posterior mean of Zi ) (b) Covariate 40 1.00 0.75 36 0.50 0.25 0.00 32 −90 SAMSI working group −80 Data fusion for ecological studies 14 Estimated detection (posterior mean of pi ) (b) Covariate 40 0.20 36 0.15 0.10 0.05 32 −90 SAMSI working group −80 Data fusion for ecological studies 15 Shared-occupancy model for EB and BBS data I Let Wi and Ei be the number of sightings and hours of effort for the EB data in cell i. I We assume the joint model Yi ∼ Binomial(Ni , pi Zi ) and Wi ∼ Poisson[Ei (Zi exp(θi2 )+q)]. I I Zi = I (θ0i > 0) is the shared occupancy indicator. I θi2 controls abundance I q > 0 is the false positive rate θ i = (θ0i , θ1i , θ2i )T is modeled with an MCAR. SAMSI working group Data fusion for ecological studies 16 Estimated occupancy (posterior mean of Zi ) (c) Shared 40 1.00 0.75 36 0.50 0.25 0.00 32 −90 SAMSI working group −80 Data fusion for ecological studies 17 Correlation model for EB and BBS data I To be more robust against bias in EB data, we also tried removing the occupancy indicator from the EB model. I We assume the joint model Yi ∼ Binomial(Ni , pi Zi ) and Wi ∼ Poisson[Ei exp(θi2 )]. I θ i = (θ0i , θ1i , θ2i )T is modeled with an MCAR. I Only the correlation of θ2i and θ0i links the data sources. SAMSI working group Data fusion for ecological studies 18 Estimated occupancy (posterior mean of Zi ) (d) Correlation 40 1.00 0.75 36 0.50 0.25 0.00 32 −90 SAMSI working group −80 Data fusion for ecological studies 19 Model comparisons Mean squared error and deviance comparing estimates based on 2012 BBS and EB data to the observed 2007-2011 BBS data. MSE Deviance Single 6.43 3714 Covariate 5.98 3692 SAMSI working group Shared 5.80 3301 Correlation 5.94 3366 Data fusion for ecological studies 20 Application – PA Bird Atlas breeding bird atlas point counts Blocks: j = 1, … , J Points: i = 1, …, I eBird – block level counts (Wj) and effort (Ej) Wj ~ Poisson(Ej*λj) log(λj) = α1 + θ1,j BBA – number of occasions seen at a point (Yj,i) p – P(detection|present) Yj,i ~ binomial(zj,i*p,5) zj,i ~ Bernouli(ψj) logit(ψj) = α2 + θ2,j Θ ~ MCAR black-throated blue warbler prairie warbler Distribution and abundance of seabirds in the Northwestern mid-Atlantic Project funded by DOE in preparation for energy development • Coast off Delaware/Maryland/ Virginia: three Wind Energy Areas (WEA) • From April 2012 - April 2014 • Ship board distance sampling surveys • 656 km – green lines • High definition aerial surveys • 3500 km - red lines Distance sampling Observation model • Detection probability p is declining function of distance to observer, 𝑝𝑝 = 𝑓𝑓(𝑑𝑑) • Detection on transect line is perfect • ‘Half-normal’: 𝑓𝑓 𝑑𝑑 = 𝑑𝑑 2 exp(− 2 ) 2𝜎𝜎 Application/Example: Loons Loons are common in the study area during the winter and frequently observed in both survey methods. Loons Boat 996 Aerial 1661