Using Bayesian Statistics, Thematic Mapper Satellite Imagery, and Breeding Bird Survey Data to Model Bird Species Probability of Occurrence in Maine

A Bayesian modeling technique was used to predict probabili ty of occurrence for 14 species of Maine land birds. The relationships between bird species survey data and the spectral values of Landsat Thematic Mapper bands 4 and 5 as well as a derived texture measure were used to build conditional probabilities for input into Bayes' Theorem. The conditional probabilities form decision rules for reclassifying the input spectral data into probability of occurrence estimates with associated estimates of error inherent in the model prediction. This methodology removed the costly and time-consuming step of creating a habitat map before modeling species occurrence. The output resolution of the species predictions is not degraded from the original 30-m TM pixel size to the coarse resolution of the wildlife survey data. Model results can be compared to results from other habitat modeling techniques and used by natural resource managers to predict the effects of land-use changes on available habitat. Introduction Determining the habitats used by different species is an important first step in managing those habitats to sustain wildlife populations. Modeling the habitat requirements of wildlife species allows wildlife managers to predict the distribution or abundance of target wildlife species (Morrison et al., 1992). Such models can take many forms, but all attempt to represent formally, through equations or decision rules, the relationships between species and their habitat. Spatially explicit relationships between wildlife species and their habitat can be systematically tested within a geographic information system (GIS) (e.g., Lyon, 1983; Lyon et al., 1987; Ormsby and Lunetta, 1987; Shaw and Atkinson, 1988; Pereira and Itami, 1991; Homer et al., 1993; Herr and Queen, 1993; Rickers. et al., 1995). To predict species occurrence in a spatially explicit manner, species-habitat models require a habitat map (e.g., Palmeirim 1988). Land-cover/ land-use (LCLU) maps are converted into habitat-type maps according to known species-habitat associations. Because LCLU classification schemes generally are not developed with the habitat requirements of specific wildlife species in mind, accurate relationships between LCLU classes and habitat types may not exist. Errors in the aggregation of habitat types that a species use and do not use or use at different rates will lead to errors in model output. Additionally, the accuracy of LCLU maps is often unMaine Image Analysis Laboratory, Department of Forest Management, 5755 Nutting Hall, University of Maine, Orono, ME 04469-5755 ([email protected]). tested, leading to the introduction of errors of unknown magnitude into habitat maps. An accuracy assessment of the LCLU map would provide a confusion matrix to allow for error simulation. Because accuracy assessment of LCLU classifications is time consuming and costly, a method that could remove the LCLU classification step entirely would be useful for spatial modeling of species occurrence. Bayesian Modeling Bayesian statistics constitute an alternative method for building predictive relationships between species and their environment. Several studies have used Bayesian statistics to predict one variable based on its statistical relationship to other variables (Tucker et al., 1997; Aspinall and Veitch, 1993; Aspinall, 1991; Bonham-Carter et al., 1988; BonhamCarter et al., 1989). Further modifications of the modeled variable based on repeated comparisons with predictor variables yields a probability map and associated errors for each location on the landscape under study. Bayes' Theorem uses a priori (subjective) and conditional probabilities to calculate the probability of an uncertain event occurring. A priori probabilities represent what the modeler believes, before testing, to be the probability of an event occurring. Conditional probabilities are probabilities that other events occur in conjunction with the original event. If species occur at a rate of 0.5 on the landscape, but occur 80 percent of the time when a closed canopy forest is present, the conditional probability of species presence for closed canopy forests is 0.80. Aspinall (1991) used classes of land cover derived from classified satellite imagery, altitude, and accumulated frost to model habitat availability for red deer (Cervus elaphus) in a region of Scotland. Aspinall and Veitch (1993) simplified the procedure by removing the classification of satellite imagery, instead using unclassified satellite imagery. They created a probability of occurrence map for the Curlew (Numenius arquata) using grouped raw digital numbers (reflectance) from selected wavebands of satellite imagery along with a digital elevation model and species presencelabsence data. Curlew survey data with coarse resolution (1-krn2 survey blocks) were used to classify the fine resolution (30-mZ) satellite image based on repeated comparisons of image pixels where Curlew were observed against image pixels where Curlew Photogrammetric Engineering & Remote Sensing, Vol. 63, No. 10, October 1997, pp. 1231-1237. 0099-1112/97/6310-1231$3.00/0 O 1997 American Society for Photogrammetry and Remote Sensing