Landscape Level Evaluation of Northern Bobwhite Habitats in Eastern Virginia Using Landsat TM Imagery
Northern bobwhite (Colinus virginianus) are important game birds associated with early successional habitats across the southeastern United States. In the past 30+ years there has been an almost universal decline in bobwhite population numbers despite a long history of management. The Virginia Bobwhite Quail Management Plan was implemented in 1996 to slow and stop the current population declines in Virginia. Virginia Department of Game and Inland Fisheries (VDGIF) personnel identified a lack of knowledge about the broad-scale, landscape level habitats in eastern Virginia. A large scale land cover map along with a detailed understanding of the spatial arrangements of bobwhite habitats will not only aid Virginia's management plan, but also allow focused efforts by our wildlife managers. I explored the possibilities of using remote sensing to map various habitats important to bobwhite. I compared several classification algorithms applied to Landsat TM imagery prior to selecting the classification method that best delineated early successional habitats. After method selection, a classified land cover map for the Coastal Plain and Piedmont of Virginia was generated.
Using the classified images available from the first part of the study and 4 years of bobwhite call count data, I studied the landscape level habitat associations of bobwhite. A number of landscape metrics were calculated for the landscapes surrounding bobwhite call count routes and were used in two modeling exercises to differentiate between high and low bobwhite populations. Both pattern recognition (PATREC) and logistic regression models predicted levels of bobwhite abundance satisfactorily for the modeled (84.0% and 96.0% respectively) and independent (64.3% and 57.1%, respectively) data sets. The models were applied to remotely-sensed habitat maps to develop prediction maps expressing the quality of a landscape for supporting a high population of bobwhite based on existing land cover.
Finally, I explored the possibility of eliminating the time consuming and very costly step of classifying a remotely-sensed image prior to examining its quality for a particular species. Using raw Landsat TM imagery and bobwhite call count data, I developed predictive logistic regression models expressing the quality of a landscape surrounding a pixel. The first model predicted the probability of the landscape supporting a high bobwhite population. Due to a number of stops with an average of zero, I was also able to generate a model that expressed the probability of the landscape supporting any number of bobwhite. This method also satisfactorily predicted high/low population and presence/absence for the modeled data (65.7% and 83.1%, respectively) and independent data (65.3% and 83.7%, respectively). The method described will allow for rapid assessment of our wildlife resources without having to classify remotely-sensed images into habitat classes prior to analyses.