Hemlocks via tree-based classification of satellite imagery and environmental data

Within the last few years, the hemlock woolly adelgid (HWA) has made significant inroads into the southern Appalachians. Since the region’s native hemlock species are not resistant, timely application of control measures is critical to minimizing hemlock mortality. Unfortunately, hemlock stands in the region are incompletely mapped, and general characteristics of their distribution present serious mapping challenges. One approach for improving classification is to integrate medium-resolution satellite imagery (Landsat, ASTER) and ancillary environmental data. We tested such an approach using images from eastern and western study areas in Great Smoky Mountains National Park. First, we created maps for masking out nonevergreen pixels via unsupervised classification (i.e., cluster busting) of winter images. We then applied the masks to corresponding summer images so we could separate hemlock and non-hemlock evergreens under optimal image conditions. We extracted a large (>14,000) random sample of points from the masked images, stratifying the sample according to an aerial photoraphy-derived vegetation map of the park. At each sample point, we recorded the vegetation label as well as image data and values for a suite of topographic, environmental, and proximity variables recorded in a geographic information system (GIS). We applied a series of tree-based classifications to this training data set to create a set of decision rules that most accurately retains the input class of sample points. Our most successful tree had 79 total “leaves” (i.e., distinct decision-rule pathways). We applied these decision rules to the images to develop hemlock maps of the study area. Thematic accuracy assessment of these maps, based on field survey and photo-derived points, indicated 85% overall accuracy in the eastern study area and 69% success at capturing hemlocks in a partial assessment of the western study area. Additional accuracy assessment may offer an opportunity to refine the rules. However, our decision rules can currently be applied elsewhere in the southern Appalachian region for management planning purposes. __________ Mapping Hemlocks via Tree-based Classification of Satellite Imagery and Environmental Data Presentations Third Symposium on Hemlock Woolly Adelgid 105