Authored By: A. Davidson, A. Hudak, W. Gould, T. Hollingsworth
A. Davidson, A. Hudak, J. Evans, W. Gould, G. González, and T. Hollingsworth
USDA Forest Service Rocky Mountain Research Station (1-3) International Institute of Tropical Forestry (4,5) and Pacific Northwest Research Station (6)
Land cover conversion, forest harvest and road construction have fragmented forests and rangelands across the United States. Fragmentation affects hazardous fuel distribution, introduction of invasive species, wildlife habitat suitability, and other ecological variables. Edge gradients between forest stands and harvested stands, agricultural fields, or roads vary in steepness, making edge mapping difficult to standardize across landscapes. This study evaluated the ability of the eCognition multiresolution segmentation tool to consistently produce image objects with optimal size and shape characteristics for managers, across temperate, tropical, and boreal forest types in the United States. The desired scale of the image objects generated was that of a forest stand, the management unit typically used by forest managers. The utility of the results was assessed using edge locations previously ground-truthed at 720 sites in Idaho, Washington, Minnesota, Puerto Rico, and Alaska. Source image data were Landsat 7 ETM+ multispectral bands (30 m resolution) alone or fused with the panchromatic band (15 m resolution); Landsat ETM+ scenes were selected because they are available across all field sites (for this study) and elsewhere (for widespread applicability). For the 30m and 15m Landsat data the eCognition multiresolution segmentation algorithm was run with three different shape factors (20, 30, and 40) against four different scale factors (0.2, 0.3, 0.4, and 0.5), while all other parameters were held constant. In general, a shape factor of 30 and scale factor of 0.3 produced the best results, in terms of maximizing the percent of ground validation edges correctly detected, while not allowing >1 % of the management units to be < 2 ha, or smaller than is useful for managers. We conclude that eCognition could be used operationally to delineate landscape units across a broad range of forest vegetation types and to help assess fragmentation effects on a variety of ecological processes.
corresponding author:
Andrew T. Hudak
USFS Rocky Mountain Research Station
1221 South Main Street
Moscow, ID 83843
208-883-2327
ahudak@fs.fed.us