An object-specific image-texture analysis of H-resolution forest imagery

A new structural image-texture technique, termed the triangulated primitive neighborhood method (TPN), is employed to investigate the variable spatial characteristics of high-resolution forest objects, as modeled by a Compact Airborne Spectrographic Imager data set. Based on current psychophysical texture theory, this technique incorporates location-specific primitives and a variable-sized and shaped moving kernel to automatically provide object- and area-specific regularized images. These object-rich, but variance-reduced images allow a traditional classifier to be used on a complex high-resolution forest data set with improved accuracy. The robustness of this technique is evaluated by comparing the maximum likelihood classification accuracy of nine forest classes generated from a combination of the grey level cooccurrence matrix method, semivariance, and customized filters, against those derived from the TPN method. By including into the classification scheme an object-specific channel that models crown density, the highest overall classification accuracy (78%)from all techniques is achieved with the TPN method.