Unsupervised training area selection in forests using a nonparametric distance measure and spatial information

Abstract

A new unsupervised technique that automatically delineates areas with a similar tone is described. The proposed algorithm grows a region of homogeneous tone around a seed pixel; membership criteria for the region is based upon a nonparametric distance measure. The thematic image output can be used to define training areas for a supervised classifier. Two commonly used unsupervised strategies for delineating training areas (viz., clustering and uniform texture mapping) are compared with the proposed technique using SPOT digital data collected over a multi-aged forest plantation in south-east Australia.