On unsupervised segmentation of remotely sensed imagery using nonlinear regression

Abstract

A novel segmentation technique for remotely sensed imagery is introduced. Here, image segmentation is posed as a regression problem. The solution is computed by generating a piecewise constant image with minimum deviation from the original input image. The regression technique avoids the problems of region merging, poor boundary localization, region boundary ambiguity, region fragmentation, and sensitivity to noise. Results generated from the nonlinear regression technique and from other traditional segmentation algorithms are given for a study of the Great Victoria Desert using Landsat Thematic Mapper (TM) imagery.