Comparison of segmentation techniques remotely sensed images for land cover features

Abstract

This study presents the comparison of edge and region-based segmentation approaches in segmenting linear and polygonal land cover features respectively in optical, single polarization SAR, and multi-polarization SAR images which covered the area of Asajaya, Sarawak and Kuala Nerus, Terengganu. To segment the linear features, the procedures included: edge detection, edge map transformation, edge thinning, and edge linking. From the results obtained, the kernel-based first derivative (i.e. Frei-Chen, Kirsch, Prewitt, and Sobel) gave the better outcomes based on the identification accuracy computed. The segmentation was, however, bounded by two factors: (l) the sensitivity of edge detectors to image texture and (2) the characteristics of input data. For polygonal features, three different region-based segmentors, namely centroid linkage region grower, split-and-merge, and morphological watershed transform, were applied to the following inputs: (1) spectral (or SAR backscattering) data alone, (2) texture data alone, and (3) combined spectral (or SAR backscattering) and textural data. In this study, it was found that the centroid linkage region growing was superior to the split-and-merge and watershed transform. The Landsat-5 TM and TOPSAR data, with their multichannel information, gave the better segmentation results. The segmentation was difficult for both ERS-l and Radarsat images due to their only single channel information. An improvement was achieved by the incorporation of the textural information where the combined spectral (or SAR backscattering) and textural input yielded lower errors than that of using spectral (or SAR backscattering) or textural data alone.