Hybrid Remote Sensing Image Segmentation Considering Intrasegment Homogeneity and Intersegment Heterogeneity

Abstract

Image segmentation plays an indispensable role in geographic object-based image analysis, but it remains difficult to obtain meaningful geo-objects from landscape scenes. In recent years, hybrid segmentation methods using the split-and-merge strategy have gained considerable attention. One of the most popular hybrid segmentation methods is the combination of the watershed transformation and full lambda-schedule algorithm (FLSA). The FLSA method only considers the intersegment heterogeneity to calculate the merging cost but ignores the impact of intrasegment homogeneity on the merging cost, thus limiting the satisfactory segmentation quality. To overcome this limitation, this letter improved the FLSA method by enhancing within-segment homogeneity and between-segment heterogeneity. In this letter, the refined method was implemented on four different study regions from Gaofen-1 images and then compared with the FLSA and local spectral angle merging (LSAM) methods. Both visual analysis and quantitative evaluation results indicated that the refined method was more accurate in generating satisfactory segmentation than the competing methods.