Remote sensing image segmentation based on statistical region merging and nonlinear diffusion

As remote sensing images are multi-sensor, multi-spectral, multi-temporal phase and multi-resolution, general image segmentation methods always can not obtain satisfactory results. In this paper, we introduce the statistical region merging (SRM) model to segment remote sensing images. And according to the characteristics and defects of SRM, the model is improved as follows: Firstly, image gradient information is added in the sort function, which can increase the differences between regions; Secondly, we combine SRM with the nonlinear diffusion which can protect borders, then the requirements of regional homogeneity are better meted, and the model´s anti-noise ability is also strengthened; Thirdly, for the issue of SRM has over merging defect, we give a predicate, by which the over merging regions are chosen and then segmented by IAC. Experiments on two color remote sensing images display the quality of the novel method.