Abstract:

Aiming at the problem of the inconsistency of bi-temporal images' boundaries and spatial correspondence in the task of region-based synthetic aperture radar (SAR) image change detection, a superpixel cosegmentation algorithm based on simple non-iterative clustering (SNIC) for bi-temporal SAR images is proposed. First, a fused image containing the features of the bi-temporal SAR images is constructed, and the pixel intensity similarity and spatial distance similarity between the pixels to be processed and the cluster center is calculated. Second, a computationally efficient multiscale edge detection algorithm is adopted and used to detect the edges of the bi-temporal SAR images respectively, and the edge detection results are fused to form an edge map. Finally, the pixel intensity similarity, spatial distance similarity and edge map information are weighted to replace the distance measure in the original SNIC algorithm and the improved SNIC is utilized to perform superpixel segmentation on the fused image to obtain the segmentation result which fits the real terrain edges in the bi-temporal SAR images. The experimental results conduct on a pair of simulated SAR images and a pair of real-world bi-temporal SAR images demonstrate that the boundary recall, under-segmentation error and achievable segmentation accuracy of the proposed method are better than those of other seven state-of-the-art methods.

Key words: bi-temporal synthetic aperture radar (SAR) image, superpixel, copperative segmentation, simple non-iterative clustering, change detection