Abstract: Aiming at the problem that the simple linear iterative clustering (SLIC) algorithm in the superpixel segmentation of multi-spectral remote sensing images underutilized the image feature information and the fixed size and number of superpixels leaded to low segmentation accuracy, we proposed to introduce manifold SLIC (MSLIC) algorithm into the task of superpixel segmentation of remote sensing images and improve it. Firstly, we proposed an improved texture feature extraction method for multi-spectral remote sensing images based on color local binary pattern (CoLBP). Secondly, we expanded the spectral space of the MSLIC algorithm so that the algorithm could adapt to high-dimensional image data. Finally, we improved the clustering distance measurement of the MSLIC algorithm, fused the multi-spectral features, spatial features and texture features of the image to perform iterative clustering of pixels to achieve content-sensitive superpixel segmentation. The experimental results show that compared with the existing methods, the proposed algorithm has more accurate superpixel segmentation results of multispectral remote sensing images, and improves the edge recall rate, under segmentation error and subdivision accuracy. It can improve the problems of low accuracy in the preprocessing method of multispectral remote sensing image segmentation.

Key words: multi-spectral remote sensing images, superpixel segmentation, local binary pattern, manifold simple linear iterative clustering