Hyperspectral Band Selection Based on Classification Reward Adjustment

Hyperspectral image band selection is a crucial link in image processing. It is necessary to screen out bands with rich information and low correlation, to achieve data dimensionality reduction and retain key information, thereby improving image processing efficiency. Aiming at the differences in reflectance characteristics of targets in different regions in hyperspectral images and the strong correlation between bands, this paper proposes a gradient-guided spatial-spectral weighted hyperspectral band selection method. Firstly, gradient information is introduced to enhance the edge texture features of different targets. Then, the entropy rate superpixel segmentation algorithm is used to segment the first principal component of the image and divide multiple uniform regions. Secondly, to more comprehensively capture the global relationship between all bands in the image, a multi-graph complementary strategy is proposed to construct a regional similarity graph, aiming to maximize the mutual contribution between superpixels with spatial-spectral similarity. Finally, a unified clustering graph is generated by spectral clustering. After normalized cutting, the band with the least noise information in each sub-cube is selected to form a new band subset. The effectiveness of the proposed method is verified by classification experimental results and data analysis.