Abstract：As an emerging technology in the field of remote sensing, hyperspectral imaging provides massive content for remote sensing image processing analysis and computer vision. The advantages of hyperspectral images lie in the breadth and high resolution of the electromagnetic spectrum, which can show the inherent spectral reflection characteristics of ground objects in a more comprehensive and discriminating manner, and are widely used in ground object classification, target recognition, anomaly detection, etc. However, due to its huge amount of data and redundant information, it causes considerable difficulties for hyperspectral image processing, storage and transmission. Band selection is a data dimensionality reduction method that can effectively reduce the amount of image data without changing the physical information of the original image. In order to achieve a better classification effect of ground objects, the visual saliency model is applied to the band selection method. Firstly, the target saliency algorithm based on image space distribution is introduced to process the band image to obtain the target saliency map. Secondly, using the target saliency map to analyze the degree of separability between ground objects in each band image is defined as band saliency. To avoid overlapping band information, spectral clustering algorithm is used to divide bands into several subspaces before band selection. Then in the subspace, the bands are sorted in descending order according to the saliency of the bands, and the bands with better target saliency performance in each subspace are selected to form the band subsets. Finally, the method is verified on the hyperspectral image data collected by GF-5, the effective target saliency algorithm is screened, and the classification accuracy is compared with the commonly used band selection algorithm. The experimental results show that the band selection subset based on the LC target saliency algorithm has excellent classification results in the SVM classifier, with the overall classification accuracy and Kappa coefficient of 87.7800% and 0.8053. This method outperforms the results of the other three band selection methods and the results of all bands.
Key words：Hyperspectral remote sensing; Data reduction; Target saliency; Band selection; Object classification
