Multi-stage Features Sparse Learning Based Scene Recognition of Hyperspectral Image

In this paper, a multi-stage features sparse learning method for hyperspectral image scene recognition is proposed. The hyperspectral image is decomposed into five scales by Gaussian function at eight angles. Gabor visual features are extracted at different scales of each angle. The visual dictionary of hyperspectral image is constructed by sparse coding, and the sparse coding coefficient matrix of multi angle and multi-scale features of hyperspectral image is formed Compared with BoVW or PLSA, the overall accuracy of SVM classifier is improved by 4.47% and 5.39% respectively, and the kappa coefficient is improved by 0.0685 and 0.0768 respectively. The results show that the proposed multi-angle and multi-scale sparse coding method for hyperspectral image scene recognition is more interpretable than the description of hyperspectral image features with BoVW or PLSA, and can effectively improve the effect of hyperspectral image scene recognition.