The spectral features in hyperspectral images are easily affected by the spatial position changes in objects. In this study, a Multi-Scale Spectral Spatial Discriminative model (MS2D) is proposed for the spectral feature changes caused by different positions. This model decomposes the original image into multi-scale low-rank components through multi-scale low-rank decomposition (MSLRD), which decouples the complex correlation between spatial structure and spectral variability. Considering the time complexity of MSLRD, we propose three optimization points: (1) using superpixel segmentation to gather pixels to form superpixels, so as to reduce the calculation of the number of pixels; (2) selecting the band with a large amount of information on the representative band of the input to reduce the cost of redundant bands; (3) replacing standard singular value decomposition (SVD) with random SVD to reduce computational complexity. After classification, the majority voting strategy is used to vote on the results to alleviate the discrimination conflict between features. The experimental results on three public datasets show that the performance of MS2D is better than that of the other models, which verifies that this model can control computational complexity and improve classification accuracy.
Wu et al. (Sun,) studied this question.