Adaptive and Rotating Non-local Weighted Joint Sparse Representation Classification for Hyperspectral Images

Abstract

In this paper, we propose an adaptive and rotating non-local weighted joint sparse representation classification (ARW-JSRC) method for hyperspectral image (HSI). The proposed method aims at avoiding misclassification of the HSI pixels located around the boundaries of class and over-smoothed classification performance caused by the window-based technique used in joint sparse representation classification (JSRC). Since the window-based technique leads to the undesired classification result, an adaptive threshold based on the spectral angle between different classes and the rotated similar window replaced the traditional rectangular window are applied to sufficiently utilize the rich spectral-spatial signatures and alleviate this problem. Furthermore, a new weight formula that accurately reflects the spectral-spatial feature in HSI is applied to obtain more appropriate weights for HSI pixels in search window. Experimental results indicate that our method achieves great improvement in HSI classification, comparing to several widely used classification methods.