Clustering Based Spatial Spectral Preprocessing for Hyperspectral Unmxing

Numerous spectral-based endmember extraction algorithms (EEAs) for hyperspectral unmixing (HU) at the price of ignoring spatial context information in recent years. In this paper, we propose a novel preprocessing module by integrating spatial-spectral information, which consists of three parts: 1) k-means algorithm based on spectral angle distance measurement criterion is used to identify hyperspectral image homogenous regions; 2) the local window is utilized to detect the anomalous pixels that hide in the scene; 3) the reconstruction weight that takes into account spatial and spectral information jointly is designed to revise the anomalous pixels to strengthen image homogeneity. The principal contribution of the proposed algorithm is to promote the homogeneity of image and lessen computational complexity while improving the accuracy of endmember extraction. The experimental results obtained by using real hyperspectral data set show a slight improvement for HU while comparing with the state-of-art spatial preprocessing framework.