Spatial-spectral combined preprocessing method for hyperspectral endmember extraction

The existence of mixed pixels is the main reason that restricts the application accuracy of hyperspectral remote sensing, so hyperspectral unmixing is necessary. As the key of hyperspectral unmixing, the endmember extraction is often susceptible to noise and outliers. In order to improve the accuracy of endmember extraction, a spatial-spectral combined preprocessing method for hyperspectral endmember extraction is proposed in this paper. Firstly, a new concept of spectral purity index (SPI) is defined, which is used to estimate the spectral purity of each pixel in hyperspectral image. Secondly, a spatial de-redundancy method based on SPI is provided, utilizing the continuity of spatial distribution of real objects in the image to judge and eliminate redundant pixels in hyperspectral image, and finally a fine set of candidate endmembers is formed. Experimental results show that after using the proposed preprocessing method, for the simulated hyperspectral image, the angle between the extracted endmembers and the original endmembers is reduced by 9.022 3 degrees on average, and the number of candidate endmembers is less than 10% of the number of original pixels. The proposed preprocessing method not only eliminates the interference of noise and outliers effectively and improves the accuracy of endmember extraction, but also reduces the time complexity greatly.