Optimization of Endmember Extraction in Mixed Pixel Unmixing in Hyperspectral Images

Hyperspectral unmixing is a process for unmixing the mixed pixels of a hyperspectral image composed of several substances and their corresponding proportions. Extracting endmembers is a major problem in hyperspectral unmixing. Owing to the goal of achieving a large search range for the endmembers, the efficiency of the traditional algorithm is usually low. Based on an improved ICA（independent component analysis）and optimized endmember extraction method, an improved method of unmixing was proposed. Firstly, endmember extraction was optimized by means of an optimized FastICA algorithm. Thereafter, spatial information and spectral information were considered and combined into a multi-core parallel processing to optimize the candidate endmembers. Lastly, the performance of the proposed algorithm was verified using real hyperspectral data. The method was shown to overcome the shortcomings of the traditional method and obtain more accurate endmembers. In particular, compared to the accuracy of the traditional N-FINDER method, the accuracy of the extracted endmembers increases by 3.55%. The object classification accuracy also improves immensely, and the overall classification accuracy is increased by 2.88%.