Improving automatic target generation process for hyperspectral endmember extraction

Although many endmember extraction algorithms (EEAs) have been proposed, the accurate identification of endmembers is still a challenging task in spectral unmixing of hyperspectral imagery. One of the EEAs, automatic target generation process (ATGP), works by iterative orthogonal projections of the data then finding the largest magnitude vector of this projection, and it will stop until reaches a predefined number of endmembers. This paper proposes an updated version of ATGP by making improvements on two aspects of the method. First, spectral and spatial redundancies are removed, and only a group of candidate endmember pixels will be processed by ATGP. Second, after an endmember pixel is found using orthogonal projection, this pixel will be used to divide the group of candidate endmember pixels into a smaller group and a cluster using similarity measure. Furthermore, a threshold criterion is set to evaluate the quantity of the cluster, which avoids the found pixel is an interfering pixel. A comparative study and the obtained experimental results show that the improved ATGP algorithm not only reduces computational complexity but also provides better performance than the four well-known published algorithms.