Robust spatial purity based endmember extraction in the presence of rare ground objects

Endmember Extraction (EE) is widely utilized to extract spectrally unique and singular spectral signatures for Spectral Mixture Analysis (SMA). The Spatial Purity based EE (SPEE) algorithm has been demonstrated to be an effective EE technique by integrating spectral similarity and spatial context. However, this algorithm tends to neglect anomalous endmembers that occur rarely but are also of interest. In this paper, a Robust SPEE (RSPEE) algorithm is proposed to extract both abundant endmembers from spatial uniform areas and rare endmembers accounting for rare ground objects. Abundant endmembers are firstly extracted by the SPEE algorithm and their spectral variabilities are alleviated by utilizing spatial information. By projecting all the pixels to the simplex determined by abundant endmembers, the rare endmembers are then identified iteratively by finding the maximum residual errors. Experimental results demonstrated that the simplex projection errors have been greatly reduced by identifying rare endmembers as well as abundant endmembers in the proposed RSPEE algorithm.