Finding Endmember Classes in Hyperspectral Imagery

Endmember finding has received considerable interest in hyperspectral imaging. In reality an endmember finding algorithm (EFA) suffers from endmember variability which causes inaccuracy, inconsistency and instability. In this case a real endmember may not exist but rather appears as its variant, referred to as virtual signature (VS). This paper presents a new approach to finding VSs by taking endmember variability into account. It first determines a required number of endmember classes by virtual dimensionality (VD), then designs an unsupervised method to find endmember classes and finally develops an iterative algorithm to find VSs. Comprehensive experiments including synthetic and real image scenes are conducted to demonstrate effectiveness of the proposed approach.