Repetitive Band Sampling for Finding Endmembers in Hyperspectral Imagery

Band sampling (BSam) is a new concept recently developed for hyperspectral imaging. It implements a two-stage process by randomly sampling a partial band subset from full bands in the first stage followed by a series of feedback loops in the second stage which captures spatial information from the processed data to augment the current data cube in an iterative manner. Its potential and promise have been demonstrated in hyperspectral anomaly detection and image classification. However, a direct application of BSam to endmember finding (EF) is not feasible because endmembers are considered as pure signatures and using spatial filters not only smears but also contaminates the signature purity of endmembers. To address this issue, this paper re-derives BSam as a new BSam without feedback, called repetitive band sampling (RBSam) which utilizes a repetitive process rather than an iterative feedback process to make BSam applicable to EF. In particular, the endmembers found by an EF algorithm are not necessarily true endmembers, in which case, such EF algorithm-found endmembers are referred to as virtual endmembers (VEs). Since virtual dimensionality (VD) used to determine number of endmembers,neis only an estimate and may not be accurate in real practical applications, to resolve this dilemma, a progressive RBSam (PRBSam) is further extended for this purpose. Extensive experiments show that PRBSam not only can perform well on EF but also can actually approximate the true value ofneas close as possible.