Effect of endmember clustering on proportion estimation: results on the SHARE 2012 dataset

Abstract

Estimating the number of endmembers and their spectrum is a challenging task. For one, endmember detection algorithms may over or underestimate the number of endmembers in a given scene. Further, even if the number of endmembers are known beforehand, result of the endmember detection algorithms may not be accurate. They may find multiple endmembers representing the same class, while completely missing some of the endmembers representing the other classes. This hinders the performance of unmixing, resulting in incorrect endmember proportion estimates. In this study, SHARE-2012 AVON data pertaining to the unmixing experiment was considered. It was cropped to include only the eight pieces of cloth and a portion of the surrounding asphalt and grass. This data was used to evaluate the performance of five endmember detection algorithms, namely the PPI, VCA, N-FINDR, ICE and SPICE; none of which found the endmember spectra correctly. All of these algorithms generated multiple endmembers corresponding to the same class or they completely missed some of the endmembers. Hence, the peak-aware N-FINDR algorithm was devised to group the endmembers of the same class so as not to over or under-estimate the true endmembers. The comparisons with or without this refinement for the N-FINDR algorithm are demonstrated.