Unmixing improvement based on bootstrap for hyperspectral imagery

Abstract

This paper presents an unmixing method for hyperspectral data containing small targets. Each pixel of the target represent the useful data embedded among a large number of background pixels. With recent technological developments of hyperspectral sensors, the spatial resolution increases, and it is possible to detect some small targets containing few pixels. We propose in this paper a new approach based on bootstrap resampling method adapted to the linear mixing model which leads to artificially increase the abundance of useful pixels corresponding to the small targets . Then, we use the non-negative matrix factorization (NMF) with these resampled data to estimate the spectra of the targets. The experimental results based on synthetic and real images demonstrate the efficiency of this new approach for the unmixing of smallscale data such as small objects in hyperspectral images.