Subpixel mapping of hyperspectral images with hybrid endmember library and optimized abundances

Abstract

Subpixel mapping of hyperspectal image is treated in this paper, to produce a classification map in subpixel level. Specifically, a framework with two paralleled branches integrated by decision fusion is proposed. In one branch, a subpixel level segmentation map is obtained by applying unsupervised clustering to the upsampled hyperspectral image. In the other one, a subpixel mapping result is obtained by combining supervised classification, spectral unmixing and subpixel spatial attraction model. To improve the subpixel mapping accuracy, a labeled-unlabeled hybrid endmember library as well as optimized abundances are employed for spectral unmixing. Experimental results illustrate that the proposed approach clearly outperforms some state-of-the-art subpixel mapping approaches, and is less dependent on the supervised classifier adopted. The improvement can be attributed to the introduction of endmember library augmentation and the following abundance optimization.