Subpixel Mapping of Hyperspectral Images Using a Labeled-Unlabeled Hybrid Endmember Library and Abundance Optimization

Classification at subpixel level for a low resolution hyperspectral image (LR HSI) is considered in this article. Using selected labeled samples as labeled endmembers and unsupervised clustering centers of LR HSI as unlabeled endmembers, a hybrid endmember library is constructed for spectral unmixing of LR HSI. The abundances of unlabeled endmembers are used to optimize the estimated fractional abundances of labeled classes within mixed pixels to improve the estimation accuracy. A more accurate subpixel mapping result is then obtained by applying subpixel spatial attraction model with the optimized fractional abundances. To incorporate spatial contextual information and further improve the subpixel mapping performance, a subpixel level segmentation map is generated by applying unsupervised clustering to the upsampled LR HSI, and integrated with the initial subpixel mapping result by decision fusion. Experimental results demonstrate that the proposed method remarkably outperforms state-of-The-Art subpixel mapping methods, including the corresponding ones with or without spatial contextual information incorporation.