A novel subpixel mapping approach based on spectral unmixing for hyperspectral images

Hyperspectal image classification in subpixel level is treated in this paper. A hybrid framework is proposed, in which two paralleled branches are integrated by decision fusion. In one branch, a subpixel level segmentation map is obtained by applying unsupervised clustering to the upsampled hyperspectral image. In the other branch, a subpixel level classification map is obtained using subpixel spatial attraction model. To improve abundance estimation accuracy, novel endmember selection and abundance estimation strategies are employed for spectral unmixing. Experimental results illustrate that, compared to some existing subpixel mapping approaches, the newly proposed one is capable of producing results with higher accuracy. The improvement in classification accuracy can be attributed to the usage of the novel endmemeber selection and abundance estimation strategies in spectral unmixing and the consideration of spatial contextual information in decision fusion.