In this paper, a new subpixel mapping approach for hyperspectral image is proposed, using a spatial-spectral endmember dictionary with collaborative representation (CR). Different from the classic approaches, the proposed approach employ several spatially closest training samples as the endmembers used for the representation of each mixed pixel, instead of the entire training set. Furthermore, the CR coefficients are derived from the CR of the mixed pixel using the entire training set. Simulative experiments illustrate its outperformance over several classic approaches.