Hyperspectral Sparse Unmixing via Adaptive Overcomplete Dictionary Learning

In the linear sparse unmixing model of hyperspectral data, large estimation error of the fractional abundances of endmembers in each mixed pixel may be caused by the incorrect identification of endmembers. A novel sparse unmixing algorithm was proposed based on adaptive overcomplete dictionary. Firstly, according to the spatial continuity of ground objects and the strong correlation between signal components of the hyperspectral data and spectral signatures in the library, the signatures with the projection coefficients of each pixels larger than the preset threshold were grouped as an optimal subset of signatures that best match the signal component of each mixed pixel. Secondly, an adaptive overcomplete dictionary of hyperspectral data was constructed by combining such subsets. Finally, the fractional abundances in this dictionary were obtained using the alternating direction method of multipliers (ADMM). Experimental results on synthetic and real hyperspectral data show that the proposed algorithm improves the accuracy of identifying endmembers, with the reduced abundance estimation error r. When the signal to noise ratio range from 15 to 35 dB, the accuracy of the abundance estimation is improved about 1 to 2 dB compared with SUnSAL (sparse unmixing by variable splitting and augmented Lagrangian).