Robust low-rank abundance matrix estimation for hyperspectral unmixing

Hyperspecral unmixing (HU) is one of the crucial steps of hyperspectral image (HSI) processing. The process of HU can be divided into end-member extraction and abundance estimation. Lots of abundance estimation methods just take some properties of abundance into consideration, such as non-negative, sum-to-one and so on but ignore the noise corruption. However, in practical applications, there are always high-noise bands in HSI due to water absorption, atmospheric transmission, and other inevitable factors, which lead to the estimation accuracy reduction. Here, we propose a new abundance estimation model which takes the mixing pattern of endmembers and low signal-to-noise ratio (SNR) bands of HSI into consideration simultaneously. The constraints considering not only the low-rank feature of abundance but also the sparsity quality of noise are imposed on the new model for more robust results. Adequate experiments both on synthetic and real hyperspectral data have confirmed the superiority of our method.