Minimum distance constrained sparse autoencoder network for hyperspectral unmixing

Abstract. Hyperspectral unmixing is an important task in the analyses and applications of hyperspectral images. Recently, the autoencoder network has been intensively studied to unmix hyperspectral image, recovering the material signatures and their corresponding abundance maps from the hyperspectral pixels. However, the autoencoder network cannot get a unique solution since the loss function is nonconvex. In addition, the data often contain a lot of noise. To address these problems, we propose an autoencoder network, referred to as MDC-SAE, that introduces two different constraints to optimize the spectral unmixing problem. Specifically, we adopt the L1/2 norm regularizer to constrict the abundance vectors, making them sparse. At the same time, we apply the minimum distance constraint on the endmember matrix to push each endmember toward its centroid. We evaluate our method on both synthetic and real data sets, and experimental results demonstrate that the proposed method can achieve the desired solutions and outperforms several state-of-the-art methods.