An improved optimization deep nonnegative matrix factorization for hyperspectral unmixing

Hyperspectral unmixing (HU) is an essential HSI processing technique for separating the spectral signatures of endmembers and estimating their corresponding abundances within mixed pixels. While nonnegative matrix factorization (NMF) has been widely used for HU with a single-layer structure, it cannot fully explore the hidden features in a HSI. In this paper, an improved optimization deep nonnegative matrix factorization (IODNMF) method is proposed, which improves the parameter optimization approach of the deep NMF model for exploring the deep feature representation of HSI to achieve unsupervised HU. With both unknown endmembers and abundances, a heuristic algorithm is utilized to initialize the parameters of each layer to speed up the convergence of the proposed method. To ensure the nonnegativity of parameters and avoid gradient vanishing and explosion during training, a multiplicative update rule based on positive-negative separation is devised to update the endmembers and abundances in the pre-training and fine-tuning stages. In addition, two layer-setting strategies and three per-layer parameter setting strategies are proposed to effectively solve the network structure setting problem in deep NMF methods. Experimental results on synthetic and real HSI datasets show that the proposed deep NMF algorithm performs more effectively than other classical unmixing methods and achieves a significant improvement in unmixing performance.