NONLINEAR BLIND HYPERSPECTRAL UNMIXING VIA GENERALIZED MIXING MECHANISM FITTING AND ENDMEMBER CONSTRAINTS

Blind hyperspectral unmixing (HU) is designed for decomposing pixels into endmembers with estimating their abundances. Current deep learning-based methods rely on autoencoders under the linear mixing model assumption, where endmembers are represented by decoder parameters. These approaches limit their applicability in nonlinear and complex scenarios, and also make them challenging to impose constraints on endmembers, potentially resulting in poor endmember estimation. In this paper, we propose a novel nonlinear blind HU method called GMMFU through decoder-based generalized mixing mechanism fitting grounded in the initial nonlinear mixing model assumption. GMMFU simultaneously estimates abundances and endmembers while fitting the mixing mechanism, expressed as a generalized form that surpasses current autoencoder-based methods. Additionally, GMMFU separates the endmembers from the decoder and constrains them to the hyperspectral image data distribution using the proposed endmember consistency and discrepancy losses. This improves the accuracy of endmember estimation, which in turn enhances successively the abundance estimation. Experimental results show that GMMFU outperforms other algorithms qualitatively and quantitatively.