An Endmember Bundles-Guided Dual-Branch Multiscale Network for Nonlinear Unmixing

In recent years, considerable attention has been devoted to addressing nonlinear mixing in hyperspectral unmixing (HU), leading to the development of numerous deep learning (DL)-based approaches. However, existing methods primarily focus on modeling nonlinearity, while spectral variability (SV) is often simplified or neglected under different modeling assumptions. In addition, a large class of autoencoder (AE)based approaches relies on conventional architectures in which decoder weights are directly interpreted as endmembers, potentially limiting their representational flexibility. To overcome these limitations, this article proposes an endmember bundles-guided dual-branch multiscale network (EBDM-Net) that jointly accounts for both nonlinear mixing and SV. The proposed framework adopts a dual-branch architecture with a nonlinear decoder. Specifically, the multiscale abundance extraction (MAE) module incorporates a spatial–spectral refinement attention (SSRA) block to adaptively enhance salient features and suppress redundancy. It further includes a multiscale separable dilated convolution (MSDC) block that captures fine-grained spectral–spatial dependencies, thereby improving feature discrimination. The endmember bundle learning (EBL) module leverages the short-time Fourier transform (STFT) to enhance endmember separability while exploiting the spatial consistency of endmember bundles to mitigate the effects of SV. In addition, the nonlinear decoder employs a dual-branch structure to model nonlinear mixing effects in a more physically interpretable manner. Experimental results demonstrate that EBDM-Net outperforms state-of-the-art unmixing methods on both synthetic and real datasets.