DTU-Net: A Multi-Scale Dilated Transformer Network for Nonlinear Hyperspectral Unmixing

Transformers have shown significant success in hyperspectral unmixing (HU). However, challenges remain. Transformer-based unmixing networks, built on vision transformer (ViT) or Swin transformer, struggle to effectively capture essential multiscale and long-range spatial correlations. In addition, these networks predominantly rely on the linear mixing model (LLM), lacking the flexibility to accommodate scenarios with significant nonlinear effects. To address these limitations, we propose a multiscale dilated transformer-based unmixing network for nonlinear HU (DTU-Net). Its encoder integrates two branches: a spatial branch mainly employing multiscale dilated attention (MSDA) to uniquely capture intricate multiscale and long-range spatial correlations via adaptive receptive fields, and a spectral branch utilizing 3-D-convolutional neural networks (CNNs) with channel attention. This design enables comprehensive extraction as well as integration of multilevel spatial and spectral features. The decoder is specifically designed to accommodate both linear and nonlinear mixing. It explicitly models the polynomial post-nonlinear mixing model (PPNMM) by learning nonlinear coefficients as pixel-wise features, which enhances interpretability by directly reflecting the pixel-level nonlinear mixing strength. Experiments on synthetic, ray tracing, and real datasets validate the effectiveness of the proposed DTU-Net, demonstrating its superior performance compared to both PPNMM-derived and advanced unmixing networks.