Interpretable Low-Rank Sparse Unmixing and Spatial Attention-Enhanced Difference Mapping Network for Hyperspectral Change Detection

Hyperspectral change detection is a key technique for recognizing differences in hyperspectral images (HSIs) of the same location over two or more periods of time, widely used in the fields of environmental monitoring, resource management, and urban planning. Although deep learning methods have shown excellent performance in such tasks, their intrinsic “black-box” nature, that is, their lack of understanding of the intrinsic physical processes of the surface system and the need for a large amount of training data to achieve good performance, limits their generalization to application scenarios that require a high degree of transparency and interpretability. To address this problem, we propose an Interpretable low-rank sparse unmixing and spatial attention-enhanced difference mapping (LSU-SADM) network. This framework effectively combines traditional optimization algorithms with deep learning mechanisms, bridging the gap between interpretable modeling and powerful data processing. Specifically, LSU-SADM incorporates two key modules. The first module, the low-rank sparse unmixing module (LSUM), which is unfolded from the low-rank sparse coding unmixing (LSCU) model, decomposes hyperspectral data into abundance matrices, optimized via the inexact augmented Lagrange multiplier (IALM) method. The two abundance matrices obtained from the LSUM are concatenated as input to the second module, the spatial attention-enhanced difference mapping module (SADMM). The SADMM, building on the principles of particle swarm optimization (PSO), integrates both global and local spatial information, refining the accuracy of the difference map for precise change detection. The final enhanced feature map is then fed into a classifier for detection results. Experimental results show that LSU-SADM outperforms current state-of-the-art methods. Codes are available at https://github.com/QingranCai/LSU-SADM