BiFA-Net: Bidirectional Focused Attention Network for Hyperspectral Image Change Detection

Hyperspectral image (HSI) change detection faces significant challenges in balancing the high computational complexity of standard Transformers with the need to capture both long-range spatiotemporal interactions and fine-grained local details. To overcome these limitations, this paper proposes a Bidirectional Focused Attention Network (BiFA-Net) that employs a dual-branch architecture consisting of a change-aware interaction branch and a Global Semantic Difference Encoder (GSDE) branch. In the interaction branch, we design a Change-Aware Focused Linear Attention (CA-FLA) module, which introduces a global difference descriptor to dynamically generate head-specific focusing and scaling factors. This mechanism enables the adaptive re-weighting of the attention distribution and utilizes bidirectional cross-attention to enhance the alignment of bitemporal features. Simultaneously, a parallel difference branch directly extracts multi-scale local spatial-spectral patterns from the difference image. The interaction discrepancy representation generated by CA-FLA and the semantic difference representation extracted by GSDE are further refined using separate Long Short-Term Memory (LSTM) networks before feature fusion. On the River, Farmland, and Hermiston datasets, BiFA-Net achieves overall accuracies of 96.69%, 97.41%, and 98.97%, respectively, showing clear advantages over existing methods in both accuracy and robustness.