Bidirectional Multi-Scale Cross-Modal Attention Network for SAR–Multispectral Semantic Segmentation

Fusing synthetic aperture radar (SAR) and multispectral (MS) imagery for semantic segmentation remains challenging because of cross-modal semantic discrepancies, cloud-induced contamination, incomplete integration of complementary details, and the loss of spatial information during upsampling. To address these challenges, we propose BMCA, a feature-level SAR–MS fusion network for semantic segmentation and land-cover mapping. BMCA integrates bidirectional multiscale attention with cross-modal feature interaction. Specifically, the bidirectional multiscale attention module (DSM) captures deep interactions between the two modalities, refines hierarchical representations, and establishes cross-scale and cross-level dependencies, thereby enhancing both structural and semantic features. The cross-modal feature fusion module (CCF) adaptively integrates complementary information while suppressing modality-specific redundancy. Subsequently, the lightweight residual upsampling module (LRU) enhances spatial representations and recovers fine-grained details for pixel-level prediction. Extensive experiments were conducted on four remote-sensing datasets with diverse spatial resolutions and scene characteristics: SEN12MS, BigEarthNet, M3LEO, and MaRS-16M. On SEN12MS, BMCA achieves an F1-score of 86.76% and an IoU of 76.62%, surpassing the second-best method, CMX, by 1.13 and 1.75 percentage points, respectively. On the ultra-high-resolution MaRS-16M dataset, BMCA attains an F1-score of 97.48% and an IOU of 95.08%, exceeding CMX by 8.22 and 14.48 percentage points, respectively. Across the four datasets, BMCA obtains mean F1-score and IOU values of 92.58% and 86.42%, respectively, demonstrating consistent improvements over competing architectures. These results confirm that BMCA effectively preserves fine-grained spatial details, promotes semantic consistency across modalities, and provides a robust and generalizable solution for feature-level SAR–MS fusion in remote-sensing semantic segmentation.