SDFTransformer: A semantic discrepancy fusion transformer for joint land cover classification with optical and SAR images

By combining optical and SAR imagery, the drawbacks of relying on a single sensor are alleviated. However, current fusion techniques often face challenges in aligning different modalities and tend to be computationally intensive. To overcome these issues, we propose the semantic discrepancy fusion transformer (SDFTransformer), a two-branch, multi-stage fusion model. In the feature enhancement stage, a dual-axis attention module captures directional context to preserve structural continuity. In the feature learning stage, a region-aware block within the Vision Transformer reduces intra-class spectral variability in optical images and suppresses SAR speckle noise via channel attention. In the fusion stage, a difference-guided cross-modal level-wise fusion module integrates optical and SAR features through cross-level interaction, enhancing complementarity while retaining optical detail and SAR structure. Experimental results across four benchmark datasets demonstrate that SDFTransformer consistently surpasses recent advanced methods, delivering mIoU improvements ranging from 0.18% to 7.46%. Notably, the proposed approach remains stable and reliable even in scenarios affected by cloud interference. Moreover, it is computationally efficient, requiring only 0.50 GFLOPs and 1.05 M parameters.