Toward More Adaptive Fusion: Sample- and Region-Aware Dynamic Fusion Network for Full- and Missing-Modalities Land Cover Classification

In recent years, research on land cover classification (LCC) utilizing optical (OPT) and synthetic aperture radar (SAR) images under full- and missing-modalities scenarios has garnered considerable attention. Nevertheless, existing methods face challenges in sufficiently fusing multimodal features and re constructing missing-modality features. Specifically, conventional fusion methods typically utilize fixed, uniform, and inflexible strategies across all sample pairs, thereby lacking the capacity to select the most targeted fusion method for distinct pairs. Furthermore, these methods rely unconditionally on generated features without considering variations in generation quality across different regions. To address these limitations, this pa per proposes the Sample- and Region-Aware Dynamic Fusion Network (SRDFNet), an architecture that enables more adaptive fusion of features at both the sample and region levels. First, the Multi-Expert Adaptive Fusion (MAF) module is proposed, where each expert functions as an independent fusion unit, allowing dynamic activation of expert combinations tailored to individual sample pairs and thus facilitating adaptive feature fusion at the sample level. Second, the Confidence-based Adaptive Feature Generation (CAFG) module is proposed, which simultaneously produces confidence maps during feature generation to evaluate the reliability of generated features across regions, thereby enabling their refinement. Ultimately, a fusion mechanism in tegrates the refined features with regular modality features, enabling adaptive fusion at the region level. Experiments on the WHU-OPT-SAR and DFC2020 datasets demonstrate that SRDFNet achieves an improvement of 2.30% to 2.90% in mean Intersection over Union (mIoU) under full-modality scenarios, and an enhancement of 0.90% to 3.30% in scenarios with missing-modality. These results substantiate the robustness and effectiveness of the SRDFNet.