DUAL-BRANCH ASYMMETRIC NETWORK FOR ARBITRARY-MODAL SEGMENTATION OF REMOTE SENSING IMAGES

The growing availability of multimodal remote sensing data holds great promise, offering diverse perspectives through modalities such as optical images, synthetic aperture radar (SAR), hyperspectral (HS), and LiDAR images. However, semantic segmentation of multi-modal remote sensing images remains a challenging task, particularly when dealing with arbitrary numbers of modalities. Existing methods typically focus on fusion techniques tailored to specific sensor combinations or fixed modality numbers, which limits their scalability and flexibility. This paper proposes an efficient and robust approach for Arbitrary-Modal Semantic Segmentation (AMSS) of remote sensing images. We introduce two novel modules: the Arbitrary-Modal Parallel Integration (AMPI) module and the Cross-Modal Semantic Selection (CMSS) module. AMPI efficiently processes features from multiple modalities and extracts complementary scene information, and CMSS module dynamically fuses arbitrary-modal features by selecting robust features through an alignment measure. Our approach ensures optimal performance without significant computational overhead when adding more modalities, offering a flexible solution for large-scale multimodal segmentation tasks. Experimental results demonstrate the effectiveness of our method in improving segmentation accuracy while maintaining computational efficiency.