M2-DINO: Parameter-Efficient and Topology-Oriented Multimodal Fusion for Remote Sensing Image Segmentation

Abstract

Vision foundation models (VFMs) provide strong transferable visual representations for remote sensing image segmentation. However, adapting these models to multimodal remote sensing data remains challenging because of heterogeneous modality gaps, cross-modal feature imbalance, and the attenuation of fine-grained structural details in deep representations. To address these issues, we propose M2-DINO, a parameter-efficient multimodal adaptation framework for remote sensing semantic segmentation. M2-DINO keeps most pretrained weights of a SAT-493M-pretrained distilled DINOv3 vision transformer-large/16 (ViT-L/16) backbone fixed and introduces three trainable components at different network stages. First, a heterogeneous multimodal mixture-of-experts (H-M3E) adapter is inserted into selected ViT residual streams to perform token-level heterogeneous adaptation over optical, auxiliary, and spatial expert responses. Second, a gated lateral module recalibrates lateral and top-down features in the decoder pyramid through spatial gates to reduce potential feature imbalance. Third, a topology-oriented auxiliary skip connection (TASC) injects gated stride-4 auxiliary structural cues into the high-resolution decoder level. In this work, the topology-oriented terminology refers to architectural sensitivity to topology-relevant structures, such as thin-object continuity and object boundaries, rather than explicit optimization of topological invariants. Experiments on the WHU-OPT-SAR, YESeg-OPT-SAR, and DW9c datasets show competitive segmentation accuracy, reaching mean intersection over union scores of 57.50%, 69.19%, and 67.12%, respectively, while optimizing only a subset of the overall model parameters.