MultiMoE: Multimodal Mixture of Experts Adapter for Semantic Segmentation of Remote Sensing Images

Multimodal remote sensing image semantic segmentation has benefited remarkably from recent progress in deep learning. However, most existing methods treat heterogeneous modalities with excessive uniformity, failing to fully exploit their complementary features. Furthermore, the absence of visual prior knowledge limits their generalization capabilities across diverse scenarios. To address these challenges, we propose a novel multimodal fine-tuning framework based on vision foundation models, namely MultiMoE. Specifically, we select the vision foundation model (DINOv3) as the frozen encoder and develop lightweight Mixture-of-Experts (MoE) adapters on the encoder to transfer general visual knowledge to task-specific semantic contexts. These MoE-based adapters dynamically route samples to dedicated expert pools, thereby processing heterogeneous modalities distinctly and promoting complementary feature representation. Extensive experiments on two widely-used benchmark datasets (ISPRS Vaihingen and Potsdam) demonstrate the competitive performance of our MultiMoE compared to state-of-the-art methods.