A Semantic-Spatial Fusion and Tri-Stream Interaction Network for Semantic Change Detection

Remote sensing semantic change detection requires not only the localization of changed regions but also the identification of semantic categories before and after change, which places higher demands on high-level semantic representation, local spatial detail recovery, and relationship modeling across bi-temporal features. To address the limitations of existing methods in balancing global semantic modeling and local structural characterization, as well as the insufficient interaction between bi-temporal semantic features and change features, this paper proposes TriSF-Net, a DINOv3-driven semantic-spatial fusion and tri-stream interaction network for remote sensing semantic change detection. Specifically, a ViT-CNN hybrid encoder based on DINOv3 and ResNet is designed to extract high-level semantic information and local spatial details from bi-temporal images. A semantic-spatial feature aggregation module is introduced to progressively integrate multi-level semantic features and fuse them with spatial features. In addition, a tri-stream interaction module is constructed to explicitly enhance the relationship modeling between bi-temporal semantic features and change features. Experimental results on the SECOND dataset demonstrate that the proposed method achieves 74.90% mIoU and 26.59% SeK, outperforming representative baseline methods in both changed-region localization and change-type discrimination. Ablation studies further verify the effectiveness of each component.