Triple-Branch Difference Fusion Network for Land Cover Change Detection With Heterogeneous Remote Sensing Images

Abstract

Land cover change detection (LCCD) from multimodal remote sensing images (MRSIs) is crucial for environmental monitoring, disaster response, and urban analysis, but it is still challenging due to the heterogeneous imaging mechanisms of different sensors. In particular, direct pixel-wise comparison across modalities is often unreliable. To address this issue, we propose a triple-branch difference fusion network (TBDFN) for multimodal LCCD. TBDFN is composed of two key components. First, a modality conversion module is designed to alleviate the modality discrepancy via cross-modal transformation, which blends content information from one modality with style characteristics from the other to improve cross-modal consistency. Second, built upon the converted representations, a triple-branch fusion architecture is developed to extract temporal difference features from synthetic aperture radar (SAR) and optical imagery, while the same-time cross-modal difference (ST-CMD) module explicitly models intermodality differences within each temporal phase. The three types of difference features are subsequently aggregated to generate the final change map. Experiments on four datasets, with comparisons against six representative methods, demonstrate that TBDFN achieves superior qualitative and quantitative results, validating its effectiveness and robustness for multimodal LCCD. Our proposed method has, respectively, increased the overall accuracy (OA) rate by 1.29%, 0.17%, 1.07%, and 0.13%. The code of the proposed approach can be available at https://github.com/ImgSciGroup/2026-TBDFN