Answers are grounded in this paper and cite their sources.
The goal of remote sensing change detection (RSCD) is to accurately distinguish land-surface changes. Recently, Vision foundation models (VFMs), such as the self-supervised DINOv3, have demonstrated strong semantic parsing capabilities across various visual tasks. However, directly applying DINOv3 to RSCD remains challenging. First, DINOv3 produces low-resolution, single-scale feature representations, which are unfavorable for pixel-wise change classification. Second, DINOv3 is not tailored for RSCD and thus struggles to model geospatial differences along the temporal dimension. To address these issues, Marrying DINOv3 with Dual-Branch Complementary Reconstruction for Remote Sensing Change Detection (DINO-CD) is proposed. Specifically, a dual-branch complementary reconstruction module is introduced to enable interaction between DINO and CNN branches for complementary feature enhancement, producing multi-scale representations that preserve global semantic consistency while strengthening local detail discrimination. In addition, a hierarchical difference aggregation module is designed to comprehensively enhance temporal discrepancies via explicit frequency learning and multi-scale directional aggregation, thereby highlighting effective change regions and reducing false alarms and missed detections. Extensive experiments on multiple datasets demonstrate that DINO-CD outperforms state-of-the-art methods, achieving gains of 6% in F1-score and 8% in IoU on the CLCD dataset.
No takes yet. Share an insight, caveat, or question.
Cao et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: