CMD Net: Cross-Modal RGB–Depth Fusion for Dam Detection in High-Resolution Remote Sensing Imagery

Accurate dam detection in high-resolution remote sensing imagery is challenging because dams exhibit substantial variations in scale and morphology, indistinct boundaries, and high visual similarity to roads, bridges, and shorelines. Existing single-modal detection methods rely primarily on spectral and texture information and therefore have limited capability to characterize the geometric structure and spatial relationships of dams. To address these limitations, this study proposes CMD Net, a cross-modal dam detection network in complex remote sensing scenes. CMD Net adopts a dual-branch backbone to independently extract RGB texture features and the pseudo-depth geometry-oriented structural prior and progressively integrates their complementary information. Within the framework, the frequency-decoupled cross-modal gated attention (FD-CMGA) module selectively fuses high- and low-frequency information via approximate spatial decomposition to address the potential negative transfer caused by indiscriminate RGB–depth fusion; the cross-stage multi-scale gated feature aggregation (CS-MGFA) module and the C2GDA module are further incorporated to enhance dam-specific multi-scale representation and geometry-aware feature modeling, respectively. In addition, a new RGB–Depth dam detection dataset (DAM RGB-D) was constructed using 2 m resolution remote sensing imagery, comprising 7712 image pairs and 8242 annotated dam instances. Experiments show that CMD Net achieves 91.8% Precision, 91.9% Recall, 95.1% mAP50, and 63.5% mAP50:95 on the DAM RGB-D dataset. On DIOR-D, it obtains the best dam-category Recall, mAP50, and mAP50:95, reaching 60.5%, 62.0%, and 32.5%, respectively. In the regional-scale experiment, CMD Net detects 108 of 123 verified dams and achieves a Recall of 87.8% and an F2-score of 79.8%, while a subsequent lightweight geographic–spectral filtering strategy further raises Precision to 88.8%. These results demonstrate that CMD Net effectively exploits the complementary texture and geometric information of RGB and the pseudo-depth prior, providing a reliable approach for high-recall dam detection and regional-scale dam candidate extraction.