ASA-DETR: Adaptive sparse attention enhanced RT-DETR for remote sensing landslide detection with multi-scale frequency-aware feature learning

Accurate landslide detection from remote sensing imagery is essential for geological hazard monitoring. However, existing methods suffer from insufficient feature extraction for complex geological textures, limited multi-scale fusion effectiveness, and poor small-target detection performance. This paper proposes ASA-DETR, a novel RT-DETR-based landslide detection algorithm that addresses these challenges through an improved lightweight backbone, a joint spatial-frequency domain feature learning module, and an adaptive multi-scale attention fusion mechanism. We also construct the RSLD-2K dataset containing 2,299 images with 6,545 landslide annotations from Google Earth, Planet Labs, and Sentinel-2. Experiments on RSLD-2K demonstrate that ASA-DETR achieves 73.2 % mAP@0.5 and 52.5 % mAP@0.5:0.95, improving 3.7 and 2.2 percentage points over baseline RT-DETR, respectively, while reducing parameters by 7.6 %. Cross-domain experiments on the DOTAv2 dataset yield 55.1 % mAP@0.5, suggesting a certain degree of transferability to broader remote sensing detection scenarios.