Enhanced Rockfall and Landslide Detection Using an Improved RT-DETR

Accurate and efficient detection of rockfalls and landslides is critical for disaster prevention. This paper proposes a novel computer vision-based model FGR-DETR (Fast Geo-Risk Detection Transformer) for detecting rock slope failure disasters. Building on the RT-DETR (Real-Time Detection Transformer) framework, FGR-DETR introduces a lightweight backbone network based on Starnet, which maps input features to a higher-dimensional space without increasing network width, reducing model size and complexity. The integration of the SimSPPF(Simplified Spatial Pyramid Pooling - Fast) module and an enhanced attention mechanism further improves performance. Experimental results on rock collapse dataset demonstrate that the FGR-DETR model achieves an accuracy of 88.7% while reducing model size by 40.3%, outperforming the conventional RT-DETR algorithm. These findings emphasize the effectiveness of the proposed model in rock slope failure disaster monitoring and its potential for comprehensive applications in disaster management.