SFD-YOLO: A novel framework for subsidence funnels detection in China based on large-scale SAR interferograms

Abstract

Accurate identification of subsidence funnels is essential for assessing surface deformation in mining areas, preventing disasters, and optimizing resource management. However, recognizing subsidence funnels of varying sizes in large-scale interferograms poses significant challenges, particularly for small-sized funnels. Their indistinct features and susceptibility to background noise interference often result in suboptimal detection accuracy. To address these challenges, this study proposes a deep learning network based on the YOLO architecture—SFD-YOLO (Sinking Funnel Detection-YOLO). The model incorporates the DWR-C2f module, which enhances multi-scale feature extraction and significantly improves the detection of small-sized subsidence funnels. Additionally, the innovative Inner-WIoU regression loss function improves the localization accuracy of detection boxes while also alleviates the imbalance between hard and easy samples. Experimental results demonstrate that the fully trained SFD-YOLO model achieves an mAP50 accuracy of 92.00% while maintaining high efficiency, significantly outperforming other advanced methods. Applying the SFD-YOLO model to interferograms across China detected a total of 3,842 subsidence funnels, with Shanxi, Inner Mongolia, Shaanxi, and Anhui identified as the four provinces with the highest funnel number. Overall, subsidence funnels are predominantly distributed in northern and northwestern China. Further analysis and experimental evaluation reveal that the SFD-YOLO model exhibits strong generalization capabilities across complex surface environments nationwide and multi-source satellite data.

Keywords

• Mining subsidence detection
• National scale
• Interferograms
• Deep learning
• Small-sized funnels