改进YOLOv5s的广域InSAR矿区地表形变检测

Improved YOLOv5s for large-scale InSAR mining deformation detection

针对传统InSAR全流程处理在广域矿区地表形变监测中效率低无法快速响应矿区地质灾害的问题,以InSAR干涉条纹为研究对象,提出一种广域观测范围下基于改进YOLOv5s的矿区形变干涉条纹检测方法。使用迁移学习方法对网络模型进行训练,引入ECA注意力机制和ASPP模块、增加小目标网络预测层以提升网络模型对形变干涉条纹的检测精度。改进后的YOLOv5s平均精度均值为94.6%,比YOLOv5s、SSD、Faster-RCNN、YOLOv3、YOLOv7分别提升2.4%、5.7%、4.4%、2.7%、3.2%;F1-Score为89.9%,分别提升3.9%、6.9%、17.7%、2.4%、2.6%;对研究区形变干涉条纹进行检测,准确率为92.7%,误检率为4.3%。实验结果表明,该方法不仅能准确识别广域矿区形变,及时发现灾害隐患并对其进行评估与预警,还能大大提升传统InSAR形变监测效率。

Regarding the challenges of low efficiency and the inability to promptly identify geological disasters in extensive mining areas through traditional InSAR full-process processing,this paper focuses on InSAR interference fringes and proposes a method for detecting deformation interference fringes in mining areas across a large-scale observation range,based on the improved YOLOv5s.The network model is trained using transfer learning methods,incorporating the ECA attention mechanism and ASPP module,and a small target network prediction layer is added to enhance the accuracy of detecting deformation interference fringes.The improved YOLOv5s achieves an average accuracy of 94.6%,surpassing YOLOv5s,SSD,Faster RCNN,YOLOv3,and YOLOv7 by 2.4%,5.7%,4.4%,2.7%,and 3.2%,respectively.The F1 Score reaches 89.9%,indicating increases of 3.9%,6.9%,17.7%,2.4%,and 2.6%,respectively.The accuracy of detecting deformation interference fringes in the research area is 92.7%,with a false detection rate of 4.3%.Experimental results demonstrate that this method can not only accurately identify deformations in extensive mining areas,promptly identify disaster hazards,and evaluate and issue warnings,but also significantly enhance the efficiency of traditional InSAR deformation monitoring.