Abstract: To address the challenges of insufficient automatic recognition accuracy of InSAR technology and difficulty in adapting complex terrain landslide features in current landslide detection,as well as the lack of existing deep learning models for InSAR landslide recognition methods,and the limited ability of recognizing fine deformation and small-scale landslides,this paper proposes a method of automatic detection of active landslides by fusing SBAS-InSAR with the improvement of YOLOv8,in order to achieve the accurate identification of active landslides in large areas.The method first generates a surface deformation rate map by processing multi-temporal SAR images with SBAS-InSAR,and then constructs the improved VRO_YOLO model: We propose variable kernel convolution to adapt to the irregular morphology of landslides,and integrate RepVGG,ShuffleNet,and One-Shot Aggregation to construct the channel one-shot shuffle (RVS-OSA)module to enhance the multi-scale identification of active landslides.Taking the Pinglu Canal area in Guangxi as the study area,49-view Sentinel-1 SAR images are used to generate the rate map and construct the dataset to carry out the experiments.Finally,the multi-dimensional validation by combining high-resolution remote sensing,UAV tilt photography and field exploration further supports the accuracy of the method.The results show that the detection precision of VRO_YOLO reaches 51.2%,the recall rate reaches 55.2%,and the mean accuracy reaches 46.8%,and the further comparisons,ablations,and generalization experiments validate the applicability and accuracy of the model accuracy.In conclusion,the method provides a strong potential for landslide detection,contributing to disaster prevention and risk reduction.

Key words: landslide, InSAR, deformation, VRO_YOLO, remote sensing