Research On Landslide Detection Method Based On Deep Learning

Frequent landslide hazards have caused serious threats to the lives and properties of the nation,and rapid localization and accurate identification of landslides are essential for timely disaster assessment and post-disaster rescue,etc.Traditional landslide detection methods rely on expert knowledge with high accuracy but low efficiency.Landslide detection methods based on Deep Learning and image recognition technology automatically learn features related to landslide hazards with the advantage of high efficiency and accuracy.However,the existing research mainly focuses on detecting landslides in specific regions,and the research on crossdomain landslide detection is insufficient.To address this problem,this paper selected many areas in China where landslides have occurred as the research objects and carried out the research of cross-domain landslide detection based on CNN(Convolutional Neural Network)for optical remote sensing images:To meet the localization and segmentation requirements of landslide detection tasks,the YOLOX target detection algorithm and U-Net semantic segmentation algorithm were introduced for the fast localization and accurate segmentation of landslides.First,the transfer learning method is introduced to alleviate the big data dependency of the model and solve the problem of lacking cross-domain landslide research samples.Then,the Google Earth software was used to extract the optical remote sensing images of 40 landslide areas in China to make landslide samples and combined with the open-source Bijie landslide dataset,a cross-domain landslide dataset with rich samples and various types was constructed.In addition,optical remote sensing images of about 28.72 square kilometers of the cluster landslide area around Mibei village in Longchuan County,Guangdong Province,were extracted for landslide detection practical application.To solve the problem of unbalanced and high similarity between landslide samples and negative samples in cross-domain landslide detection research and the difficulty of identifying special landslides,the improved YOLOX and Improved U-Net networks were proposed.For the Improved YOLOX network,the original target score loss function is replaced by the Varifocal function,and the weights are adaptively adjusted according to the importance and recognition difficulty of each type of sample to improve the classification performance of the network and enhance the robustness.The CA attention module is added to the network’s backbone to focus on important information and improve the localization accuracy of landslides.For the Improved U-Net network,the Res Net network with higher model complexity is used as the encoding structure of the model to enhance the recognition ability of landslide features.The Focal function replaces the original classification loss function to suppress the loss of easily classified samples and focus on difficult samples to improve the accuracy of landslide recognition.The experimental results of cross-domain landslide detection show that the Improved YOLOX network shows stronger recognition ability and more accurate localization for small and complex landslides compared with YOLOX,Faster R-CNN,SSD,and YOLOv5 networks.The ablation experiments using VGG16,Res Net50,and Focal loss function show that the Improved U-Net network performs better overall segmentation with enhanced recognition of multiple types of complex landslides.The detection results of the Mibei landslide region show that the two landslide detection methods proposed in this paper are effective,generalizable,robust,and can achieve fast localization and accurate segmentation of landslides.In addition,it is verified that the transfer learning method can significantly shorten the training period of the network and improve the generalization and robustness of the model.This research work solves part of the problems in cross-domain landslide detection methods,confirms the importance of the reasonable selection of sample scaling in landslide segmentation methods,discusses the advantages and shortcomings of the detection methods of landslides based on optical remote sensing images,and provides a technical reference for the research of deep learning cross-domain landslide detection methods.