Landslide Susceptibility Evaluation By Integrating InSAR Correction And Deep Learning

With the complex geological environment,frequent earthquakes and variable climate,the southwest region are the most serious landslide-prone and hazards-threatened region of China.To provide scientific data for landslide hazard warning and risk management,researcher often predict the probability of landslide occurrence by landslide susceptibility evaluation models.However,the evaluation models in previous studies were only based on static geographic information data and used quantitative statistical modeling to construct the relationship between occurrence probability and controlling factors,without take timeliness into account.In addition,as geological hazard data is characterized by multi-source,heterogeneity and non-linearity,it may be difficult for the traditional model to fully explore the hidden relationships in landslide systems.In recent years,multiple temporal interferometric synthetic aperture radar(MTInSAR)has become a necessary tool for wide-area geohazard identification and deformation monitoring with the advantages of high accuracy,large scale and short revisit time.Therefore,this paper takes the Wudongde hydropower reservoir area at the junction of Sichuan and Yunnan in China as the study case,and uses MT-InSAR to invert the deformation characteristics and effectively integrates it with the static landslide susceptibility evaluation results,aiming to refine the landslide susceptibility rank levels.On the other hand,the study explores the applicability and superiority of convolutional neural network(CNN)coupled with 3D samples in susceptibility evaluation.The details of the study are as follows:1)Investigate the regional landslide geological environment and construct a database of landslide inventory and disaster-causing factors by combining visual interpretation and geographic information system;2)Extract the surface deformation rate map of the study area from January 2021 to December 2022 by using MT-InSAR technology;3)Construct a CNN model incorporating 3D information and conduct landslide susceptibility evaluation;4)Fusing MT-InSAR deformation map to correct the susceptibility zoning levels.The main findings of the paper are as follows:1)3126 landslides in the study area were visually deciphered based on Google Earth,and 17 types of landslide causative factors were obtained through spatial analysis and a database was constructed for damage assessment and susceptibility evaluation.2)Based on Sentinel-1A data,the timeseries surface deformation rate map from January 2021 to December 2022 was extracted using Small Baseline Subset method.The results show that the whole area is relatively stable,with the annual average deformation rate ranging from-80 to 38 mm/y.The classification of the detected deformation areas was carried out by combining the time-series optical images,then the 229 deformation areas were classified into six different categories;3)By comparing the performance of the landslide susceptibility evaluation model based on CNN and random forest,the designed model improved 3.7%,9.2%and 3.1%in accuracy,precision and F1-score,respectively;4)The correction matrix was used to fuse the landslide susceptibility results with the deformation rate map,thus introducing time information and further improving the accuracy of landslide susceptibility evaluation rating.The overall results show that the CNN-based landslide susceptibility evaluation model can obtain smoother and more accurate results than the machine learning model,and the incorporation of deformation information into the susceptibility evaluation results can further improve the rating accuracy.The research results can provide more scientific and timely information for the reservoir operation and maintenance,and disaster prevention and control.