Research On Deep Learning Methods For Identifying Potential Landslides And Assessing Susceptibility In Luding County

China is one of the countries where landslides are most developed.Landslides are characterized by many hiddenness,complexity,suddenness,substantial spatiotemporal uncertainty,dynamic variability,and many potential dangers,seriously threatening people’s lives and property.Luding County is located at the junction of the eastern edge of the Qinghai-Tibet Plateau and the Sichuan Basin,with complex geological structures,intense terrain cutting,developed active fault structures,and fragmented rock and soil bodies.It is the frequent landslide area of China’s high mountains and canyons.The area is prone to frequent major and catastrophic earthquakes and has prominent regional and locally heavy rainfall,making landslides a long-term and frequent occurrence,with a serious situation in terms of prevention and control.Currently,deep learning algorithms and ground observation techniques are gradually being applied to identify potential landslide dangers and predict susceptibility,achieving some results.However,problems still need to be solved,such as noise or missing data,scarce sample label data in the data set,difficulty fusing multi-modal data,and weak model interpretability.How to deeply mine the multi-modal data of landslides and combine various Synthetic Aperture Radar Interferometry(InSAR)technologies,image processing techniques,and deep learning methods to achieve intelligent identification of landslide hidden dangers and susceptibility prediction in complex geological and environmental scenarios is an urgent problem that needs to be solved.This thesis presents a comprehensive approach that combines multi-temporal highresolution optical remote sensing images and synthetic aperture radar(SAR)data to investigate the technology of multi-temporal interferometric SAR(InSAR)deformation monitoring,as well as convolutional neural network(CNN)and recurrent neural network(RNN)algorithms.The aim is to explore methods for early identification and dynamic prediction of landslide,and to establish an integrated research model that spans from data acquisition and mining,algorithm development and testing,to system design and development.This model will aid in advancing the identification and risk assessment of landslide hazards in Luding County,and in constructing a dual-control pattern for geological disaster risks.The main research content and achievements include the following:(1)Conducting a comprehensive analysis of the factors influencing landslides and their spatial differentiation characteristics.Geographic detectors were used to investigate the spatial differentiation characteristics of landslides and assess the impact of single and double-factor interactions on landslides.The optimal order of landslide influencing factors was determined,including slope,aspect,curvature,NDVI,distance to rivers,landform,Topographic Wetness Index(TWI)Slope Length and Slope Steepness(LS),lithology,elevation,land use type,annual average rainfall,distance to faults,distance to epicenters,and distance to roads.(2)Combining InSAR technology and high-resolution optical remote sensing images to create landslide remote sensing interpretation labels.The types of landslides and their stages of deformation were clarified,and multi-modal optical remote sensing data and multi-temporal InSAR technology were used to conduct InSAR monitoring of landslide hazards in Luding County.The characteristics of landslide hazard changes were understood,a multi-type remote sensing interpretation system was constructed,and the landslide hazard area in Luding County was demarcate.(3)Constructing a hybrid labeled landslide dataset and designing a weight transfer function for active learning mask features to address the issue of insufficient training sample data with masked annotations and the difficulty of identifying landslide hazards with small scales and unclear features in remote sensing images.The Mask scoringRCNN algorithm was improved,and intelligent identification of landslide hazards was conducted at different scales,multiple levels,long time series,high accuracy,and intelligence.The experimental results indicate that the algorithm can accurately identify and segment landslide hazards with only a small number of landslide mask label samples,particularly in small and medium-sized landslides.(4)Proposing a bi-directional long-short term memory(BiLSTM)neural network algorithm based on landslide density fusion of high-dimensional features and sampling optimization strategy to achieve dynamic and accurate prediction of landslide susceptibility and mapping of landslide zones.The high-dimensional features fusion is based on the contribution value order of the interaction of landslide impact factors to the spatial differentiation of landslides.A sampling optimization strategy is designed to independently sample the training sample area and the model prediction area.The landslide hazard density is used as the label for the positive samples in the dataset,which are gradually expanded in batches.The results show that compared with the BiLSTM model with landslide objects as sample labels,the BiLSTM algorithm based on landslide hazard density performs the best when the number of positive and negative samples tends to be the same(accuracy=0.903,recall=0.899,F1 score=0.901,AUC=0.940),and performs better than the BiLSTM algorithm based on landslide objects.At the same time,the performance of the algorithm(AUC=0.9407)is significantly better than that of the information value model and random forest algorithm,and it can effectively extract landslide features and accurately predict landslide susceptibility in complex scenarios.(5)Designing and developing an integrated system for landslide hazard identification and dynamic susceptibility prediction.The system is developed using the Python language’s PyQt5 framework and Qt Designer software,providing an integrated display interface.The system integrates processes such as landslide surface deformation monitoring,landslide hazard identification,and susceptibility analysis,and provides an intuitive human-computer interaction interface for users to display and analyze the results of various tasks,perform real-time comparison,and annotation processing.