Landslide Identification And Prediction With The Application Of Time Series InSAR

Landslides is one of the most important geological disasters.It threatens the safety of human being's lives,as well as environment,resources and property.In China,huge economic losses and casualties were caused by frequent geological disasters every year,for example,the Maoxian landslide happened in 2017 caused more than 100 deaths.Affected by the periodic scheduling of reservoir water level and heavy rainfall,a large number of landslide disasters have developed in the Three Gorges Reservoir area(TGRA).According to the demand of the disaster prevention and reduction work in the TGRA,it's of great theoretical and practical significance to carry out effective early identification,hazard assessment and prediction and early warning of landslides.Synthetic Aperture Radar Interferometry(InSAR)is a novel technique for crustal deformation monitoring.Compared with conventional methods,it has the advantages of wide range,high density and accuracy,and can work in all weather conditions.InSAR provides a new technique for landslide research.However,how to better combine landslide research with InSAR,and effectively improve the accuracy and timeliness of landslide early warning and risk assessment is still an urgent problem to be solved.Landslides in the TGRA was taken as case studies.The time series InSAR technique was applied to extract the crustal deformation velocity of TGRA using Sentinel-1 images.With the integration of engineering geology and machine learning techniques,the research of landslide early identification,dynamic hazard assessment and deformation prediction were carried out.The main contents and research results are shown as follows:(1)Landslide early identification using time series InSAR.The crustal deformation velocity of Wanzhou-Wushan section in TGRA was obtained applying time series InSAR and Sentinel-1 images.The velocity in line of sight(LOS)direction was projected into the maximum slope direction,and points with abnormal velocity value were eliminated.The Getis–Ord Gi* statistic and kernel density estimation methods were used to analyze the hotspot of deformation to achieve the initial positioning of landslides.The final landslide identification was realized by combining with the InSAR deformation velocity value.In this study,50 active slopes were identified,including 40 known landslides and 10 unknown landslides.Through landslide catalog data analysis and field investigation,it was found that InSAR is a reliable deformation monitoring technique,which can effectively identify landslides.(2)Regional landslide susceptibility assessment applying machine learning technique.Reservoir bank area of Wushan was taken as a case study,the relationship between the landslide influencing factors and its development was quantitatively analyzed using information value model.The variance inflation factors and tolerances were applied to test the multicollinearity among the factors,and the unimportant factors were selected and eliminated using the information gain ratio method.Support vector machines(SVM),artificial neural networks,logistic regression and classification regression tree models were used for susceptibility modeling.The results show that,the SVM performed the best.The factors of distance to river and Lithology affect the most in the development of landslides.Eliminating the unimportant factors can improve the accuracy of landslide susceptibility modeling.(3)Landslide dynamic risk assessment with integration of time series InSAR technique.The landslide catalog and rainfall data was statistically analyzed,and the probability of landslide occurrence in specific rainfall-inducing events was calculated.The hazard assessment was conducted based on the result of susceptibility modeling.The crustal deformation velocity of the study area in different periods were extracted with application of InSAR.With the combination of the initial hazard assessment results and the extracted deformation velocity information,the evaluation of dynamic hazard was carried out.The results show that,landslide hazard varied apparently in different periods.Through dynamic monitoring of crustal deformation with InSAR,the dynamic landslide hazard assessment was achieved.Moreover,the false positive and false negative errors can be reduced by the proposed method in landslide hazard assessment.(4)Spatiotemporal deformation analysis of the reservoir bank landslides based on time series InSAR technique.Taking Shuping landslide and Muyubao landslide as examples,the time series InSAR was applied to extract the deformation velocities of the two landslides during images acquisition period using Sentinel-1 images from March 2016 to September 2017.It is found that,the Shuping landslide deformed mainly during the decline period of the Three Gorges reservoir(May-July)with step-like cumulative displacements,the larger deformation area of which is located in the middle and right side of the landslide,and the displacements of the main sliding zone is synchronous;the cumulative displacement of the Muyubao landslide is linear in time,the deformation of which is larger in the right rear zone,followed by the middle sliding body,and the left sliding body deformed the least.(5)Landslide displacement prediction based on machine learning technique and triggering factor analysis.In this study,the Shuping and Muyubao landslide were taken as examples.The displacement sequence extracted by InSAR was decomposed into trend term displacement,periodic term displacement and noise using wavelet analysis,and the influencing factors of each item were selected.Artifcial bees colony-kernel extreme learning machine(ABC-KELM),extreme learning machine and support vector machine models were used for prediction respectively,the predicted total displacement was obtained by adding both the predicted items.From the prediction accuracy assessment,it can be found that,the three models show good prediction performance and can accurately predict the deformation of different landslide types.The ABC-KELM with consideration of triggering factors has the highest prediction accuracy and stable performance.