Research On Meteorological Early-Warning For Landslide Risk In Wanzhou District, Three Gorges Reservoir

China has complex geological formations and diverse topography,with landslides and other geological disasters showing a long-term high-incidence trend and causing severe casualties and property loss.Meanwhile,climate change with corresponding extreme rainfall and rapid urban expansion can foster landslide disasters and cause severe damage in areas that are traditionally considered relatively safe from such hazards.In response to the practical needs of disaster prevention and mitigation,a breakthrough should be made from the current research on landslide stability,mechanisms and control measures,and carry out landslides early warning through observation and analysis of extreme meteorological conditions.However,the main drawback of the regional early warning models based on empirical rainfall thresholds is poor spatial resolution.Meanwhile,the empirical and probabilistic thresholds obtained from specific rainfall events ignore the physical background of slope instability,leading to a relatively lower performance with high false alarms.In addition,few studies have attempted to extend the purpose of early warning from landslide detection to potential landslide risk prediction.The thesis takes rainfall-induced landslide in the Wanzhou District of the Three Gorge Reservoir area as the research object.Landslide development patterns and their corresponding rainfall event were analyzed first.And then,based on geological engineering principles,statistical theory,machine learning methods,GIS technology and Python programming,depth exploration on landslides susceptibility mapping,rainfall threshold analysis,landslide spatiotemporal probability modelling,vulnerability analysis of the element at risk and value estimation has been conducted step by step.The research aims to integrate the result of landslide risk assessment into the early warning process and establish a landslide meteorological early-warning model with risk evaluation capability.The main contents and results are as follows:(1)Characterizing the development pattern of landslides and their corresponding rainfall event.A landslide inventory map,including 1448 landslides affected by rainfall,was compiled to analyze their spatial and temporal distribution patterns and noncumulative density-size distribution.Meanwhile,geological and topographical factors related to widespread deformation and failure were studied.Then,the characteristics of rainfall events related to landslides were summarized.The main results are shown as follows: Landslides mainly occurred in the central,north-western and southeastern areas of the Wanzhou District,in which the base was mainly composed of Jurassic sand and mudstone interbedded.The elevation and mean slope also contributed to the occurrence of landslides.In addition,compared with rainstorms on the day of landslide initiation(0 ～ 20 mm and 20 ～ 40 mm),the antecedent rainfall with different time intervals was more significant,with an average value of 25 mm,65 mm,85 mm and 200 mm,respectively.(2)Landslide susceptibility mapping based on improved sample selection strategy.In this chapter,the general process related to landslide susceptibility mapping was summarized first.Based on that,an improved sample selection strategy named slope-unit control sampling(SCS)was proposed to address the subjective and random selection of non-landslide samples.The slope-unit control sampling and Random Forest algorithm(SCS-RF)were combined for landslide susceptibility evaluation.Then,the uncertainty during susceptibility analysis was identified,and the landslide susceptibility confidence map(LSCM)based on the coefficient of variation and susceptibility was proposed to quantify its magnitude.The main results are shown as follows: Areas with high and very high susceptibility were mainly located in Dazhou,Xiaozhou and Tiancheng towns in the north part of the study area and Xintian to Changping towns along the Yangtze River.Moreover,the prediction accuracy of the SCS-RF model is higher than that of the integration of the normal sample selection strategy and random forest model(NS-RF).In addition,the robustness of the SCS-RF model tended to be higher in the area with high and very high susceptibility,and the prediction areas with higher uncertainty were primarily concentrated in the low and moderate susceptibility areas.The proposed LSCM can effectively identify the uncertainties due to the difference from input datasets,and the evaluation results may be especially useful for decision-makers,providing support for territorial spatial planning and landslide risk mitigation.(3)Integration of antecedent rainfall to improve the performance of 3D rainfall threshold for landslides early warning.A total of 788 landslide records and hourly data from 78 rainfall gauges spanning 2014 to 2020 were used for statistical analysis.The automated algorithm Massive Cumulative Brisk Analyzer(Ma Cum BA)was utilized to establish the intensity–duration(I-D)threshold.To extend the thresholds from plane to spatial,the mean effective areal rainfall(MEAR),representing the average rainfall amount infiltrated into the ground,was introduced as the third dimension.The main results are shown as follows: The accuracy of the I-D thresholds ranges from 58% to 92%,while the number of false detections is high with low precision(from 11% to 39%).The participation of MEAR satisfies this shortcoming with a consistent decrease in false alarms(from 24% to 95%).Moreover,compared with the 3D thresholds without reduction,the I-D-MEAR thresholds showed good performance and robustness that could ignore the negative influence of long-interval antecedent rainfall.(4)Research on meteorological early-warning for landslide risk in Wanzhou District.Based on the rainfall threshold proposed in chapter 4,the temporal probabilities of landslide occurrence in different I-D-MEAR intervals were calculated via the Bayesian approach,and the critical rainfall condition for different warning levels was determined.Then,landslide risk with different warning levels was evaluated.For this purpose,landslide hazard zoning(spatiotemporal probability modelling)under different warning levels was conducted,and the potential loss was quantified by combining the vulnerability and value distribution of the element at risk.Finally,some specific rainfall events were used as an example for model validation,which also clearly illustrated the operation process of the warning model.The main results are shown as follows: The population casualties and economic risks increase with the warning level.In detail,alert zone F includes the urban territorial of the study area,and alerts zone B and D,in which the base was mainly composed of Jurassic sandstone and mudstone or thin shale interbedded with typical dip slope,have a high potential for population casualties and economic losses.The meteorological early-warning model based on the Bayesian approach and risk evaluation theory has a good performance in practice.