Landslide Susceptibility Assessment In Large Range Based On Deep Learning

Landslides are a common geological hazard,which often causes huge casualties and property losses.Landslide susceptibility assessment is to determine where landslides are most likely to occur within a certain range based on the analysis of historical landslide occurrence conditions,and is a key technical tool for geological disaster prevention and mitigation.In recent years,more and more research has been conducted on landslide susceptibility assessment,but there are still some problems:(1)landslide susceptibility assessment involves many data types,and the processing process is complex and time-consuming.(2)Landslide susceptibility assessment based on deep learning is still in development,especially lacking of application in complex geological scenarios spanning a wide range of multiple geomorphological units.At the same time,due to the different landslide influencing factors have different contributions to landslide occurrence.How to adaptively weighted to improve the performance of the model.(3)The landslide susceptibility assessment based on deep learning often depends on raster units,while the slope units are the basic units for the development of geological hazards.There are few studies on how to obtain the results of landslide susceptibility assessment based on the slope units.The Qinghai-Tibet plateau transportation corridor is an important part of the 13 th Five-Year Plan of China.The planning and construction of the Qinghai-Tibet plateau transportation corridor is of great and far-reaching significance to the economic and social development of the Tibet Autonomous Region,Sichuan Province and other regions in western China.Therefore,the Qinghai-Tibet plateau transportation corridor was selected as the study area.On the basis of comprehensive analysis of historical landslide data and geological environment conditions in the study area,the landslide susceptibility in this area was evaluated by using deep learning technology.The corridor passes through the most complex geological,topographical and geomorphic areas in the world.Landslides are one of the most serious natural disasters in this area.It is of great significance for railway engineering safety to carry out landslide susceptibility assessment in this area.The main research results are as follows.(1)A toolbox for multi-source data preprocessing and dataset production that can be integrated into Arc GIS platform has been developed(https://github.com/Huang WBill/SVM-LSM-Toolbox).The toolbox basically covers the pre-preparation of landslide susceptibility assessment.It includes the most commonly used influencing factor production tool such as topographic factor calculation,line vector data and nc4 rainfall data format conversion,as well as non-landslide data generation,data sample production and split,and influencing factor selection.The Pearson Correlation Coefficient(PCC)and Information Gain Ratio(IGR)results show that there is a high correlation between elevation and rainfall,slope and relief amplitude,slope and surface roughness,relief amplitude and surface roughness,and rainfall,profile curvature,plane curvature and NDVI are of high importance for landslide occurrence.(2)A deep learning model that can take into account the importance of influencing factors,namely the Conv-SE-LSTM model,is proposed and compared with traditional support vector machine(SVM),convolutional neural network(CNN),and long and short-term memory network(LSTM)models.The results show that CNN model has good performance but poor stability,SVM and LSTM models have good stability but poor model performance,and the proposed Conv-SE-LSTM model has the best performance and robustness for a large range of complex geological conditions.The high and very high landslide susceptibility areas along the Qinghai-Tibet plateau transportation corridor are mainly located on both sides of the Yalong River,Jinsha River,Lancang River,Nu River,Yarlung-Zangbo River and the highway near Chengdu.(3)The slope units are divided based on r.slopeunits,and the most reasonable slope unit division result was selected by using the comprehensive evaluation index S(a,c).The index is calculated from terrain segmentation metric F(a,c)and the susceptibility map accuracy metric U(a,c).The average of the susceptibility index of all raster units contained in each slope unit was used as the susceptibility index of the slope unit to generate the optimal landslide susceptibility map based on the slope unit.The results show that when the minimum area threshold a is 150000 and the minimum square error threshold c is 0.5,the result of slope units is the optimal,with a total of 552,259 slope units.The susceptibility map based on the slope unit can more accurately represent the landslide susceptibility in the study area.