Landslide Susceptibility Assessment Based On Machine Learning At Long Time Scale ——A Case Study Of Lixian County

The phenomenon of geological disasters in Gansu Province is relatively serious,especially landslide disasters(including debris flow,collapse),which have the characteristics of large quantity,wide distribution,frequent occurrence and strong destructiveness.Longnan area is a region with frequent geological disasters in Gansu Province,and Lixian County is one of the counties with the most developed landslide disasters,so it is necessary to study the landslide disasters.The assessment of landslide susceptibility is the basic work of landslide prevention and control.Nowadays,the application of machine learning model in this field is more and more in-depth,but the suitability degree and assessment results of different models in landslide susceptibility assessment are different.Therefore,this study took Lixian County,Longnan City,Gansu Province as the research area,and took elevation,aspect,slope,distance to river,road and fault zone,geology lithology,soil type,land use and precipitation as influencing factors.Four kinds of machine learning models(Decision Tree,Random Forest,XGBoost and Cat Boost)were used to evaluate the landslide susceptibility in the study area,and the accuracy of assessment results of each model was improved through parameter adjustment.Since the assessment of a single time node cannot reflect the change of landslide susceptibility on the time scale,and the impact of land use on landslide susceptibility on the time scale is particularly significant,this paper selects the land use data with obvious changes in three periods as a time benchmark to assess landslide susceptibility on a long time scale.The main conclusions are:(1)The classification results of the Decision Tree model with the lowest accuracy on the susceptibility of the whole study area are more inclined to the dichotomies,and when the accuracy is improved by means of parameter tuning,the generalization degree of the model results is reduced.However,when the accuracy of the random forest model with the highest accuracy is improved,the zoning results are more generalized,and the area proportion of the moderately prone zoning increases,which is more consistent with the actual situation of the study area.Combined with the results of the test,this paper believes that the random forest model is more suitable for the situation of the study area,and the results obtained also show better spatial consistency compared with the actual situation.(2)Feature is a very important part of machine learning classification and prediction process.By judging the performance of feature,it can be found that the decision tree model underestimates the importance of distance to road.In XGBoost model,the effect of elevation factor is emphasized,while the effect of slope direction and distance to river is weakened.Elevation factor is the common factor that the landslide susceptibility index is overestimated in the three models.Cat Boost model weakens the effect of distance to road and emphasizes the contribution of elevation and rainfall factors to the model.(3)The assessment results show that the very high and high susceptibility areas of landslide disaster in Lixian county are mainly distributed around the low-altitude watershed(1073--1611m)where the western han river is the main stream,and the low-susceptibility area is highly coincident with the high-altitude area in the west of the study area.The overall susceptibility index was higher in the east and lower in the west,higher in the north and lower in the south.From the perspective of time change,the landslide susceptibility index in the study area decreased from 1980 to 2018,and the area of extremely high and high landslide susceptibility area decreased.However,the regional index with low landslide susceptibility index increased from 2000 to2018.(4)Natural vegetation in the study area is not conducive to the development of landslide disaster,such as patches of forest land,shrub land,open forest land and grassland with medium and high coverage.However,the possibility of landslide in the bare rock land and desert tundra area with high altitude,sparse population and single land use type and lithologic structure is also low.The low coverage grassland with low coverage and poor root strength falls in the areas with high and high susceptibility.Similarly,the land use categories mainly located in the areas with high and high susceptibility include river canals,beaches,reservoirs,ponds,urban and rural residential land,paddy fields and dry lands.The slope stability of these land use areas is poor,Landslide is more likely than other areas.(5)In Lixian county,human activities have a two-way effect on the occurrence of landslides.In order to ensure the safety of people's life and property,disaster prevention and mitigation measures were implemented in the township areas near the main stream of the Western Han River(Yanhe Town,Shiqiao Town,Wangba Town,etc.)in the region.Therefore,the area of the zone with higher landslide susceptibility index decreased.In other areas,the transformation of surface and slope by human activities increases the susceptibility index of landslide and the possibility of landslide.