Landslide Susceptibility Evaluation Based On Spatial Database And Machine Learning

China has a vast territory and is a country with frequent geological disasters,especially in mountainous areas.Among them,landslide disasters are the main part of many geological disasters,and they pose a threat to human life and property.They also cause great damage to infrastructure and seriously affect economic and social activities.The study area is located in Xinyuan county of Xinjiang,China.The special and complex topography of Xinyuan County,coupled with abundant rainfall and widespread distribution of loess,make geological disasters frequent.Xinyuan County mainly focuses on agriculture,animal husbandry,and tourism services.Frequent geological disasters not only endanger the lives and property of local residents,threaten the safety of visitors,but also seriously affect the county’s economic development.Therefore,the study of the landslide susceptibility in Xinyuan County can provide the necessary scientific foundation for monitoring of landslide disasters and reducing economic losses.In this paper,we discussed and analyzed the distribution of different landslide influence factors and the comparison of landslide susceptibility with different machine learning models,and obtained the following conclusions:(1)We picked 16 landslide factors,including slope,aspect,elevation,plane curvature,profile curvature,standard curvature,distance to faults,distance to rivers,distance to roads,NDVI,lithology,annual average rainfall,loess distribution,land use,TWI and TPI.Due to the imbalance problem,we used random sampling to make the number of non-landslide samples as same as the number of landslide samples.The relationship between landslide influence factors and landslides and the distribution rules of landslides are summarized: landslides are concentratedly distributed at elevations of 1395-1732 meters,slopes of 13.759 °-27.38 °,annual average rainfall of550-650 mm,and NDVI in the range of 0.231-0.352;and are distributed In the loess region,woodland,grassland,and clastic rock groups dominated by weak interbedded sandstone,mudstone,and conglomerate.For distances from roads,distances from faults,and distances from rivers,the closer the distance,the more likely a landslide will occur.(2)Based on the three models of logistic regression,random forest and Cat Boost,the contribution of landslide impact factors in Xinyuan County was analyzed.In the analysis results of the logistic regression model,NDVI,distance to rivers,slope,elevation,distance to faults,annual rainfall and profile curvature had an important effect on landslides,of which NDVI and distance to rivers had the greatest influence.For the landslide susceptibility evaluation results of random forests,elevation was of great importance to the occurrence of landslides,followed by distance from faults,slopes,loess distribution,and average annual rainfall respectively;and in the Cat Boost landslide susceptibility evaluation results,factor importance rankings were: elevation>slope> distance to faults> distance to roads> average annual rainfall.In summary,slope,elevation,distance to faults,and average annual rainfall were identified as important contributing factors in the three models,however the contributions of NDVI,distance to rivers,profile curvature,loess distribution,and distance to roads were different.(3)This paper compared and analyzed the landslide susceptibility in Xinyuan County of logistic regression,random forest,and Cat Boost,and draw the following conclusion: the three models had good prediction ability.On the training sets,the AUC values of the logistic regression,random forest,and Cat Boost model success rate curves were 0.917,0.939 and 0.997,respectively.For the testing sets,the AUC values of the logistic regression,random forest,and Cat Boost model prediction rate curves were0.900,0.918 and 0.920,respectively.Cat Boost was slightly better than logistic regression and random forest model.The three models were reliable in the classification of susceptibility mapping,and the results of high and very high levels of the total area percentage were basically the same.The result of logistic regression in the high and very high grades as a percentage of the total area was 23%,and the other two models were 22%.The logistic regression model was the most common model for landslide susceptibility evaluation.The Cat Boost model has a simple and fast process and reliable classification,which not only ensured the prediction accuracy of the susceptibility analysis,but also retained the details of the regional susceptibility mapping.Therefor the Catboost model provides a new approach for regional landslide susceptibility mapping(4)SAVI,MDNWI,and NDBI were calculated based on high-resolution Sentinel-2 remote sensing data,and new index image are combined.Further,the new index image combined with POI data to extract building land information.The research results showed that the combination of POI data and index images can effectively extract building land information.In addition,the extracted construction land information and the landslide susceptibility map were performed an overlay analysis,we found that the tourist attractions of Nalati Scenic Spot,Kansu Village,Karasu Village,Karasu Village,Dongfanghong Farm,and other tourist attractions in Xinyuan County,Villages and farms were potential landslide impact areas.Provide scientific basis and support for the monitoring of local potential landslide affected areas,and provide decision-making reference for disaster response departments.