Evaluation Of Landslide Susceptibility In Southeast Tibet By Machine Learning Algorithms

Landslide is a major geological disaster,its risk level is slightly lower than that of earthquake,more often than not developed in mountainous areas,and landslide failures take place often in southeast Tibet,in particular in the Yarlung Tsangpo River and Niyang River banks are the most concentrated,the Yarlung Tsangpo River and Niyang River banks are geologically active,landslides regularly occur,which will now not solely motive irreparable have an impact on the property and protection of neighborhood residents,however additionally frequently motive gorgeous harm to the surrounding countries.Landslide susceptibility evaluation can correctly minimize the damage to human existence and property brought about via the incidence of disasters.In this paper,we find out about the topography and geomorphology,meteorology and hydrology of the southeast Tibetan vicinity for analysis,particularly to explore the landslide improvement traits on each aspect of Yarlung Tsangpo River and Nyanang River,where landslide mess ups are most frequent.Based on the historic landslide information and subject survey results,we combine the computing device gaining knowledge of mode mannequin in Arc GIS to habits the following studies.(1)Landslide data,topography and geomorphology,meteorology and hydrology,stratigraphic lithology,geological structure,human things to do and different records had been accrued and processed in southeast Tibet,and landslide dangers and formation stipulations had been analyzed in southeast Tibet.The consequences confirmed that landslide dangers had been determined in elevation 32-1544 m,topographic alleviation 0-20 m,slope 80-100°,slope route southwest,curvature-2-2,distance from river0-1000 m,The wide variety of occurrences is excessive between 0.43-0.65 NDVI,inside 2000 m distance from the fracture zone,inside 500 m distance from the road.(2)Based on Python 3.6 and R language,the weighted random forest,XGBoost,and Light GBM algorithms primarily based on the Gini coefficient were used.188 landslide samples were considered,and seven influencing elements of elevation,slope,stratigraphic lithology,land use,fracture quarter and fault,river,and avenue were selected.The function selection algorithm used Bayesian,and after optimizing the hyperparameters via grid search and five-fold cross-validation method,the prediction model was developed based on 70%(131 samples)of the training set.A test set of 30%(57 samples)was used to confirm the feasibility of the training models.(3)The landslide susceptibility effects of the three fashions confirmed some differences,however there was an average convergence.The GiniRF,XGBoost,and Light GBM models all had the highest share values in the very low category,with very high susceptibility areas of 11.99%,12.05%,and 12.14%,respectively.(4)respectively.The outcomes of every category were analyzed using the precision,recall,and F1 metrics,and the overall performance was evaluated by accuracy.The results showed that the Gini-RF,XGBoost,and Light GBM algorithms had higher overall performance,with prediction accuracies of 0.8026,0.8251,and 0.8256,respectively.In particular,the XGBoost model performed the best among all models with an AUC rate of0.8432,accuracy of 0.8531,F1 score of 0.8345,and very high and high sensitivity of 12.14% and 12.41%,respectively.The validation with two landslides in the study area shows that the model is highly reliable and the landslide zoning map can provide guidance for disaster prevention and mitigation activities of relevant local departments.(5)Two landslide web site surveys,Qiangna Baga landslide and Murdoch County street landslide,have been chosen for contrast and validation.Both landslides are in high landslide susceptibility areas,which again verifies the accuracy of the zoning of the machine learning model in this paper.The research results can be used as reference for regional landslide control related departments.