EVALUATION OF LANDSLIDE SUSCEPTIBILITY BASED ON FREQUENCY RATIO AND ENSEMBLE LEARNING—TAKING THE BATANG-DEGE SECTION IN THE UPSTREAM OF JINSHA RIVER AS AN EXAMPLE

The Batang-Dege section in the upstream of Jinsha River is located in the east of Qinghai-Tibet Plateau,where the geology,terrain and landform are extremely complex and the landslide hazards are developed well. The analysis of the regional landslide disasters susceptibility is significant to the landslide disaster prevention and mitigation. Taking the Batang-Dege section in the upstream of Jinsha River as the research area. It is based on the landslide record and field surveys. The 11 impact factors include the elevation,slope,aspect,curvature,relief amplitude,degree of surface cutting,surface roughness,stratum lithology,fault,road and water system. They are used to construct a landslide susceptibility evaluation system by analyzing the distribution law and influencing factors. The Pearson coefficient is calculated to remove the high-correlation impact factors. The frequency ratio method is used to analyze the relationship between each factor and the landslide development quantitatively. The frequency ratio model is applied to selecting non-landslide samples and the ensemble learning model is used to evaluate landslide susceptibility. According to the index of susceptibility,the search area is divided into five levels including extreme-highly susceptible area,highly susceptible area,moderately susceptible area,low susceptible area,and extreme-low susceptible area. As shown in the susceptibility divisional graph and ROC curve,the extreme-highly susceptible and highly susceptible area areas are mainly distributed along the banks of Jinsha River and ravines. The area under curve of success ratio of the Random Forest model is 0.84,the number of disasters located in extreme-highly and highly susceptible areas accounted for 84.8 percent of the total landslides. The area under curve of success ratio of the Gradient Boost Tree model is 0.79,the number of disasters located in extreme-highly and highly susceptible areas accounted for 79.3 percent of the total landslides. By the value of AUC and the distribution of historical disasters,it can be observed in the research area that the Random Forest model has better evaluation accuracy and higher prediction ability in landslide susceptibility evaluation than the Gradient Boost Tree model.