An optimal sample selection-based logistic regression model of slope physical resistance against rainfall-induced landslide

Abstract

Due to the insufficient ability of the slope to resist deformation and frequent rainfall, geological disasters often occur on the mountain slopes, especially rainfall-induced landslide. However, the method of assessing the ability of mountain slopes to resist rainfall-induced landslide is still not available. Therefore, using a typical example, this study mainly focuses on developing a mapping model for disaster resistance of rainfall-induced landslide in mountain slopes taking data mining of historical slope damage records and conditioning factors into account. To be specific, specifically, 18 assessment factors, elevation, slope gradient, slope aspect, slope position, plane curvature, section curvature, comprehensive curvature, terrain roughness, CRDS, terrain wetness index, sediment transport index, stream power index, micro-landform, lithology, average annual rainfall, NDVI, distance from faults, and distance from streams, are selected as the evaluation factors for the mountain slope disaster resistance. Afterward, a geospatial database is built to assess the disaster resistance of mountain slopes by combining 477 historical rainfall-induced landslides in the study area. Then, a sample dataset is selected with the negative and positive sample ratio of 1:1. An optimal training sample is selected by tenfold cross-validation, while the testing sample is randomly selected from the sample dataset comprising 30% of the sample dataset. A logistic regression (LR) model is subsequently obtained by the optimal training sample, which contributes to the simulation analysis of the disaster resistance of mountain slopes throughout the study area. Finally, the simulation results are classified into five disaster resistance classes: low, relatively low, medium, relatively high, and high. It should be noted that the area of low and relatively low resistance zones accounts for only 25.06% of the total area, but 83.07% of the historical rainfall-induced landslides are located in this region. While the area of relatively high and high resistance zones accounts for 59.26% of the total area, only 7.77% of the historical rainfall-induced landslides are located there. In addition, an examination based on the receiver operating characteristic (ROC) curve shows that the area under curve values of the testing sample, training sample, sample dataset, and geospatial cells are 0.8552, 0.8924, 0.8605, and 0.8527, respectively. It is found that over the last 1.5 years, the location of 10 rainfall-induced landslides was mainly located in the relatively low and low resistance zones. Thus, it can be concluded that the LR assessment model for disaster resistance of mountain slopes, which is based on a data mining analysis of historical data has high stability and reliability in the assessment of mountain slope disaster resistance against rainfall-induced landslides.