Assessment of the effects of characterization methods selection on the landslide susceptibility: a comparison between logistic regression (LR), naive bayes (NB) and radial basis function network (RBF Network)

Abstract

Landslides are natural disasters that are difficult to control without continuous monitoring. Xiji County is located in the southern mountainous area of Ningxia Hui Autonomous Region, where geological and ecological conditions are complex and the number and extent of landslides hinder local economic development. To address this, a comprehensive landslide inventory was created, comprising 529 historical landslides and an equal number of non-landslide points. Thorough analysis of these datasets ensured an unbiased assessment. The data was randomly divided into training (70%) and validation (30%) sets. Using 15 spatial datasets, including elevation, slope, curvature, distance to various features, rainfall, land use, lithology, and maximum ground acceleration, a system for landslide susceptibility evaluation was established with 12 influential indices. The frequency ratio method was applied to analyze the relationship between landslides and each index. Three evaluation models (LR, NB, and RBF Network) were built, utilizing different landslide characterization methods (landslide point and landslide polygon), resulting in six result maps for landslide susceptibility evaluation. Statistical analysis of frequency ratios in susceptibility class intervals ensured model rationality. The NB model based on landslide polygons showed optimal performance with high success rate (AUC = 0.965), prediction rate (AUC = 0.886), consistency (FRA = 0.873). This methodology and landslide susceptibility map provide decision-making support for researchers and local governments in mitigating future geological hazards.