Landslide Susceptibility Evaluation Based on the Combination of Environmental Similarity and BP Neural Networks

Abstract

Landslides represent a widespread global geological hazard, presenting significant risks to both human populations and critical infrastructure. The accuracy of landslide susceptibility evaluation models serves as a critical prerequisite for landslide hazard prediction and risk management, while insufficient landslide sample data may constrain the reliability of susceptibility modeling and evaluation results. To address the challenge of limited landslide samples in complex mountainous areas, this study proposes a novel landslide susceptibility evaluation method integrating environmental similarity theory with a backpropagation neural network (Environmental Similarity Model–BP Neural Network, ESM-BP). Taking the Baihetan reservoir area as the study region, the environmental similarity degrees between potential prediction points and historical landslide samples were calculated using eight environmental factors. A normal distribution approach was employed to classify similarity thresholds, thereby constructing an enhanced landslide sample dataset. The BP neural network model was subsequently applied for susceptibility assessment, with comparative validation against support vector machine (SVM) and random forest (RF) models. The experimental results demonstrate that (1) the integration of environmental similarity theory effectively expanded the dataset by 4398 samples with distinct susceptibility levels, resolving data scarcity issues and significantly enhancing the model’s generalization capabilities. (2) Among the three models tested with supplemented samples, the BP neural network achieved optimal performance, showing improvements in the accuracy values by 0.02 and 0.14 compared to SVM and RF, respectively, Kappa coefficient enhancements of 0.02 and 0.18, and RMSE reductions of 0.04 and 0.21. This methodology enhances the applicability and reliability of landslide susceptibility evaluation models in complex mountainous environments, providing innovative insights for related research in landslide susceptibility assessment.