Random Forest Landslide Susceptibility Assessment Based on Weighted Factor and Bayesian Mean Particle Swarm Optimization

Abstract

To address the overfitting problem of the random forest (RF) model in landslide susceptibility assessment, and to improve both the model's prediction accuracy and hyperparameter optimization efficiency, a random forest model combining weighted factors and Bayesian mean particle swarm optimization (BW-MPSO-RF) is proposed. First, the contribution of landslide influencing factors is quantified by random forest feature importance and SHAP value, and a weighted evaluation index system is constructed, with higher contribution factors given higher weights. Second, mean particle swarm optimization (MPSO) and Bayesian optimization algorithm are integrated. MPSO is used to complete the global coarse search of hyperparameters, and Bayesian optimization is used to achieve the local fine search, forming a two-layer hyperparameter adaptive optimization strategy. Finally, taking Lanzhou as the study area, the research is carried out from four aspects: model construction, construction of weighted index system, accuracy verification, and evaluation result analysis. The results show that the AUC value of the BW-MPSO-RF model reaches 0.9427, and the prediction accuracy is 94.16%. Compared with MPSO-RF, PSO-RF, GA-RF and traditional RF models, the prediction accuracy and stability are significantly improved. This model can effectively solve the overfitting problem of RF, and the hyperparameter optimization efficiency is 32.6% higher than that of MPSO alone, providing more efficient and accurate technical support for landslide disaster risk management.