Evaluation of landslide susceptibility and contribution analysis of key driving factors on the Loess Plateau
摘要
Objective The Loess Plateau is a region critically susceptible to landslides, posing significant risks to human life and infrastructure. Accurate identification of prone areas is vital for disaster mitigation. However, current coupled models often suffer from limitations: they rely on simplistic combinations or default parameters without systematic hyperparameter optimization and fail to achieve deep integration at the feature level, resulting in suboptimal performance and interpretability. This study aims to overcome these shortcomings by developing systematically optimized models and leveraging interpretability tools to elucidate the underlying mechanisms of landslide occurrence. Methods Based on a multicollinearity analysis, thirteen evaluation factors were selected for model construction. We developed three landslide susceptibility models: an extreme gradient boosting (XGBoost) model, a frequency ratio-coupled XGBoost (FR–XGBoost) model, and a frequency ratio-coupled random forest (FR–RF) model. A key advancement in our methodology was the employment of the Optuna framework for the systematic and automated optimization of model hyperparameters to enhance predictive performance. Furthermore, to overcome the “black-box” nature of machine learning models and gain mechanistic insights, we applied shapley additive explanations (SHAP) and partial dependence plots (PDPs) to interpret the models, identify key driving factors, and reveal their interaction effects. Results The results demonstrated significant performance differences among the three models. The coupled models, FR–XGBoost and FR–RF, substantially outperformed the single XGBoost model, with AUC values of 0.968 and 0.963, respectively, compared to 0.805 for the single model. This not only confirms the superior predictive capability achieved by integrating the frequency ratio but also validates the effectiveness of the systematic hyperparameter optimization using the Optuna framework and the selection of evalu