Application of gradient boosting for landslide risk assessment in mountainous areas

Introduction. Reliable prediction of hazardous slope processes remains a significant challenge in engineering geology due to the complex nonlinear relationships between topographic, geological, and hydrological factors. Conventional statistical approaches often demonstrate limited predictive capability when applied to heterogeneous mountainous environments. This study investigates the applicability of a gradient boosting algorithm for improving the accuracy of landslide susceptibility assessment using a limited but informative set of terrain and geological characteristics. Particular attention is given to identifying the most influential conditioning factors and evaluating the predictive performance of the developed classification model. Methods. The study was performed using a digital elevation model, geological information, and records of documented landslide occurrences collected within a local mountainous area. Nine conditioning factors describing terrain morphology, hydrological conditions, geological structure, and anthropogenic influence were selected for analysis. Spatial data preprocessing included coordinate harmonization, missing value treatment, feature standardization, and correlation analysis. A balanced dataset consisting of 1,240 observations was divided into training and testing subsets with five-fold cross-validation. The prediction model was developed using the XGBoost implementation of the gradient boosting algorithm. Model performance was evaluated using Accuracy, Precision, Recall, F1 score, Cohen’s Kappa coefficient, ROC-AUC, confusion matrix analysis, and SHAP-based feature importance assessment. Results. The developed model demonstrated high predictive performance, achieving an overall classification accuracy of 93.8%, Precision of 94.8%, Recall of 92.4%, an F1-score of 0.936, a Cohen’s Kappa coefficient of 0.912, and an ROC-AUC value of 0.961. Feature importance analysis revealed that slope angle, topographic wetness index, distance to the river network, and lithological characteristics contributed most significantly to landslide susceptibility prediction. The generated susceptibility map successfully identified the majority of documented landslide locations, while additional sensitivity analysis confirmed the stability of the model under varying training sample sizes and moderate levels of data uncertainty. Conclusions. The obtained results demonstrate that gradient boosting provides an effective approach for landslide susceptibility assessment in mountainous areas while requiring only a moderate amount of input data and standard computational resources. The proposed methodology enables reliable identification of potentially hazardous slopes, supports spatial prioritization of risk zones, and can be integrated into engineering geological investigations, regional hazard mapping, and preliminary decision-making for infrastructure planning and land-use management. The study also confirms the practical value of combining gradient boosting with SHAP analysis to improve both prediction accuracy and model interpretability.

Confusion matrix · Digital elevation model · Gradient boosting · Kappa · Landslide · Preprocessor · Terrain · Landslides and related hazards · Rock Mechanics and Modeling · Soil erosion and sediment transport