Landslide Susceptibility Analysis Based on Dataset Construction of Landslides in Yiyang Using GIS and Machine Learning

This study aims to explore the methodology for assessing landslide susceptibility by using machine learning techniques based on a geographic information system (GIS) in an effort to develop landslide susceptibility maps and assess landslide risk in the Yiyang region. A landslide dataset in Yiyang was constructed after 16 landslide predisposing factors were identified across four categories, topography, geology, environment, and hydrometeorology, through factor state determination and multicollinearity analysis. A Blending ensemble model was created and achieved higher prediction accuracy by fusing predictions from Random Forest, CatBoost, and XGBoost with logistic regression used as the meta-learner, thus deriving the importance coefficients of the landslide predisposing factors and their contribution rates. The Blending ensemble model achieved high predictive accuracy with an AUC value of 0.8784, demonstrating balanced and stable performance characteristics. With the addition of the rainfall factor, the AUC value of the Blending ensemble model has increased by 0.1199. In combination with the information value method, this model was applied to assess landslide susceptibility and rainfall-induced landslide risks in Yiyang City, demonstrating its validity. In addition, experimental validation confirmed the prediction and evaluation accuracy of the GIS-based Blending ensemble model. Results showed that the frequency ratio (FR) of historical landslide occurrences in high-susceptibility and extremely high-susceptibility zones in Yiyang City exceeded 1, indicating strong consistency between the landslide risk classification and actual distribution of historical landslides. The landslide susceptibility maps created for Anhua County, Heshan District, and Taojiang County in Yiyang City may provide support for the early warning and prevention of landslides and land-use planning in this region. The proposed methodology may be of reference value for improving natural disaster prevention and risk management.