Integrates information value with machine learning methods to enhance landslide susceptibility assessment, suggesting improved predictive accuracy in varied terrains.
Key Points
The aim is to enhance predictive accuracy while maintaining model interpretability in landslide susceptibility mapping using machine learning.
Developed a hybrid framework integrating information value with machine learning algorithms.
Implemented a three-stage feature selection protocol to retain significant geological variables.
Validated models using statistical metrics and spatial distribution analysis based on 372 GPS-documented sites.
XGBoost achieved the highest predictive performance with 96% accuracy and an AUC-ROC of 0.991.
RF and SVM followed with 82% accuracy (AUC-ROC 0.888) and 91% accuracy (AUC-ROC 0.944), respectively.
The proposed models allocated greater portions to high susceptibility zones compared to traditional methods.