Machine learning-based susceptibility assessment of geological hazards in Guizhou Province

High-precision assessment of geological hazard susceptibility is crucial for disaster prevention and mitigation initiatives as well as land use planning. This study aims to enhance the accuracy of geological hazard susceptibility assessments in Guizhou Province by integrating a statistical modeling approach with machine learning technique. An evaluation index system was constructed with 15 impact factors, including engineering geological rock formations and others. Susceptibility levels were predicted by the models of weights of evidence(WOE), random forest(RF), and coupled WOE-RF, whose accuracies were systematically evaluated and compared. The results indicate a significant improvement in the predictive precision of both the RF and WOE-RF models compared to the WOE model alone:The AUC value increased by 24.32%-32.43%, and the number of registered geological hazards captured within the moderate- and high-susceptibility zones rose by 14.5%. The coupled WOE-RF model, which utilizes the WOE values of input feature classes as predictors and selects non-hazard samples from non-susceptible zones, demonstrated the best predictive performance and generalization capability. High-susceptibility zones in Guizhou Province are predominantly located in the western and northwestern regions, where topographic factors and stratigraphic lithology exert a dominant control on hazard development. The findings may provide a valuable reference for enhancing the precision of regional susceptibility assessments.