Performance Assessment of Individual and Ensemble Learning Models for Gully Erosion Susceptibility Mapping in a Mountainous and Semi-Arid Region

Abstract

High-accuracy gully erosion susceptibility maps play a crucial role in erosion vulnerability assessment and risk management. The principal purpose of the present research is to evaluate the predictive power of individual machine learning models such as random forest (RF), decision tree (DT), and support vector machine (SVM), and ensemble machine learning approaches such as stacking, voting, bagging, and boosting with k-fold cross validation resampling techniques for modeling gully erosion susceptibility in the Oued El Abid watershed in the Moroccan High Atlas. A dataset comprising 200 gully points, identified through field observations and high-resolution Google Earth imagery, was used, alongside 21 gully erosion conditioning factors selected based on their importance, information gain, and multi-collinearity analysis. The exploratory results indicate that all derived gully erosion susceptibility maps had a good accuracy for both individual and ensemble models. Based on the receiver operating characteristic (ROC), the RF and the SVM models had better predictive performances, with AUC = 0.82, than the DT model. However, ensemble models significantly outperformed individual models. Among the ensembles, the RF-DT-SVM stacking model achieved the highest predictive accuracy, with an AUC value of 0.86, highlighting its robustness and superior predictive capability. The prioritization results also confirmed the RF-DT-SVM ensemble model as the best. These findings highlight the superiority of ensemble learning models over individual ones and underscore their potential for application in similar geo-environmental contexts.