Environmental Erosion Research Journal

Landslides are natural disasters that cause destructive effects on infrastructure and the environment. The main objective of this study is to evaluate the accuracy and reliability of various machine learning models—including Maximum Entropy, Random Forest, and Support Vector Machines—in predicting landslide occurrence using 16 influential factors in the Chesb watershed of Zanjan province. Using 70% of landslide occurrences for training and 30% for validation, the models' accuracies were assessed. The Maximum Entropy model results showed that 49.83% of the area falls into very low and low susceptibility classes, 30.17% in the moderate susceptibility class, and 20% in the high and very high susceptibility classes, with an area under the ROC curve (AUC) of 0.911, indicating good model performance. The Random Forest model indicated that 74.31% of the area belongs to the very low and low susceptibility classes, 12.94% to the moderate class, and 12.76% to the high and very high susceptibility classes, with an AUC of 0.934. Likewise, the Support Vector Machine model provided similar results, with 73.05% of the area in very low and low susceptibility classes, 8.15% in the moderate class, and 18.83% in the high and very high classes, with an AUC of 0.873. Based on these evaluations, the Random Forest model was identified as the best model, offering high accuracy in predicting landslide hazards and strong alignment with the real conditions of the region.