<p>This study assesses the susceptibility of forest fires in the Middle Atlas Forest (Sefrou province) of Morocco using four classification models of machine learning, Extreme Gradient Boosting (XGBoost), K Nearest Neighbor (KNN), Naïve Bayes (NB), and Random Forest (RF). Twelve predictors are derived from environmental factors, climatic variables, human factors, topographical parameters, and fuel and vegetation indices. To evaluate the suitability of the models and estimate the variance and bias of the estimation, the training dataset obtained from the National Water and Forests Agency of the Sefrou province was divided into a ratio of 70% for model training and 30% for testing. It was then subjected to resampling techniques such as cross-validation (CV). The importance of each predictor variable was determined using the average gain in model accuracy (Gain) when it is used for making splits in decision trees, assigning a weight to the spatial relationship with fire occurrence. The models were trained and validated using the Area Under the Curve (AUC) and Receiver Operating Characteristics (ROC) curve. Results from the 10-fold cross-validation with three repetitions showed average AUC values of approximately 0.85 for XGBoost, 0.89 for KNN, 0.87 for NB, and 0.88 for RF, with KNN outperforming the other models with a maximum AUC value of 0.89. It was reported that land surface temperature emerged as the most significant factor in wildfire susceptibility mapping for the four models, whereas slope and curvature had the most minor influence. The importance of this study lies in its ability to predict forest fire risk areas, which will provide crucial support to forest ecosystem managers in the concerned region. Overall, the resampling process improved the prediction performance of all models.</p> Graphical Abstract <p>Twelve factors derived from environmental, climatic, human, topographical, and vegetation parameters were analyzed to assess forest fire susceptibility in the Middle Atlas Forest of Morocco These factors were modeled using ArcGIS for spatial analysis and processed in R for statistical evaluation. The fire inventory data, provided by the National Water and Forests Agency, was divided into 70% for training and 30% for testing, ensuring a robust evaluation framework. Four machine learning models, Extreme Gradient Boosting, K Nearest Neighbor, Naïve Bayes, and Random Forest were employed to predict fire susceptibility. Model performance was rigorously evaluated using 10-fold cross-validation with three repetitions, and validated through the Area Under the Curve (AUC) and Receiver Operating Characteristics (ROC) curve. The susceptibility maps generated by these four models were classified according to the intensity of fire risk, ranging from very low to very high, providing a clear spatial representation of areas prone to wildfires. Results showed average AUC values of 0.85 (XGBoost), 0.89 (KNN), 0.87 (NB), and 0.88 (RF), with KNN emerging as the top-performing model. Among the predictors, land surface temperature had the most significant impact on fire susceptibility, while slope and curvature showed minimal influence. The study highlights the critical role of these factors in shaping predictive accuracy and underscores the effectiveness of resampling techniques in enhancing model performance.</p>