Assessing and predicting forest fires has long been an arduous task. Nowadays, the rapid advancement of artificial intelligence and machine learning technologies have provided a novel solution to forest fire occurrence assessment and prediction. In this research, we developed a novel hybrid machine-learning-technique algorithm to improve forest fire prediction based on random forest (RF), gradient-boosting decision tree (GBDT), support vector machine (SVM), and other machine learning models. The dataset we employed was satellite fire point data from 2010 to 2018 from the Chinese Department of Fire Prevention. The efficacy and performance of our methods were examined by validating the model fit and predictive capability. The results showed that the ensemble model LR (logistic regression)-RF-SVM-GBDT outperformed the single RFSVMGBDT model and the LR-RF-GBDT integrated framework, displaying higher accuracy and greater robustness. We believe that our newly developed hybrid machine-learning algorithm has the potential to improve the accuracy of predicting forest fire occurrences, thus enabling more efficient firefighting efforts and saving time and resources.

Forest-fire occurrence susceptibility prediction remains challenging in mountainous regions because fire occurrence is controlled by nonlinear interactions among climatic, vegetation, topographic, and anthropogenic factors. Developing interpretable prediction models that integrate multi-source environmental information is essential for improving susceptibility assessment and understanding predictor contributions. Using 2003–2022 MODIS active-fire records and multi-source environmental data from the Dabie Mountains, this study developed an explainable XGBoost-based forest-fire occurrence susceptibility prediction framework. The predictive performance of XGBoost was compared with Random Forest and radial basis function support vector machine (RBF-SVM) using five-fold cross-validation, and TreeSHAP was applied to interpret predictor contributions, nonlinear responses, and interactions. KDE, standard deviational ellipses (SDEs), and GeoDetector were further used to characterize historical fire patterns and complementary spatial associations, while an independent climate-scale XGBoost model was applied to evaluate future susceptibility sensitivity under SSP126 and SSP585 scenarios. XGBoost achieved an ROC-AUC of 0.850, and log-transformed population showed the largest global SHAP contribution. GeoDetector identified precipitation and relative humidity as the strongest spatial associations with historical fire-density patterns across KDE bandwidths. The climate-scale model achieved a pooled five-fold ROC-AUC of 0.830 and indicated stronger late-century susceptibility-score redistribution under SSP585 than under SSP126. Together, these results demonstrate an explainable susceptibility modeling approach that integrates predictive modeling, model interpretation, spatial analysis, and future climate assessment. The outputs represent relative susceptibility and scenario-specific sensitivity responses rather than unconditional fire probabilities or deterministic forecasts.