A machine learning-based wildfire susceptibility mapping framework for China's three-north shelterbelt region

The Three-North Shelterbelt Program (TNSP) region, a key ecological barrier in China, has recently faced more frequent wildfires, while comprehensive evaluations of its fire risk are still lacking. Here, we integrate multisource remote sensing and socioeconomic datasets representing climatic, vegetative and human-activity factors and employ five machine learning classifiers — Random Forest (RF), Gradient-Boosting Decision Tree (GBDT), eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and Logistic Regression (LR) — together with explainable-AI methods to model and analyze regional fire susceptibility levels. Model comparisons show that tree-based ensemble models generally outperform linear models, with the LightGBM achieving the highest performance (cross-validated AUC > 0.98); the small differences among the ensemble models suggest that the data quality and feature dimensionality factors have greater effects on performance than the selected algorithm does. Seasonal fire hazard maps reveal a clear bimodal pattern, with primary peaks in the spring and autumn, a secondary peak during the summer, and minimal activity in the winter. High-risk areas are concentrated in the northeastern forest zone, the Inner Mongolian grasslands, the North China Plain and the Loess Plateau—with spatial patterns that closely match historical fire records. Independent validations conducted using three active-fire datasets (MODIS C6.1, J1 VIIRS C2 and SUOMI VIIRS C2) for 2021–2024 produce average AUC values of 0.8576, 0.8160 and 0.8174, respectively, supporting the predictive ability of the map across multiple fire products. The results of a driver analysis indicate that vegetation (based on the normalized difference vegetation index and leaf area index) and meteorological factors (based on the fire weather index) dominate in the spring and autumn (with autumn amplified by drought); summer risk is more strongly regulated by human activity (human footprint); and winter risk is correlated most strongly with temperature, drought indicators (drought code) and socioeconomic variables (gross domestic product). SHAP-based explainability tests further reveal that driver effects change gradually within typical value ranges but intensify sharply once certain extreme thresholds are exceeded. This study presents a scalable, robust framework for fire hazard prediction and interpretation that can support sustainable management and risk mitigation strategies for the TNSP region and similar large-scale ecological projects.