Wildfire susceptibility mapping supports proactive forest management, and estimated predictive performance may vary with spatial dependence and the control-point sampling strategy. We developed an interpretable random-forest framework to map wildfire occurrence probability across Hubei Province, China, by integrating multi-source environmental (meteorological, topographic, and vegetation) and socio-economic predictors. To enhance methodological robustness and address high-dimensional data complexity, the Boruta algorithm was employed for rigorous feature selection, identifying the most significant drivers while filtering out random noise. The model showed strong discrimination on held-out data (AUC = 0.942, accuracy = 87.9%), and variable importance highlighted sunshine duration, elevation, relative humidity, and maximum temperature as dominant predictors. Predicted wildfire probability exhibited a clear east–west gradient; high and very high susceptibility classes covered 22% of forested land while containing 82% of historical fires, indicating priority zones for targeted prevention and resource allocation. These results demonstrate that combining multi-source predictors with machine-learning interpretability can produce actionable susceptibility maps for regional fire-risk management.