This study develops an explainable machine-learning framework to investigate recorded wildfire occurrence in the Aurès Mountains of north-eastern Algeria using multi-source geospatial data. A total of 584 wildfire occurrence locations extracted from NASA’s Fire Information for Resource Management System (FIRMS) for 2015–2025 were combined with 1,926 randomly generated background locations, yielding 2,510 observations. Ten explanatory variables representing vegetation condition, climate, topography, and anthropogenic accessibility were derived from Sentinel-2 imagery, WorldClim, the USGS Digital Elevation Model, and OpenStreetMap. A Random Forest model was developed following hyperparameter optimisation. Performance was evaluated using conventional random train–test partitioning and five-fold stratified cross-validation, complemented by five-fold spatial-block cross-validation using 10 × 10 km blocks to assess geographic generalisation. Under random evaluation, the model achieved an accuracy of 90.84% and an ROC–AUC of 0.939, with a mean five-fold ROC–AUC of 0.937 ± 0.014. Spatial-block cross-validation produced a substantially lower mean ROC–AUC of 0.714 ± 0.111, indicating reduced performance across spatially separated areas. Altitude and temperature were consistently identified among the leading predictors, while SHAP analysis indicated that higher temperature and NDVI generally contributed to higher modelled likelihood of recorded wildfire occurrence, whereas higher precipitation and NDMI contributed to lower likelihood. These findings highlight the importance of climatic, vegetation, topographic, and accessibility-related spatial gradients. However, predictor collinearity and the use of background locations rather than confirmed fire-absence observations require cautious interpretation. The framework provides an interpretable basis for wildfire susceptibility assessment and spatially targeted prevention and forest management, while geographic transferability requires further evaluation.