Wildfire occurrence has intensified across Mediterranean ecosystems under the combined effect of rising temperatures and prolonged summer droughts, posing a critical challenge for fire risk management in data-scarce regions such as North Africa. This study proposes a machine learning framework for predicting wildfire occurrence in northern Morocco using integrated meteorological and fire weather indicators. A multi-source dataset covering the Tangier–Tétouan–Al Hoceima region was constructed by combining ERA5 meteorological reanalysis variables and Fire Weather Index (FWI) components with wildfire occurrence data derived from NASA FIRMS for the period 2019–2024. Two gradient boosting algorithms, XGBoost and LightGBM, were trained using a temporal hold-out strategy (train: 2019–2022, test: 2023–2024), and a structured ablation study was conducted to quantify the contribution of each feature group. Results show that XGBoost trained on the full feature set achieved the best performance (Accuracy = 0.737, ROC-AUC = 0.813, PR-AUC = 0.739). The ablation analysis demonstrates that FWI components significantly improve predictive performance by capturing multi-day fuel moisture dynamics not represented by instantaneous meteorological conditions, while meteorological and FWI variables alone retain meaningful predictive skill in the absence of geographic coordinates (ROC-AUC = 0.733). Post-hoc SHAP analysis identifies longitude, surface pressure, temperature, and solar radiation as the dominant predictors, with longitude reflecting spatially structured fire occurrence patterns consistent with multiple co-varying geographic factors across the TTA region. The findings provide an interpretable and temporally validated machine learning framework that can support early warning and fire risk management in Mediterranean North Africa.