Advancing Wildfire Susceptibility Mapping through Machine Learning and SHapley Additive exPlanations-Integrated Geospatial Analysis in Northern Morocco's Mediterranean Region

Adil Moumane^1*Adnane Al Karkouri¹Abdessamad ELMOTAWAKKIL²Wafa Saleh Alkhuraiji³Nazih Y. Rebouh⁴Youssef M. Youssef^5,6*

• ¹Department of Geography, Faculty of Humanities and Social Sciences, Ibn Tofail University, 14000 Kenitra, Morocco, Kenitra, Morocco
• ²3Department of Computer Science, Faculty of Sciences, University Ibn Tofail, 14000 Kenitra, Morocco, Kenitra, Morocco
• ³Department of Geography and Environmental Sustainability, College of Humanities and Social Sciences, Princess Nourah bint Abdulrahman University, P.O. Box 84428, Riyadh 11671, Saudi Arabia, Riyadh, Saudi Arabia
• ⁴Department of Environmental Management, Institute of Environmental Engineering, RUDN University, Miklukho-Maklaya St., 117198 Moscow, Russia, Moscow, Russia
• ⁵Suez University, Suez, Egypt
• ⁶Geological and Geophysical Engineering Department, Faculty of Petroleum and Mining Engineering, Suez University, Suez 43518, Egypt, Suez, Egypt

Wildfires pose a major environmental threat to Mediterranean ecosystems, intensified by climate change and growing human pressures. Yet, limited research has combined machine learning (ML) and SHapley Additive exPlanations (SHAP) to jointly assess predictive accuracy and interpret wildfire-driving mechanisms, particularly in data-scarce regions such as northern Morocco's Tangier–Tétouan– Al Hoceima (TTA) area—a recognized wildfire hotspot requiring advanced predictive tools for effective risk mitigation. This study applied a multi-model ML framework to map wildfire susceptibility by integrating environmental, climatic, and topographic variables with historical fire records. Remote sensing indices (NDVI, LST, wind speed) from summer 2022 were combined with topographic parameters (elevation, slope, aspect, TWI) and proximity measures (distance to roads, settlements, streams) derived from regional datasets. Five ML algorithms—CART, k-NN, SVM, LightGBM, and XGBoost—were tested, with SHAP employed to interpret model behavior. Among these, XGBoost achieved the highest performance (accuracy = 0.920; F1-fire = 0.926; F1-nonfire = 0.912), followed by LightGBM (accuracy = 0.905; AUC = 0.965), confirming the superiority of gradient boosting techniques over conventional models. SHAP analysis identified NDVI as the most influential predictor, underscoring vegetation density as the primary driver of fire susceptibility through its contribution to fuel load. Secondary predictors varied: LightGBM emphasized elevation and wind speed, whereas XGBoost highlighted LST and wind speed. Interaction effects revealed that concurrent high temperatures and strong winds during Chergui events, as well as interactions between vegetation density and terrain position, substantially increase fire likelihood. Overall, wildfire susceptibility in Mediterranean landscapes arises from complex, non-linear interactions among vegetation, topography, and meteorological extremes. The resulting susceptibility maps deliver actionable insights for targeted fire prevention, resource allocation, and early warning, providing a robust framework to enhance adaptive wildfire management in Morocco's most vulnerable ecosystems.