Geospatial and Machine Learning-Based Assessment of Wildfire Risk in Arid Oasis Ecosystems: A Case Study of the Middle Ziz Valley, Morocco

The increasing frequency of wildfires in arid ecosystems, driven by climate change and growing anthropogenic pressures, highlights the need for robust susceptibility and risk assessment approaches. This research evaluates wildfire susceptibility in the oases of the Middle Ziz Valley (Morocco) through an integrated framework combining Geographic Information Systems (GIS), multi-source Remote Sensing (RS) data, and machine learning (ML) techniques. A wildfire probability map was developed using 130 historical fire occurrences (2010–2023) derived from NASA FIRMS data (MODIS and VIIRS). Nine conditioning factors were considered, including topographic, climatic, environmental, and anthropogenic variables. Four machine-learning algorithms Random Forest (RF), Logistic Regression (LR), Support Vector Machine (SVM), and Extreme Gradient Boosting (XGBoost) were implemented and compared. Data preprocessing included feature standardization and multicollinearity assessment using the Variance Inflation Factor (VIF). Variable importance was evaluated using Information Gain (IG), while model interpretability was enhanced through SHapley Additive exPlanations (SHAP). In addition, ecological and urban vulnerabilities were quantified using the Remote Sensing Ecological Index (RSEI) and the Night-Time Lights Index (NTLI), enabling the construction of a composite ecological–urban vulnerability index All models showed strong predictive performance (AUC = 0.91–0.94), with RF achieving the highest accuracy (AUC = 0.94). Key influencing factors include temperature, wind speed, NDVI, and proximity to roads. High-risk zones are mainly concentrated in peripheral oasis areas, particularly in Aoufous, where ecological fragility and anthropogenic pressures intersect. The proposed integrated framework provides a transferable and effective tool for wildfire risk assessment in arid environments.