Understanding Fire Return Intervals in a Protected Savanna Ecosystem Using Machine Learning and Explainable Artificial Intelligence (XAI)

Fire is an important ecological process in savannah ecosystems, regulating vegetation structure, biodiversity and ecosystem functioning. However climate variability and anthropogenic pressures are altering fire regimes, threatening the ecological resilience of protected landscapes. Despite advances in remote sensing, the drivers of fire return intervals (FRIs) remain poorly understood because of the complex and nonlinear interactions between climatic, topographic, vegetation, and human factors. The study used the Moderate Resolution Imaging Spectroradiometer (MODIS) MCD64A1 burned-area data, machine learning, and Explainable Artificial Intelligence (XAI) to quantify the drivers of FRI in the Chewore Safari Area, Zimbabwe from 2001 to 2024. Fire return intervals were derived from MODIS observations and temporal trends in burned area were evaluated using the Seasonal Mann–Kendall test. We built Random Forest, XGBoost, and CatBoost models to predict FRI and SHAP (Shapley Additive Explanations) to interpret model predictions. The burned area did not show a significant long-term temporal trend, while FRIs were spatially heterogeneous. Most fire hotspots (FRI < 5 years) were in the south and northeast areas, with the northwest areas being fire-resistant. XGBoost achieved the best prediction performance (R2=0.639), and SHAP analysis revealed that elevation, NDVI, and temperature were the primary controls, contributing approximately 55% to the total model importance. This study integrates remote sensing, machine learning and explainable AI to provide a transparent and transferable framework to understand fire regimes and support evidence-based fire management and conservation planning across African savanna ecosystems.