Machine Learning Prediction of Early-Season Wildfire Risk in Ghana’s Guinea Savannah Using Multi-Source Data and CMIP6 Climate Scenarios

Abstract Wildfires in Ghana’s Guinea Savannah zone burn 2–3 million hectares annually, yet no operational early-season wildfire danger index exists. We develop a machine learning-based early-season (November–January) wildfire danger index integrating remote sensing, multi-depth soil moisture, climate projections, and human factors. Using a consistent 1 km grid, we assembled monthly fire occurrence data (2001–2024) from NASA FIRMS and 21 predictor variables, including ERA5-Land, CHIRPS, FLDAS soil moisture, MODIS vegetation indices, SRTM topography, and population and infrastructure data. Features included lag variables (1–12 months), rolling statistics (3–12 months), and drought indices (SPI, SPEI). Five machine learning models Random Forest, XGBoost, LightGBM, CatBoost, and a Multi-Layer Perceptron were developed with hyperparameter tuning and spatial cross-validation. LightGBM achieved the highest performance (AUC = 0.962, F1 = 0.869), outperforming the Canadian Fire Weather Index by 41%, highlighting the value of regionally calibrated models. SHAP analysis identified surface soil moisture, vegetation indices, temperature, and distance to roads as dominant predictors, while deeper soil layers reflected antecedent moisture effects. Human factors ranked among the top 10 predictors, confirming that anthropogenic ignitions drive early-season fires. CMIP6 projections indicate substantial increases in fire risk by 2080–2099: +15.6% under SSP2-4.5 and + 35.6% under SSP5-8.5, with high-risk areas expanding northward. These results demonstrate the potential for emissions mitigation to moderate future fire danger. This study provides the first Ghana-specific early-season wildfire danger index, offering a critical tool for wildfire management, early warning, and climate adaptation strategies in West Africa.