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

This study develops and evaluates a machine learning fire-risk prototype for Ghana's Guinea Savannah zone, addressing the lack of a locally calibrated early-season danger tool for November-January. Using a 1 km gridded dataset (2001-2024) combining MODIS burned area, FIRMS active fires, ERA5-Land, CHIRPS, FLDAS soil moisture, MODIS vegetation, SRTM topography, and human-access variables, 21 base predictors were expanded into 123 features using lags, rolling statistics, anomalies, drought indices, and seasonal indicators. Five algorithms were compared: Random Forest, XGBoost, LightGBM, CatBoost, and a multilayer perceptron. Tree-based models achieved high discrimination (AUC 0.960-0.962) on stratified test samples. Threshold-dependent metrics are interpreted with prevalence-adjusted estimates due to balanced sampling. The full-feature LightGBM exceeded the monthly Canadian Fire Weather Index by 0.28 AUC, though this reflects both algorithm and feature advantages. Feature attribution revealed no dominant soil-moisture control. Instead, temperature range, population density, surface pressure, precipitation memory, vegetation state, and soil-moisture memory all contributed to predicted risk. Static variables like surface pressure are interpreted cautiously as potential spatial/elevation proxies. Bias-corrected CMIP6 experiments suggest higher late-century fire risk under SSP2-4.5 and SSP5-8.5, but these are illustrative due to tree ensemble extrapolation limits and constant human driver assumptions. The revised index is presented as a research prototype supporting future operational testing, not a fully validated warning system. This work provides a foundational framework for improved fire danger assessment in data-sparse savannah regions, with potential for adaptation to other West African contexts.