A machine learning model for early season wildfire risk in Ghana’s Guinea savannah assessing discrimination, operational prevalence, spatial transferability, and CMIP6 climate scenarios

Tropical savannas account for most of the world’s burned area, and West Africa is a centre of both current fire activity and projected future risk, yet the northern Guinea Savannah of Ghana still lacks a locally calibrated early season fire risk tool. This study develops and critically evaluates a machine learning fire risk model for the November to January window on a 1 km grid over 2001 to 2024, using MODIS fire products, ERA5-Land reanalysis, CHIRPS precipitation, FLDAS multi-depth soil moisture, MODIS vegetation indices, SRTM topography, drought indices, and human access layers into 123 engineered features. Five algorithms were compared under a temporal split with spatial cross-validation. Discrimination was high on the held-out period, with tree ensemble AUC of 0.960 to 0.962 and overlapping bootstrap confidence intervals, and pairwise DeLong tests showing the leading gradient boosting models to be statistically comparable. Three methodological findings frame the contribution. First, because fire is rare at the cell month scale, threshold-dependent skill measured on a class balanced sample overstates operational performance: a base rate correction to the operational prevalence of about 1.7% reduces precision to roughly 0.12 to 0.14, so the model is a screening rather than an alerting tool. Second, feature attribution does not identify a single dominant driver; temperature range, population density, precipitation and soil moisture memory, and vegetation state jointly govern predicted risk, and a spatial ablation experiment shows that removing the five most influential static predictors lowers AUC by only about 0.013, indicating that most skill derives from time varying environmental variables rather than spatial memorisation. Third, CMIP6 scenario experiments suggest higher late century risk under SSP2-4.5 and SSP5-8.5 but are reported as illustrative because tree ensembles cannot extrapolate beyond their training envelope and human drivers were held fixed. The framework offers a transferable template for evidence based, honestly evaluated fire risk screening in West African savanna landscapes.