Where Do Geospatial Foundation Models Add Value for Wildfire Mapping?

Wildfire mapping from satellite imagery is increasingly important for monitoring the growing impacts of wildfires across diverse environments. Meanwhile, Geospatial Foundation Models (GFMs) promise transferable representations learned from large-scale Earth Observation data, yet when and to what extent they offer advantages over task-specific models for wildfire mapping remains unclear. We therefore evaluate representative GFMs across multiple wildfire datasets under two complementary settings: in-distribution (ID) learning and out-of-distribution (OOD) cross-dataset transfer. Under ID evaluation, GFMs consistently outperform the same architectures trained from scratch, but remain behind strong task-specific models. This gap narrows as labeled data become scarce, while GFMs can also be adapted with fewer parameters. Under OOD evaluation, selected GFMs retain stronger performance when transferred across datasets exhibiting compound shifts in wildfire characteristics, spectral bands, geography, and label sources. Overall, our results suggest that the main value of GFMs for wildfire mapping lies less in peak in-distribution accuracy than in data-efficient adaptation and transferability across heterogeneous wildfire distributions.