WILDFIRE-FM: First Foundation Model for Wildfire Prediction
A team of researchers has unveiled WILDFIRE-FM, the inaugural foundation model specifically designed for predicting wildfires by utilizing data on weather, active fires, topography, vegetation, and static environmental factors. This development tackles the shortcomings of current Earth foundation models (Earth FMs), which are focused on broader atmospheric and geophysical tasks instead of wildfire prediction. Additionally, the study introduces a fixed-contract evaluation framework to mitigate the impact of matching rules and evaluation settings on transfer conclusions, particularly due to the infrequency of wildfire occurrences. This framework features checks for matching-rule and head-selection effects. WILDFIRE-FM is evaluated against ten Earth-FM baselines under matched contracts.
Key facts
- WILDFIRE-FM is the first foundation model pretrained specifically for wildfire prediction.
- It uses weather, active-fire observations, topography, vegetation, and static environmental data.
- Existing Earth FMs are pretrained for general atmospheric and geophysical objectives.
- Wildfire events are sparse in space and time, making evaluation sensitive to matching rules.
- A fixed-contract evaluation framework is introduced with two controlled checks.
- The fixed-output check addresses matching-rule effects.
- The fixed-feature check addresses head-selection effects.
- WILDFIRE-FM is compared with ten Earth-FM baselines under matched contracts.
Entities
—