FireWx-FM Model Card

Summary

FireWx-FM is a compact U-Net model trained for 12-hour active-fire occupancy prediction on a 5 km EPSG:5070 grid over the Lower 48 United States. The final serving checkpoint is intended for gridded wildfire-risk screening and for downstream aggregation to county or sub-county units.

Inputs

The model uses a fixed 16-channel tensor containing HRRR weather variables, validity masks, LANDFIRE fuel and canopy layers, Wildfire Risk to Communities housing density, and LandScan population. The final no-exposure serving run zeros the housing and population channels.

Output

The output is a probability map with the same spatial grid as the input. Each value represents the model-estimated probability of active-fire occupancy at a 12-hour lead.

Training Scope

The released final model family was trained with chronological splitting over 2024 fire-season time maps. Training uses tile sampling to reduce empty-cell dominance, per-channel input normalization for continuous variables, sparse-label-aware loss weighting, and an auxiliary spatial-support objective.

Intended Use

Research evaluation of wildfire-focused gridded prediction.
CONUS-scale active-fire occupancy screening.
County or sub-county summaries produced from the native 5 km grid by transparent aggregation.

Not Intended For

Emergency response decisions without expert review.
Direct replacement of operational fire-weather or incident-management products.
Interpretation as a calibrated forecast without local validation.
Inference from inputs that do not follow the documented channel order and normalization contract.

Data Boundary

This repository does not redistribute raw source data. Users must obtain NOAA HRRR, NASA FIRMS, LANDFIRE, Wildfire Risk to Communities, and LandScan resources from their original providers.

Known Limitations

The native model is gridded at 5 km resolution; finer administrative summaries are aggregations, not new fine-scale predictions.
Output quality depends on correct reprojection, normalization, mask handling, and source-variable selection.