Data Sources
This repository documents source-data access for the resources used by the paper. The table below records each resource, its role in the experiments, and public access entry points. Users should obtain each source under its own terms.
| Source | Role in paper | Public access entry point |
|---|---|---|
| NOAA High-Resolution Rapid Refresh (HRRR) | Dynamic weather fields for the California regional gridded occupancy inputs. | NOAA/NCEI HRRR product page: https://www.ncei.noaa.gov/products/weather-climate-models/high-resolution-rapid-refresh; AWS Open Data archive: https://registry.opendata.aws/noaa-hrrr-pds/. |
| NASA FIRMS active-fire detections | Active-fire detections used to derive gridded occupancy labels. | FIRMS download and API services: https://firms.modaps.eosdis.nasa.gov/download/ and https://firms.modaps.eosdis.nasa.gov/api/. |
| LANDFIRE 40 Fire Behavior Fuel Models | Static fuel layer used in the FireWx-FM gridded input. | LANDFIRE data access portal: https://landfire.gov/data. |
| LANDFIRE Forest Canopy Cover | Static canopy layer used in the FireWx-FM gridded input. | LANDFIRE data access portal: https://landfire.gov/data. |
| Wildfire Risk to Communities housing-unit density | Static exposure layer used in the FireWx-FM gridded input. | Wildfire Risk to Communities data access: https://wildfirerisk.org/download/. |
| LandScan Global 2024 | Static population layer used in the FireWx-FM gridded input. | Oak Ridge National Laboratory LandScan access: https://landscan.ornl.gov/. |
| WFIGS incident/perimeter attributes | Event-level incident metadata for supporting burned-area and analog tasks. | NIFC Open Data portal for WFIGS layers: https://data-nifc.opendata.arcgis.com/. |
| MTBS burned area and burn severity | Event-scale burned-area and burn-severity records for supporting tasks. | MTBS data access and direct download pages: https://www.mtbs.gov/ and https://www.mtbs.gov/direct-download. |
| Earth-FM/backbone sources | Frozen feature sources for transferred Earth-FM comparisons. | Original model providers and their terms. Examples include Hugging Face model cards, model-provider GitHub repositories, and provider-hosted model files. |
Pretrained Input Channels
The released checkpoints expect a fixed 16-channel tensor in [channel, y, x]
order. Channel order matters. The pretrained 5 km, 12-hour-lead occupancy input
uses the following contract. A machine-readable copy is stored at
models/wildfire_fm/input_channels.json.
No mean/std normalization, min/max scaling, or unit conversion is applied before inference. HRRR fields remain in native units after reprojection to the model grid. Invalid or missing values are zero-filled, with validity masks included as channels 10 and 11.
| Channel | Name | Dataset/source | Level or selection | Units | Normalization | Role |
|---|---|---|---|---|---|---|
| 0 | t2m |
NOAA HRRR | 2 m above ground | K | none | 2 m temperature |
| 1 | d2m |
NOAA HRRR | 2 m above ground | K | none | 2 m dew point |
| 2 | u10 |
NOAA HRRR | 10 m above ground | m/s | none | 10 m east-west wind component |
| 3 | v10 |
NOAA HRRR | 10 m above ground | m/s | none | 10 m north-south wind component |
| 4 | cape |
NOAA HRRR | surface, instant | J/kg | none | Surface instantaneous CAPE |
| 5 | sp |
NOAA HRRR | surface, instant | Pa | none | Surface pressure |
| 6 | blh |
NOAA HRRR | surface diagnostic, instant | m | none | Boundary-layer height |
| 7 | vis |
NOAA HRRR | surface diagnostic, instant | m | none | Visibility |
| 8 | prate |
NOAA HRRR | surface, instant | kg m^-2 s^-1 | none | Precipitation rate |
| 9 | tp |
NOAA HRRR | surface, accumulated | kg m^-2 | none | Accumulated precipitation |
| 10 | firewx_valid |
Cache validity channel | input presence | binary mask | none | Cache-level dynamic input presence mask; 1.0 everywhere in this release |
| 11 | static_valid |
Static reprojection mask | static layers | fraction | none | Fraction of static layers valid at the grid cell after reprojection |
| 12 | fuel_fbfm40 |
LANDFIRE FBFM40 | static layer | categorical code | none | Fire-behavior fuel model |
| 13 | canopy_cover |
LANDFIRE CC | static layer | provider-native value | none | Canopy cover |
| 14 | housing_density |
Wildfire Risk to Communities | static layer | provider-native value | none | Housing-unit density |
| 15 | population |
LandScan Global 2024 | static layer | provider-native value | none | Population exposure |
FIRMS detections are used to derive the occupancy target, not as an input channel. WFIGS and MTBS are event-level resources for supporting burned-area and analog tasks and are not part of the 16-channel pretrained occupancy input.
The cape channel is the HRRR surface instantaneous CAPE field selected with
typeOfLevel=surface and stepType=instant. It is not the most-unstable CAPE
or a layer CAPE variable.
The two validity channels have different definitions. firewx_valid is a
cache-level dynamic input presence mask; in the released California regional
cache it is 1.0 everywhere and is not a per-variable HRRR missing-data mask.
static_valid is the fraction of the four static layers that are valid after
reprojection at each grid cell, with possible values 0.0, 0.25, 0.5, 0.75, and
1.0.
Notes
- The paper places gridded resources on a projected 5 km EPSG:5070 grid.
- The bundled artifacts contain summary values for release checking.
- Full raw-data reruns require users to obtain each source independently and to construct the intermediate grids/features described in the paper.
- Access mechanisms and licensing can change. The links above point to the original providers.