license: mit
tags:
- wildfire
- geospatial
- weather
- earth-observation
- pytorch
- reproducibility
pipeline_tag: image-segmentation
library_name: pytorch
pretty_name: FireWx-FM
FireWx-FM
FireWx-FM is a wildfire-focused gridded model for 12-hour active-fire occupancy prediction over the Lower 48 United States. The released code maps a fixed 16-channel tensor on a 5 km EPSG:5070 grid to an active-fire occupancy probability map.
This Hugging Face release is the clean ICLR/technical-report companion package: it includes model code, checkpoints, inference utilities, training/cache templates, public data-source notes, example outputs, the ICLR manuscript source, and the technical report source.
Contents
| Path | Purpose |
|---|---|
firewxfm/ |
Model definition and tiled CONUS inference code. |
models/checkpoints/ |
Released PyTorch checkpoints. |
models/metadata/ |
Input-channel contract, normalization statistics, and run metadata. |
training/ |
Cache-building, training, and metric utilities. |
training/configs/ |
Path-free templates for CONUS and multi-year reruns. |
scripts/ |
Data preparation, release checks, live-stack builder, seam audit, and smoke tests. |
examples/final_prediction/ |
Small final probability/preview example for inspection. |
data_sources/ |
Source-data provenance and access notes. |
paper/iclr_source/ |
ICLR manuscript LaTeX source. |
technical_report/source/ |
Technical report LaTeX source. |
artifacts/results/ |
Compact public result summaries without internal filesystem paths. |
Quick Start
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python3 scripts/check_release_files.py
python3 scripts/smoke_inference.py --device cpu
The smoke test uses a synthetic tensor. Full scientific reproduction requires
the public source datasets listed in data_sources/DATA_SOURCES.md.
Model Input Contract
The model expects a channel-first tensor with shape [16,H,W].
| Channels | Source | Role |
|---|---|---|
| 0-9 | NOAA HRRR | Dynamic weather fields. |
| 10-11 | Input builder | Dynamic and static validity masks. |
| 12 | LANDFIRE | Fire-behavior fuel model. |
| 13 | LANDFIRE | Canopy cover. |
| 14-15 | WRC / LandScan | Housing density and population; zeroed in the no-exposure serving run. |
See models/metadata/input_channels.json and
models/metadata/input_normalization_stats.json for the exact machine-readable
contract.
Inference
Run inference from a prebuilt 16-channel stack:
python3 -m firewxfm.serve_conus \
--input-npy /path/to/input_stack_16chw.npy \
--checkpoint models/checkpoints/firewxfm_2024_staticfix_region_balanced_bce_seed42.pt \
--normalization-stats models/metadata/input_normalization_stats.json \
--output-npy outputs/firewxfm_probability.npy
For full-map serving, use overlap-window inference with phase averaging. The
default settings are documented in docs/serving_notes.md.
Reproducibility
Use REPRODUCIBILITY.md for the exact verification path:
- artifact presence check,
- checkpoint smoke inference,
- ICLR source compilation,
- technical report source compilation,
- real-data cache rebuild and retraining commands.
Raw source data are not redistributed. This package documents where each source comes from and provides path-free templates for local reproduction.
Citation
@misc{firewxfm2026,
title = {FireWx-FM: A Wildfire Foundation Model Technical Report},
author = {Xu, Yangshuang and Dai, Yuyang and Chang, Liling and Wang, Qi and Dong, Yushun},
year = {2026},
note = {FireWx-FM code, checkpoints, and reproducibility package}
}