Wildfire Ignition Prediction — Model Weights
⚠️ Paper under peer review. Weights shared for transparency; link to follow.
Trained weights for Multi-Sensor Active-Fire Fusion and Deep Learning for Multi-Day Wildfire Ignition Prediction Across Four Fire Regimes (Gottfriedsen et al., 2026).
Pixel-wise daily ignition prediction across Australia, Brazil, California, Greece,
with three architectures (UNet2D, UNet3D, RandomForestClassifier) at daily and
weekly resolution. Ground truth is fused active-fire detection from 26 LEO/GEO platforms;
inputs are 37 environmental and socio-economic predictors on a 0.1° daily grid.
Performance (PR-AUC, best model per region)
| Region | Model | PR-AUC | vs FWI |
|---|---|---|---|
| Greece | 2D U-Net | 25.0 ± 4.3% | 36× |
| California | 3D U-Net | 20.8 ± 14.8% | 42× |
| Brazil | 2D U-Net | 16.0 ± 3.1% | 5× |
| Australia | 3D U-Net | 7.8 ± 0.9% | 16× |
Input / output
- 2D U-Net
(B, 37, H, W)· 3D U-Net(B, 37, 7, H, W)· Random Forest(N, 37) - Output: per-pixel ignition probability
[0, 1]
License
CC BY-NC 4.0 — research use with attribution; commercial use requires permission from OroraTech GmbH.
Contact
Julia Gottfriedsen — julia.gottfriedsen@campus.lmu.de · LMU, Munich
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support