juliag0/wildfire-ignition-prediction

Wildfire Ignition Prediction — Model Weights

⚠️ This paper is currently under peer review. Results and model weights are shared for transparency; the manuscript has not yet been accepted for publication.

This repository contains trained model weights for the paper:

Multi-Day Wildfire Ignition Prediction: Fusing Satellite Observations, Environmental Drivers, and Socio-Economic Factors Across Four Fire Regimes
Gottfriedsen et al., JGR: Machine Learning and Computation, 2026
DOI: 10.57967/hf/8739

---

Overview

We benchmark three architectures for pixel-wise daily wildfire ignition prediction across four climatically distinct regions:

Architecture	Description
UNet2D	Single-day spatial model
UNet3D	7-day spatiotemporal model
RandomForestClassifier	Per-pixel tabular baseline

across four regions: Australia, Brazil, California, Greece, under two temporal resolutions: daily and weekly.

All models are evaluated under 6-fold (daily) or 3-fold (weekly) temporal cross-validation. Each fold corresponds to a held-out 12-month test period.

---

Repository Structure

daily/
├── australia/
│   ├── UNet2D/k=1 … k=6     # model.pth, ml_config.yaml, metrics/, plots/
│   ├── UNet3D/k=1 … k=6
│   └── RandomForestClassifier/k=1 … k=6   # model.pkl
├── brazil/
│   ├── UNet2D/k=1 … k=6
│   ├── UNet3D/k=1 … k=6
│   └── RandomForestClassifier/k=1 … k=6
├── california/
│   ├── UNet2D/k=1 … k=6
│   ├── UNet3D/k=1 … k=6
│   └── RandomForestClassifier/k=1 … k=6
└── greece/
    ├── UNet2D/k=1 … k=6
    ├── UNet3D/k=1 … k=6
    └── RandomForestClassifier/k=1 … k=6

weekly/
├── australia/UNet3D/k=1 … k=3
├── australia/RandomForestClassifier/k=1 … k=3
├── brazil/UNet2D/k=1 … k=3
├── brazil/RandomForestClassifier/k=1 … k=3
├── california/UNet3D/k=1 … k=3
├── california/RandomForestClassifier/k=1 … k=3
├── greece/UNet2D/k=1 … k=3
└── greece/RandomForestClassifier/k=1 … k=3

Each fold directory contains:

• model.pth or model.pkl — trained weights
• ml_config.yaml / rf_config.yaml — full training configuration
• metrics/ — per-epoch PR-AUC, F1, precision, recall
• plots/ — PR-AUC, ROC, confusion matrix curves

---

Performance (PR-AUC, 6-fold mean ± std)

Region	Best model	PR-AUC	vs FWI baseline
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×

---

Installation

Requires Python ≥ 3.12.

# with uv (recommended)
uv pip install -r requirements.txt

# with pip
pip install -r requirements.txt

---

Loading a Model

UNet (PyTorch):

import mlflow

model = mlflow.pytorch.load_model(
    "daily/greece/UNet2D/k=6"  # best fold for Greece
)
model.eval()

Random Forest:

import pickle

with open("daily/greece/RandomForestClassifier/k=6/model.pkl", "rb") as f:
    model = pickle.load(f)

---

Input Format

Models expect a predictor tensor of shape:

• 2D U-Net: (B, 37, H, W)
• 3D U-Net: (B, 37, 7, H, W) — 7-day sequence
• Random Forest: (N, 37) — flattened pixels

37 predictor channels: 14 meteorological (ERA5-Land), 9 land cover (ESA CCI), 5 terrain (GLO-30 DEM), 5 socioeconomic (WorldPop, OpenStreetMap), 4 spatiotemporal encodings.

Output: per-pixel fire ignition probability [0, 1].

---

Cross-Validation Folds

Fold	Test period
k=1	Apr 2024 – Mar 2025
k=2	Apr 2023 – Mar 2024
k=3	Apr 2022 – Mar 2023
k=4	Apr 2021 – Mar 2022
k=5	Apr 2020 – Mar 2021
k=6	Nov 2019 – Mar 2020

---

License

Model weights are released under CC BY-NC 4.0 — free for research use with attribution; commercial use requires permission from OroraTech GmbH.

---

Citation

@article{Gottfriedsen2026wildfire,
  author  = {Gottfriedsen, Julia and Strebl, Johanna and Forster, Linda
             and Pascolini-Campbell, Madeleine and Muller, Sybrand
             and Tresp, Volker},
  title   = {Multi-Day Wildfire Ignition Prediction: Fusing Satellite
             Observations, Environmental Drivers, and Socio-Economic
             Factors Across Four Fire Regimes},
  journal = {JGR: Machine Learning and Computation},
  year    = {2026},
}

@misc{Gottfriedsen2026weights,
  author    = {Gottfriedsen, Julia},
  title     = {Wildfire Ignition Prediction Model Weights},
  year      = {2026},
  publisher = {Hugging Face},
  doi       = {10.57967/hf/8739},
  url       = {https://huggingface.co/juliag0/wildfire-ignition-prediction}
}

---

Contact

Julia Gottfriedsen — julia.gottfriedsen@campus.lmu.de
OroraTech GmbH, Munich, Germany