FireScope: Wildfire Risk Prediction with a Chain-of-Thought Oracle

Official code repository for the paper:

FireScope: Wildfire Risk Prediction with a Chain-of-Thought Oracle
Accepted to CVPR 2026

📄 Paper: https://arxiv.org/abs/2511.17171 🤗 Dataset: https://huggingface.co/datasets/INSAIT-Institute/FireScope-Bench 🔥 Website: www.firescope.ai

---

Overview

FireScope introduces a framework for wildfire risk prediction that leverages Chain-of-Thought (CoT) reasoning via an Oracle model to improve predictive performance and interpretability.

This repository contains the full codebase used for:

• Oracle training
• Subsequent Encoder-Decoder trainings
• Ablations
• Evaluation and benchmarking

The dataset used in the experiments (FireScope-Bench) combines Sentinel‑2 imagery, climate data, expert-defined wildfire risk rasters, and real wildfire events to enable reasoning‑based wildfire risk modeling and cross‑continental generalization experiments.

---

Dataset

The FireScope-Bench dataset is available on Hugging Face:

https://huggingface.co/datasets/INSAIT-Institute/FireScope-Bench

Please download the dataset and place it inside your configured DATA_DIR (see configuration section below).

---

Repository Structure

.
├── custom_datasets/     # datasets used for training and evaluating models
├── data_generation/     # code for downloading alphaearth embeddings and generating Oracle predictions
├── evaluation/          # generation and evaluation of results
├── models/              # models used for all experiments
├── prompts/             # prompts used for all experiments
├── rewards/             # GRPO reward components
├── training/            # all trainings
└── config.py            # configuration (dataset paths etc.)

---

Configuration

The repository uses a hardcoded data directory.

In config.py you will find:

DATA_DIR = ...

Some parts of the codebase also directly reference the following path:

/work/wildfirerisk/

Important

To reproduce the experiments you will likely need to:

1. Modify DATA_DIR in config.py
2. Update any hardcoded references to /work/wildfirerisk/
3. Ensure dataset folder names match those expected by the code

Note that some dataset folders may have different naming conventions than the official huggingface data, so you may need to adjust paths accordingly after downloading the datasets.

---

Citation

If you find this work useful in your research, please cite:

@article{markov2025firescope,
  title={FireScope: Wildfire Risk Prediction with a Chain-of-Thought Oracle},
  author={Markov, Mario and Ailuro, Stefan Maria and Van Gool, Luc and Schindler, Konrad and Paudel, Danda Pani},
  journal={arXiv preprint arXiv:2511.17171},
  year={2025}
}

---

Contact

For questions or issues, please open a GitHub issue in this repository.