PyroSense

Wildfire prediction using foundation models and multi-source feature fusion.

PyroSense explores whether foundation models can improve wildfire prediction by combining satellite imagery embeddings from NASA's Prithvi model with weather data and Google's AlphaEarth embeddings.

Pipeline

Research Questions

1. Can Prithvi embeddings provide predictive signal for fire risk?
2. Does multi-source fusion improve predictions over single-source models?
3. Can Vision-Language Models provide interpretable fire risk explanations?

Future Work

• Can finetuning foundation models like Prithvi improve fire prediction further?
• Does adding multi-temporal imagery (multiple dates per location) to Prithvi improve feature quality?
• Can this approach generalize to other regions beyond the Mediterranean?
• Can we integrate real-time satellite feeds for operational deployment?

What We Built

• Multi-Source Feature Fusion: Combines 1024-dim Prithvi embeddings (NASA's Prithvi-EO-2.0-300M), weather data from Open-Meteo, and 64-dim AlphaEarth embeddings from Google Earth Engine
• Stacking Ensemble: Two-level architecture with source-specific base models and a meta-learner for optimal feature combination
• EarthDial VLM Integration: Vision-language model that generates human-readable fire risk reports with terrain and vegetation analysis

Why This Approach?

Benefits of Multi-Source Fusion

Traditional fire prediction systems often rely on a single data source. PyroSense combines three complementary data types:

By fusing these sources, we capture both immediate conditions (weather, current vegetation) and structural factors (terrain, land use patterns) that influence fire behavior.

Practical Advantages

• Robustness: If one data source is unavailable or noisy, others compensate
• Real-time Ready: Weather and satellite data update frequently; the pipeline can run on fresh inputs
• Scalable: Feature extraction is parallelizable; inference is fast (<1 second per location)

The Role of Vision-Language Models

Fire prediction alone isn't enough—emergency responders need actionable insights. EarthDial provides:

• Terrain Analysis: Identifies slopes, valleys, and features affecting fire spread
• Vegetation Assessment: Describes fuel types and density visible in imagery
• Strategy Recommendations: Suggests firefighting approaches based on visual analysis

This transforms a probability score into an explainable report that supports decision-making.

Training Data

Training set: 200 events (100 fire, 100 no-fire) from Greece, 2015-2021

Results

Key Findings:

1. Prithvi embeddings alone achieve strong performance (0.95 AUC), validating foundation models for fire prediction
2. Multi-source fusion improves accuracy — the stacking ensemble outperforms any single source
3. EarthDial successfully generates interpretable reports with vegetation, terrain, and strategy analysis

Getting Started

The easiest way to explore PyroSense is through the Jupyter notebook:

# Clone and install
git clone https://github.com/georgepap23/pyrosense.git
cd pyrosense
pip install -e .

# Run the notebook
jupyter notebook notebooks/01_prithvi_value_test.ipynb

The notebook walks through the complete pipeline: data loading, feature extraction, model training, and evaluation.

Documentation

• Installation & CLI Guide - Installation and all CLI commands
• Data Download Guide - How to obtain the datasets
• Project Structure - Codebase organization
• Contributing - Development guidelines

Quick Install

pip install -e .

See Installation & CLI Guide for Earth Engine, EarthDial setup, and CLI commands.

Citation

@software{pyrosense2025,
  title = {PyroSense: Wildfire Prediction with Foundation Models},
  author = {Papakonstantinou, George},
  year = {2025},
  url = {https://github.com/georgepap23/pyrosense}
}

Acknowledgments

• Mesogeos Dataset - Orion AI Lab
• Prithvi-EO-2.0 - NASA IMPACT & IBM
• AlphaEarth - Google
• EarthDial - Vision-Language Model

License

MIT License - see LICENSE for details.