GDG @ Penn State Solution Challenge Hackathon 2026 UN Sustainable Development Goals: SDG 13 (Climate Action)
Live Demo: https://firesight-tau.vercel.app Backend API: https://firesight-api-278574018619.us-central1.run.app
FireSight is an AI-powered wildfire forecasting platform that uses historical U.S. wildfire-related data and machine learning to predict future wildfire risk and visualize those forecasts through an interactive map overlaid with live NASA satellite fire detections.
Wildfires are becoming more frequent, more destructive, and more costly — threatening ecosystems, displacing communities, and accelerating climate change through massive CO2 emissions. Most existing wildfire tools are reactive, showing fires after they occur. FireSight takes a proactive approach by forecasting where wildfire risk will be elevated, helping with preparedness, resource allocation, and long-term risk understanding.
This directly supports:
- SDG 13 (Climate Action): Early wildfire risk identification helps reduce response times and the carbon impact of uncontrolled fires
- SDG 15 (Life on Land): Protecting forests and biodiversity by enabling preemptive action in high-risk areas
-
Data Collection — ERA5 atmospheric reanalysis data from ECMWF (2020–2025) provides temperature, wind speed, and precipitation at 0.25-degree resolution. NASA FIRMS satellite fire detections provide ground-truth labels for where fires actually occurred.
-
Feature Engineering — Daily grid-cell features include mean/max temperature, wind speed, precipitation deficit, month, day of year, and 3/7-day rolling averages. Lag features capture whether nearby cells had recent fire activity.
-
Model Training & Comparison — Four models were trained and compared: XGBoost, Random Forest, Extra Trees, and CatBoost. XGBoost was selected as the best performer. Class weighting handles the severe imbalance (~0.1% of grid cells have fire events).
-
Forecasting — The trained model generates fire probability projections for June–October 2026 across the entire US grid. Predictions are interpolated to 0.125-degree resolution for smoother visualization.
-
Live Fire Overlay — The backend fetches real-time fire detections from NASA FIRMS (VIIRS SNPP satellite) every request, allowing users to see model predictions alongside actual current fires.
The datasets are too large for GitHub but are documented here for reproducibility:
| Dataset | Source | Size | Description |
|---|---|---|---|
| ERA5 Reanalysis (2020–2025) | ECMWF/Copernicus | ~2 GB | Hourly 2m temperature, 10m u/v wind, total precipitation. Accumulated and instantaneous variables in separate NetCDF files. |
| ERA5 Forecast (2026 June–Oct) | ECMWF/Copernicus | ~200 MB | Same variables for the 2026 projection period |
| NASA FIRMS VIIRS (2020–2025) | NASA FIRMS | ~500 MB | Suomi NPP and NOAA-20 VIIRS active fire detections with latitude, longitude, brightness, confidence, FRP, and acquisition timestamps |
To reproduce: download from the sources above, place ERA5 .nc files in data/raw/era5/ and FIRMS .csv files in data/raw/firms/.
FireSight/
├── backend/ # FastAPI backend (Google Cloud Run)
│ ├── main.py # API: /api/risk-grid + /api/live-fires
│ ├── generate_predictions.py # Converts model CSV → predictions.json
│ ├── predictions.json # Pre-computed risk grid (95k cells, 0.125°)
│ ├── Dockerfile # Cloud Run container
│ └── requirements.txt
├── frontend/ # React + Vite (Vercel)
│ ├── src/
│ │ ├── App.jsx # Data fetching + layout
│ │ ├── components/
│ │ │ ├── FireMap.jsx # Leaflet map with grid + fire markers
│ │ │ ├── RiskGridLayer.jsx # Colored 0.125° grid cells
│ │ │ ├── CursorInfo.jsx # Lat/lon + fire risk % on hover
│ │ │ └── Legend.jsx # Color scale legend
│ │ └── main.jsx
│ └── package.json
├── src/ # ML training pipeline
│ ├── app/
│ │ ├── xgboost_forecast_us.py
│ │ ├── random_forest_forecast_us.py
│ │ ├── extra_trees_forecast_us.py
│ │ └── catboost_forecast_us.py
│ └── visualization/
│ ├── firesight_streamlit_app.py
│ └── visualize_model_comparison.py
└── data/ # Datasets (not included — see above)
cd backend
python3 -m venv venv && source venv/bin/activate
pip install -r requirements.txt
# Generate predictions from model output:
python generate_predictions.py --from-csv ../data/processed/xgboost_projection_2026_jun_oct.csv
# Or use synthetic demo data for development:
python generate_predictions.py --demo
uvicorn main:app --reload --port 8000cd frontend
npm install
npm run dev| Variable | Where | Description |
|---|---|---|
NASA_API_KEY |
Backend | Free NASA FIRMS API key — get one here |
VITE_API_URL |
Frontend | Backend URL (default: http://localhost:8000) |
cd backend
gcloud run deploy firesight-api \
--source . \
--region us-central1 \
--allow-unauthenticated \
--set-env-vars NASA_API_KEY=your_key \
--port 8080cd frontend
vercel --prodSet VITE_API_URL in Vercel project settings to your Cloud Run URL.
- Backend: Python, FastAPI, httpx, Google Cloud Run
- Frontend: React (Vite), Leaflet, react-leaflet, axios, Vercel
- ML: XGBoost, Random Forest, Extra Trees, CatBoost, scikit-learn, pandas, xarray
- Data: ERA5 reanalysis (ECMWF), NASA FIRMS VIIRS satellite fire detections
AI tools (Claude, GitHub Copilot) were used during development for code generation, debugging, and deployment assistance.