This repository contains the AI components for a real-time wildfire spread prediction system, powered by geospatial data pipelines, reinforcement learning, and satellite-based fire detection feeds.
High resolution (300m) wildfire spread forecasting
RL based propagation model (A3C) trained on 10 years of data
Real time monitoring mode integrated with KFS(산림청) fire reports
Demo mode with synthetic ignition events
Systemd service deployment for background inference and 24/7 monitoring
WildfirePrediction
├── inference/
│ ├──rl/ # RL model inference
│ ├──sl/ # SL model inference
│ ├──demo_rl/ # Demo mode for RL model
│ ├──demo_rl_multi/ # Demo mode for RL model (Multi-step)
│ ├──demo_sl/ # Demo mode for SL model
│ ├──demo_sl_multi/ # Demo mode for SL model (Multi-step)
│ └──fire_monitor/ # KFS API monitoring
├── rl_training/ # Reinforcement learning training
│ ├──a3c_10ch/ # 10-channel A3C model
│ ├──a3c_16ch/ # 16-channel A3C model (Production)
│ └──...
├── sl_training/ # Supervised learning training
│ ├──ag_unet_16ch/ # Spatial Attention U-Net model (Production)
│ ├──unet_16ch/
│ ├──unet_16ch_v2/
│ ├──unet_16ch_v3/
│ └──...
├── src/ # Common utility scripts
├── deployment/ # systemd service files
├── embedding_src/ # Data embedding scripts
├── tilling_src/ # Environment tiling scripts
│
├── README.md # This file
├── requirements.txt # Python dependencies
├── download_data.sh # Data download script (~1.6GB)
├── install_env.sh # CUDA + NVIDIA driver install script
├── start_demo.sh # Start demo mode script
└── start_monitoring.sh # Start production monitoring mode script
- Renaming repo to
WildfirePredictionis optional, but recommended for clarity
git clone https://github.com/WildFirePrediction/ai.git WildFirePrediction
cd WildFirePrediction- script to download embedding data to construct environment tiles for inference
- google drive (wget)
./download_data.sh
Tested Environment
- Ubuntu 24.04.3 LTS
- CUDA 13.0
- NVIDIA Driver 580.95.05
- Optional, but recommended to match tested environment
./install_env.sh
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt./start_demo.sh
- Generates synthetic ignition every 120 seconds and runs full inference.
- Creates html visualization and JSON output for each inference.
WildfirePrediction
└── inference/
└──demo_rl/
└──outputs/
├──*.html
└──*.json
./start_monitoring.sh
- Polls KFS API for new fire detections
- Runs wildfire spread inference
- Sends results to production backend
# Configure backend URL in .env
EXTERNAL_BACKEND_URL=https://api.example.com/wildfire/predictionssudo cp deployment/wildfire-api.service /etc/systemd/system/ sudo cp deployment/wildfire-monitor.service /etc/systemd/system/ sudo systemctl daemon-reload
sudo systemctl start wildfire-api sudo systemctl start wildfire-monitor
sudo systemctl stop wildfire-api sudo systemctl stop wildfire-monitor
- This repository contains only the AI inference engine.
- Due to file size limits, training data is maintained elsewhere.