WildFirePrediction/ai

AI Module for Real-Time Wildfire Spread Prediction

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.

Features

• 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

Quick Start

1) Clone the Repository

• Renaming repo to WildfirePrediction is optional, but recommended for clarity

git clone https://github.com/WildFirePrediction/ai.git WildFirePrediction
cd WildFirePrediction

2) Download Required Data (~1.6GB)

• script to download embedding data to construct environment tiles for inference
• google drive (wget)

./download_data.sh

Running the Wildfire Prediction

Tested Environment

• Ubuntu 24.04.3 LTS
• CUDA 13.0
• NVIDIA Driver 580.95.05

---

0. (Recommended) Install CUDA + NVIDIA Driver

• Optional, but recommended to match tested environment

./install_env.sh

1. Create Virtual Environment & Install Dependencies

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

2. Run Inference

2-1. Demo Mode (Fake Fire Data)

./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

2-2. Production Mode (Real KFS API Monitoring)

./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/predictions

2-3. Interactive Web Demo

./start_web.sh     # start engine + web server
./stop_web.sh      # stop everything

• Browser UI on port 8080 (--port to change). The start script prints the Tailscale / LAN / local URLs.
• Click anywhere on the map of South Korea to place one or more ignition points, set the ignition time and prediction horizon (10 min steps, up to 2 hours), then run inference.
• Shows predicted spread per timestep on the 400m grid with a timeline player, plus live KMA weather, terrain and spread statistics per fire.
• Saves a JSON result for every prediction.

WildfirePrediction
 └── webdemo/
        ├──server.py       # Flask app + JSON API
        ├──engine.py       # RL inference wrapper
        ├──static/         # map UI
        └──outputs/*.json

Background Deployment (systemd)

1. Install Services

sudo cp deployment/wildfire-api.service /etc/systemd/system/
sudo cp deployment/wildfire-monitor.service /etc/systemd/system/
sudo systemctl daemon-reload

2. Start Services

sudo systemctl start wildfire-api
sudo systemctl start wildfire-monitor

3. Stop Services

sudo systemctl stop wildfire-api
sudo systemctl stop wildfire-monitor

Development Notes

• This repository contains only the AI inference engine.
• Due to file size limits, training data is maintained elsewhere.