chaseungjoon/wildfire-prediction-A3C-LSTM

Wildfire Spread Prediction Model (A3C-LSTM, Korea 300m Grid)

This repository provides a reinforcement-learning–based wildfire spread prediction model trained on a 300 m spatial grid covering the Korean Peninsula.

The model uses an A3C agent with an LSTM policy, optimized to learn spatiotemporal wildfire propagation patterns using real wildfire episodes.

---

Model Overview

• Architecture: CNN Encoder → LSTM Policy → Spatial Upsampling Policy Head
• RL Algorithm: A3C
• Input Resolution: 300 m × 300 m grid tiles
• Input Channels:
  • DEM (Elevation)
  • RSP (Slope)
  • LCM (Land Cover Map)
  • FSM (Forest Structure Model)
  • NDVI (Vegetation Index)
  • Fire mask history (previous timesteps)
• Output: Probability distribution over 8 directional spread actions + stay
• Training Episodes: Derived from 10 years of NASA FIRMS wildfire detections (VIIRS/MODIS), clustered temporally into fire episodes.

---

Included in This Model Repo

• v3.pt — trained model weights
• config.json — architecture & input configuration
• README.md — model card (this file)
• model.py - A3C model architecture
• inference_engine.py - Inference logic (must use wildfire-korea-embedded-300m for inference)
• wildfire_env_inference.py - Environment interface
• requirements.txt - All dependencies
• visualizer.py - Visualization utilities

Github Repo

Inference/Training code, dataset preparation scripts, and environment setup are hosted on GitHub.

---

DISCLAIMER

The model is not intended as an operational early-warning system without additional validation.