chaseungjoon/wildfire-korea-embedded-300m

Wildfire Korea Embedded 300m (16-Channel Environmental Tiles)

A geospatial dataset containing 300 m embedded environmental feature tiles for the Korean Peninsula, designed for wildfire spread prediction and reinforcement-learning research.

This dataset provides model-ready 16-channel tensors representing static and quasi-static environmental conditions needed for wildfire simulation and RL inference.

It is paired with the wildfire episode dataset (wildfire-korea-episodes-300m) and supports the A3C-LSTM wildfire spread model.

---

Overview

This dataset contains preprocessed, aligned, and normalized multi-layer environmental tiles covering the Korean Peninsula at 300 m spatial resolution.

Each tile encodes the following:

• Topography (DEM-derived slope/aspect)
• Landcover
• Forest structure
• Vegetation index (NDVI)
• Weather features (temperature, humidity, wind, precipitation, etc.)
• Additional static/categorical layers

All layers have been merged into a consistent 16-channel feature tensor, ready to be fed directly into wildfire spread models.

---

Feature Channels (16 total)

Channel Range	Description
0–1	DEM-derived slope & aspect
2–10	Weather features (temp, humidity, wind speed/dir, pressure, cloud, visibility, dew point, precipitation)
11	NDVI
12–15	Forest Structure Model (FSM) or forest-type one-hot channels

All layers are normalized and spatially aligned.

---

DISCLAIMER

It is not intended for operational, real-time fire forecasting without additional validation.

---