chaseungjoon/wildfire-korea-episodes-300m

South Korean Wildfire Episodes Dataset (300 m Grid)

A geospatial dataset of South Korean wildfire progression combining NASA FIRMS (VIIRS/MODIS) fire detections with static environmental layers (DEM, landcover, slope, NDVI, forest structure, etc.).

Tiles are preprocessed into a 300 m spatial grid and segmented into temporal wildfire episodes suitable for reinforcement-learning–based wildfire spread prediction research.

---

Overview

This dataset provides a unified, model ready representation of wildfire behavior across the Korean Peninsula from 2015–2025. Each episode contains:

• A temporal sequence of environmental feature tensors (16 channels)
• Corresponding fire spread masks at each timestep
• Pre-normalized and aligned geospatial layers
• Episode segmentation based on FIRMS temporal-spatial clustering

This dataset directly supports training RL models such as A3C-LSTM wildfire spread predictors.

---

Dataset Structure

Each episode is stored as a compressed .npz file:

episode_XXXX.npz
├── states # (T, 16, 30, 30) environmental grids over time
└── fire_masks # (T, 30, 30) binary wildfire progression masks

Where:

• T = number of timesteps in the episode
• 16 channels include:
  • DEM (elevation, slope, aspect)
  • Weather (temp, humidity, wind speed, wind direction, precip, pressure, cloud, visibility, dew point)
  • NDVI
  • FSM (forest susceptibility or categorical forest-type channels)

---

Data Fields

states

Shape: (T, 16, 30, 30)
Contains the full spatiotemporal feature tensor used by RL agents.

fire_masks

Shape: (T, 30, 30)
Binary wildfire mask where:

• 1 = active fire
• 0 = non-burning cell

---

Preprocessing Pipeline

1. FIRMS VIIRS/MODIS fire detections collected (2015–2025)
2. Sliding Window Scan clustering applied to group detections into separate wildfire episodes
3. All geospatial layers resampled to 300 m grid resolution
4. Static layers embedded as multi-channel raster tiles
5. Episode .npz constructed from:
   • FIRMS-derived fire mask sequence
   • Environmental tensors per timestep

---

Intended Use

This dataset is designed for:

• Reinforcement-learning wildfire spread models (A3C, A2C, LSTM-based agents)
• Spatiotemporal forecasting research
• Geospatial ML and uncertainty modeling
• Comparative experiments with CNN/LSTM/U-Net wildfire models

---

Limitations

• FIRMS detections have temporal gaps and positional uncertainty
• Environmental features are aggregated to 300 m resolution
• Episode segmentation depends on clustering thresholds
• Dataset is not meant for operational wildfire forecasting without additional calibration

---

Citation

If you use this dataset, please cite:

Cha, S.J. (2025). Korean Peninsula Wildfire Episodes Dataset (300 m Grid).
WildFirePrediction Project, Chung-Ang University.