Data and code for "Deep learning prediction of wildfire burned-area extent and burn probability from ignition conditions using topography, fuel, and meteorology in South Korea"

Data, code and trained models accompanying: Choi, J., Chae, H. (2026). Deep learning prediction of wildfire burned-area extent and burn probability from ignition conditions using topography, fuel, and meteorology in South Korea. Ecological Informatics (manuscript ECOINF-D-26-00783R1).

The deposit reproduces every number, table and figure of the manuscript from derived data and trained weights on a CPU. It contains the satellite-derived burned-area masks and per-event metadata for 118 wildfire events in South Korea (2018–2025), the 102-event operating-envelope subset with its cross-validation and hold-out split, all model input layers (SRTM terrain, ridge distance, canopy height, pre-fire NDVI, station and ERA5 meteorology), the trained weights of the proposed self-attention U-Net and of every comparison and ablation arm, all evaluation outputs, and the complete training, evaluation and figure-generation code with a pinned Python environment. docs/CLAIM_MAP.md maps each manuscript claim to the file, field and script that regenerate it; docs/REPRODUCTION.md gives the evaluation, retraining and preprocessing procedures.

Archives: code/docs/results/figures/logs (1 zip), data (1 zip), weights (7 zips, one per model arm). All unpack into a single folder wildfire-burn-probability-korea/. Start with README.md.

Licences: code MIT; data, weights, results and figures CC BY 4.0. Contains modified Copernicus Sentinel data (2017–2025); ERA5 data from the Copernicus Climate Change Service (CC BY 4.0); SRTM (NASA, public domain); Korea Forest Service and Korea Meteorological Administration data under the Korea Open Government Licence Type 1; canopy height from Potapov et al. (2021), CC BY 4.0.

Code repository: github.com (tag v1.0.0)