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

Committing suppression, evacuation, and post-fire recovery decisions the moment a wildfire is dispatched requires knowing, before any perimeter exists, how large it will grow and where it will most likely burn. We present a deep learning framework that forecasts, from geospatial and meteorological conditions available at ignition and without the detailed fuel-load and moisture maps that process-based models require, a calibrated georeferenced burn-probability field for the small- to intermediate-extent fires that dominate the record. A U-Net maps a twelve-channel input stack—ignition prior, topography, fuel, fire duration, and meteorology—over a fixed 4.5 km ignition-centred window (256 × 256 grid, 17.6 m per pixel) to a unit-integral spatial allocation, and a gradient-boosted regression predicts burned-area size from fire duration and aggregated meteorology; the two combine into one coherent field p(x) = E[A]·a(x). We trained and evaluated the framework on South Korean fires, taking ground truth from Sentinel-2 burned-area masks validated against official records (untransformed-area R 2 = 0.872, log-space R 2 = 0.752, median detected-to-official ratio 0.82) and predictors from SRTM terrain, satellite fuel proxies, and station and reanalysis meteorology. Over the operating envelope (102 fires; 87 cross-validation, 15 hold-out test), the field was well calibrated in aggregate (expected calibration error 0.025) yet overconfident in its sparse high-probability bins. Its advantage over six reference baselines rested on precision for the rare burned class: per-fire precision–recall AUC reached 0.429 against 0.321 for the strongest baseline (paired Wilcoxon p = 0.008), whereas ranking skill (per-fire ROC-AUC 0.950) exceeded the climatological 0.904 only modestly. Burned-area size was moderately predictable under leave-one-out validation ( r = 0.475, under-dispersed), and the finest perimeter remained limited by inputs that cannot represent suppression and spotting. For the routine fires that dominate the record, ignition-time prediction from public data can deliver a calibrated burn-probability field and event-scale size forecast to support fire management and ecological risk assessment.