Wildfire Time-of-Arrival Reconstruction from Satellite Active-Fire Detections: External Evaluation of Variance-Weighted Kriging and Application to South Korea

Abstract

Large wildfires cause casualties, ecosystem damage, and property loss. Reducing them requires predicting spread; validating a spread model requires pixel-level time of arrival (ToA) from real fires. Few regions have both such data and an independent reference. We combine polar-orbiting (VIIRS) and geostationary (GOES, Himawari) active-fire detections by kriging to produce ToA inside a known final perimeter with a per-pixel kriging standard deviation. For 29 U.S. wildfires with aerial infrared perimeters, we compared configurations differing in whether geostationary data were added and whether pixel size entered the variance, against an inverse-distance-weighting baseline. With all observations available, differences among configurations were small. In polar-orbiting gaps, configurations using geostationary data reconstructed ToA better, and variance-weighted kriging was best by median. Weighing reconstruction performance and the per-pixel standard deviation, we selected variance-weighted kriging. Applying it to 21 South Korean wildfires ≥ 100 ha, we built a 375 m arrival-time and standard-deviation dataset for 15 events with event-level evaluation information. This is an application, not an independent accuracy validation, and reconstruction stays inside the final perimeter. Even so, the dataset supports pixel-level validation of spread models and analysis of spread drivers where aerial or field observations are unavailable.