Machine Learning-Based Near-Real-Time Monitoring of Wildfire Spread Extent Using GK2A and VIIRS

Rapid and reliable assessment of wildfire spread is critical for minimizing ecological and socioeconomic damage.Polar-orbiting satellites have high spatial resolution but low temporal resolution, limiting their ability to capture the rapid dynamics of wildfire expansion.To address this limitation, we propose a near-real-time framework for estimating wildfire extent using high-frequency (2-minute) observations from the GEO-KOMPSAT-2A (GK2A) geostationary satellite, employing Visible Infrared Imaging Radiometer Suite (VIIRS) VNP14IMG products as reference data.A 10-minute temporal averaging scheme was introduced to mitigate single-observation noise and enhance detection stability.Model performance was evaluated across six large wildfires in South Korea, with negative samples down-sampled at a ratio of 1:5 relative to positive fire pixels.In repeated random-split (7:3) and region hold-out evaluations, the Extreme Gradient Boosting (XGBoost) model achieved a mean F1 score of 0.958, slightly higher than that obtained by Random Forest (RF; 0.950).For the Uljin (2022) wildfire, XGBoost achieved an F1 score of 0.948, whereas RF achieved a score of 0.741.The superiority of XGBoost was further confirmed via independent full-pixel validation for the Uiseong (2025) and Uljin (2022) wildfires, obtaining precisions of 0.812 and 0.682, respectively, and F1 scores of 0.729 and 0.699, respectively.For both wildfires, RF yielded higher recall but generated a greater number of false positives.These differences may be attributed to the inherent characteristics of the models, with XGBoost's gradient-boosting approach emphasizing precision and overall accuracy, and RF tending to favor recall, often at the cost of increased false positives.The timeseries analysis demonstrated that, with 10-minute averaging, wildfire growth can be reliably tracked from approximately 10 minutes after ignition onward at 2-minute intervals.This suggests that GK2A observations can be exploited not only for wildfire detection but also for early-stage monitoring of fire spread, thereby supporting rapid decision-making for resource allocation and initial suppression strategies.