Estimating 10-h Dead Fuel Moisture in Gangwon State, South Korea, using GK-2A Satellite Data and Ground Observations with Tree-Based Machine Learning

Dead fuel moisture (FM) strongly influences wildfire ignition and spread, but direct measurements remain sparse and difficult to obtain over large mountainous regions. This highlights the need for spatially continuous FM estimation using satellite and ground-based observations. Gangwon State, located in northeastern South Korea, is characterized by predominantly forested and mountainous terrain with complex topographic variability, making high-resolution FM estimation particularly important. We developed tree-based machine learning (ML) models to estimate 10-h FM in Gangwon State, South Korea, using AWS meteorological observations, GK-2A satellite radiance, and temporal and topographic variables. These input variables were designed to reflect atmospheric conditions, local characteristics, and diurnal variations relevant to FM dynamics. Random forest, extreme gradient boosting, and light gradient boosting models were trained and optimized using five-fold cross-validation. The models achieved high accuracy for 10-min interval estimates, with RMSE values of 1.20–1.41% and R² values of 0.94–0.96. The extreme gradient boosting model showed the best overall performance, while the ensemble of the three models provided stable and competitive estimates. The models also performed well in identifying the wildfire-risk threshold of FM < 10% (equitable threat score = 0.78; accuracy = 0.95). The classification results further indicate that the models can reliably distinguish critically dry fuel conditions associated with elevated wildfire risk. These findings suggest that integrating geostationary satellite data with ground observations and tree-based ML can support high-resolution FM estimation and wildfire risk applications in complex terrain.

 

Acknowledgments: This work was funded by the Korea Meteorological Administration Research and Development Program “Research on Weather Modification and Cloud Physics” under Grant (KMA2018-00224).