Machine-Learning Based Downscaling of GK-2A Land Surface Temperature: A study of Seoul, South Korea

Downscaling Land Surface Temperature (LST) data from geostationary satellites is crucial for high-resolution city-scale environmental monitoring and localized applications. However, the accuracy of such downscaling is often hindered by spatial heterogeneity in land cover and topography. In this study, we present, to our knowledge, the first machine-learning based downscaling of GK-2A LST from 2 km to Landsat-like 30 m resolution over complex Korean terrain. Our framework trains models using Random Forest (RF) and eXtreme Gradient Boosting (XGB) with predictors drawn from the GK-2A LST, Landsat NDVI, and terrain variables like DEM-derived slope and aspect. Model performance was then evaluated against Landsat-derived LST. We analyze 40 cloud-free Landsat dates spanning 2022-2025 in the Seoul metropolitan region. Across all dates, RF achieved RMSE = 1.10-5.27 K, MAE = 0.83-3.93 K, and R2 = 0.43-0.85; XGB was similar with RMSE = 1.09-5.31 K, MAE = 0.83-1.01 K, and R2 = 0.43-0.85. Performance was typically strongest in spring/early summer scenes with clear conditions. For comparison, we implemented a thermal sharpening (TsHARP) downscaling approach. TsHARP showed markedly lower performance, with R2 generally 0.00-0.36, RMSE = 1.70-8.01 K, and MAE = 1.28-6.32 K across the same dates, reflecting limitations of a purely linear NDVI-LST relationship in heterogeneous urban-mountain area. The proposed ML framework reliably produces 30 m LST grids from 2 km GK-2A inputs using widely available ancillary data (NDVI and DEM). Compared with the TsHARP, RF/XGB consistently yield lower errors over diverse Seoul, while remaining computationally practical for routine production. These downscaled LST data provide actionable inputs for urban heat-risk assessment and climate-adaptation planning in Korea and can be integrated into energy modeling approaches for improving efficiency and thermal comfort.