Long-Term (2013–2024) Reconstructed All-Sky Land Surface Temperature from MODIS Integrating Machine Learning and Cloud Conditions for South Korea

Abstract

Long-term, spatially continuous land surface temperature (LST) data are crucial for monitoring surface conditions and understanding thermal dynamics across regions. Although the Moderate Resolution Imaging Spectroradiometer (MODIS) provides one of the most extensively used satellite-based LST products, its applicability is limited by data gaps under cloudy conditions. Despite various efforts to reconstruct MODIS LST, existing datasets are still constrained by temporal discontinuities and limited representation of cloud effects. In this study, we reconstruct an all-sky 1 km LST dataset based on MODIS for South Korea spanning 2013–2024 by applying a machine learning approach, providing continuous spatiotemporal coverage for all seasons. A light gradient boosting machine model was trained with Local Data Assimilation and Prediction System (LDAPS) analysis data, MODIS-derived variables including cloud fraction, and other auxiliary variables to account for cloud effects on LST. Clear-sky MODIS LST and in situ LST observations collected under cloud cover were used as target variables. Under clear-sky conditions, the model showed robust performance against MODIS LST, achieving a coefficient of determination (R²) of 0.96–0.97 and a root mean square error (RMSE) of 1.79–1.94°C during daytime, and an R² of 0.95–0.98 with an RMSE of 1.16–1.87°C at night in a spatial five-fold cross-validation. For cloudy-sky conditions, leave-one-out cross-validation using in situ LST from 25 weather stations yielded an R² of 0.91–0.93 with an RMSE of 2.26–2.76°C during daytime, and an R² of 0.96–0.98 with an RMSE of 1.38–1.76°C at night. Seasonal evaluations further revealed that the proposed model consistently outperformed LDAPS across all seasons. The reconstructed all-sky LST dataset effectively captured long term LST trends across South Korea and aligned well with extreme weather phenomena such as heatwaves and coldwaves. These results highlight its strong potential for diverse applications in LST monitoring, urban climate assessment, and climate change studies.