Generation of global 1 km daily land surface–air temperature difference and sensible heat flux products from 2000 to 2020

Abstract

Accurate estimation of land surface sensible heat flux (H) is crucial for comprehending the dynamics of surface energy transfer and the cycles of water and carbon. Yet, existing H products mainly are meteorological reanalysis datasets with coarse spatial resolutions and high uncertainties. FLUXCOM is the sole remotely sensed product with its 0.0833° spatial and 8- temporal resolution spanning from 2001 to 2015, so there is still a need for accurate and high spatial resolution global product based on satellite data. To address these issues, we generated the first global high resolution (1 km) daily H product from 2000 to 2020 using long short-term memory (LSTM) deep learning models, incorporating data from the Global LAnd Surface Satellite (GLASS) product suite. Furthermore, considering that the difference between land surface temperature and air temperature (Ts-a) is a key driver of H, we introduce the first global accurate satellite-based Ts-a product. This product refines the uncertainty compared with obtaining Ts-a directly from existing products by subtracting air temperature from land surface temperature. Our model, distinct from previous models that estimate H per pixel through physically-based models requiring parameters that are not readily accessible, can conveniently derive global values and efficiently capture nonlinear interactions. Additionally, it accounts for the temporal variation of H. Validation against independent in-situ measurements yielded a root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R2) of 25.54 W m−2, 18.649 W m−2, and 0.54 for H, and 1.46 K, 1.073 K, and 0.52 for Ts-a, respectively. The estimated H and Ts-a values are more accurate than current products such as MERRA2, ERA5-Land, ERA5, and FLUXCOM under most conditions. Additionally, the new H product offers more detailed spatial information in diverse landscapes. The estimated global average land surface H from 2000 to 2020 is 35.29 ± 0.71 W m−2. These high-resolution H and Ts-a products are invaluable for climatic researches and numerous other applications. The daily mean values for the first three days of each year can be freely downloaded from https://doi.org/10.5281/zenodo.14986255 (Liang et al., 2025a), and the complete product is freely available at https://www.glass.hku.hk/ (last access: 8 September 2025).