A global daily dataset of surface heat fluxes produced by integrating in-situ observations and machine learning

The study utilized observation data from FLUXNET2015 TIER 2 and meteorological stations distributed globally to generate a global daily scale surface LE and H dataset using an artificial neural network (ANN) model. A deep neural network model was constructed by integrating multiple sources of information including GLASS atmospheric net radiation, MODIS vegetation types, GLASS leaf area index, meteorological station data, and GLAS surface data to generate a global daily scale surface energy flux dataset covering 4933 meteorological stations from 2001 to 2024. Our validation shows that the dataset has good performance at the site level: the root mean square errors (RMSE) of predicted surface net radiation (LE) and surface latent heat (H) are 14.23 watts/square meter and 16.72 watts/square meter, respectively, with determination coefficients (R ²) of 0.86 and 0.82, respectively. This dataset provides more reliable data support for validating climate model predictions and conducting global climate change research, especially at specific site locations.