This dataset was generated following a two-step process. First, a machine-learning emulator with 18 years of JULES data (2000-2018) was trained.

The emulator was evaluated and validated using data from 2019 and 2020, respectively. The input features for the emulator are: surface downward short-wave (SW) radiation, minimum and maximum daily land surface temperature, Fraction of Photosynthetically Active Radiation (FAPAR), soil moisture content of the top layer, and a soil moisture availability factor.

These inputs are listed in Table 1 of the data product user guide. The output of the emulator is GPP in kg m-2 s-1. The project developed 13 emulators, which correspond to the 13 different Plant Functional Types (PFTs) from JULES (listed in Table 2), where each emulator has inputs and output specific to a particular PFT. The outputs of all 13 emulators are then combined using the PFT fraction to provide the total gridbox GPP.

The second step involved running the emulators using inputs from equivalent EO datasets to generate the final GPP dataset. The EO datasets used in this project are listed in Table 3 of the data product user guide, including two existing GPP datasets used for comparison to validate the product. There are two exceptions of inputs that do not come from EO, but we take them from JULES: surface downward SW radiation, which is part of the TRENDY forcing data used to drive JULES, and the soil moisture availability factor (fsmc), which is generated by JULES and does not currently have an EO equivalent. Further work is required to replace fsmc with another related variable that is observable.