Addressing sample distribution imbalances through data augmentation, then using multi-source data and machine learning algorithms to generate long-term inter annual 250m global terrestrial ecosystem net productivity products.This product was developed using multiple data sources, including NEP observations from global flux networks (AmeriFlux, Fluxnet, ChinaFlux, and ICOS), MODIS NDVI and land cover data, meteorological variables (temperature, precipitation, and downward solar radiation) from the ERA5 reanalysis dataset produced by the European Centre for Medium-Range Weather Forecasts (ECMWF), soil organic carbon content at 0.3 m depth from OpenLandMap, and global atmospheric CO₂ concentration data. The methodology for product production involved several steps: First, flux tower-observed NEP data underwent quality control and temporal aggregation to generate annual NEP. Other multi-source datasets were preprocessed and extracted according to flux tower locations to establish an initial sample set. The spatial distribution and land cover representativeness of the samples were then analyzed, with data augmentation techniques employed to address sample imbalance. Finally, a random forest model was trained using the augmented dataset to produce a global 250-meter resolution NEP product covering the period from 2001 to 2022.