Estimating gross primary production (GPP) over space and time is fundamental for understanding the feedbacks of the terrestrial biosphere under climate change. Eddy-covariance flux towers provide in situ measurements of GPP at ecosystems levels, but their representativeness remains geographically constrained and allows only limited conclusions for deviating conditions. Machine learning (ML) techniques have been used to address this issue by modeling land-atmosphere fluxes based on local GPP measurements and remotely sensed data.
Recent advances in Automated ML (Auto-ML) offer a new automated way to select and synthesize different ML models which outperform classical ML approaches in diverse problems. In this work, we trained different Auto-ML models on eddy-covariance measurements of GPP at 243 globally distributed locations. We assessed the models’ capabilities to predict GPP and its spatial and temporal variability based on different sets of remote sensing variables. We found that Auto-ML architectures typically outperform classical ML models, especially in estimating trends and interannual variabilities. Furthermore, the best-performing Auto-ML model was deployed to generate global wall-to-wall maps showing spatiotemporal patterns consistent with dynamic vegetation models. This research benchmarks the application of Auto-ML in GPP estimation and assesses the potential and limitations in quantifying long-term variabilities of global photosynthesis.