Projection of Global Gross Primary Production using Machine Learning approaches

An accurate estimation of the amount of carbon uptake by the terrestrial ecosystem is crucial for informing future global carbon budget assessments. Terrestrial gross primary productivity (GPP) plays a crucial role in carbon-cycle dynamics and climate feedback mechanisms. This study aims to forecast future changes in GPP using a range of Earth system models from the Intercomparison Project Phase 6 (CMIP6). Firstly, we developed a machine learning-based model for global terrestrial GPP estimation, incorporating three key meteorological variables: 2-meter air temperature (T2M), diurnal temperature range (DTR), and precipitation (PRE). Monthly meteorological data from the Climatic Research Unit gridded Time Series (CRU-TS v4.07) and GPP datasets from the Trends and Drivers of the Regional-Scale Sources and Sinks of Carbon Dioxide (TRENDY v.10) ensemble were utilized at a spatial resolution of 0.5° × 0.5° (720 × 360 grids). Three machine learning algorithms—Random Forest (RF), Support Vector Regressor (SVR), and Artificial Neural Networks (ANN)—were employed and compared, in company with a multi-linear regression (MLR) baseline model. Training spanned 1981–2014 (408 months), with validation conducted for 2015–2020 (72 months). All three machine learning models demonstrated robust estimation accuracy, achieving R2 values of 0.99 and root-mean-square-errors (RMSEs) below 2.88 gCkm^-2s^-1, except for MLR (R2 of 0.99 and RMSE of 3.26 gCkm^-2s^-1). Subsequently, we projected future global annual GPPs using these machine learning models and CMIP6 Earth system models under various climate scenarios from 2015 to 2100. While our models generally showed an under-estimation pattern compared to existing GPP projections based on CMIP6 scenarios, the ANN-based GPP projections displayed an only overestimation trend in the SSP585 scenario. The proposed GPP estimation models and projections would expand our understanding regarding carbon uptakes under current Earth system models.