An effective machine learning approach for predicting near real-time ecosystem carbon cycle

Tropical forests store about half of the world’s above ground carbon and act as critical climate regulators as they absorbed one third of the global CO2 emissions over the past decades. These estimates of the present-day (and future) land carbon sinks are primarily obtained from land surface models (LSM) which are mechanistic tools that simulate the processes occurring at the interface between the atmosphere, the biosphere and the pedosphere. LSM are hence critical tools for understanding and predicting the dynamics of the land surface, its role in a changing Earth, and the impact of future climate and disturbances on its functioning. However, LSM have become increasingly complex and slow machinery that require heavy expert knowledge and computational tools to run. In this study, we tested whether data-driven (black box) models could efficiently reproduce process-based (mechanistic) models. To do so, we trained machine learning algorithms (gradient-boosted decision trees) with the model outputs of TrENDYv11 that were initially generated to estimate the global land carbon sink. Data-driven models performed extremely well in reproducing the long-term trends and the seasonality of the carbon sink over the Tropics, with an average accuracy of 91% and could further be used to make predictions, including near real-time forecasting of the carbon cycle of forests. We illustrate the latter by quantifying the impacts of last-year El-Niño on tropical ecosystem productivity, with a specific focus on the severe drought in the Amazon. While the simulations of the process-based models will only emerge in a year or so when the different teams will have run their own models, our tool could simulate in near real-time that the 2023 drought was for the largest reduction of Amazon GPP in recent history.