Interpretable Machine Learning to Understand Multi-Scale Meteorological Impacts on Ecosystem Carbon Uptake

Terrestrial carbon uptake constitutes the main driver of interannual variations in atmospheric CO₂ and thus one of the least understood parts of the global carbon cycle. Meteorological factors, such as variations in weather patterns and extreme climate events, are the main drivers of interannual variations in the land carbon uptake. Assessing these impacts of meteorological events on the annual carbon balance, in terms of their timing and duration and their interaction with ecosystems, remains a challenging problem.

 

Here we propose a data-driven approach to shed light on the meteorological drivers of terrestrial carbon variability. We use a convolutional neural network to predict carbon and water fluxes in forest ecosystems, which is trained on wavelet-transformed key meteorological variables to explicitly represent a wide spectrum of time-scales in the input. We curate a dataset conflating eddy covariance data from 15 deciduous broadleaf forest sites from the FLUXNET network, meteorological measurements gap-filled with reanalysis data and a random walk variable for validation. The application of an explanatory machine learning technique provides insights into the importance of the different meteorological events in terms of their length and timing in controlling anomalies in the annual terrestrial carbon balance. Additionally, we test our approach trained on carbon and water fluxes output from a comprehensive land-surface model to evaluate the validity of the observation-driven model.

 

The model shows that water availability is the dominant factor of local variations in the carbon balance. In particular, vapour pressure deficit events lasting 20-40 days in summer are one of the most important drivers for the model to predict a lower annual carbon uptake. Furthermore, we compare these quantitative results and results of a case study of the 2003 heatwave with the model setup trained on land-surface model output. Such studies are important to demonstrate the potential of interpretable machine learning methods to improve our understanding of land-atmosphere interactions, and crucially, to learn the complex responses of ecosystems to meteorological variability from data.