EXCITED: an open machine learning workflow for estimating terrestrial carbon exchange

Despite its importance to the global carbon budget, the exchange of carbon dioxide between atmosphere and vegetation is currently not accurately quantified. To model the global biospheric CO2 exchange, data-based (machine learning) models have been developed using training data from eddy-covariance measurement sites [1, 2]. While these models are widely used and can successfully predict carbon fluxes on short timescales, they can severely overestimate the annual carbon uptake by many ecosystems.

In the EXCITED project, we aim to better constrain the CO2 exchange by terrestrial ecosystems on longer timescales using estimates from inverse models (i.e., CarbonTracker) as additional input data. The workflow consists of first training two models: one on the site-based data, and one on the CarbonTracker data. While the site-based model can produce fluxes on small temporal and spatial scales, the CarbonTracker-based model will be more accurate on long time scales. From the machine learning models we can then produce (global) datasets. Finally, these datasets can be merged to produce a dataset which has the best of both worlds.

Aside from the produced datasets, we will make the trained machine learning models available, as well as the full workflow that generates the model and data. The workflow consists of the (Python) code, Jupyter notebooks, along with documentation to guide new users. Having the code and documentation openly available makes it easier for others to adapt it to their needs or to further extend it. With the open workflow we aim to build a community around these tools for modeling and forecasting terrestrial carbon exchange.