Enhancing quantification of local carbon sinks through Eddy Covariance CO2 Flux and Machine Learning

Under the Paris Agreement, countries are encouraged to preserve and enhance existing carbon sinks. Europe, in particular, has committed to achieving climate neutrality—attaining a balance between anthropogenic emissions from sources and removals by sinks—by 2050. Achieving these ambitious goals requires accurate and credible estimation of CO2 fluxes. However, discrepancies between observations and global models hinder the tracking of collective progress towards climate neutrality. Efforts to improve transparency and data comparability are crucial to better align national mitigation strategies with global pathways. In particular, effective climate mitigation policies increasingly depend on local-level actions where detailed data on CO₂ removals from forests and other land uses are traditionally lacking. Addressing the uncertainty in land-sector mitigation potential and enhancing the availability and comparability of data are critical for achieving climate goals by cities and regions. Different models, including process-based and data-driven approaches, exist to estimate land carbon fluxes, but their application and accuracy often vary significantly depending on the scale and quality of input data.

In this study, we tested a data-driven method based on eddy covariance (EC) data to quantify the current role of the regional carbon sink of the Aosta Valley Region (Italy) through the integration of various approaches. Our model relies on FLUXCOM-X framework specifically trained to achieve robust results at the regional scale. An XGBoost model was developed using global hourly meteorological data from sites across the global eddy covariance networks paired with remote sensing data from MODIS. The algorithm was optimized through feature selection analysis and best training subset selection, identifying the ensemble of experimental sites that provided the most accurate predictions while avoiding overfitting. The optimal training subset was obtained via partitioning the full range of sites into subsets based on key characteristics (Plant Functional Type, geographical zone, biogeographical region, elevation). This approach ensured the biophysical comparability of the sites with the target region (Aosta Valley) while maintaining a balance between generalizability and specificity. Model evaluation focused on how the model performed on the local eddy covariance measurements. The resulting model was subsequently upscaled to the regional level. This was achieved using eddy covariance measurements of CO2 fluxes, MODIS NDVI (250 m resolution), daily gridded meteorological data at 100 m resolution, and a land cover map at 250 m resolution. Moreover, the methodology demonstrated potential for replication in other local realities such as regions, providing a flexible framework for assessing local carbon budgets and supporting climate-smart management strategies. Our results were finally compared with independent data from the National Forest Inventory (NFI) available for the target area (Aosta Valley). Discrepancies between methods will be analyzed, considering their strengths, weaknesses, and spatio-temporal variability.