Global Carbon Fluxes Using Multioutput Gaussian Processes Regression and MODIS Products

Abstract

The quantification of carbon fluxes (CFs) is crucial due to their role in the global carbon cycle having a direct impact on Earth's climate. In the last years, considerable efforts have been made to scale CFs from eddy covariance (EC) data to the globe. In this work, a data-driven approach that exploits a multioutput Gaussian processes regression algorithm ($\mathcal{G}$-model) is proposed to jointly estimate gross primary production (GPP), terrestrial ecosystem respiration (TER), and net ecosystem exchange (NEE) at a global scale. The $\mathcal{G}$-model not only provides an estimate of the CFs but also an uncertainty. Moreover, it derives the three fluxes jointly preserving their physical relationship. The predictors are selected from a set of the moderate-resolution imaging spectroradiometer (MODIS) products available on Google Earth engine. The performance of the model revealed high accuracies (R2 reaching 0.82, 0.69, and 0.80 in the case of GPP, NEE, and TER, respectively), and low root mean square errors (1.55 g m−2 d−1 in the case of GPP, 1.09 g m−2 d−1 for the NEE, and 1.14 g m−2 d−1 for TER) over the FLUXNET2015 data set at eight-day time scale. The GPP estimates provided by $\mathcal{G}$-model outperformed the MOD17A2 product, and a state-of-the-art GPP product (PML_V2) without using meteorological forcing data sets. The results reported mean annual amounts of 133.7, 114.8, and 18.9 Pg yr−1 for GPP, TER, and NEE, respectively, during the 2002–2023 period. The proposed approach paves the way for the development of multioutput strategies that preserve the physical relationships among CFs in upscaling processes.

Keywords

• Data-driven
• gross primary production (GPP)
• machine learning (ML)
• multioutput Gaussian process regression
• net ecosystem exchange (NEE)
• terrestrial ecosystem respiration (TER)