Tower-to-global upscaling of terrestrial carbon fluxes driven by MODIS-LAI, Sentinel-3-LAI and ERA5-Land data

Recent efforts in upscaling terrestrial carbon fluxes (TCFs) from eddy covariance (EC) flux towers have gained momentum with machine learning, capturing complex relationships between TCFs and their driving variables. We applied Gaussian process regression (GPR) models to upscale TCF products from tower-to-global scale and studied the predictive capacity of climate variables and leaf area index (LAI) across biomes (2004-2023). The developed GPR models (EC-GPR-TCFs) were trained with FLUXNET data, including gross primary productivity, ecosystem respiration, and net ecosystem exchange. LAI field measurements were combined with climate variables observed at EC towers, including soil variables at varying depths. Upscaling was realized in Google Earth Engine at varying spatial resolutions from 300 m to 5 km, employing the EC-GPR-TCFs models with the MCD15A3H-LAI product from MODIS, the S3-TOA-GPR-LAI product derived from Sentinel-3 (S3) and ERA5-Land climate data. Tower data availability for training and validating the model is reduced when combining multiple variables and measurements, a challenge addressed by the EC-GPR-TCFs models. Bidecadal temporal correlation analysis (2004-2023) between TCFs and predictor variables highlighted three key variables for estimating TCFs: LAI, shortwave incoming solar radiation (SW), and latent heat flux (LE), with the strongest correlations in forests of temperate and cold climates. Multi-year (2019-2023) intercomparison of EC-GPR-TCFs estimations against EC towers data for validation revealed consistent results with $R^2$ and $rmse$ generally around 0.6 and below 3 μmol m^-2 s^-1, respectively. At the global scale, decadal (2010-2020) intercomparison against four benchmark TCF products (LPJ-GUESS, MOD17A2H, FLUXCOM, and SCOPE-GPR-TCFs) resulted in varying fits with corresponding median $R$ and $rmse$ values in the ranges 0.7-0.85; 1.97-4.36 μmol m^-2 s^-1 for GPP and 0.81-0.85; 2.20-3.38 μmol m^-2 s^-1 for RECO. Our analysis offers a deeper understanding of how climate predictors influence TCFs across biomes, providing a new perspective on TCF upscaling methodologies.