Global patterns of land-atmosphere fluxes of carbon dioxide, latent heat, and sensible heat derived from eddy covariance, satellite, and meteorological observations

We demonstrate progress in upscaling FLUXNET observations to the global scale using a machine learning technique called Model Tree Ensembles (MTE). We trained MTE to predict gross primary productivity (GPP), terrestrial ecosystem respiration (TER), net ecosystem exchange (NEE), latent energy (LE), and sensible heat (H) based on vegetation indexes retrieved by remote sensing, climate, and land use information. We applied the trained MTEs to generate global fields of these fluxes at 0.5° spatial and monthly resolution from 1982-2008. Our estimated land surface-atmosphere fluxes of LE (39 ± 2 W/m2), H (41 ± 4 W/m252 ), and GPP (119 ± 6 PgC/yr) are within the range of independent global estimates. We inferred that our estimate of global TER is 96 ± 6 PgC/yr appears to be low by 5-10 %, but direct independent global quantification of TER is lacking. We identified hot spot regions of interannual variability of carbon fluxes with observations located in semi-arid to semi-humid regions, where variations in moisture supply control the interannual variability of biosphere-atmosphere fluxes. This pattern is corroborated by simulations of four dynamic global vegetation models. We find that interannual variability of NEE is driven by GPP in most places. Our results stress the significance of the hydrological cycle in driving global biogeochemical cycles. The empirically-derived products of biosphere-atmosphere fluxes that integrate a large body of Earth observation data are well suited for calibration and evaluation of process-oriented Land Surface Models, as well as for explorative analysis and diagnostic assessments of the biosphere.