B32A-06
FLUXCOM - Overview and First Synthesis

Wednesday, 16 December 2015: 11:35
2006 (Moscone West)
Martin Jung1, Kazuhito Ichii2, Gianluca Tramontana3, Gustau Camps-Valls4, Christopher R Schwalm5, Dario Papale6, Markus Reichstein1, Fabian Gans1, Ulrich Weber1 and FLUXNET Team, (1)Max Planck Institute for Biogeochemistry, Jena, Germany, (2)JAMSTEC Japan Agency for Marine-Earth Science and Technology, Kanagawa, Japan, (3)Tuscia University, Viterbo, Italy, (4)University of Valencia, Burjassot, Spain, (5)Federal GEOS Funding, Phoenix, AZ, United States, (6)Tuscia University, Department for Innovation in Biological, Agro-food and Forest systems (DIBAF), Viterbo, Italy
Abstract:
We present a community effort aiming at generating an ensemble of global gridded flux products by upscaling FLUXNET data using an array of different machine learning methods including regression/model tree ensembles, neural networks, and kernel machines. We produced products for gross primary production, terrestrial ecosystem respiration, net ecosystem exchange, latent heat, sensible heat, and net radiation for two experimental protocols: 1) at a high spatial and 8-daily temporal resolution (5 arc-minute) using only remote sensing based inputs for the MODIS era; 2) 30 year records of daily, 0.5 degree spatial resolution by incorporating meteorological driver data. Within each set-up, all machine learning methods were trained with the same input data for carbon and energy fluxes respectively. Sets of input driver variables were derived using an extensive formal variable selection exercise. The performance of the extrapolation capacities of the approaches is assessed with a fully internally consistent cross-validation.

We perform cross-consistency checks of the gridded flux products with independent data streams from atmospheric inversions (NEE), sun-induced fluorescence (GPP), catchment water balances (LE, H), satellite products (Rn), and process-models. We analyze the uncertainties of the gridded flux products and for example provide a breakdown of the uncertainty of mean annual GPP originating from different machine learning methods, different climate input data sets, and different flux partitioning methods. The FLUXCOM archive will provide an unprecedented source of information for water, energy, and carbon cycle studies.