[1] We design a Bayesian inversion method (gradient-based) to optimize the key functioning parameters of a process-driven land surface model (ORganizing Carbon and Hydrology In Dynamic EcosystEms (ORCHIDEE)) against the combination of prior information upon the parameters and eddy covariance fluxes. The model calculates energy, water, and CO 2 fluxes and their interactions on a half-hourly basis, and we carry out the inversion using measurements of CO 2 , latent heat, and sensible heat fluxes as well as of net radiation over a pine forest in southern France. The inversion method makes it possible to assess the reduction of uncertainties and error correlations of the parameters. We designed an ensemble of inversions with different set ups using flux data over different time periods, in order to (1) identify well-constrained parameters and loosely constrained ones, (2) highlight some model structural deficiencies, and (3) quantify the overall information gained from assimilating each type of CO 2 or energy fluxes. The sensitivity of the optimal parameter values to the initial carbon pool sizes and prior parameter values is discussed and an analysis of the posterior uncertainties is performed. Assimilating 3 weeks of half-hourly flux data during the summer improves the fit to diurnal variations, but merely improves the fit to seasonal variations. Assimilating a full year of flux data also improves the fit to the diurnal cycle more than to the seasonal cycle. This points out to the key importance of timescales when inverting parameters from high-frequency eddy-covariance data. We show that photosynthetic parameters such as carboxylation rates are well-constrained by the carbon and water fluxes data and get increased from their prior values, a correction that is corroborated by independent measurements at leaf scale. In contrast, the parameters controlling maintenance, microbial and growth respirations, and their temperature dependencies cannot be robustly determined. The CO 2 flux data could not discriminate between the different respiration terms. At face value, all the parameters controlling the surface energy budget can be safely determined, leading to a good model-data fit on different timescales.Citation: Santaren, D., P. Peylin, N. Viovy, and P. Ciais (2007), Optimizing a process-based ecosystem model with eddy-covariance flux measurements: A pine forest in southern France, Global Biogeochem. Cycles, 21, GB2013,

We investigate the benefits of assimilating in situ and satellite data of the fraction of photosynthetically active radiation (FAPAR) relative to eddy covariance flux measurements for the optimization of parameters of the ORCHIDEE (Organizing Carbon and Hydrology in Dynamic Ecosystem) biosphere model. We focus on model parameters related to carbon fixation, respiration, and phenology. The study relies on two sites-Fontainebleau (deciduous broadleaf forest) and Puechabon (Mediterranean broadleaf evergreen forest)-where measurements of net carbon exchange (NEE) and latent heat (LE) fluxes are available at the same time as FAPAR products derived from ground measurements or derived from spaceborne observations at high (SPOT (Satellite Pour l′Observation de la Terre)) and medium (MERIS (MEdium Resolution Imaging Spectrometer)) spatial resolutions. We compare the different FAPAR products, analyze their consistency with the in situ fluxes, and then evaluate the potential benefits of jointly assimilating flux and FAPAR data. The assimilation of FAPAR data leads to a degradation of the model-data agreement with respect to NEE at the two sites. It is caused by the change in leaf area required to fit the magnitude of the various FAPAR products. Assimilating daily NEE and LE fluxes, however, has a marginal impact on the simulated FAPAR. The results suggest that the main advantage of including FAPAR data is the ability to constrain the timing of leaf onset and senescence for deciduous ecosystems, which is best achieved by normalizing FAPAR time series. The joint assimilation of flux and FAPAR data leads to a model-data improvement across all variables similar to when each data stream is used independently, corresponding, however, to different and likely improved parameter values.