Optimizing a process-based ecosystem model with eddy-covariance flux measurements: A pine forest in southern France

摘要

[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 CO2 fluxes and their interactions on a half-hourly basis, and we carry out the inversion using measurements of CO2, 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 CO2 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 CO2 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.