Calibration of global scale gross primary productivity simulated with the ORCHIDEE model by assimilating OCO-2 SIF products

Our ability to anticipate the evolution of the spatial and temporal distribution of sources and sinks of carbon in terrestrial ecosystems under a changing climate largely relies on dynamic global vegetation models (DGVMs). Gross primary productivity (GPP) simulated by DGVMs show large spreads. They results from structural differences in the process representation and from the different values assigned to the associated parameters. The question of optimizing the parameter values in DGVMs from in situ measurements or space-borne observations through data assimilation has grown in importance over the years. Satellite retrievals of sun-induced fluorescence (SIF) over land surfaces have been available for nearly a decade. SIF and GPP are functionally related through the photosynthetically absorbed radiation (APAR), and SIF is often considered as a proxy of GPP: At large spatial and temporal scales, they exhibit a strong linear relationship, with scaling factors between the two depending however on biome type and climate. DGVMs are hence legitimately evolving to take advantage of SIF observations to optimize their parameters. This requires the development of an observation operator that functionally relates the model variables and parameters to observed SIF. In this study, we assimilate SIF retrievals from OCO-2 within the ORCHIDEE model. The modelling of SIF in ORCHIDEE relies on i) the modelling of the fluorescence yield at the leaf scale to simulate the dynamic regulation of SIF with meteorological conditions, and ii) the calculation of the total canopy fluorescence based on a parametric representation of the SCOPE model. A parametric model has been developed to estimate the Non Photochemical Quenching, and further the fluorescence yield. With this approach, we assimilate OCO-2 SIF products in ORCHIDEE in order to optimize some of its parameters related to photosynthesis and phenology. We evaluate the calibrated model simulations of the spatial and temporal distribution of GPP relative to the independent FLUXCOM GPP products, and comparatively to the constraints brought by other data-streams. We then discuss the limits of the approach and propose strategies aiming at making the best of the use of SIF observations for reducing uncertainties in DGVM parameterization.