ASSIMILATION OF BIOGEOCHEMICAL OBSERVATIONS IN OCEAN MODELS

Biogeochemical models of the ocean are characterised by several features which make their calibration/validation critical. First, the relationships between different trophic levels are very often based on semi empirical formulation, which are associated with badly known parameters, possibly varying in time and space. Secondly, these relationships are strongly non-linear. Finally, considering the difficulty of in situ biogeochemical observations, oceanic data are rather sparse. Nevertheless, satellite (sea colour) data now available allows a quasi synoptic global coverage of surface chlorophyll distribution. Moreover, during the last twenty years, large progress was made in computing capacity and in data assimilation techniques in environmental sciences. Therefore, together with in situ high frequency observations obtained from new automised devices (sensors on buoys or moorings), this new type of satellite data can be used to better constrain biogeochemical ocean models. Contrary to more classical applications of assimilation methods, aiming at obtaining the state (initial conditions) of a system, the major issue in ocean biogeochemistry deals with parameter estimation or inversion. The functioning of the marine ecosystems strongly depend upon hydrodynamical and environmental conditions : therefore, many model parameters should be "regionalised" in order to take into account the variability of those conditions. Taking into account the strong non-linearity of biogeochemical models, it is also important to wonder which process or trophic level can be constrained by a given data set : as an example, one can wonder how the information of sea colour data is transferred toward high trophic levels, in order to estimate the flux of ocean organic matter into the deep ocean ? The use of data assimilation in ocean biogeochemistry is at its infancy. In order to present some major scientific issues, as well as more methodological questions, several studies already achieved will be discussed. New developments in progress in our laboratory will also be shown : these works use mostly classical methods based on the adjoint model, but considerations on neural networks will be presented as well.