Towards probabilistic analyses and predictions of the Green Ocean using a stochastic NEMO-PISCES modelling system

The Copernicus Marine service (CMEMS) in operation today routinely delivers information about the Green Ocean based on satellite and in situ data combined with numerical models. The aim is to provide users with “best estimate” representations of the state of marine ecosystems and biogeochemical indicators of interest. A key strategic evolution at Copernicus 2 horizon will be to consolidate the service with more robust information about product uncertainties, whether in real time, in delayed mode (reanalyses) and in forecast mode with a few days of lead time. In that perspective, the transition to probabilistic analysis and prediction methodologies is a necessary step, e.g. to provide more actionable information to help in decision-making and management of marine ecosystems.

In the frame of the H2020 SEAMLESS project, ensemble generation methods are being developed with the aim to improve the service through better data assimilation / inversion methods. A stochastic version of the NEMO-PISCES model has been developed and implemented in a global ocean configuration at ¼° inherited from the CMEMS global Monitoring and Forecasting Centre.

A 40-member ensemble is generated using 2019 unperturbed ERA5 atmospheric forcings and assuming uncertainties associated to (i) 7 critical biogeochemical model parameters of the PISCES formulation; (ii) sub-grid scale effects associated to the eddy-permitting resolution, and (iii) misplacement of mesoscale structures and associated advective/diffusive fluxes. The resulting 40-member ensemble represents a probabilistic view of the 2019 seasonal cycle in the global and North Atlantic ocean.

The ensemble is analysed in terms of spread, median, min and max distributions of model state variables related to surface chlorophyll concentration, as well as on a variety of targeted indicators (e.g. NPP, phenology, trophic efficiency). In order to evaluate the relevance of the ensemble pdfs with respect to observed data, verification statistics have been produced to check the consistency against daily L4 ocean colour products from the CMEMS catalogue. The computed metrics include rank histograms, CRPS (decomposed into reliability and resolution skill scores) and RCRV.

We will present a synthesis of the ensemble scores obtained in the different regions, highlighting situations where the prior ensemble is consistent with uncertainty hypotheses made in the stochastic NEMO-PISCES model. Further, we will show how to take into account irreducible uncertainties in the verification data products to compute the scores. We will discuss the sensitivity of the computed metrics against these uncertainties, underlying the importance of properly accounting for error propagation in the CMEMS TAC production chains. We will finally describe first applications of a new 4D Bayesian inversion scheme aimed at delivering probabilistic analyses and predictions with a few days of lead time.