Assessment of a regional physical–biogeochemical stochastic ocean model. Part 2: Empirical consistency

Abstract

In this Part 2 article of a two-part series, observations based on satellite missions were used to evaluate the empirical consistency of model ensembles generated via stochastic modelling of ocean physics and biogeochemistry. A high-resolution Bay of Biscay configuration was used as a case study to explore the model error subspace in both the open and coastal ocean. In Part 1 of this work, three experiments were carried out to generate model ensembles by perturbing only physics, only biogeochemistry, and both of them simultaneously. In Part 2 of this work, empirical consistency was checked, first by means of rank histograms projecting the data onto the model ensemble classes, and second, by pattern-selective consistency criteria in the space of “array modes” defined as eigenvectors of the representer matrix. Rank histograms showed large dependency on geographical region and on season for sea surface temperature (SST), sea-level anomaly (SLA), and phytoplankton functional types (PFT), shifting from consistent model-data configurations to large biases because of model ensemble underspread. Consistency for SST array modes was found to be verified at large, small and coastal scales soon after the ensemble spin-up. Array modes for the along-track sea-level showed useful consistent information at large scales and at the mesoscale; for the gridded SLA was verified only at large scale. Array modes showed that biogeochemical model uncertainties generated by stochastic physics, were effectively detected by PFT measurements at large scales, as well as at mesoscale and small-scale. By contrast, perturbing only biogeochemistry, with an identical physical forcing across the ensemble, limits the potential of PFT measurements at detecting and possibly correcting small-scale biogeochemical model errors. When an ensemble was found to be inconsistent with observations along a particular direction (here, an array mode), a plausible reason is that other error processes must have been active in the model, in addition to the ones at work across the ensemble.