Data assimilation through an ensemble
Kalman filter (EnKF) is not exempt from deficiencies, including the
generation of long-range unphysical correlations that degrade its
performance. The covariance localization technique has been proposed and
used in previous research to mitigate this effect. However, an
evaluation of its performance is usually hindered by the sparseness and
unsustained collection of independent observations.
This
article assesses the performance of an ocean prediction system composed
of a multivariate EnKF coupled with a regional configuration of the
Regional Ocean Model System (ROMS) with a covariance localization
solution and data assimilation from an ocean glider that operated over a
limited region of the Ligurian Sea. Simultaneous with the operation of
the forecast system, a high-quality data set was repeatedly collected
with a CTD sensor, i.e., every day during the period from 5 to 20 August
2013 (approximately 4 to 5 times the synoptic time scale of the area),
located on board the NR/V Alliance for model validation.
Comparisons between the validation data set and the forecasts provide
evidence that the performance of the prediction system with covariance
localization is superior to that observed using only EnKF assimilation
without localization or using a free run ensemble. Furthermore, it is
shown that covariance localization also increases the robustness of the
model to the location of the assimilated data. Our analysis reveals that
improvements are detected with regard to not only preventing the
occurrence of spurious correlations but also preserving the spatial
coherence in the updated covariance matrix. Covariance localization has
been shown to be relevant in operational frameworks where short-term
forecasts (on the order of days) are required.