A coastal ocean data assimilation system is being developed. The goal is to combine large and disparate datasets with ocean numerical models, producing accurate analyses, forecasts, and respective uncertainty estimates for any littoral region. A modular interface combines the Estuarine and Coastal Ocean Model (ECOM) and the Local Ensemble Transform Kalman Filter (LETKF) into a highly scalable, portable and efficient ocean data assimilation system. The ECOM is a state-of-the-art, three-dimensional, hydrodynamic ocean model developed as a derivative of the Princeton Ocean Model [Blumberg et al., J. Hydrologic Eng., 125, 799–816, 1999]. The LETKF, a recent adaptation of ensemble Kalman filtering techniques, works particularly well for very large non-linear dynamical systems in both sparse and dense data regimes, and provides efficient algorithms for error estimation and quality control [Szunyogh et al., Tellus A, 60, 113–130, 2008]. In simulation experiments for highly idealized data distributions in the New York Harbor Observing and Prediction System (NYHOPS) the filter quickly converges, eliminating bias and greatly reducing rms errors [Hoffman et al., JAOT, in press, 2008]. This behavior is robust to changes in ensemble size, data coverage, and data error.
Ensemble data assimilation provides a promising path for making use of remotely sensed ocean data such as sea surface temperature, ocean color, turbidity, surface currents, free surface elevation, and sea surface salinity. In theory, the ensemble approach provides the best way for observations of one variable to affect the analysis of other correlated variables, either collocated or nearby, through the depth of the water column. With many sub-models available in the ECOM for biogeochemistry, sediment transport, water quality, waves, and particle tracking, there are opportunities to extend the assimilation to non-standard data such as ocean color and turbidity, chemical tracers, wave energy, and locations of drifting buoys and autonomous underwater vehicles. These opportunities exist because the LETKF method is completely general in the sense that when the observation errors can be assumed to be Gaussian, any observation of a physical parameter that has a known functional dependence on the variables of the dynamical model, can potentially be usefully assimilated.
In ensemble data assimilation due to a limited sample we expect many distant correlations to appear to be significant. But they are not. Localization limits the region considered, eliminating these spurious correlations. But how big should the localization region be? In this approach we will increase the sample size by considering samples over the length of the data assimilation experiment (at least a few days, and maybe a month). For these very large samples correlations are expected to die down with distance in a smooth manner and the e-folding distance can be used to set the localization domain. We will explore how these parameters vary with depth, location, and variable. Because the NYHOPS geometry includes rivers, bays, coastal areas, and open ocean shelf regions, the correlation structures are quite varied. The method uses simulated data only, and can be applied in situations with few observations. Several extensions to this work are possible. Of particular interest would be application to unusual observing systems that integrate over space or time where the best approach to localization is not straightforward. Example observing systems of this type include for example stream flow, GPS soundings, and radiances.
Supplementary URL: