Science

A draft review of advanced data assimilation techniques is found in D3.1

Some interesting presentations have been discussed during the first annual meeting and progress meeting. Research papers will appear hopefully soon in the Documents

Some highlights in research in new data assimilation techniques

A particule filter method trying to avoid the curse of dimensionality: The equivalent weight particle filter (EWPF) was applied to a 2D case and cases with nonlinear observing operators. (Presentation at progress meeting and nonlinear H )

A Multivariate Rank Histogram Filter (MRHF) exploits joint probability density functions to infer unobserved variables from observed ones. When working in a low dimensional system, the approach performs very well for non-gaussian problems, but in higher dimensions some strong hypotheses need to be formulated. By neglecting unobserved variables in some of the conditional statements a feasible MRHF is within reach but good insight in the effect of the hypotheses is needed (Presentation at progress meeting and Poster)

Stochastic perturbations applied to the state equation providing heat and salt conserving perturbations leading to nice and stable ensemble spread (Presentation at progress meeting and publication)

Localization is often applied for ensemble methods. Two problems which occur can now be tackled: how to build the localisation function objectively and how to maintain global convervation constraints when adding localisation. (Presentation at progress meeting )

Ensemble methods heavily exploit the reduced rank of the covariance matrix for efficient matrix inversions. A new method combining such reduced rank matrices with local parametric covariances was developped. It still leads to efficient computations but in addition allows for scale separations in the analysis (Publication)

Anamorphosis is a way to deal with data having a non-gaussian distribution. Its effect in a coupled physical-biochemical model indicates better spatial correlations when using the transformation (Publication)

Some highlights in research in diagnostics and benchmarks

Selecting among different locations and combinations of observing systems is simplified by characterizing the incremental information via the spectrum of the observed part of the error covariance of the forecast. This leads to selection criteria which allow to distinguish the added value of ferroboxes vs gliders for example (Presentation at progress meeting )

The medium size benchmark was successfully implemented by a post-doc new to the assimilation toolbox and the NEMO model within 3 month. Adding then different assimilation strategies (difference incremental assimilation updates IAU with different levels of implicitness) allowed to show the benefits of using IAU(0) compared to standard intermittend updates. (Publication)

The large scale benchmark was sucessfully implemented by two groups with different complementary perturbations for the first time in such a context with 4D observing operators (Poster at GODAE and Poster et EGU)