It is all a matter of the assumed background error covariance hypothesis and its parameters. What do you trust more the model or the observations? HF radar velocity observations are very tricky since it is a very noisy data, and depends on the processing: averaging, detiding, super-observations, errors of representation, etc. I am not an expert on HF radar, but our group is actively working on their impact in the circulation 4D-Var analysis. We are assimilating the radials instead of the biased gridded product. ROMS observation operator can assimilate the original HF radar radials. We are also working on quality control and standardization of this observation asset within the U.S. In general, assimilating velocity observations are very tricky because it requires sophisticated processing and error models. I know of several people that question how useful are the HF radar observations in data assimilation. As you have found out, it is a difficult dataset that requires special considerations. We have been carrying out observation impacts and observation sensitivities about these observations that we publish shortly.
One thing that we recommend is to use the dual (
W4DPSAS plus
RPCG) instead of the primal (
IS4DVAR) formulation. The primal formulation is becomming deprecated everywhere. The dual formulation is much better because it allows strong and weak constraint, and several useful analysis tools to examine the analysis and forecast observations impacts and observations seinsitivities
(4D-Var)T. See
WikiROMS for more information.
