Data Assimilation Meetings at Reading

Date Meeting type Speakers
3 October 2012 DARC seminar Jochen Broecker (UoR)
Statistical Assessment of Data Assimilation Algorithms
Suppose we have some observations as well as a model, and suppose we have assimilated the observations into it. Now, how do we know that our solution is any good? Launching forecasts and evaluating them against real observations is a hard check, but we would have to do this many times in order to get significant results, and we would really only evaluate the terminal point of the assimilated solution. I am thinking more of assimilating a long trajectory, for example for reanalysis purposes.
Another option is to compare the solution with the original observations. First of all, this suggestion should set off the alarm bells, as we are using the observations to fit the solution AND to evaluate it. This in-sample evaluation, taken at face value, might give overly optimistic results. As I will show in this talk, this problem can be taken care of, though. It is possible to compute (and substract) the optimisim in in-sample evaluations at least approximately. Of course, terms and conditions apply. I will show this for some very simple examples in order to give you the idea, and then discuss possible applications to data assimilation if time permits.

Page navigation