EFSO and DFS diagnostics for JMA’s global Data Assimilation system:
their caveats and potential pitfalls
Daisuke
Hotta
JMA
noon May 20 in Room 2155
Abstract:
Diagnostic methods have recently been devised that allow estimation of
observational impact on a Data Assimilation System (DAS). Such
diagnostics not only enable quantification of the “value” of
observations but can also provide clues to improving the diagnosed DAS.
Two such diagnostic methods, Ensemble Forecast Sensitivity to
Observations (EFSO; Kalnay et al. 2012) and Degrees of Freedom for
Signal (DFS; Liu et al. 2009), have been implemented to the EnKF
component of JMA’s pre-operational global hybrid DAS. This talk will
present findings achieved through these diagnostics. Particular
emphasis will be placed on caveats and potential pitfalls in
interpreting their results.
The first part of the talk will present EFSO implemented at JMA. It was
found that the forecast error reduction estimated by EFSO accounts for
only 20% of the actual forecast error reduction. In order to understand
mechanism behind this underestimation, we conducted diagnosis where the
forecast error vector is decomposed into the column and null space of
the ensemble forecast perturbations, recognizing that portions of
forecast errors that are in the null space are discarded during EFSO
computation. The result indicated that 80% of the forecast errors are
in fact in the null space, explaining the mechanism of the
underestimation. Reasons for why so little of the forecast error is
accounted for by the ensemble will also be discussed.
In the second part, I will show that that the information content
extracted from observations by EnKF as quantified by DFS is an order of
magnitude smaller than that by 4D-Var. This underestimation is
particularly conspicuous to dense observations. Theoretical
consideration reveals that, this is because, in EnKF, DFS can never
exceed the size of the background ensemble, limiting the amount of
information extractable from observations if the number of observations
is much larger than the ensemble size. Implications of this DFS
underestimation on broad aspect of EnKF, including localization,
inflation and observation thinning, will also be discussed.