 |
Center for Nonlinear Studies |
A good representation of state error statistics in square root or ensemble Kalman filters is often difficult to obtain. Firstly, because the nature of errors is complex, poorly known and multi-source. Secondly, because these statistics must be represented with a limited number of error modes or ensemble members. What can easily lead to the filter divergence. So, an adaptive parameterization of the errors statistics is often necessary to avoid the drift of the filter and also to prevent the error estimates inconsistent with the reality. A recipe consists in adjusting the statistics online during the assimilation process, a procedure commonly referred to as adaptive filtering. The classical strategy (Dee, 1995) is to introduce a dependence of error statistics on some random parameters. These parameters are then estimated to get the best possible coherence between the state error statistics and the innovation statistics. It used the real differences between predictions and observations and error statistics are then improved by a temporal discrepancies between predictions and observations. The objective of the present study is to apply and evaluate the formulation developed in a recent paper (Brankart et al., 2010) in a more realistic oceanographic context. The benefits of adaptive filtering is illustrated with results from twin experiments with a high resolution ocean model (NEMO).
Host: Balu Nadiga, CCS-2, balu@beasley.lanl.gov