Observation sensitivity based on ensembles for storm surge forecasting

In addition to observation sensitivity experiments, either with simulated data (OSSE) or with existing observations (OSE), the observation sensitivity method introduced by Langland and Baker [3, 2,1] is quickly becoming a standard tool for assessing the value of parts of the observing system. The value of this adjoint based sensitivity method is that, with little additional computation, the impact of observations can be grouped in many dierent ways. Recently, Liu and Kalnay [5, 4] extended this approach to ensemble based data-assimilation methods. In meteorological applications the impact score used for evaluation of the observation impact is very often an energy norm of the dierence between forecast and the 'verifying' analysis. This measure requires that the analysis is signicantly more accurate than the forecast almost everywhere. However, in some oceanographic applications the assimilation of observations may have a negative impact on the accuracy in data-sparse regions. Therefore, we propose to use the dierences between forecast and observations as a measure. As this measure is needed in many data-assimilation methods, it is likely that this requires little additional programming effort. The ensemble based observation sensitivity method is applied to a storm surge model for the North Sea. Since the observations in this case consist of time-series with a time-step of typically 10 minutes, the impact of observations at only one time is quite small. This implies that the impact of earlier observations can no longer be ignored. We extend the method to include the impact of previous observations. The implications of this modication are illustrated with a simple prototype model. The method is implemented in the generic data-assimilation framework OpenDA (www.openda.org). Future work will aim to deign an improved observation network for tide-gauges in this area.