A Simulation Study Using a Local Ensemble Transform Kalman Filter for Data Assimilation in New York Harbor

Author: Ross N. Hoffman, Rui M. Ponte, E.J. Kostelich, A.F. Blumberg, I. Szunyogh, Sergey Vinogradov and John M. Henderson
September 1, 2008 - April 1, 2010
Journal Article
Journal of Atmospheric and Oceanic Technology

Hoffman, R.N., R.M. Ponte, E.J. Kostelich, A. Blumberg, I. Szunyogh, S.V. Vinogradov, and J. M. Henderson, 2007. A simulation study using a local ensemble transform Kalman filter for data assimilation in New York Harbor, J. of Atmos. and Oceanic Tech., Vol. 25 (September 2008), pp. 1638--1656, doi:10.1175/2008JTECHO565.1.

Data assimilation approaches that use ensembles to approximate a Kalman filter have many potential advantages for oceanographic applications. To explore the extent to which this holds, the Estuarine and Coastal Ocean Model (ECOM) is coupled with a modern data assimilation method based on the local ensemble transform Kalman filter (LETKF), and a series of simulation experiments is conducted. In these experiments, a long ECOM “nature” run is taken to be the “truth.” Observations are generated at analysis times by perturbing the nature run at randomly chosen model grid points with errors of known statistics. A diverse collection of model states is used for the initial ensemble. All experiments use the same lateral boundary conditions and external forcing fields as in the nature run. In the data assimilation, the analysis step combines the observations and the ECOM forecasts using the Kalman filter equations. As a control, a free-running forecast (FRF) is made from the initial ensemble mean to check the relative importance of external forcing versus data assimilation on the analysis skill. Results of the assimilation cycle and the FRF are compared to truth to quantify the skill of each. The LETKF performs well for the cases studied here. After just a few assimilation cycles, the analysis errors are smaller than the observation errors and are much smaller than the errors in the FRF. The assimilation quickly eliminates the domain-averaged bias of the initial ensemble. The filter accurately tracks the truth at all data densities examined, from observations at 50% of the model grid points down to 2% of the model grid points. As the data density increases, the ensemble spread, bias, and error standard deviation decrease. As the ensemble size increases, the ensemble spread increases and the error standard deviation decreases. Increases in the size of the observation error lead to a larger ensemble spread but have a small impact on the analysis accuracy.