Parameter Estimation for Observation Bias with an Ensemble Kalman Filter

Abstract:

In this work we evaluate a method to estimate systematic errors of individual in-situ observations. The approach is based on parameter estimation using an augmented state in an ensemble adjustment Kalman filter. Biased observations are assimilated in the highly chaotic Lorenz (2005) model that combines small and large scales. For this purpose, synthetic observations are created by introducing a grid-point dependent bias and random errors to a nature run (truth). The sensitivity of the methodology to model error, the number of ensemble members, the number of parameters, and the parameter variance is evaluated.

Results demonstrate that the methodology is able to estimate observation bias for a both a perfect model and an imperfect model if the model error is estimated. The parameter estimation and the rms errors are significantly deteriorated if model error is ignored. Errors are qualitatively independent of model forcing when model error is estimated. By contrast, parameter estimation depends on model error when model error is not estimated, especially if observation biases are estimated. Errors increase with the number of estimated parameters, but they are independent of ensemble size as long as the number of ensembles members is large enough. There is an optimum value of parameter variance that minimizes the errors and improves the parameter estimation, but this value depends on observation bias and model error. Overall, the results suggest more accuracy in observation bias estimates than model error estimates.