Observation and Model Bias Estimation in the Presence of Either or Both Sources of Error

Abstract

In numerical weather prediction and in reanalysis, robust approaches for observation bias correction are necessary to approach optimal data assimilation. The success of bias correction can be limited by model errors. Here, simultaneous estimation of observation and model biases, and the model state for an analysis, is explored with ensemble data assimilation and a simple model. The approach is based on parameter estimation using an augmented state in an ensemble adjustment Kalman filter. The observation biases are modeled with a linear term added to the forward operator. A bias is introduced in the forcing term of the model, leading to a model with complex errors that can be used in imperfect-model assimilation experiments. Under a range of model forcing biases and observation biases, accurate observation bias estimation and correction are possible when the model forcing bias is simultaneously estimated and corrected. In the presence of both model error and observation biases, estimating one and ignoring the other harms the assimilation more than not estimating any errors at all, because the biases are not correctly attributed. Neglecting a large model forcing bias while estimating observation biases results in filter divergence; the observation bias parameter absorbs the model forcing bias, and recursively and incorrectly increases the increments. Neglecting observation bias results in suboptimal assimilation, but the model forcing bias parameter estimate remains stable because the model dynamics ensure covariance between the parameter and the model state.