An Approach to Reduce Systematic Representativeness Errors of Surface Observations in Ensemble Data Assimilation

Assimilation of surface observations is challenging. One reason is the potentially large and systematic representativeness errors that can affect them, especially over complex terrain. For instance, the discretized grid-cells of models are unable to capture terrain details such as surface slope and aspect, and differences in model and terrain elevation can be large. These errors can introduce a bias in the observation increments, making the assimilation sub-optimal. To mitigate this issue an approach was developed using the ensemble Kalman filter. The method relies on including a bias parameter, unique to each observation, in the forward operator. The bias parameters are estimated by augmenting the state vector in the filter, and simultaneous to the state update. Positive results from earlier experiments in a low-order model motivate observing system simulation experiments (OSSEs) with the Weather Research and Forecasting (WRF) model.

To evaluate the efficacy of the bias estimation and correction, OSSEs are constructed for the conventional observation network including radiosondes, aircraft observations, atmospheric motion vectors, and surface observations. Three experiments are created by adding different kinds of bias to temperature from synthetic METAR observations: (1) a spatially invariant bias, (2) a spatially changing bias proportional to height differences between the model and the observations, and (3) bias that is proportional to the temperature. The target region characterized by complex terrain is the western U.S. on a domain with 30-km grid spacing. Observations are assimilated every 3 hours using an 80-member ensemble during September 2012. Results demonstrate that the approach is able to estimate and correct the bias when it is spatially invariant (1); errors in estimating the known parameters are small, but similar to the systematic errors in innovations for unbiased observations. More complex bias structure in experiments (2) and (3) are more difficult to estimate, but still possible. The results demonstrate that parameter estimation for surface observation bias may be a viable approach for real-data problems.