Submitted to: American Geophysical Union
Publication Type: Abstract Only
Publication Acceptance Date: 11/30/2007
Publication Date: 12/12/2007
Citation: Ryu, D., Crow, W.T., Zhan, X. 2007. Correcting unintended perturbation biases in hydrologic data assimilation using the ensemble Kalman Filter [abstract]. American Geophysical Union Fall Meeting. 2007 CDROM. Interpretive Summary:
Technical Abstract: Recent advances in hydrologic data assimilation have demonstrated the value of remotely sensed surface soil moisture in improving forecasts of key hydrologic variables such as root-zone soil moisture and surface runoff. In hydrologic data assimilation, the ensemble Kalman filter (EnKF) provides a robust framework to optimally update model predictions using observations based on the uncertainties of the model and observations. In the EnKF, model uncertainty is obtained using a Monte Carlo approach where the states, parameters, or forcing data for the model are perturbed to create an ensemble of states, and the covariance of the ensemble members represents the model uncertainty. However, although adding mean-zero perturbations to the model should not affect the mean performance of the underlying model, it often causes systematic model state biases due to nonlinear land surface model physics and the bounded nature of key land surface states (e.g. soil moisture). Here we show an example of systematic biases caused by land surface model perturbations, present a diagnosis for these biases, and propose a simple method to reduce them. The Noah land surface model within the Land Information System (LIS) is applied to a medium-scale basin in the United States Southern Great Plains for the demonstration. Our proposed method removes virtually all the bias caused by the mean-zero ensemble perturbations and significantly improves the performance of the EnKF. Compared to non-corrected EnKF results, combining the bias-correction scheme with EnKF updating reduces forecast errors in evapotranspiration by 51%, and surface runoff by 74%. Additional effects of the nonlinear model physics and bounded states on model performance will be discussed and alternative methods to optimally summarize the nonlinear diagnostic variables will also be presented.