EXPLORING APPROPRIATE INFLATION AND LOCALIZATION METHODS TO STABILIZE ENSEMBLE DATA ASSIMILATION OF A RAINFALL-RUNOFF-INUNDATION MODEL

Abstract

 Data assimilation can improve forecast accuracy of dynamical models by combining model state variables and real-world observations. This study applied the ensemble Kalman filter (EnKF) for a rainfall-runoff-inundation (RRI) model to adjust model state variables with operational water-level observations. In contrast to atmospheric models, model state errors do not propagate in the non-chaotic RRI model. Therefore, it is important to explore error inflation methods for providing appropriate background error covariance for the EnKF. For that purpose, this study perturbed rainfall intensity for ensemble members as a way of the covariance inflation.

 A series of experiments with and without EnKF were employed in Omono River in Akita Prefecture. Our experiments showed that predicted water level was improved at both observed and unobserved stations compared to the RRI simulations without assimilation. This study also investigated effective localization methods for the RRI model. The application of localization along the river channel was found to perform as well as traditional localization based on Euclid distances commonly used in atmospheric data assimilation.