Assimilating water level observations with the ensemble optimal interpolation scheme into a rainfall-runoff-inundation model: A repository-based dynamic covariance matrix generation approach

Although conceptually attractive, the use of ensemble data assimilation methods, such as the ensemble Kalman filter (EnKF), can be constrained by intensive computational requirements. In such cases, the ensemble optimal interpolation scheme (EnOI), which works on a single model run instead of ensemble evolution, may offer a sub-optimal alternative. This study explores different approaches of dynamic covariance matrix generation from predefined state vector repositories for assimilating synthetic water level observations with the EnOI scheme into a distributed rainfall-runoff-inundation model. Repositories are first created by storing open loop state vectors from the simulation of past flood events. The vectors are later sampled during the assimilation step, based on their closeness to the model forecast (calculated using vector norm). Results suggest that the dynamic EnOI scheme is inferior to the EnKF, but can improve upon the deterministic simulation depending on the sampling approach and the repository used. Observations can also be used for sampling to increase the background spread when the system noise is large. A richer repository is required to reduce analysis degradation, but increases the computation cost. This can be resolved by using a sliced repository consisting of only the vectors with norm close to the model forecast.