How the state vector configuration matters in multivariate data assimilation for streamflow predictions of snow-fed rivers

Abstract:

Hydrological modelling and streamflow prediction for watersheds over which multiple data sets are available can benefit from data assimilation. For example, updating modelled upstream flows and snow water equivalent (SWE) via existing correlations with downstream flow and SWE observations can positively impact short-term (days) and mid-term (weeks) streamflow forecast, respectively. Other variables can be updated indirectly if they are included in the state vector, which will further affect results. In order to fully benefit from existing correlations between variables, one may be tempted to augment the state vector to include all related variables and parameters, or choose to include a very limited number of variables in order to prevent erroneous correlations from deteriorating other model states. Localizing the correlations on the spatial level or between variables can also affect results. This makes it unclear as to how to configure the state vector, especially when multivariate observations are assimilated. This study presents a sensitivity analysis of the state vector configuration for synthetic multivariate data assimilation using an Ensemble Kalman filter. A spatially distributed hydrological model is used to simulate streamflow predictions for the mountainous Nechako River located in British-Columbia, Canada. Synthetic data includes daily snow cover extent, daily measurements of snow water equivalent (SWE) at three different locations and daily streamflow data at the watershed outlet. Results show a large variability of the Nash-Sutcliffe efficiency and streamflow bias over a wide range of prediction horizons (days to weeks) depending on the state vector configuration and the type of observations assimilated. Some configurations are shown to improve the accuracy of streamflow predictions while others offer worse results than the open loop simulation. These results serve as a first step toward comparing streamflow prediction performance of various real multivariate data assimilation scenarios.