Variational Assimilation of Streamflow Observations in Improving Monthly Streamflow Forecasting

Abstract. Uncertainties associated with the initial conditions (e.g. soil moisture content) of a hydrologic model have been recognized as one of the main sources of errors in hydrologic predictions, specifically over a rainfall-runoff regime. Apart from the recent advances in Data Assimilation (DA) for improving hydrologic predictions, this study explores variational assimilation (VAR) of gauge-measured daily streamflow data for updating initial state of soil moisture content of Variable Infiltration Capacity (VIC) Land Surface Model (LSM) in order to improve streamflow simulations as well as monthly streamflow forecasting. The study is conducted for the Tar River basin in North Carolina over 20-year period (1991–2010). The role of two critical parameters of VAR DA – the frequency of DA application and the length of assimilation window – in determining the skill of DA-improved streamflow predictions is also assessed. We found that correcting VIC model's initial conditions using a 7-day assimilation window results in the highest improvement in the skill of streamflow predictions quantified by Kling-Gupta Efficiency (KGE) and Nash-Sutcliffe Efficiency (NSE) metrics. In addition, the potential gain from VAR DA framework is quantified and compared under two 1-month ahead streamflow forecasting schemes: 1) deterministic forecasts developed by using ECHAM4.5 GCM 1-month ahead precipitation forecasts and 2) Probabilistic forecasts from Ensemble Streamflow Prediction (ESP) approach. This study also examines the persistence of the DA impact in the monthly predictions by quantifying the enhanced accuracy in daily flows over extending forecast lead time blocks. Analyses show that the the corrected initial state conditions continually enhance the 7–8 days ahead predictions, but after that the errors in forcings dominate the DA effects. Still, the overall impact of VAR DA in monthly streamflow forecasting is positive.