Data Assimilation for Streamflow Forecasting: State–Parameter Assimilation versus Output Assimilation

Abstract

This paper compares two data assimilation methods: state–parameter assimilation and output assimilation in improving streamflow forecasting using the Soil and Water Assessment Tool (SWAT) model. The state–parameter assimilation is performed by updating the stored water content and soil curve number with the extended Kalman filter (EKF); the output assimilation is carried out by updating the model output errors with autoregressive (AR) models. The performances of the two data assimilation techniques are compared for a dry year and a wet year, and it is found that whereas both methods significantly improve forecasting accuracy, their performances are influenced by the hydrological regime of the particular year. During the wet year, the average root-mean-square error (RMSE) for seven days forecasts is improved from 670.46 to 420.42 m3/s when output assimilation is used, and to 367.60 m3/s when state–parameter assimilation is used. The Nash–Sutcliffe coefficient (NSC) is improved from 0.63 to 0.85 and 0.88, respectively; the mean error (ME) is improved from −375.83 m3/s to −131.68 m3/s and −129.11 m3/s, respectively. For shorter forecast leads (1–4 days), the state–parameter assimilation outperforms output assimilation in both dry and wet years. For longer forecast leads (5–7 days), the output assimilation could provide better results in the wet year. A hybrid method that combines state–parameter assimilation and output assimilation performs very well in both dry and wet years according to all three indicators.