Evaluating Simulations of Daily Discharge from Large Watersheds Using Autoregression and an Index of Flashiness

Abstract

<abstract><title><italic>Abstract. </italic></title> Watershed models (e.g., the Soil and Water Assessment Tool, or SWAT) are routinely calibrated to provide reliable estimates of stream discharge. In modeling studies of large watersheds where long-term records of daily discharge are available, model performance statistics are often reported for monthly intervals. With monthly data, guidelines are available to directly evaluate model performance. However, assessing model performance on a daily time step is more appealing because the hydrologic processes being simulated are better expressed through changes in daily discharge. Yet it is difficult to assess simulations at a daily time step due to large variance of daily data. Transformation reduces data variance and preserves data detail; however, assessing daily model performance with transformed data raises separate issues (e.g., accurate simulation across all flow conditions becomes important). One possibility for assessing daily simulations is to apply autoregression, which could provide a model performance target using transformed daily data. The objective of this study was to evaluate autoregressive models as an aid to assess simulations of river basin hydrology at a daily time step. Autoregressive models were fitted to natural log-transformed daily discharge records (2001-2009) from four watersheds in central Iowa to generate a statistical replica of each record. Results provided a realistic target for SWAT model performance for the three watersheds that were the least flashy (i.e., had Richards-Baker flashiness index (RBI) values <0.3). Discharge at the fourth gauge exhibited greater flashiness (RBI > 0.3), and therefore weaker autocorrelation, which caused the autoregressive model to fail to generate a performance target for SWAT. The RBI stream flashiness index could be used as a simple parameter to assess watershed model performance at a daily time step using non-transformed data. For watersheds with RBI < 0.3, autoregression on transformed data can provide an independent, unbiased estimate of observed daily stream data, which can offer a target for model performance using measured data alone.