Improved guidance for calibration of deterministic watershed models at intermittent streamflow sites

Deterministic Watershed Models (DWMs) have been widely used as a tool to analyze and predict streamflow traces. While flood and average streamflow skill statistics dominate DWM calibration decisions, it often comes at the expense of limiting model skill during low streamflow events. It is important that DWMs provide more accurate low streamflow estimates to support drought planning and management. Low streamflows are a specific concern for: agricultural, industrial, and municipal water supply; hydropower generation; ecosystem health and services; and recreation. Intermittent streams pose a particularly acute low streamflow model calibration and simulation challenge. Here we use WRF-Hydro to demonstrate new calibration methodologies to potentially improve DWM low streamflow skill scores at intermittent streamflow sites. Our study sites are two unregulated watersheds in the Catchment Attributes and MEteorology for Large-sample Studies (CAMELS) database with varying percentages of daily streamflows recorded as zero. We calibrate WRF-Hydro at these two sites using a number of different calibration metrics including real- and log-space NSE, square root and inverse NSE, and a censored maximum likelihood estimator (CMLE) technique. We quantify WRF-Hydro model skill using a variety of performance metrics over different time scales, such as annual d-day minimum streamflows, annual flow duration curve quantiles, as well as Area Under the Curve (AUC) confusion matrix classification metrics. Initial results indicate that calibration metrics have a large impact on a model’s ability to reproduce low flow series, and that commonly used calibration metrics such as NSE generally result in DWMs that over-predict low-flow series. Calibration by log transformed or inverse NSE leads to improved representation of low flows, but depends highly on the constant value added to zero flows prior to data transformation. The use of the CMLE may improve how DWM represent low flow series at intermittent streamflow sites, as it captures not only the fit to non-zero observations, but also the frequency of streamflows recorded as zero.