A strategy for quality controlling hourly rainfall observations and its impact on hourly streamflow simulations

Abstract

Real-time streamflow forecasts are often produced at hourly and shorter time steps. To calibrate the hydrological models used to generate forecasts, archives of streamflow and rainfall observations are essential. The collection and archiving of sub-daily rain gauge observations is typically automated. Errors in rainfall observations can arise for many reasons and may manifest as anomalously high or low values for a single observation or longer periods of time. Quality control of sub-daily rainfall is onerous because of the large volume of data, and sub-daily rainfall data are frequently not quality-controlled when the data are archived in real-time. Errors in rainfall observations used to calibrate hydrological models can lead to poor out- of-sample model simulations and contribute to poor streamflow forecasts. In this paper we describe a simple automated strategy for quality controlling archives of rain gauge observations. The strategy compares running totals of rain gauge observations with a reference rainfall data set. Where the differences between the observations and the reference dataset exceed a threshold, rain gauge observations are considered to be of poor quality and set to a missing value. The sensitivity of the quality control can be manipulated by adjusting the period over which the running totals are computed, and the definition of the difference threshold. We apply the quality control strategy to hourly rain gauge observations from five catchments across Australia. For these applications we use the daily Australian Water Availability Project rainfall data set as the reference and quality control the hourly rain gauge observations at daily time steps. We seek only to remove gross errors in the rain gauge observations, where 5-day observed totals are greater than five times the reference or smaller than one fifth of the reference. We demonstrate the efficacy of the quality control strategy for hydrological models run at an hourly time-step with cross-validation experiments. The calibration and validation performance of hydrological simulations forced by the quality-controlled data are vastly superior to those forced by the raw rainfall observations. In some instances, improvements in validation Nash-Sutcliffe Efficiency values greater than 0.7 are achieved by using the quality-controlled rainfall observations. The performance of the hydrological simulations also tends to be more consistent between calibration and validation periods when quality-controlled rainfall observations are used. The method allows rapid quality-control of large sub-daily rainfall datasets, allowing new streamflow forecasting systems to be established quickly.