Gauging ungauged catchments – Active learning for the timing of point discharge observations in combination with continuous water level measurements

Highlights

• We explore the value of water levels and discharge observations for calibration.

• We test an active learning approach to select informative discharge observations.

• Active learning leads to a hydrologically meaningful selection of observation times.

• Combining water levels with a few selected discharge observations is encouraged.

Abstract

Hydrological models have traditionally been used for the prediction in ungauged basins despite the related challenge of model parameterization. Short measurement campaigns could be a way to obtain some basic information that is needed to support model calibration in these catchments. This study explores the potential of such field campaigns by i) testing the relative value of continuous water-level time series and point discharge observations for model calibration, and by ii) evaluating the value of point discharge observations collected using expert knowledge and active learning to guide when to measure streamflow. The study was based on 100 gauged catchments across the contiguous United States for which we pretended to have only limited hydrological observations, i.e., continuous daily water levels and ten daily point discharge observations from different hypothetical field trips conducted within one hydrological year. Water level data were used as a single source of information, as well as in addition to point discharge observations, for calibrating the HBV model. Calibration against point discharge observations was conducted iteratively by continually adding new observations from one of the ten field measurements. Our results suggested that the information contained in point discharge observations was especially valuable for constraining the annual water balance and streamflow response at the event scale, improving predictions based solely on water levels by up to 50% after ten field observations. In contrast, water levels were valuable to increase the accuracy of simulated daily streamflow dynamics. Informative discharge sampling dates were similar when selected with either active learning or expert knowledge and typically clustered during seasons with high streamflow.