Item

Abstract

Our understanding and modeling of hydrologic processes relies on long-term streamflow monitoring. However, streamflow records can suffer from flaws that lead to streamflow values reflecting non-natural hydrology, for example due to instrument failure, interpolation of missing data, or anthropogenic influences. Visual inspection of flow time series by humans is highly recommended, although it is time consuming. Thus, no study has examined the proportion of flaws, their temporal distribution, and their influence on hydrologic indicators on a large data set. We summarized the results of a large visual inspection campaign of 674 streamflow time series in France by 43 evaluators, who were asked to identify flaws belonging to five categories: linear interpolation, drops, noise, point anomaly, and others. We examined the individual behavior of evaluators in terms of severity and consistency with other evaluators, as well as the temporal distributions of flaws and their influence on commonly used hydrological indicators.We found that agreement among evaluators was surprisingly low, with an average of 12% of overlapping periods reported as flaws. The most common types of flaws identified were linear interpolation and noise, and they were most often reported during low-flow periods in summer. The impact of cleaning the reported flaws from the time series was higher for low-flow indicators than for high-flow indicators. We conclude that flaws identification in streamflow time series is highly dependent on the goals and skills of individual evaluators, raising the need for better practices for data cleaning that could benefit from future advances in machine learning tools.