Detecting gradual and abrupt changes in hydrological records

Hydrological time series increasingly exhibit non-stationarity, e.g., variables such as precipitation and streamflow values do not maintain a consistent mean over long periods, due to natural and anthropogenic changes. Detecting whether such shifts are gradual or abrupt is a growing concern for water resources planning and management. This paper shows that conventional trend and change-point tests do not adequately enable these two types of change to be distinguished. We propose a method for combining the rank correlations of the Mann-Kendall and Pettitt statistics to extract an indicator whose value determines whether a shift observed in a given time series is gradual or abrupt. This method allows the success rate to be independent of the length of record, and it is validated with Monte-Carlo experiments. The limitations caused by the short and noisy nature of hydroclimatic time series are discussed. As an application, the proposed method provides useful insights on changes in hydroclimatic variables in the United States during 1910-2009 using time series from 1217 stations in the United States Hydroclimatic Data Network (USHCN).