Non-stationarity in long hydrological time series: a new theoretical technique for detecting multiple changes in mean and variance

Nicholas Howden1, Tim Burt2, Fred Worrall3, Magdalena Bieroza1 1. Dept. of Civil Engineering, University of Bristol, Queen’s Building, University Walk, Bristol, BS8 1TR, UK. 2. Dept. of Geography, Science Laboratories, South Road, Durham, DH1 3LE, UK. 3. Dept. of Earth Sciences, Science Laboratories, South Road, Durham, DH1 3LE, UK. We present a theoretical framework to detect changes in the mean and variance of long hydrological time series. This enables multiple change points to be identified, assigns a significance level to each, and provides a detailed description of how changes in the time series are manifested. The framework has the following useful characteristics: (1) Unlike conventional parametric tests, the method is independent of temporal context (i.e. starting or finishing point); (2) The technique is highly sensitive and vastly increases our ability to detect trends compared with non-parametric (i.e. Mann-Kendall) and parametric (i.e. conventional time series analysis) techniques; (3) Where change points or trends are identified, these can be traced to particular forms of driver (i.e. continuous, or discrete) and the time at which these drivers operate can be identified; We show how the technique can be used to explore current water quality problems, including carbon loss from the northern peatlands and nitrate export from agricultural land. We show that progressive increases in dissolved organic carbon (DOC) loss from UK peatlands were actually caused by severe drought conditions and not by rising temperatures, changing atmospheric deposition or enriched atmospheric CO2 as previously suggested.