Detecting dominant changes in irregularly sampled multivariate water quality data sets

Abstract

Time series of catchment water quality often exhibit substantial temporal and spatial variability which can rarely be traced back to single causal factors. Numerous anthropogenic and natural drivers influence groundwater and stream water quality, 20 especially in regions with high land use intensity. In addition, typical existing monitoring data sets, e.g. from environmental agencies, are usually characterized by relatively low sampling frequency and irregular sampling in space and / or time. This complicates the differentiation between anthropogenic influence and natural variability as well as the detection of changes in water quality which indicate changes 25 of single drivers. Detecting such changes is of fundamental interest for water management purposes as well as for scientific analyses.We suggest the new term 'dominant changes' for changes in multivariate water quality data that concern 1) more than a single variable, 2) more than one single site and 3) more than short-term fluctuations or single events and present an exploratory 30 framework for the detection of such 'dominant changes' in multivariate water quality data sets with irregular sampling in space and time. Firstly, we used a non-linear dimension reduction technique to derive multivariate water quality components. The components provide a sparse description of the dominant spatiotemporal dynamics in the multivariate water quality data set. In addition, they can be used to derive 35 hypotheses on the dominant drivers influencing water quality. Secondly, different sampling sites were compared with respect to median component values. Thirdly, time series of the components at single sites were analysed for seasonal patterns and linear and non-linear trends. Spatial and temporal heterogeneities are efficiently used as a source of information rather than being considered as noise. Besides, non-40 linearities are considered explicitly. The approach is especially recommended for the exploratory assessment of existing long term low frequency multivariate water quality monitoring data.