Reconciling monitoring and modeling: An appraisal of river monitoring networks based on a spatial autocorrelation approach - emerging pollutants in the Danube River as a case study

Highlights: • Our river perception is based on monitoring network measurements. • Autocorrelation models for 235 emerging pollutants in the Danube River are set. • Correlation lengths are derived from the spatial variation of autocorrelation indexes. • 27% compounds out of 235 have a sub-optimal monitoring network. • Neighbors vs. local relative contributions of monitored variables are quantified. • For 92% compounds local contributions dominate over neighbors influence. Rivers extend in space and time under the influence of their catchment area. Our perception largely relies on discrete spatial and temporal observations carried out at certain sites located throughout the catchment (monitoring networks, MN). However, MNs are constrained by (a) the distribution of sampling sites, (b) the dynamics of the variable considered and (c) the river hydrological conditions. In this study, all three aspects were captured and quantified by applying a spatial autocorrelation modeling approach. We exemplarily studied its application to 235 emerging contaminants (pesticides, pharmaceuticals, and personal care products [PPCP], industrial and miscellaneous) measured at 55 sampling sites in the Danube River. 22 out of the 235 compounds monitored were present at all sites and 125 were found in at least 50%.We first calculated the Moran Index (MI) to characterize the spatial autocorrelation of the compound set. 59 compounds showed MI ≤ 0, which can be interpreted as 'no spatial correlation'. Next, spatial autocorrelation models were set for each compound. From the autocorrelation parameter ρ, catchment average correlation lengths were derived for each compound. MN optimality was examined and compounds were classified into three groups: (a) those with ρ ≤ 0 [25%]; (b) those with ρ > 0 and correl. length < average distance between consecutive sites [ 2%] and (c) those with ρ > 0 and correl. length > average distance between consecutive sites [73%]. The MN was considered optimal only for the latter class. Networks with the larger average distance between consecutive sites resulted in a decreasing number of optimally monitored compounds. Furthermore, neighbors vs. local relative contributions were quantified based on the spatial autocorrelation model for all the measured compounds. The results of this study show how autocorrelation models can aid water managers to improve the design of river MNs, which are a key aspect of the Water Framework Directive.