Separation of geochemical signals in fluvial sediments: New approaches to grain-size control and anthropogenic contamination

Abstract

A compositional data analysis (CoDA) in fluvial sediments is performed to achieve separation of the geochemical signals (SGS) of grain size, anthropogenic contamination, and possible post-depositional alteration. The SGS is demonstrated and developed in the study of the sediments from the Skalka Reservoir (Czechia) and the floodplain of its tributary rivers, which have been impacted by pollution from the Chemical Factory Marktredwitz (Bavaria, Germany) brought through temporary sinks in the channels and floodplains to the reservoir. This paper compares CoDA tools with standard empirical approaches based on using deeper strata as uncontaminated or pre-industrial (examination of element concentration depth profiles), scatterplots with risk elements (mainly Zn in this study) as dependent variables and lithogenic reference elements as independent variables to construct background functions and to calculate local enrichment factors (LEF), and a principal component analysis performed on raw and geochemically normalised elemental concentrations. The utilised CoDA tools include classical and robust methods using the log-ratio approach that fully respects the mathematical specificity of the compositional data (data closure, or more generally scale invariance, and further related aspects like non-Gaussian distribution, and commonly polymodality) like the robust PCA with centred log-ratio (clr) transformation of concentrations; consequently, histograms of the raw and normalised concentrations and contamination scores were compared. The multivariate CoDA was considerably facilitated by a novel tool for understanding the grain-size control of sediment composition, i.e. a functional data analysis of particle size distributions (densities) based on Bayes spaces. Also, the robust correlation analysis was efficient using a (log-)ratio methodology. Several elements can be used for the geochemical normalisation and LEF calculations, of which Al, Fe, and Ti can definitely be recommended, while Cr, Mg, and even Si also produced comparable results. A more critical factor is a proper selection of the background functions. We demonstrated the limits of using some popular tools for the compositional data mining: the ordinary PCA failed or performed worse than LEF in the separation of grain-size and contamination signals. Some log-ratio methods performed well, in particular robust regression with selected (lithogenic elements at explaining side) and robust PCA with clr transformation. Even for apparently simple tasks, such as the separation of anthropogenic contamination signals, knowledgeable decisions during data preparation for the CoDA are still indispensable.