The use of methods of compositional data analysis for the separation of geochemical signals in fluvial sediments

Fluvial sediments datasets, similarly as other types of a concentration based data, are typical by their relative nature and therefore they need preprocessing or normalization prior to the main statistical analysis. In the geochemical practice, several normalization methods are used, like a simple normalization of the target element concentration with the concentration of the reference (conservative, lithogenic) one, double normalization or concentration conversion to local enrichment factor. As an alternative to these methods, the approach using the principles of compositional data analysis (CoDA) can be considered.  Instead of the standard statistical analytical methods, like ordinary least squares regression, correlation of principal component analysis (PCA), applied on the raw or the target element normalized concentrations, the CoDA methods consider the relative structure of the whole dataset. CoDA together with the use of robust statistical methods, which are down weighting the influence of the outlying observations, have a potential to provide more accurate results. This property is demonstrated and discussed on the base of dataset from mapping the sediments from the Skalka Reservoir in the Ohře River, Czech Republic, and its tributaries. Mainly the performance of the robust versions of regression, correlation and principal components analysis, respecting the CoDA principles, will be presented and the way to them will be explained.