Geochemical pattern recognition through matrix decomposition

Abstract

This study compares the use of joint singular value decomposition and semi-discrete decomposition (JSS) and non-negative matrix factorisation with univariate analysis of raw data, to detect multi-element patterns in soils related to geochemical dispersion from Mississippi Valley-Type Pb-Zn deposits in the Irankuh area of central Iran. Joint singular value decomposition and semi-discrete decomposition clustering of the data identified a suite of mineralisation-related variables and corresponding sample clusters that are spatially associated with mineralisation or variations in parent lithology. Non-negative matrix factorisation generated three main factors and sample clusters that relate to the main zones of sulphide mineralisation, variations in clay mineralogy and the composition of unaltered host rocks. These two matrix decomposition techniques deliver similar results, though JSS delivers a significantly higher sample classification accuracy. Both methods result in delineation of a contiguous cluster of samples above an extensive zone of blind mineralisation at the Tappe Sorkh Pb-Zn deposit and define new potentially mineralised targets between this deposit and the Gushfil Pb-Zn deposit. More conventional univariate methods to detect anomalous populations of Pb and Zn in the soil geochemical data, such as number-size fractal analysis, were less effective at defining these targets. Joint application of practical matrix decomposition techniques, utilising noise removal and modelling of the underlying geochemical patterns associated with ore-forming processes, has led to more reliable prospectivity mapping of the Pb-Zn deposits in Irankuh.