Erik R Christensen
University of Wisconsin–Milwaukee, USA
Title: Trace element data for surface soil and moss in Norway analyzed by PMF and PCA with optional CLR transformation
Biography
Biography: Erik R Christensen
Abstract
Statement of the Problem: Classical principal component analysis (PCA) is useful to separate anthropogenic and geogenic sources of trace elements in surface soil and moss. However, while mass input (µg/g) of trace elements can be determined by positive matrix factorization (PMF), this cannot readily be done by PCA. Also, classical PCA can be sensitive to minor changes in input data. We considered here datasets for 464 stations in mainland Norway and more than 21 trace elements including Pb, Cd, Ag, As, and Hg and several rare earth elements. PMF scores for individual stations were plotted, and PCA with CLR transformation was tested to check robustness and to see if significant factors were missed.
Methodology & Theoretical Orientation: Surface soil and moss (Hylocomium splendens) samples were collected all over mainland Norway from open air sites. Chemical and statistical analyses were conducted as described previously.
Findings: PMF factor contributions (scores) of anthropogenic factors from dated moss samples with high levels of Pb, Cd, Cr, Co, As, Hg at Svanvik near the Russian border confirm the influence of a Russian copper - nickel smelter. A much smaller increase is seen in the soil samples. Using CLR transformation with PCA, increased robustness is reflected by the fact that all five factors each have near-equivalent PMF factors, which is not the case for PCA without CLR transformation. However, scores are noisy and have significant negatives. Classical PCA on soil samples produces an Ag, Hg factor at Ulefoss, along with a distinct Eu factor free of Ag and Hg (PC5), both not seen with the CLR transformation.
Conclusion & Significance: PMF and PCA factor scores at individual stations such as Svanvik or Ulefoss can be an important adjunct to score maps in identifying pollution sources and PCA with CLR transformation provide more robust factors, however, at the possible expense of missing significant and more distinct factors.
