Multi-Block Chemometric Approaches to the Unsupervised Spectral Characterization of Geological Samples

Abstract

As an example for the potential use of multi-block chemometric methods to provide improved unsupervised characterization of compositionally complex materials through the integration of multi-modal spectrometric data sets, we analysed spectral data derived from five field instruments (one XRF, two NIR, and two FT-Raman), collected on 76 bedrock samples of diverse composition. These data were analysed by single- and multi- block latent variable models, based on principal component analysis (PCA) and partial least squares (PLS). For the single-block approach, PCA and PLS models were generated; whilst hierarchical partial least squares (HPLS) regression was applied for the multi-block modelling. We also tested whether dimensionality reduction resulted in a more computationally efficient muti-block HPLS model with enhanced model interpretability and geological characterization power using the variable influence on projection (VIP) feature selection method.

The results showed differences in the characterization power of the five spectrometer data sets for the bedrock samples based on their mineral composition and geological properties; moreover, some spectroscopic techniques under-performed for distinguishing samples by composition. The multi-block HPLS and its VIP-strengthened model yielded a more complete unsupervised geological aggrupation of the samples in a single parsimonious model. We conclude that multi-block HPLS models are effective at combining multi-modal spectrometric data to provide a more comprehensive characterization of compositionally complex samples, and VIP can reduce HPLS model complexity, while increasing its data interpretability. These approaches have been applied here to a geological data set, but are amenable to a broad range of applications across chemical and biomedical disciplines.