Deploying chemometric techniques to guide provenance analysis with field portable instrumentation
Date
2023
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
The emergence of handheld spectrometers over the past few decades has allowed the analytical laboratory to expand and solve problems that were previously unattainable. One such application is forensic provenance analysis, or the determination of sample origin, be it growing region, manufacturing location, creator/producer, etc. Since the spectral features associated with provenance tend to be subtle, tasks that aim to establish provenance almost always require the application of chemometric analysis. In the studies presented herein, two forms of portable spectrometers, handheld laser induced breakdown spectroscopy (LIBS) and fiber optic reflectance spectroscopy (FORS) are used to establish the provenance of tropical hardwoods and historical textiles, respectively, through both established classification algorithms and novel applications of network graphs. Chapters 2 and 3 collectively analyze 15 species of Dalbergia spp. and its lookalikes, assessing the impact of various preprocessing treatments, data transformations, variable selection methods, and classification algorithms, including Principal Component Analysis (PCA), Partial Least Squares Discriminant Analysis (PLS1-DA, PLS2-DA), k-Nearest Neighbors (k-NN), Classification and Regression Trees (CART), Random Forests (RF), and Support Vector Machines (SVM). These analyses show that, despite the poorer analytical figures of merit associated with portable spectrometers, handheld LIBS shows promise for field use as a first-pass screening method for tropical hardwood identification at ports of entry, with multi-class classification performance ranging between 0.85-1.00, as assessed by Cohen’s Kappa statistic. Chapter 4 addresses the novel implementation of network graphs as visualization tools for multi-class confusion matrices, and how this technique can be used to aid the analyst in making decisions throughout the modelling process. Lastly Chapter 5 shows cluster analysis of FORS spectra collected from 204 historical textiles suspected to be dyed with cochineal and madder in varying “recipes” as a potential non-destructive alternative to the current state of the art for textile analysis, high performance liquid chromatography (HPLC). The results from this study indicate that the primary influence in FORS spectra is the color information in the visible region, while dye concentrations in textiles are too low for meaningful information to be present in the near infrared region, though there is reasonable correlation between HPLC-identifications and FORS spectral features. The combination of these studies shows that, generally speaking, chemometric classification algorithms can be a viable tool for assessing the provenance of samples from an array of fields and that coupling chemometric analysis to handheld spectrometers shows significant promise for the future of field analysis.
Description
Keywords
Chemometrics, Machine learning, Portable instrumentation, Spectroscopy, Textile analysis