Multicriteria wavenumber selection in cocaine classification

Abstract

Cocaine ATR-FTIR spectra consist of a large number of wavenumbers that typically decreases the performance of exploratory and predictive multivariate techniques. This paper proposes a framework for selecting the most relevant wavenumbers to classify cocaine samples into two categories regarding chemical composition, i.e. salt and base. The proposed framework builds a wavenumber importance index based on the Bhattacharyya distance (BD) followed by a procedure that removes wavenumbers from the spectra according to the order suggested by the BD index. The recommended wavenumber subset is chosen based on multiple criteria assessing classification performance, which are recalculated after each wavenumber is eliminated. The method was applied to ATR-FTIR spectra from 513 samples of cocaine, remarkably reducing the percent of retained wavenumbers and yielding near to perfect classifications in the testing set. In addition, we compared our propositions with other methods tailored to wavenumber selection; we found that the proposed framework, which relies on simple mathematical fundamentals, yielded competitive results.

Keywords: Cocaine; KNN; LDA; PNN; Wavenumber selection.