Classification of highly colinear infrared spectral data

S. Lubbe${}^{a,b}$, N. J. le Roux${}^a$, R. J. Cornelissen${}^c$ and H. H. Nieuwoudt${}^d$

${}^a$MuViSU (Centre for Multi-dimensional Data Visualisation), Department of Statistics and Actuarial Science, Stellenbosch University, Stellenbosch, South Africa, ${}^b$NITheCS (National Institute of Theoretical and Computational Sciences), Stellenbosch University, Stellenbosch, South Africa, ${}^c$Namaqua Wines, Vredendal, South Africa, ${}^d$South African Grape and Wine Research Institute, Department of Viticulture and Oenology, Stellenbosch University, Stellenbosch, South Africa

This study addresses a classification problem involving variable selection from a large set of highly correlated predictors. In spectral data, measurements at neighbouring wavenumbers are typically strongly correlated, resulting in severe multicollinearity. We propose a variable selection approach that first clusters highly correlated variables and then selects a single representative from each cluster for potential use in the classification model. The method also incorporates a constraint to prevent non-contiguous variables from being grouped within the same cluster. The approach is demonstrated using mid-infrared (MIR) spectral data from grape samples collected to assess grapevine bunch rot, an issue of economic importance to wineries. The objective is to classify new samples either as rot-affected (Yes/No) or into multiple categories reflecting the severity (%) of rot. From the smaller set of potential variables, those important for classification are selected and biplots are used to visualise the separation between classes.

Keywords: biplots, classification, multicollinearity, clustering, bunch rot