Differentiating five agrochemicals used in the treatment of intact olives by means of NIR spectroscopy, discriminant analysis and compliant class models

• NIR spectroscopy to differentiate pesticide-treated versus untreated olives. • Comparative analysis between ECOC-PLS2-CM and a hierarchical model. • More sensitive and specific class models to identify food contamination. • Evaluating the best sensitivity/specificity in a multiclass problem. • How to evaluate the sensitivity/specificity in a multiclass problem using DMCEN. This paper deals with the application of near infrared spectroscopy (NIR) in a classification problem involving multiple classes in order to differentiate contaminated olives. A total of 452 samples, ripe and unripe, were treated with five different agrochemicals reproducing the traditional fumigation process in the olive tree. The main objective was to differentiate through a classification if the samples were or were not treated, but also, which chemical was used for each olive. Firstly, Partial Least Squares-Discriminant Analysis (PLS-DA) was performed to differentiate between untreated and treated samples. Then, two novel chemometric approaches, a classification one and a modelling one, were applied for ripe and unripe olives, achieving good results and determining with which chemical were the olives sprinkled with. For the classification of the samples in the six different classes (untreated olives, or treated with one of the five agrochemicals), an Automatic Hierarchical Model Builder (AHIMBU) was used, applying sequential binary PLS-DAs. Nevertheless, for the modelling approach, a compliant model, PLS2-CM, also based on PLS, was used with two different codifications for the classes: i) the classic and well-known One Versus All (OVA), and ii) the Error Correction Output Code (ECOC) optimal matrix. The final global results were evaluated using the Diagonal Modified Confusion Entropy (DMCEN) index, which ranges between 0 and 1, and is very sensitive to changes in the sensitivity–specificity matrices (note that the lower the DMCEN, the better the classification is). The best DMCEN value in prediction for unripe olives, 0.4898, was obtained for the PLS2-CM-ECOC, while 0.6937 and 0.7705 DMCEN values were obtained for AHIMBU and PLS2-CM-OVA, respectively. For the case of the ripe samples, the DMCEN values in prediction were better than the ones for the unripe olives: 0.6016, 0.5051, and 0.4166, for AHIMBU, PLS2-CM-OVA and PLS2-CM-ECOC, respectively. In every case, the best DMCEN has been obtained with the PLS2-CM-ECOC procedure.