Near-Infrared Spectroscopy Machine-Learning Spectral Analysis Tool for Blueberries () Cultivar Discrimination.

is one of the main sources of commercialized blueberries across the world. This species has a large number of distinct cultivars, leading to significantly different berries characteristics such as size, sweetness, production rate, and growing season. In this context, accurate cultivar discrimination is of significant relevance, and currently, it is mostly performed through berries examination. In this work, we developed a method to discriminate 19 cultivars from the species through their leaves' near-infrared spectra. Spectra were acquired from fresh blueberry leaves collected from two geographic regions and across three seasons. Machine-learning-based models, selected from a pool of 10 classifiers based on their discrimination power under a twofold stratified cross-validation process, were trained/tested with 1 to 20 components obtained by the application of data dimensionality reduction (DDR) techniques (dictionary learning, factor analysis, fast individual component analysis, and principal component analysis) to different near-infrared (NIR) spectra regions' data, to either analyze a single spectral region and season or combine spectral regions and/or seasons for each side of the blueberry leaf. The percentage of correct cultivar discrimination ranged from 52.27 to 100%, with the best spectral results obtained with the adaxial side of the leaves in the fall (100% Accuracy) and the abaxial side of the leaves in the winter (100% Accuracy); the fast ICA DDR technique was present in 83.33% of the best analyses (five out of six); and the LinearSVC was present in 66.67% (four out six best analyses). The results obtained in this work denote that near-infrared spectroscopy is a suitable and accurate technique for cultivar discrimination.