Feature discovery in NIR spectroscopy based Rocha pear classification

• Published in: Expert systems with applications Vol. 177; p. 114949
• Main Authors: Daniel, Mariana, Guerra, Rui, Brázio, António, Rodrigues, Daniela, Cavaco, Ana Margarida, Antunes, Maria Dulce, Valente de Oliveira, José
• Format: Journal Article
• Language: English
• Published: New York Elsevier Ltd 01-09-2021; Elsevier BV
• Subjects: Classification; Classifiers; Data analysis; Diffuse reflectance spectroscopy; Feature extraction; Feature selection; Fruits; Machine learning; Physical chemistry; Statistical methods; Statistical tests; Feature extraction; Feature selection; Data analysis; Classification; Machine learning
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2021.114949

•The problem of Vis/NIR based fruit classification is considered.•An innovative method of feature generation and selection is proposed.•A total of 3050 Rocha Pear heterogeneous sample is analyzed.•The proposed method outperforms existing methods while reduces the feature number.•Selected features correspond to chemically meaningful wavelength bands.
Non-invasive techniques for automatic fruit classification are gaining importance in the global agro-industry as they allow for optimizing harvesting, storage, management, and distribution decisions. Visible, near infra-red (NIR) diffuse reflectance spectroscopy is one of the most employed techniques in such fruit classification. Typically, after the acquisition of a fruit reflectance spectrum the wavelength domain signal is preprocessed and a classifier is designed. Up to now, little or no work considered the problem of feature generation and selection of the reflectance spectrum. This work aims at filling this gap, by exploiting a feature engineering phase before the classifier. The usual approach where the classifier is fed directly with the reflectances measured at each wavelength is contrasted with the proposed division of the spectra into bands and their characterization in wavelength, frequency, and wavelength-frequency domains. Feature selection is also applied for optimizing efficiency, predictive accuracy, and for mitigating over-training. A total of 3050 Rocha pear samples from different origins and harvest years are considered. Statistical tests of hypotheses on classification results of soluble solids content – a predictor of both fruit sweetness and ripeness – show that the proposed preliminary phase of feature engineering outperforms the usual direct approach both in terms of accuracy and in the number of necessary features. Moreover, the method allows for the identification of features that are physical chemistry meaningful.