Open AccessFeasibility of FT-NIR spectroscopy and Vis/NIR hyperspectral imaging for sorting unsound chestnuts
Author(s) -
Giacomo Bedini,
Giorgia Bastianelli,
Swathi Sirisha Nallan Chakravartula,
Carmen MoralesRodrìguez,
Luca Rossini,
Stefano Speranza,
Andrea Vannini,
Roberto Moscetti,
Riccardo Massantini
Publication year - 2020
Publication title -
italus hortus
Language(s) - English
Resource type - Journals
ISSN - 1127-3496
DOI - 10.26353/j.itahort/2020.1.0318
Subject(s) - hyperspectral imaging , partial least squares regression , near infrared spectroscopy , linear discriminant analysis , sorting , smoothing , mathematics , pattern recognition (psychology) , artificial intelligence , categorical variable , remote sensing , computer science , statistics , geology , optics , algorithm , physics
Authors explored the potential use of Vis/NIR hyperspectral imaging (HSI) and Fourier-transform Near-Infrared (FT-NIR) spectroscopy to be used as in-line tools for the detection of unsound chestnut fruits (i.e. infected and/or infested) in comparison with the traditional sorting technique. For the intended purpose, a total of 720 raw fruits were collected from a local company. Chestnut fruits were preliminarily classified into sound (360 fruits) and unsound (360 fruits) batches using a proprietary floating system at the facility along with manual selection performed by expert workers. The two batches were stored at 4 ± 1 °C until use. Samples were left at ambient temperature for at least 12 h before measurements. Subsequently, fruits were subjected to non-destructive measurements (i.e. spectral analysis) immediately followed by destructive analyses (i.e. microbiological and entomological assays). Classification models were trained using the Partial Least Squares Discriminant Analysis (PLS-DA) by pairing the spectrum of each fruit with the categorical information obtained from its destructive assay (i.e., sound, Y = 0; unsound, Y = 1). Categorical data were also used to evaluate the classification performance of the traditional sorting method. The performance of each PLS-DA model was evaluated in terms of false positive error (FP), false negative error (FN) and total error (TE) rates. The best result (8% FP, 14% FN, 11% TE) was obtained using Savitzky-Golay first derivative with a 5-points window of smoothing on the dataset of raw reflectance spectra scanned from the hilum side of fruit using the Vis/NIR HSI setup. This model showed similarity in terms of False Negative error rate with the best one computed using data from the FT-NIR setup (i.e. 15% FN), which, however, had the lowest global performance (17% TE) due to the highest False Positive error rate (19%). Finally, considering that the total error rate committed by the traditional sorting system was about 14.5% with a tendency of misclassifying unsound fruits, the results indicate the feasibility of a rapid, in-line detection system based on spectroscopic measurements.