Deep neural networks (DNNs) chemical compositions estimation of fresh durian in-line via near infrared spectroscopy

Simple item page
dc.contributor.authorPosom, Jetsada
dc.contributor.authorSaenphon, Chirawan
dc.contributor.authorDitcharoen, Sirirak
dc.contributor.authorPitak, Lakkana
dc.contributor.authorSirisomboon, Panmanas
dc.contributor.authorMaraphum, Kanvisit
dc.date.accessioned2026-08-06T10:51:17Z
dc.date.available2026-08-06T10:51:17Z
dc.date.issued2025-06-01
dc.description.abstractNear infrared (NIR) spectroscopy and deep neural network (DNN) models were adopted for evaluating the nutritional compositions in durian pulp. The quality inspection of durians through online channels remains challenging because consumers cannot directly touch or smell the fruit. This leads to issues with substandard durians, such as unripe ones or those infested with pests. One hundred and sixty durian samples of Mon Thong varieties were used in this experiment. Then, the collected NIR spectra were augmented to improve the generalization ability of regression models. Deep neural network (DNNs) regressions were developed. The results showed that deep neural network (DNN) regression model possessed the best prediction performance, which were provided performance index. Almost all the best performance models were developed from the second derivative, except for the fat content model, which was developed from raw spectra. The best total soluble solids (TSS) model provided the coefficient of determination of calibration (R<sup>2</sup>c) and root mean square error of calibration (RMSEc) were 0.84 and 2.25 % and coefficient of determination of calibration (R<sup>2</sup>p), root mean square error of prediction (RMSEp) and ratio of performance to deviation (RPD) were 0.72 2.92 % and 1.92, respectively, and those of dry matter content (DMC) were provided R<sup>2</sup>c and RMSEc were 0.997, 0.58 %, and for prediction set provided R<sup>2</sup>p, RMSEp and RPD were 0.94, 3.13 %, and 4.21, respectively. For fat content (FC), they have also provided R<sup>2</sup>c and RMSEc were 0.84, 0.36 (g/100 g), while it provided R<sup>2</sup>p, RMSEp and RPD were 0.86, 0.50 (g/100 g), and 2.72, respectively. Moreover, for total sugar content (TSC) value, it also gave high accuracy, which provided R<sup>2</sup>c and RMSEc were 0.93, 0.49 (g/100 g) and R<sup>2</sup>p, RMSEp, and RPD were 0.91, 0.81 (g/100 g), 3.44, and for starch content (SC) also provided R<sup>2</sup>c and RMSEc were 0.93, 0.49(g/100 g) and R<sup>2</sup>p, RMSEp, and RPD were 0.76, 2.79 (g/100 g), 2.09, respectively. Therefore, the proposed method offers an ultrasensitive and effective strategy for estimation of nutritional compositions.
dc.identifier.citationJournal of Food Composition and Analysis, 142, 2025
dc.identifier.doi10.1016/j.jfca.2025.107410
dc.identifier.issn08891575
dc.identifier.other2-s2.0-85218872416
dc.identifier.urihttps://dspace.kmitl.ac.th/handle/123456789/17008
dc.sourceJournal of Food Composition and Analysis
dc.subjectDurian
dc.subjectFat content
dc.subjectNutrition
dc.subjectSugar content
dc.titleDeep neural networks (DNNs) chemical compositions estimation of fresh durian in-line via near infrared spectroscopy
dc.typeArticle

Files

Collections

Research