Benchmarking of a new portable NIR sensor for application in post-harvest technology

  • Bertotto, M. (Contributor)
  • Yannick Weesepoel (Contributor)
  • El Harchioui, N. (Contributor)
  • Syed Sarfaraz (Contributor)
  • Nithin Manuel (Contributor)
  • Kees van Gorsel (Contributor)

Activity: Talk or presentationInvited talkAcademic

Description


The primary aim of this study was to conduct a comprehensive evaluation of a novel portable Near-Infrared (NIR) sensor, designated as "Prototype A," with a focus on its potential applications for postharvest technology. This assessment encompassed a comparative analysis against an existing NIR device, commonly used in this field. Our investigation was concentrated on the analysis of unprocessed olives and table grapes. These particular matrices were chosen due to their well-established characterization in terms of both wet-chemical attributes and NIR spectral properties.
We evaluated dry matter (DM) in raw olives and total soluble solids expressed as degrees Brix in table grapes using Prototype A. We compared the performance of the prototype to Golden Standard device ‘B’, being a VISNIR spectrometer. 141 olives were acquired from a Spanish olive company in 2023. 200 table grapes (164 white and 36 red) were purchased in local supermarkets in The Netherlands in 2023. Brix values were determined with an accuracy of 0.1 oBx using a digital refractometer. Moisture was determined by gravimetry with an accuracy of 0.01g and drying at 80°C for 3 days in a drying oven.

Data exploration was performed by Principal Component Analysis (PCA), in order to detect and remove outliers. Raw spectra was pretreated by Standard Normal Variate (SNV), Savitzky Golay (2,17,2). Each data set was divided into calibration and validation sets 70/30, randomly. Regression models were trained in the calibration sets, and evaluated in the validation sets, in terms of Root Mean Square Error of Prediction (RMSEP) and R2. In order to compare RMSEPs between devices, two F. statistical tests were carried out: “If two calibrations have equal performance, then the ratio of their two variances will have an F distribution” (Tom Fearn, 2023).a

Prototype A models presented a lower predictive performance when compared to the models developed using the device ‘B’ but at the cost of incorporating a greater number of latent variables in the models (greater complexity). For the oBx results: The upper 5% point of F 29,29 is 1.840, a ratio greater than this (3.704) was evidence that the observed difference in variances was unlikely due to chance. Olives, %DM: The upper 5% point of F 41,41 is 1.690, a ratio lower than this (1.138) was evidence that the observed difference is not big enough to be statistically significant.

From the results obtained in this work, it can be concluded that both devices allowed to predict the parameters studied (% DM and oBx) in table grapes and % DM in olives, with high accuracy and high correlation coefficients.

Differences between calibration models of both devices were not significant for olive samples, and significant for grapes, as shown by the F-test study.

Period20 Aug 202324 Aug 2023
Event titleInternational Conference of Near Infrared Spectroscopy
Event typeConference
Degree of RecognitionNational