Affiliation:
1. Department of Arboricolture and pomology, University of Turin, Via Leonardo da Vinci 44, Grugliasco (TO), Italy
2. National Food Research Institute, 2-1-12 Kannondai, Tsukuba 305-8642, Japan
Abstract
In a practical near infrared (NIR) sweetness sorting facility, there is a high possibility that the variety of incoming fruit will be different from the one(s) used to develop a calibration model. In this kind of situation, significant bias could occur and result in a high price difference. In this work, a method to identify whether the pre-installed calibration models could be used to predict Brix values of the unknown samples or not was established. Two kinds of calibration model for Brix determination of nectarines were examined, one developed from spectra of the single nectarine variety called Big Bang Maillarà (BB), the other developed from spectra of two nectarine varieties, BB and Nectaross (NT). Another variety called Sweet Red (SR) was used to demonstrate the situation of an unknown sample or a different variety. For the general biases evaluation, significant biases occurred in the SR samples whether the prediction was done using the BB calibration (bias = −6.74°Brix) or the BB+NT calibration (bias = −3.09°Brix). A survey using score plots of principle component analysis (PCA) indicated that the characteristics of the BB and the SR samples were quite different from each other, while the NT samples were located in between. Even though the trend of differences between the PCA scores of the three varieties could be observed, a clear classification result could not be obtained. Another classification attempt was made using soft independent model for classification analogy (SIMCA). For each calibration (BB and BB+NT), a SIMCA model was developed using the samples used to develop the calibration model. A calculation was performed to evaluate whether the unknown samples (BB, NT and SR) were in the same class as the calibration sample or not. It was found that, in the case of the single variety calibration where the SIMCA model was also developed from the single variety (BB), clear identification of unfitted samples could be obtained with classification accuracy more than 95% (false negative 0%). However, in the case of the two varieties calibration, the use of a SIMCA box developed from two varieties would make the in-class distance become too wide and this reduced the classification accuracy of unfitted samples (false negatives = 74%). The solution was to develop two SIMCA boxes, one for BB and one for NT, then examine the compatibility of unknown samples twice. Using this technique, satisfactory classification results with false negatives of 10% could be obtained.
Cited by
10 articles.
订阅此论文施引文献
订阅此论文施引文献,注册后可以免费订阅5篇论文的施引文献,订阅后可以查看论文全部施引文献