Journal of Spectral Imaging,   Volume 6   Article ID a4   (2017)

Peer reviewed Paper

Application of calibrations to hyperspectral images of food grains: example for wheat falling number

  • Nicola Caporaso
  • Martin B. Whitworth
  • Ian D. Fisk
Campden BRI, Station Road, Chipping Campden, Gloucestershire, GL55 6LD, UK

 Search for papers by this author
Division of Food Sciences, The University of Nottingham, Sutton Bonington Campus, Loughborough LE12 5RD, UK

 https://orcid.org/0000-0001-8448-3123
 Search for papers by this author
 Corresponding Author
Campden BRI, Station Road, Chipping Campden, Gloucestershire, GL55 6LD and Division of Food Sciences, The University of Nottingham, Sutton Bonington Campus, Loughborough LE12 5RD, UK
[email protected]
 https://orcid.org/0000-0002-1826-2565
 Search for papers by this author

The presence of a few kernels with sprouting problems in a batch of wheat can result in enzymatic activity sufficient to compromise flour functionality and bread quality. This is commonly assessed using the Hagberg Falling Number (HFN) method, which is a batch analysis. Hyperspectral imaging (HSI) can provide analysis at the single grain level with potential for improved performance. The present paper deals with the development and application of calibrations obtained using an HSI system working in the near infrared (NIR) region (~900–2500 nm) and reference measurements of HFN. A partial least squares regression calibration has been built using 425 wheat samples with a HFN range of 62–318 s, including field and laboratory pre-germinated samples placed under wet conditions. Two different approaches were tested to apply calibrations: i) application of the calibration to each pixel, followed by calculation of the average of the resulting values for each object (kernel); ii) calculation of the average spectrum for each object, followed by application of the calibration to the mean spectrum. The calibration performance achieved for HFN (R2 = 0.6; RMSEC ~ 50 sRMSEP ~ 63 s) compares favourably with other studies using NIR spectroscopy. Linear spectral pre-treatments lead to similar results when applying the two methods, while non-linear treatments such as standard normal variate showed obvious differences between these approaches. A classification model based on linear discriminant analysis (LDA) was also applied to segregate wheat kernels into low (<250 s) and high (>250 s) HFN groups. LDA correctly classified 86.4% of the samples, with a classification accuracy of 97.9% when using an HFN threshold of 150 s. These results are promising in terms of wheat quality assessment using a rapid and non-destructive technique which is able to analyse wheat properties on a single-kernel basis, and to classify samples as acceptable or unacceptable for flour production.

Keywords: hyperspectral imaging, partial least squares calibration, food grains, whole wheat kernels, Hagberg Falling Number (HFN)


Metrics

Downloads:

1,418

Full-Text HTML Views:

121

Abstract Views:

1,632