An automated classification model based on NIR spectra to predict the technological quality of common wheat

Empirical testing methods on wheat flour have been performed since the early days of cereal science: they provide useful information and are widely accepted for wheat flour characterization and quality control. However, they frequently involve a large number of experimental measurements, generally expensive and time-consuming, and characterized by a rather complex and prone to subjectivity interpretation of results. In this context, the aim of this research was to develop a new and automated method based on FT-NIR spectra, able to predict - in short times and with low uncertainty - the technological quality of common wheat flours “as a whole”, differently from the usual quantification of single physical or chemical properties. An extensive preliminary work was done to objectively define the instrumental measuring conditions leading to the highest information and the lowest noise, considering the nature of the sample (kernels, whole meal flour, flour, different sampling tools (fiber optic, integrating sphere,transmission unit), and 2 independent laboratories using 2 Bruker MPA FT-NIR spectrometers. Results confirmed that optimal measuring conditions can vary depending on the type of analyzed matrix, and allowed in the selection of the NIR sampling tools for the subsequent acquisition of a wide spectral dataset. According to these findings, the FT-NIR spectra of 303 bread wheat samples, harvested in 2007-2008 years and representative of the different Italian wheat varieties and of the different soil and climate areas, were collected in parallel in the 2 laboratories and submitted to data processing. Automated classification models, based on PLS-Discriminant Analysis, were developed to predict new common wheat quality classes, based on the reference experimental parameters and the FT-NIR spectra acquired on the same samples.