The Challenge of Single Estimates in near Infrared Calibration and Prediction: The Measurement of Durum Vitreousness Using Receival Instruments

Abstract

The development of a near infrared (NIR) calibration for the estimation of percentage hard vitreous kernel (%HVK) in whole durum at grain receival points has been driven by the needs of international trade. This application presented some special challenges due to the nature of the available reference data, which were obtained using the Farinator method. The Farinator (transverse grain cutter) method is subjective and is thus subject to large errors, leading to single estimates being very poor reference data. This problem can be overcome by making a number of replicate measurements and using the average of these as a more accurate estimate of the true value. However, this is at the expense of substantially increased analysis time, hence cost. In addition, the HVK methodology involves a grain by grain assessment in which each grain can have only one of two possible classifications, i.e. “vitreous” and “not fully vitreous”. A binomial test such as this introduces an additional complication arising from sampling error, thus making the accuracy of the method sensitive to the number of grains examined. The accuracy also varies across the range of %HVK values so that the greatest errors are observed in the middle of the range. Fortunately, however, the high end of the range corresponding to the lowest errors is also the most commercially relevant. This paper describes a procedure in which, by restricting the calibration range to the optimum for a binomial test, an NIR model suitable for use at receival points could be obtained. The compromise was to calibrate using single Farinator estimates on a large number of samples but to validate using average Farinator estimates on a small number of samples. The results have confirmed that the NIR model yields a more reliable estimate of the true value of %HVK than would a single Farinator estimate on which the calibration was based.