Evaluation of sampling strategies in grains using spatial distribution surfaces from NIR spectral data

The aim of feed and food regulations is an integrated system of controls across the entire food chain so as to achieve safe, traceable and quality products. An essential element of this is the need to sample and analyse bulk products, such as feed ingredients, in an efficient and cost-effective manner. The ability of near-infrared spectroscopy (NIRS) to provide rapid analytical measurements at low cost make it an ideal tool in this context. The possibility of making many more measurements on a bulk sample, with a result for each sampling point rather than a single result for a combined sample, is a motivation for rethinking existing sampling strategies and investigating how we can exploit the extra spatial information.
As a first step towards this, the present study investigated, on a laboratory scale, the sampling and measurement by NIRS of mixtures of wheat and barley grains with varying degrees of homogeneity. For this purpose, twelve lots containing 18 kg sample/lot were prepared, two of them from pure samples of wheat (W) and barley (B) with the rest being mixtures with percentages 90W-10B, 75W-25B, 50W-50B, 25W-75B, 10W-90B, using two mixing times, 5 and 15 minutes in a V type mixing machine, in each case. Each lot was placed into an infrared-transparent glass box (0.5 x 0.3 x 0.35 m) and analysed by the Corona 45 instrument, a diode array Vis-NIR reflectance spectrometer with range 380-1690 nm and a sampling interval of 2 nm. The spectrometer measured through the base of the box with a fixed distance of 13 mm to the bottom surface of the sample. The diameter of the window of the instrument is 7 cm, giving a measurement area of 38.48 cm2 per measurement point, and measurements were made over a grid of 20 points/lot.
The spectral data were used to predict wheat percentages for each measuring point in each lot. After this, a routine in R (3.2.0 version, R Project) provided trend surfaces by least-squares from a polynomial interpolation, using some or all of the NIR predictions as an input, for the evaluation of three different sampling protocols. In the first procedure all sampled locations were used by the model to obtain the continuous maps for each lot, while 75% and 50% of points were randomly selected from the whole data set as input to the fitting procedure for the second and the third sampling protocols respectively. Then, the Root Mean Square Error of Prediction (RMSEP) was calculated over the entire data set in each case, to assess how well the model predicted unsampled points and described the spatial distribution for each sampling strategy. The RMSEP for the first protocol was 6%, smaller than the values for the second (11%) and the third (18%) strategies, indicating a declining performance of the model in the description of the spatial distribution as the number of points became lower and the spatial distance increased.