NIR Instruments and Prediction Methods for Rapid Access to Grain Protein Content in Multiple Cereals

Abstract

Achieving global goals for sustainable nutrition, health, and wellbeing will depend on delivering enhanced diets to humankind. This will require instantaneous access to information on food-source quality at key points of agri-food systems. Although laboratory analysis and benchtop NIR spectrometers are regularly used to quantify grain quality, these do not suit all end users, for example, stakeholders in decentralized agri-food chains that are typical in emerging economies. Therefore, we explored benchtop and portable NIR instruments, and the methods that might aid these particular end uses. For this purpose, we generated NIR spectra for 328 grain samples from multiple cereals (finger millet, foxtail millet, maize, pearl millet, and sorghum) with a standard benchtop NIR spectrometer (DS2500, FOSS) and a novel portable NIR-based instrument (HL-EVT5, Hone). We explored classical deterministic methods (via winISI, FOSS), novel machine learning (ML)-driven methods (via Hone Create, Hone), and a convolutional neural network (CNN)-based method for building the calibrations to predict grain protein out of the NIR spectra. All of the tested methods enabled us to build relevant calibrations out of both types of spectra (i.e., R² ≥ 0.90, RMSE ≤ 0.91, RPD ≥ 3.08). Generally, the calibration methods integrating the ML techniques tended to enhance the prediction capacity of the model. We also documented that the prediction of grain protein content based on the NIR spectra generated using the novel portable instrument (HL-EVT5, Hone) was highly relevant for quantitative protein predictions (R² = 0.91, RMSE = 0.97, RPD = 3.48). Thus, the presented findings lay the foundations for the expanded use of NIR spectroscopy in agricultural research, development, and trade.

Keywords: Convolution Neural Network (CNN); Hone Create; cereals; near-infrared spectroscopy (NIRS); prediction methods; protein; winISI.

Figure 1

Graphical overview of the methodology visualizing the process used for testing the NIR instruments and methods for prediction of protein content in multiple cereal grains.

Figure 2

Box plots depicting variation and distribution of protein content (%, (g·100 g⁻¹)) in the grains of five cereals, as estimated through laboratory analyses. Legend: Each box represents one crop species; different crops are distinguished by color (finger millet = brown; maize = yellow; sorghum = green; foxtail millet = orange; pearl millet = grey; and the entire set of 328 multiple cereals = blue); solid line within the box (–) represents the mean of each crop.

Figure 3

Histograms depicting the distribution of (A) the protein content (%, (g·100 g⁻¹)) in samples, and (B) the number of samples used within each of the crop species belonging to the calibration (80%) and validation (20%) datasets.

Figure 4

Mean of the near-infrared (NIR) spectra of all grain samples extracted from the benchtop FOSS-DS2500 (400–2498 nm; solid line (–) in grey colour) and the portable HL-EVT5 (1350–2550 nm; dashed line (---) in red colour) instruments.

Figure 5

Means of the near-infrared (NIR) spectra of the grain samples of five cereal species produced using (A) FOSS-DS2500, 400–2498 nm; solid line (–), and (B) HL-EVT5, 1350–2550 nm; dashed line (---) instruments. Different crops are distinguished by color (Legend: finger millet = brown; foxtail millet = orange; maize = yellow; pearl millet = grey; sorghum = green).

Figure 6

Matrix of scatter plots showing protein predicted for the calibration and validation datasets of FOSS-DS2500 and HL-EVT5 via methods available in (I) WinISI software, (II) Hone Create soft-ware, and (III) CNN-based customized method. Detailed metrics for comparison with other methods are shown in Table 2.