Development of a near‐infrared spectroscopy calibration for Hagberg falling number assessment of barley (Hordeum vulgare): A comparison of methods

Abstract

Climate change leads to increased risks of reduced Hagberg falling numbers (HFN).The objectives of this study were to (i) compare partial least square (PLS) with deep learning methods with respect to establishing near-infrared spectroscopy (NIRS) calibrations for HFN in spring barley, (ii) compare the accuracy of NIRS calibrations for metric versus categorical response variables and (iii) discuss the usefulness of the developed NIRS calibrations in the context of barley breeding programmes. This study was based on 560 samples from the preapplication trial of a commercial spring barley breeding programme and the double round robin population of barley. Across all the examined preprocessing combinations, the lowest root mean square error of prediction (RMSEP) in the calibration set was observed with 59.95 s for PLS regression where that of the best deep learning method was higher. The accuracies to predict HFN from NIR spectra observed in our study indicated that they cannot replace the HFN reference method, but their use for selection in early generations of barley breeding programmes are promising.

Malt quality profile of barley predicted by near‐infrared spectroscopy using partial least squares, Bayesian regression, and artificial neural network models

“…Parameter estimates based on the calibration dataset demonstrated that the best model was PLSR, outperforming Bayesian and PCA‐ANN models (Table 4) likely due to its ability to handle multicollinearity in our spectral data. The superiority of the PLSR model compared to other advanced models is consistent with earlier works, with the PLSR model outperforming ANN and GBM regression deployed in the rapid detection of the microbial spoilage of beef fillets, 26 rapid determination of moisture content in rubber sheets, 27 and the prediction of Hagberg's falling number in barley 28 . This has been attributed to a linear relationship between the spectral data (predictors) and reference data 29 .…”

“…If this is achieved, a “walk back” approach can be used to select superior barley lines using the developed barley malt NIR spectroscopy calibration as a “proxy,” an idea we are currently exploring. We suspected that the high predictive ability obtained for WP compared to the rest of the malt quality traits may have been influenced by the sensitivity of the NIR spectral data to proteins, especially seed storage proteins 28 . NIR spectroscopy is based on measurements arising from inter‐atomic bonds between atoms of low mass such as C, N, O, and H. Although it is more sensitive to proteins through a conglomeration of the N‐H bond vibrations, the sensitivity of the NIR response to individual amino acids is quite challenging 12 …”

“…The superiority of the PLSR model compared to other advanced models is consistent with earlier works, with the PLSR model outperforming ANN and GBM regression deployed in the rapid detection of the microbial spoilage of beef fillets, 26 rapid determination of moisture content in rubber sheets, 27 and the prediction of Hagberg's falling number in barley. 28 This has been attributed to a linear relationship between the spectral data (predictors) and reference data. 29 In order to explore the factors contributing to the low prediction performance observed in our study, we utilized the pre-treated data as predictors.…”

“…We suspected that the high predictive ability obtained for WP compared to the rest of the malt quality traits may have been influenced by the sensitivity of the NIR spectral data to proteins, especially seed storage proteins. 28 NIR spectroscopy is based on measurements arising from inter-atomic bonds between atoms of low mass such as C, N, O, and H. Although it is more sensitive to proteins through a conglomeration of the N-H bond vibrations, the sensitivity of the NIR response to individual amino acids is quite challenging. 12 The PLSR and Bayesian model were used to identify the informative waveband regions from the full spectrum that are directly or indirectly linked to the malt quality parameters.…”