Multiblock spectral imaging for identification of pre-harvest sprouting in Hordeum vulgare

Abstract

A novel data fusion method based on the use of visible/near-infrared (VNIR) and shortwave infrared (SWIR) imaging sensors, to distinguish between pregerminated and ungerminated barley grain is proposed. Spectral imaging was used to fingerprint germinated and ungerminated barley grain from a total of 5640 average spectra representing single barley kernels varying with respect to germination time. Chemometric approaches utilising partial least squares-discriminant analysis (PLS-DA) and multiblock sequential and orthogonalized partial least squares-linear discriminant analysis (SO-PLS-LDA) and sequential and orthogonalized covariance selection-linear discriminant analysis (SO-CovSel-LDA) were used to build classification models. SO-PLS-LDA achieved a total classification rate of 99.88%, while SO-CovSel-LDA resulted in a classification accuracy of 97.46% when a maximum of 8 variables were selected from each data block (VNIR and SWIR) – models were validated on an independent test set. The use of multiblock approaches led to increased prediction accuracy, compared to PLS-DA, and a viable solution to address the industry problem to detect pregerminated malting barley in a rapid, non-destructive manner. This represents a significant advance with respect to the current dated methods which are hindered by time-consuming wet chemistry techniques and human subjective bias. The potential of the proposed new technique also has the further advantage of moving toward multispectral systems which can be used to detect pre-harvest germinated barley using an even more computationally rapid and affordable online sorting machine incorporating the wavebands of importance selected by SO-CovSel-LDA. The study highlights how sequential and orthogonalised data fusion approaches, in the food and agricultural sector, are powerful solutions to real world problems.