Hierarchical classification pathway for white maize, defect and foreign material classification using spectral imaging

Abstract

This study aimed to present the South African maize industry with an accurate and affordable automated analytical technique for white maize grading using near infrared (NIR) spectral imaging. The 17 categories and sub-categories stipulated in South African maize grading legislation were simultaneously classified (1044 samples; 60 kernels of each class) using 25 partial least squares discriminant analysis (PLS-DA) models. The models were assembled in a hierarchical decision pathway that progressed from the most easily classified classes to the most difficult. The full NIR spectrum (288 wavebands) model performed with an overall accuracy of 93.3% for the main categories. Three waveband selection techniques were employed, namely waveband windows (48 wavebands), variable importance in projection (VIP) (21 wavebands) and covariance selection (CovSel) (13 wavebands). Overall, the VIP set based on only 7.3% of the original spectral variables was recommended as the best trade-off between performance and expected cost of a reduced waveband system.