Classification of white maize defects with multispectral imaging.

Area of Science:

• Agricultural Science
• Image Analysis
• Spectroscopy

Background:

• Accurate grading of white maize kernels is crucial for quality control and regulatory compliance.
• Traditional methods for kernel defect detection can be labor-intensive and subjective.
• Developing objective and efficient grading technologies is essential for the grain industry.

Purpose of the Study:

• To evaluate multispectral imaging combined with object-wise multivariate image analysis for grading whole white maize kernels.
• To develop and validate models for distinguishing sound maize from various classes of defective materials.
• To identify key spectral wavelengths indicative of specific kernel defects and composition.

Main Methods:

• Utilized a multispectral imaging instrument capturing data across UV, visible, and near-infrared (NIR) regions (375-970nm).
• Classified defective materials into 13 distinct categories based on grading legislation.
• Developed and validated object-wise partial least squares discriminant analysis (PLS-DA) models using independent datasets.

Main Results:

• Achieved high performance in differentiating sound maize from undesirable materials, with classification accuracies ranging from 83% to 100%.
• Cross-validated coefficients of determination (Q²) for the models ranged from 0.35 to 0.99.
• Identified specific wavelengths (e.g., 505-590nm for pigments, 890-970nm for fat and water content) as critical features for classification.

Conclusions:

• Multispectral imaging with object-wise multivariate analysis is a promising technology for automated grading of white maize kernels.
• The study successfully demonstrated the potential for accurate defect detection and classification based on spectral properties.
• Key wavelengths in the visible and NIR regions provide valuable information for assessing maize kernel quality and composition.