Classification of maize kernels using NIR hyperspectral imaging.

Area of Science:

• Agricultural Science
• Spectroscopy
• Machine Learning

Background:

• Maize kernel hardness is a critical quality trait affecting processing and end-use.
• Accurate and efficient classification methods are needed for large-scale production.

Purpose of the Study:

• To evaluate near-infrared (NIR) hyperspectral imaging for classifying maize kernels into hardness categories (hard, medium, soft).
• To compare pixel-wise and object-wise classification approaches for accuracy and feasibility.

Main Methods:

• NIR hyperspectral imaging data were acquired from maize kernels.
• Pixel-wise classification was attempted, with improvements using a threshold-based approach.
• Object-wise classification utilized score histograms and mean spectra for feature extraction.

Main Results:

• Pixel-wise classification yielded high misclassification initially; thresholding improved sensitivity to 0.75 and specificity to 0.97.
• Object-wise classification with score histograms achieved 0.93 sensitivity and 0.97 specificity for hard kernels.
• Object-wise classification with mean spectra achieved 0.95 sensitivity and 0.93 specificity for medium kernels.

Conclusions:

• Object-wise classification approaches using NIR hyperspectral imaging are highly effective for maize kernel hardness classification.
• Both score histograms and mean spectra are suitable feature extraction methods for production-scale applications.