Pixel-level aflatoxin detecting in maize based on feature selection and hyperspectral imaging.

Area of Science:

• Agricultural Science
• Food Safety
• Spectroscopy

Background:

• Aflatoxins are highly toxic fungal metabolites commonly found in maize, posing a significant risk of liver cancer.
• Accurate detection of aflatoxin contamination is crucial for food safety and public health.

Purpose of the Study:

• To develop and evaluate hyperspectral imaging models for pixel-level aflatoxin classification in maize.
• To compare the effectiveness of different feature extraction and selection methods for improving classification accuracy.

Main Methods:

• Utilized hyperspectral data with 600 bands per pixel, labeled as 'clean' or 'contaminated'.
• Applied three feature extraction/selection approaches: 4 specific bands, Principal Component Analysis (PCA), and algorithms (Fscnca, Fscmrmr, Relieff, Fishier) for top 10 band selection.
• Classified pixels using Random Forest (RF) and K-nearest neighbor (KNN) algorithms with selected/extracted features.

Main Results:

• The Relieff feature selection method achieved the highest accuracy (99.38% with RF, 98.77% with KNN).
• PCA feature extraction with RF yielded 93.83% accuracy.
• Using all 600 bands without feature extraction resulted in 100% accuracy, though potentially computationally intensive.
• Bands selected from prior studies achieved 89.51% accuracy.

Conclusions:

• Feature extraction and selection methods significantly impact aflatoxin classification accuracy using hyperspectral data.
• The Relieff algorithm combined with RF offers a highly accurate and efficient approach for aflatoxin detection.
• Full-band hyperspectral data provides the highest accuracy for aflatoxin classification when computational time is not a constraint.