Identification of Defective Maize Seeds Using Hyperspectral Imaging Combined with Deep Learning

Area of Science:

• Agricultural Science
• Computer Vision
• Spectroscopy

Background:

• Traditional maize seed quality assessment is destructive and time-consuming.
• Developing non-destructive, rapid methods is crucial for agricultural productivity.

Purpose of the Study:

• To establish a fast, non-destructive, and effective approach for maize seed defect detection using hyperspectral imaging (HSI) and deep learning.
• To develop and validate novel deep learning models for spectral data analysis.

Main Methods:

• Pre-processing of hyperspectral data using detrending (DE) and multiple scattering correction (MSC).
• Development of a convolutional neural network with feature selection (CNN-FES) for wavelength importance.
• Design of a convolutional neural network with an attentional mechanism (CNN-ATM) for spectral classification.
• Comparison with traditional machine learning algorithms like LDA, RF, and SVM.

Main Results:

• CNN-FES effectively selected 24 feature wavelengths, outperforming SPA and CARS.
• CNN-ATM achieved high classification accuracy (>90%) on full spectral data.
• CNN-ATM utilizing feature wavelengths reached up to 97.50% accuracy, 98.28% sensitivity, and 96.77% specificity.

Conclusions:

• Hyperspectral imaging is a feasible and effective technology for maize seed defect detection.
• The proposed deep learning methods demonstrate significant potential for analyzing complex hyperspectral data in agriculture.