Wheat Kernel Variety Identification Based on a Large Near-Infrared Spectral Dataset and a Novel Deep Learning-Based...

Area of Science:

• Agricultural Science
• Spectroscopy
• Machine Learning

Background:

• Nondestructive sensing technologies are crucial for crop seed inspection.
• Near-infrared (NIR) hyperspectroscopy is an emerging technology for analyzing seed characteristics.
• Analyzing large spectral datasets requires effective feature selection methods.

Purpose of the Study:

• To develop a novel convolutional neural network-based feature selection method (CNN-FS) for spectral data.
• To design a convolutional neural network with attention (CNN-ATT) framework for classifying wheat kernels.
• To evaluate the performance of the proposed methods against conventional techniques.

Main Methods:

• A large dataset of over 140,000 wheat kernels was analyzed using NIR hyperspectroscopy.
• A CNN-FS was developed to identify relevant spectral channels.
• A CNN-ATT model was implemented for classification, compared with SVM and PLS-DA.

Main Results:

• The CNN-ATT model achieved 93.01% accuracy using full spectra.
• Using features selected by CNN-FS, the CNN-ATT maintained high accuracy (90.20%).
• The proposed CNN-FS outperformed conventional feature selection algorithms in representing the data.

Conclusions:

• The combination of NIR hyperspectroscopy and the proposed CNN-ATT model enables accurate, automatic, nondestructive classification of single wheat kernels.
• The CNN-FS method effectively screens spectral channels, enhancing classification performance.
• These methods show significant potential for analyzing other large spectral datasets in agriculture.