Identification of Soybean Varieties Using Hyperspectral Imaging Coupled with Convolutional Neural Network

Area of Science:

• Agricultural Science
• Spectroscopy
• Machine Learning

Background:

• Soybean variety influences stress resistance, nutritional content, and commercial value.
• Accurate soybean variety classification is crucial for agricultural management and breeding programs.

Purpose of the Study:

• To classify three soybean varieties (Zhonghuang37, Zhonghuang41, Zhonghuang55) using near-infrared hyperspectral imaging.
• To investigate the efficiency of convolutional neural networks (CNNs) for soybean variety identification.
• To explore the possibility of achieving high classification accuracy with fewer samples.

Main Methods:

• Near-infrared hyperspectral imaging was employed to capture spectral data from soybeans.
• Pixel-wise and average spectra were extracted and preprocessed.
• Convolutional neural networks (CNNs) were developed using both pixel-wise and average spectral data.
• A majority vote strategy was applied to pixel-wise CNN models for single soybean identification.

Main Results:

• CNN models achieved high classification accuracy for soybean varieties, exceeding 90%.
• Pixel-wise CNN models demonstrated strong performance in predicting both pixel-wise and average spectra.
• Classification performance improved with an increased number of soybean samples.
• Models utilizing pixel-wise spectra from 60 soybeans showed comparable performance to models using average spectra from 810 soybeans.

Conclusions:

• Near-infrared hyperspectral imaging combined with CNNs offers an effective method for soybean variety classification.
• Acquiring pixel-wise spectral data enables accurate soybean discrimination using significantly fewer samples compared to traditional methods.
• This approach reduces the time and labor required for hyperspectral data acquisition and analysis in soybean breeding and quality control.