Self-supervised learning for aflatoxin B1 detection using masked spectra.

Area of Science:

• Agricultural Science
• Analytical Chemistry
• Machine Learning

Background:

• Aflatoxins are toxic metabolites commonly found in peanuts and maize.
• Accurate detection of Aflatoxin B1 is crucial for food safety.
• Existing detection methods may require extensive labeled data.

Purpose of the Study:

• To develop an accurate and efficient method for Aflatoxin B1 detection in peanuts and maize using self-supervised learning (SSL).
• To leverage spectral data for improved detection accuracy with limited samples.

Main Methods:

• Constructed an SSL backbone network for spectral reconstruction on masked spectra.
• Utilized spectral reconstruction error as a wavelength attention mechanism.
• Developed a classification model by integrating the SSL-tuned encoder with the attention mechanism.

Main Results:

• Achieved high mean accuracy: 0.9805 ± 0.0019 for peanuts and 0.9478 ± 0.0060 for maize.
• Outperformed other deep learning models in Aflatoxin B1 detection.
• Identified 40 key wavelengths using spectral reconstruction error, enhancing spectral feature capture.

Conclusions:

• The proposed SSL strategy provides an efficient solution for few-sample spectral detection of Aflatoxin B1.
• The novel wavelength attention mechanism improves detection accuracy and feature extraction.
• This approach offers a promising tool for food safety and quality control.