Quantitative detection of Aflatoxin B1 by subpixel CNN regression

Area of Science:

• Agricultural Science
• Analytical Chemistry
• Computer Science

Background:

• Aflatoxins, particularly aflatoxin B1 (AFB1), are highly toxic contaminants found in peanuts, posing significant risks to human and animal health.
• Accurate and rapid detection methods for AFB1 are crucial for ensuring food safety and preventing widespread contamination.
• Traditional detection methods can be time-consuming and may lack the sensitivity required for early-stage contamination assessment.

Purpose of the Study:

• To develop and validate a novel quantitative detection method for aflatoxin B1 (AFB1) in peanuts.
• To leverage quantitative remote sensing principles and deep learning for sub-pixel AFB1 content analysis.
• To assess the performance of modified deep learning models, including ResNet18, for accurate AFB1 quantification.

Main Methods:

• Subpixel decomposition techniques, including endmember extraction and nonnegative matrix decomposition, were employed to determine AFB1 content at the sub-pixel level.
• Modified transfer learning models (LeNet5, AlexNet, VGG16, ResNet18) were adapted into a deep learning regression network for quantitative AFB1 detection.
• Hyperspectral data (415-799 nm) from 67,178 training pixels and 67,164 testing pixels were utilized for model development and validation.

Main Results:

• The modified ResNet18 deep learning model demonstrated superior performance in AFB1 detection, achieving a coefficient of determination (R²) of 0.8898.
• The ResNet18 model yielded a Root Mean Square Error (RMSE) of 0.0138 and a Ratio of Performance to Inter-Quartile Range (RPD) of 2.8851, indicating high accuracy and robustness.
• The developed sub-pixel model and deep learning regression approach proved effective for quantitative AFB1 detection.

Conclusions:

• The proposed sub-pixel analysis combined with deep learning offers a robust and accurate method for quantitative AFB1 detection in peanuts.
• The modified ResNet18 model exhibits high predictive ability and robustness, making it suitable for practical applications.
• This innovative approach provides a new technological scheme for designing automated sorting machines, enhancing food safety inspection capabilities.