Multi-Scale Feature Attention Network for Rapid and Non-Destructive Quantification of Aflatoxin B₁ in Maize Using...

Area of Science:

• Agricultural Science
• Analytical Chemistry
• Computer Science

Background:

• Maize is a vital global crop vulnerable to aflatoxin contamination, necessitating precise detection methods.
• Aflatoxin B1 (AFB1) poses a significant health risk, making its accurate quantification in maize essential for food safety.

Purpose of the Study:

• To develop and evaluate a deep learning model (MSFNet-ECA) for accurate AFB1 quantification in maize using near-infrared hyperspectral imaging.
• To investigate the impact of different data augmentation techniques and preprocessing strategies on model performance.

Main Methods:

• Near-infrared hyperspectral imaging combined with a deep learning Multi-Scale Feature Network with Efficient Channel Attention (MSFNet-ECA) model.
• Comparison of data augmentation methods: multiplicative random scaling, bootstrap resampling, and Wasserstein generative adversarial networks (WGAN).
• Evaluation of preprocessing strategies, including second derivative (D2), in conjunction with augmentation techniques.

Main Results:

• The MSFNet-ECA model, enhanced with multiplicative random scaling and D2 preprocessing, achieved superior predictive performance.
• Achieved a root mean square error of prediction (RMSEP) of 2.3 μg·kg⁻¹, coefficient of determination for prediction (Rp2) of 0.99, and residual predictive deviation (RPD) of 9.
• Demonstrated significant accuracy improvements over conventional models (PLSR, SVR, ELM, 1D-CNN), ranging from 42.5% to 86.4%.

Conclusions:

• Data augmentation techniques substantially enhance the predictive capabilities of deep learning-based hyperspectral chemometric models.
• The proposed MSFNet-ECA model, coupled with data augmentation, offers an efficient and reliable tool for real-time AFB1 detection in maize.
• This approach supports improved food quality and safety monitoring in hyperspectral applications.