Rapid determination of potential mycotoxin-producing fungi contamination in post-harvest blueberries based on near infra-red spectroscopy and enhanced prototypical network

Abstract

Blueberries are susceptible to fungal infection that may lead to mycotoxin contamination. Accurate identification of mycotoxin-producing fungi in blueberries is critical, as processed products derived from mould-contaminated fruits pose significant health risks due to the persistence of mycotoxins. However, classifying fungal diseases in blueberries remains challenging because infections beneath the dark-coloured epidermis are often visually undetectable. Near-infra-red (NIR) spectroscopy presents a promising alternative, as it exhibits distinct absorption characteristics for different fungal pathogens and enables non-contact measurement. This study investigates the use of NIR spectra to classify diseases caused by three mycotoxin-producing fungi: Alternaria alternata, Fusarium fujikuroi, and Penicillium expansum. Blueberry samples were individually inoculated with one of these pathogens or maintained as non-inoculated controls. NIR spectra collected from fruits across two independent harvests were analysed 1–6 days post-inoculation. To address the classification challenge with limited data, this study proposes an enhanced prototypical network to classify blueberry fungal infections, training on the first harvest and validating on the second harvest. A customised convolutional neural network with spatial attention improves feature extraction, while data normalisation and MetaMix augmentation enhance model performance. Through stratified cross-validation, the five-class classifier with 18-shot prototypes achieved an average accuracy of 97.9%, sensitivity of 97.9%, specificity of 99.5%, precision of 98.1% and F1-score of 97.9%. These findings highlight the potential of NIR spectroscopy combined with prototype-based learning for early detection of mycotoxin risks in blueberries.