Non-destructive detection of mould in Fructus Aurantii via Vis/NIR spectroscopy coupled with machine and deep learning models

Abstract Fructus Aurantii (FA) is widely used as a food-medicine homologous substance, but it is vulnerable to mould contamination during storage and transport. Rapid and non-destructive detection of mould is critical to ensure safety and preserve bioactive constituents. In this study, visible and near-infrared spectra were collected from FA samples classified as normal and mouldy. Spectral preprocessing methods, including Savitzky–Golay smoothing, multiplicative scatter correction, and standard normal variate, were applied to reduce noise and scattering effects. Two analytical strategies were evaluated. The first was traditional machine learning, which combined principal component analysis, linear discriminant analysis, and classifiers including support vector machine, XGBoost (eXtreme Gradient Boosting), and random forest. The second was deep learning, which employed a one-dimensional convolutional neural network (1D-CNN) and multilayer perceptron (MLP) directly on the preprocessed spectra. Results indicated that both pipelines could effectively discriminate mouldy from normal FA, with the 1D-CNN achieving the highest accuracy more than 98%, highlighting its superior feature extraction capability under limited-band conditions. This integrated framework provides a rapid, non-destructive, and scalable approach for FA mould detection. It offers practical value for quality control in production and storage.