Deep learning-based multimodal fusion for quality prediction of chili paste using hyperspectral imaging and...

A deep learning-based intelligent multimodal system was developed to non-destructively evaluate chili paste quality by fusing color features extracted from hyperspectral images acquired by Hyperspectral Imaging (HSI), spectral features derived from HSI and Near-Infrared Spectroscopy as well as physicochemical indicators. Among the five preprocessing methods, Multiplicative Scatter Correction performed best in the data preprocessing. Uninformative Variable Elimination exhibited the best performance to identify wavelengths significantly correlated with color value, capsaicin, dihydrocapsaicin and volatiles among the three feature extraction methods. After feature-level fusion, Mixup augmentation was employed to expand the dataset from 160 to 800 samples to alleviate overfitting in deep learning. Convolutional Neural Network-Long Short-Term Memory hybrid model demonstrated excellent prediction performance in all the quality parameters (test set R² = 0.9554-0.9826) among the six predictive models. This study established the first non-destructive and rapid detection framework for chili paste fermentation, providing technical support for intelligent real-time quality monitoring.