A general deep learning model for predicting and classifying pea protein content via visible and near-infrared spectroscopy

Abstract

Rapid and accurate detection of pea protein content is crucial for breeding and ensuring food quality. This study introduces the PeaNet model, which employs an improved convolutional neural network architecture to predict and classify pea protein content. The model was developed using 156 visible and near-infrared spectral datasets from 52 varieties cultivated under varied conditions. The data were preprocessed with Savitzky-Golay smoothing and multiplicative scatter correction to improve quality. The results revealed that the model achieved an R² of 0.84 for predicting protein content and a classification accuracy of 85.33 % on the test set. On an independent validation set comprising different pea varieties, the model maintained an R² above 0.80 and a classification accuracy of 83.33 %. It significantly outperformed traditional machine learning models and conventional deep learning architectures. This study introduces a universal, accurate, and efficient method for detecting pea protein content, thereby advancing food nutrition assessment and quality control.