Simultaneous quantification of rice amylose and protein content with an optimized convolutional neural network model via near-infrared spectroscopy

Rice amylose and protein contents are pivotal quality indicators influencing rice physicochemical properties. Near-infrared (NIR) spectroscopy offers rapid, non-destructive analysis, yet conventional convolutional neural network (CNN) models often suffer from overfitting and limited consideration of global spectral features, hindering simultaneous multi-target prediction. In this study, we introduce an advanced Convolutional Neural Network Regression (CNNR) model for the simultaneous determination of amylose and protein content. Calibration models of partial least squares regression (PLSR) and support vector machine regression (SVMR) were first optimized through spectral preprocessing and wavenumber selection. While PLSR achieved superior performance in amylose prediction, CNNR significantly outperformed PLSR in protein prediction. To further enhance robustness, a deep CNNR model integrated with data augmentation (DA) and a convolutional block attention module (CBAM) was developed. This enhanced model achieved high predictive accuracy, with coefficients of determination (R P ²) of 0.972 for amylose and 0.992 for protein. In addition, a user-friendly application, RiceQuant-NIR, was developed to enable efficient data upload, prediction, and visualization, facilitating practical large-scale quality evaluation. These findings demonstrate that the DA + CBAM + CNNR model provides a robust and precise approach for simultaneous rice quality assessment, advancing the efficiency of food analysis research. • Developed DA+CBAM+CNNR model for rice amylose and protein prediction. • Achieved high accuracy: R²=0.972 (amylose), R²=0.992 (protein). • Built RiceQuant-NIR app for rapid, user-friendly rice quality analysis.