High-performance prediction of protein content in brown rice via multi-spectral fusion and deep learning: a comparative study of visible, near infrared, mid infrared spectroscopy and hyperspectral imaging

Abstract

This original work explored the potential of near-infrared (NIR), mid-infrared (MIR) spectroscopy, and hyperspectral imaging (HSI) in visible-near-infrared (Vis-NIR-HSI) and short-wave infrared (SWIR-HSI) range for non-destructive prediction and visualization of protein content in brown rice from 138 rice varieties. Feature selection and predictive modeling were integrated to identify protein-related wavelengths and enable pixel-level protein mapping. Using full-spectrum data, the MIR-based convolutional neural network (CNN) model achieved the highest predictive accuracy (R_p² = 0.96, RPD = 4.84), confirming the superior chemical sensitivity of MIR spectroscopy. Following feature band selection, the NIR-based uninformative variable elimination–support vector machine (UVE–SVM) model exhibited optimal performance among wavelength-reduced models (R_p² = 0.90, RPD = 3.16), surpassing Vis-NIR-HSI and SWIR-HSI-based approaches. For multi-spectral feature fusion integrating selected wavelengths from all four spectral modes, the competitive adaptive reweighted sampling–CNN (CARS–CNN) model yielded the best overall performance, demonstrating synergistic advantages of combining complementary spectral information. Furthermore, protein distribution maps from SWIR hyperspectral images displayed superior spatial resolution compared to Vis-NIR-HSI images. Overall, this study establishes a high-performance and transferable protein prediction framework based on multi-spectral fusion and deep learning, and provides a systematic comparison of visible, near-infrared, mid-infrared spectroscopy, and hyperspectral imaging for brown rice quality assessment.