Instrument-invariant near-infrared spectroscopy via domain adaptation for robust quantitative analysis of barley protein content

Abstract

Near-infrared (NIR) spectroscopy has emerged as a rapid and non-destructive tool for food quality assessment, yet variations among instruments remain a major barrier to its widespread deployment in grain protein prediction. This study introduces a joint domain adaptation–spectral reconstruction–prediction model (JDA-SR-PM) to achieve reliable instrument transfer for online monitoring of protein content in barley kernels. Spectral data were collected from a master and a slave instrument, and the dataset was expanded via unsupervised augmentation. The model was optimized using the Adam algorithm and benchmarked against three state-of-the-art transfer learning methods: TCASVR, PRETL, and PEAMATL. Results demonstrated that JDA-SR-PM achieved superior predictive performance, with R² = 0.923, RMSE = 0.553, and RPD = 3.607 on the test set, outperforming competing approaches. Bland–Altman analysis further confirmed close agreement between predicted and reference protein values, validating the potential of JDA-SR-PM as an alternative to commercial analyzers. SHAP-based feature attribution highlighted critical spectral regions (920 nm, 1200 nm, and 1400–1550 nm), consistent with known chemical overtones, and provided a foundation for simplified instrument design. These findings suggest that NIR spectroscopy combined with domain adaptation can extend rapid protein monitoring beyond laboratory settings, supporting practical applications in grain quality control.