Explainable AI in Near-Infrared Spectroscopy: A Case Study of Rice Protein Content

Accurate determination of rice protein is essential for quality control. This study evaluated machine learning models (partial least squares regression, PLSR; support vector machine, SVM) combined with feature selection algorithms (random frog, RF; competitive adaptive reweighted sampling, CARS; Monte-Carlo uninformative wavelength elimination, MCUVE) and explainable artificial intelligence method (SHapley Additive exPlanations, SHAP) analysis to predict protein content. Results indicated that CARS-PLSR achieved the highest accuracy (RMSEP = 0.266, R²P = 0.976, residual prediction deviation [RPD] = 6.612). Statistical analysis (F-test and t-test) verified the model's reliability. Furthermore, SHAP analysis revealed that wavelengths at 1218, 1688, and 1209 nm made the highest contributions, corresponding to N-H and C-H vibrations. This study not only provides a rapid detection method but also elucidates the chemical basis of the model's high predictive performance. PRACTICAL APPLICATIONS: This study provides a high-speed, non-destructive method for accurately measuring rice protein content, allowing food processors to monitor grain quality in real-time. By using "explainable" AI to reveal the specific chemical markers driving the results, this technology offers a transparent and reliable alternative to slow, expensive laboratory testing.