Non-destructive quantification of egg yolk ratio using visible–near-infrared hyperspectral imaging, machine learning and explainable AI

BACKGROUND: Accurate, non-destructive quantification of egg yolk ratio holds considerable significance for the food industry, nutritional assessment and egg grading. Conventional approaches are limited by destructive testing and insufficient throughput for commercial applications. This study investigates the potential of visible–near-infrared (Vis-NIR; 374–1015 nm) hyperspectral imaging (HSI) combined with machine learning (ML) and explainable artificial intelligence (AI) techniques for rapid and non-destructive yolk ratio prediction. Multiple regression models, spectral preprocessing and feature selection techniques were comprehensively evaluated to develop robust and interpretable predictive solutions. RESULTS: Regression models including partial least squares regression (PLSR), random forest, extreme gradient boosting and support vector regression were assessed for yolk ratio prediction. The PLSR model combined with Savitzky–Golay first-derivative spectral preprocessing demonstrated superior and stable predictive performance, achieving coefficients of determination (R²) of 0.79, 0.73 and 0.68 in calibration, validation and independent test datasets, respectively. Additionally, a simplified PLSR model using a few important variables selected based on regression coefficients achieved robust predictive results. Shapley additive explanations analysis provided clear insights into the wavelength regions significantly contributing to model predictions, primarily linked to water, lipid and protein contents in eggs. CONCLUSION: This research highlights the effectiveness of Vis-NIR HSI integrated with ML and explainable AI as a rapid, reliable and non-destructive approach for egg yolk ratio assessment. The developed method offers significant advantages for egg quality monitoring, providing practical, interpretable and scalable solutions beneficial for food industry applications.