FT-NIR Spectroscopy and Explainable Machine Learning for Eggshell Based Classification of Organic and Conventional Market Eggs

Abstract The premium pricing of organic eggs creates incentives for product misrepresentation and motivates the development of rapid, non-destructive screening tools to differentiate labeled organic and conventional eggs. Fourier transform near-infrared (FT-NIR) spectroscopy coupled with machine learning was evaluated for classifying retail eggs marketed as organic versus conventional. A total of 147 eggs were measured and spectra were analyzed using partial least squares–discriminant analysis (PLS-DA), along with gradient-boosted tree models (XGBoost and LightGBM). The dataset was partitioned into training (70%) and testing (30%) sets. Among the evaluated models, PLS-DA achieved the best test performance with an overall accuracy of 91.0%. To improve interpretability, SHapley Additive exPlanations (SHAP) were used to identify wavelengths that most strongly contributed to model predictions. Influential regions included bands near ~1400 nm, consistent with O-H absorptions commonly associated with moisture-related spectral features, and regions near ~2300 nm, which are often associated with crystalline calcium-carbonate-related absorptions in the near-infrared and may reflect variation in mineral-associated spectral response of eggshell. These results indicate that FT-NIR spectroscopy combined with machine learning can differentiate retail eggs labeled as organic versus conventional under the conditions of this study, supporting further development of rapid screening approaches for market classification and quality assurance applications.