80. Non-destructive Prediction of Egg Yolk Content Using Hyperspectral Imaging and Machine Learning.

Abstract Nondestructive assessment of internal egg traits could improve quality evaluation for both table eggs and hatching eggs in precision poultry systems. This study evaluated the use of hyperspectral imaging and machine learning to predict yolk content in chicken eggs and to identify spectral regions that contributed most to model performance. A dataset of 1,228 eggs, including both fertile and infertile eggs, was analyzed. Hyperspectral images were acquired in the visible to near-infrared range (400 to 1,000 nm), and spectra were processed using standard normal variate, multiplicative scatter correction, first-derivative Savitzky-Golay, and second-derivative Savitzky-Golay methods. Regression models included partial least squares regression (PLSR), XGBoost, LightGBM, and CatBoost. The dataset was split into training (70%), validation (20%), and testing (10%) subsets. Hyperparameters for the tree-based models were optimized with Optuna. Model interpretation was performed using SHapley Additive exPlanations (SHAP) to identify influential wavelengths. Among all tested approaches, PLSR with first-derivative preprocessing achieved the best performance on the testing set, with R2 = 0.63, RMSEP = 1.27, and RPD = 1.64. Among the nonlinear models, XGBoost performed best, with R2 = 0.58, RMSEP = 1.44, and RPD = 1.45, while LightGBM and CatBoost showed lower predictive performance. SHAP analysis indicated that wavelengths around 740 to 765 nm contributed strongly to prediction, suggesting that this spectral region contains useful information for estimating yolk content. These results showed that hyperspectral imaging combined with regression modeling can provide a moderate, nondestructive estimate of yolk content in chicken eggs. In this dataset, the PLSR model outperformed the tested nonlinear approaches, indicating that simpler chemometric methods may remain competitive for this application. The identified wavelength regions may support future efforts to simplify sensing systems and improve model efficiency. Additional work should assess whether integrating spectral data with egg physical traits or image-based deep learning methods can improve predictive accuracy and generalizability across egg types and production contexts. For image description, please refer to the figure legend and surrounding text. For image description, please refer to the figure legend and surrounding text.