Objective To address the low efficiency of manual screening and the high cost of existing detection methods for infertile eggs and dead-embryo eggs during goose egg incubation, this study proposed a machine learning-based ternary classification method (dead-embryo eggs, fertilized eggs, infertile eggs) for detecting goose egg viability.

Method A customized machine vision system was constructed to acquire candling images of 504 goose eggs from day 3 to day 9 of incubation. A total of 37 image features were extracted, including RGB and HSI dual-model color features, as well as gray-level co-occurrence matrix and Tamura texture features. Through analysis of variance and LightGBM-based feature importance evaluation, combined with Spearman correlation analysis, 25 image features were finally selected. A voting ensemble learning model was then built with support vector machines, random forests, and gradient boosting machines as base learners using a soft voting mechanism, and the model decision logic was interpreted by the SHAP algorithm.

Result The ensemble model achieved an overall accuracy of 77.23% and an F1 score of 77.50%, outperforming each single base classifier. After 6 days of incubation, the classification accuracy stabilized at 85.29–87.76%, and after 7 days, the precision for unfertile eggs reached 100%.

Conclusion This method is low-cost and well suited to practical production workflows, providing an efficient and reliable technical solution for large-scale screening of hatching eggs in conservation hatcheries and small-scale hatcheries.