Abstract The traditional candling methods are very hard to discriminate the infertile chicken eggs from the fertile ones before hatching on the poultry farm. The purpose of this study was to detect infertile chicken eggs before hatching by use of the nondestructive technique of visible near infrared transmission spectroscopy. The Naïve-Bayes method was adopted to link spectral data and qualitative observation values of chicken eggs to build a model. Five preprocessing methods were applied to spectral data to improve the robustness of the models. The internal test method of leave one cross validation and an external test method were employed to test the predictive ability of the established models. Naïve-Bayes with a second derivative by the Savitzky-Golay preprocessing method was selected as the optimal prediction model by comparing with four evaluation parameters of other models. The overall accuracy of the model using the second derivative—Savitzky-Golay preprocessing method was 91.67%, while the sensitivity and the specificity were 95% and 100%, respectively. The prediction ability of the optimal model was assessed by use of unknown new egg samples. The average overall prediction accuracy of the prediction set containing 111 eggs was 89.12%, while the sensitivity and specificity were 88.14 and 90.38%, respectively. Results showed that the prediction model had good predictability and stability. Therefore, visible near infrared transmission spectroscopy would be an appropriate technology to detect infertile chicken eggs before hatching. Highlights The infertile chicken eggs from the fertile ones before hatching were discriminated Visible near infrared spectroscopy first detected infertile chicken eggs before hatching. Naive bayes classifier with five different pre-processing methods were compared. The unknown eggs were predicted by the optimal Naive bayes prediction model.

中文翻译：

---

朴素贝叶斯法结合可见光近红外透射光谱预测孵化前不育鸡蛋

摘要 家禽养殖场孵化前，传统的烛光检测方法很难区分不育鸡蛋和可育鸡蛋。本研究的目的是利用可见近红外透射光谱的无损技术检测孵化前的不育鸡蛋。采用朴素贝叶斯方法将鸡蛋的光谱数据和定性观测值联系起来建立模型。对光谱数据应用了五种预处理方法，以提高模型的鲁棒性。采用留一交叉验证的内部测试方法和外部测试方法来测试所建立模型的预测能力。通过与其他模型的四个评估参数进行比较，选择具有Savitzky-Golay预处理方法的二阶导数的朴素贝叶斯作为最优预测模型。使用二阶导数Savitzky-Golay预处理方法的模型总体准确率为91.67%，敏感性和特异性分别为95%和100%。最优模型的预测能力是通过使用未知的新鸡蛋样本来评估的。包含 111 个鸡蛋的预测集的平均总体预测准确率为 89.12%，而敏感性和特异性分别为 88.14 和 90.38%。结果表明，该预测模型具有良好的可预测性和稳定性。因此，可见近红外透射光谱将是一种在孵化前检测不育鸡蛋的合适技术。亮点 区分孵化前的不育鸡蛋和孵化前的不育鸡蛋 可见近红外光谱首先检测孵化前的不育鸡蛋。对具有五种不同预处理方法的朴素贝叶斯分类器进行了比较。未知鸡蛋通过最优朴素贝叶斯预测模型进行预测。