该研究采用偏最小二乘法建立鸡胸肌冻干粉赖氨酸近红外定量预测模型，并通过光谱预处理、建模特征光谱筛选、异常样本剔除、建模和验证样本选择4个方面对模型进行优化，旨在提高模型的预测精度和稳健性。以263个鸡胸肌冻干粉为研究对象，研究7种不同光谱预处理方法、4种特征光谱筛选方法、MCCV异常样本剔除方法，SPXY和鸡种2种建模验证样本选取方法对鸡胸肌冻干粉赖氨酸近红外定量预测模型的影响。结果表明：在1000~2502 nm建模谱区，剔除68个异常样本后，使用SPXY方法选取156个校正样本、39个外部验证样本，使用原始光谱所建模型最优；没有进行异常样本剔除、建模特征谱区选择等处理时，1000~2502 nm谱段建模，光谱需进行SNV+gapsegment (1#,15,7)预处理；异常样本剔除和建模样本选择对建模谱段1386~1379 nm、1329~1323 nm、1289~1283 nm、1276~1258 nm、1240~1235 nm、1194~1184 nm、1173~1168 nm、1142~1137 nm、1103~1099 nm、1080~1076 nm、1058~1054 nm、1012~1009 nm、1004~1001 nm（540个光谱），使用原始光谱所建模型精度影响最小。研究显示，鸡肉赖氨酸近红外定量预测模型的建立受异常样本和建模样本、验证样本的影响较大，建模特征光谱对模型的适用性影响较大，光谱预处理方法仅在全谱段不做其它处理时对建模精度影响较大。

In this study, anear infrared (NIR) quantitative prediction model for the lysine in the freeze-dried chicken breast muscle powder was developed using the partial least squares method. The model was optimized through four aspects, spectrum preprocessing, model characteristic spectrum screening, outlier samples elimination, and sample selection for modeling and verification, in order to improve the prediction accuracy and robustness of the model. Taking a total of 263 freeze-dried chicken breast muscle powders as the research object, the effects of seven different spectral preprocessing methods, four characteristic spectrum screening methods, MCCV outlier samples elimination methods, and two samples selection methods for modeling and verification, SPXY and chicken breeds, on the NIR quantitative prediction model for the lysine of in freeze-dried chicken breast muscle powder. The results showed that (1) the best model was established on the basis of 156 calibration samples and 39 external verification samples after the elimination of 68 outlier samples from the original spectra in the 1000~2502 nm; (2) The spectral region in the 1000~2502 nm used for modeling should be pretreated by SNV+gapsegment (1#,15,7) method when there was no outlier spectrum or model characteristic spectral region for selection; (3) Least impact on the accuracy of the model established using the original spectra, was exerted by outlier sample elimination and samples selection for modeling in 1386~1379 nm, 1329~1323 nm, 1289~1283 nm, 1276~1258 nm, 1240~1235 nm, 1194~1184 nm, 1173~1168 nm、1142~1137 nm, 1103~1099 nm, 1080~1076 nm, 1058~1054 nm, 1012~1009 nm, 1004~1001 nm. Research showed that the establishment of a near-infrared quantitative prediction model for the lysine of in freeze-dried chicken breast muscle powder is greatly affected by outlier samples, samples for modeling and external sample selection for verification. The applicability of the model is greatly affected by characteristic spectra regions for modeling. The preprocessing of the whole spectrum without other pretreatments has a greater impact on the accuracy of model.

国家自然科学基金项目（31760487）；云南省重大科技专项（2016ZA008）；云南省现代农业禽蛋产业技术体系项目（2017KJTX0017）