[Objective]To identify the gender of Bombyx mori chrysalis in the cocoon by rapid and non-destructive method based on near infrared transmission spectroscopy,improve breeding efficiency and reduce labor cost.[Method]We used four silkworm varieties including Fu 9, 9 Fu, Xiang 7 and 7 Xiang, and compared their diffuse transmission spectra between 450-950 nm and 900-1700 nm. Partial least squares discrimination analysis (PLSDA), back propagation neural network (BPNN) and support vector machine (SVM) discrimination models were established and compared among different varieties. The robustness of the models was studied through the receiver operating characteristic(ROC) curve. Characteristic wavelengths were extracted by difference method and genetic algorithm.[Result]The identification accuracy rates for Fu 9, 9 Fu, Xiang 7 and 7 Xiang varieties were 95.20％, 95.65％, 88.80％ and 87.50％ respectively using 450-950 nm spectra, and were 100％, 96.00％, 92.22％ and 94.21％ respectively using 900-1700 nm spectra. Using PLSDA, BPNN and SVM models resulted in good identification of male and female silkworm pupae, the true female rates were 95.96％, 95.83％ and 100％, the true male rates were 98.98％, 96.04％ and 82.18％, and the accuracy rates were 97.46％, 95.94％ and 90.86％, respectively. Based on the analysis of ROC curve, the PLSDA model was the optimal, followed by the BPNN model. Twenty bands were extracted manually as the equipment input, and the true female rate, true male rate and accuracy rate were 93.75％, 95.45％ and 94.57％ respectively based on the PLSDA model.[Conclusion]Diffuse transmission spectra in the near infrared (900-1700 nm) contains more classification information of male and female pupae compared with the visible-near infrared (450-950 nm). The PLSDA model is the optimal one among three models. After extracting the characristic bands, the accuracy rate can meet the requirements of actual production.%[目的]利用近红外漫透射光谱快速无损鉴别家蚕Bombyx mori种茧茧壳内蚕蛹的雌雄,以提高育种效率、降低人工成本.[方法]以芙9、9芙、湘7和7湘4个蚕品种为研究对象,采集比较了样本在可见和近红外区间的漫透射光谱,建立比较了各品种偏最小二乘判别分析(PLSDA)、后向传播神经网络(BPNN)以及支持向量机分类(SVM)判别模型,通过分类器特性(ROC)曲线研究了各模型的鲁棒性,采用差值法和遗传算法提取了特征波长.[结果]芙9、9芙、湘7和7湘品种利用450～900 nm光谱建模的雌雄鉴别准确率分别为95.20％、95.65％、88.80％和87.50％,利用900～1700 nm光谱建模的准确率分别为100％、96.00％、92.22％和94.21％;采用PLSDA、BPNN和SVM模型都能够对蚕蛹雌雄做出较好的无损鉴别,3种模型真雌性率分别为95.96％、95.83％和100％,真雄性率分别为98.98％、96.04％和82.18％,准确率分别为97.46％、95.94％和90.86％,进一步通过ROC曲线分析,PLSDA模型效果最优,BPNN模型次之;手动提取20个波段建立PLSDA模型,鉴别真雌性率为93.75％,真雄性率为95.45％,准确率为94.57％.[结论]近红外波段900～1700 nm的漫透射光谱比可见–近红外波段450～950 nm含有更丰富的蚕蛹雌雄分类信息;3种鉴别模型中,PLSDA模型效果最优;提取特征波段后,准确率能达到生产需要.
展开▼
