SVM自助重加权采样的蚕茧雌雄特征波长选择  

Feature Selection Algorithm for Identification of Male and Female Cocoons Based on SVM Bootstrapping Re-Weighted Sampling

作　　者：陈楚汉 钟杨生[2] 王先燕[3] 赵懿琨 代芬 CHEN Chu-han;ZHONG Yang-sheng;WANG Xian-yan;ZHAO Yi-kun;DAI Fen(College of Electronic Engineering,South China Agricultural University,Guangzhou 510642,China;College of Animal Science,South China Agricultural University,Guangzhou 510642,China;Guangdong Sericulture Technology Promotion Center,Guangzhou 510640,China)

机构地区：[1]华南农业大学电子工程学院,广东广州510642 [2]华南农业大学动物科学学院,广东广州510642 [3]广东省蚕业技术推广中心,广东广州510640

出　　处：《光谱学与光谱分析》2022年第4期1173-1178,共6页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金项目(61675003);广州市科技计划项目(201707010346);广州市科技计划项目(202103000090)资助;广东省现代农业产业技术体系创新团队项目(2019KJ124)。

摘　　要：使用近红外光谱鉴别蚕茧雌雄设备成本较高,挑选有用特征可以减少成本。雌雄蚕茧的近红外光谱存在着共线性的关系,因此提出了一种包裹式的特征选择方法,基于支持向量机的自助重加权采样(BRS-SVM)的特征选择方法。使用NirQuest512近红外光谱仪采集了蚕茧的漫透射近红外光谱。用试验集的全波段建模得到特征重要度热图,并通过热图得到重要特征波段的范围。然后在重要特征波段范围内,分别用BRS-SVM、基于SVM的特征排序方法(MBR-SVM)、基于逻辑回归的特征排序方法(MBR-LR)、递归特征消除法(RFE)、连续投影算法(SPA)和遗传算法(GA)挑选单波段特征和连续波段面积特征,再分别用支持向量机(SVM)和逻辑回归(LR)建立雌雄分类模型。通过特征重要性热力图发现,蚕茧雌雄分类重要区域在900~1 399 nm内,用此波段范围建立SVM模型,试验集准确率为99.40%。用BRS-SVM挑选5个单波段特征,然后再用SVM建模,验证集准确率为93.88%,高出其他特征选择方法5%~12%,测试集准确率为89.56%,测试集准确率高出其他特征选择方法2%~4%。用BRS-SVM挑选27个单波段特征,建立SVM雌雄分类模型测试集准确率为94.97%,准确率达到生产条件要求。用BRS-SVM挑选的14个连续波段面积特征,再用SVM建模,测试集准确率为94.43%。在挑选少量特征情况下,我们提出的BRS-SVM要优于其他方法。用BRS-SVM挑选少量的特征,可以建立性能良好的蚕茧雌雄分类模型,有效减少了成本,具有重要的现实意义。The cost of identifying male and female cocoons by NIR is high,and the cost can be reduced by selecting useful features.Since there is a nonlinear relationship between the NIR spectra of female and male cocoons,a wrapper feature selection method,Bootstrapping Re-weighted Sampling Support Vector Machines(BRS-SVM),was proposed.The diffuse transmission NIR spectra of silkworm cocoons were collected by NirQuest512 NIR spectrometer.The heat map of characteristic importance was obtained by modeling the whole band of the test set,and the heat map obtained the range of important characteristic bands.Then,in the range of important characteristic bands,the single band features and continuous band area features were selected by BRS-SVM,Model-based ranking support vector machines(MBR-SVM),Model-based ranking Logistic Regression feature sorting method(MBR-LR),Recursive feature elimination(RFE),successive projections algorithm(SPA),Genetic Algorithm(GA),and then the support vector machines(SVM) and Logistic Regression(LR) sex classification models were established respectively.According to the characteristic importance heat map,it is found that the important area of male and female classification of silkworm cocoon was within 900~1 399 nm.We used this band to build the SVM model,and achieved 99.40% accuracy.BRS-SVM was used to select 5 single-band features.The accuracy of the test set is 89.56%,which is 2%~4% higher than other feature selection methods.RS-SVM was used to select 27 single-band features,and the accuracy of the test set of the SVM gender classification model was 94.97%,which reached the requirements of production conditions.The accuracy of modeling test set by BRS-SVM was 94.43% for 14 continuous band features.In the case of selecting a small number of features,our proposed BRS-SVM is superior to other methods.Using BRS-SVM to select a small number of features,we can establish a good performance of the female and male cocoon classification model,effectively reduce the cost,has important practical significance.

关 键 词：蚕茧 近红外光谱 特征选择 

分 类 号：G307[文化科学]