用于昆虫分类鉴定的人工神经网络方法研究:主成分分析与数学建模  

Research on Artificial Neural Network Method Used for Insects Classification and Identification:Principal Component Analysis and Mathematical Modeling

作　　者：蔡小娜[1,2] 黄大庄[1] 沈佐锐[2] 高灵旺[2] 

机构地区：[1]河北农业大学,河北保定071000 [2]中国农业大学IPMist实验室,北京100193

出　　处：《生物数学学报》2013年第1期23-33,共11页Journal of Biomathematics

基　　金：河北省自然科学基金项目(C2012204008)

摘　　要：为探讨人工神经网络(ANN)在昆虫分类上的可行性,本文提出利用主成分分析和数学建模等方法相结合改进ANN,并以鳞翅目夜蛾科6种蛾类昆虫为样本进行了验证.首先利用Bugshape1.0特征提取软件获取6种蛾180个右前翅样本的13项数学形态特征数据,再运用主成分分析对蛾翅数学形态特征变量重新组合生成新的综合变量,然后结合主成分分析建立BP神经网络分类器.主成分分析结果表明,前5个主成分的累积贡献率为85.52%,已基本包含了全部特征变量具有的信息.在主成分分析的基础上,建立具有5个输入层节点,12个隐含层节点和1个输出层节点的三层BP神经网络分类器.每种蛾20个样本共120组特征数据对分类器进行训练和仿真,其余60组特征数据对分类器进行验证,仿真输出值与目标值的相关系数R=0.997,分类正确率达到了93.33%.较之未经过主成分分析而单独使用BP神经网络建立的分类器,基于主成分分析的BP神经网络分类器具有更优的性能和更准确的分类能力.研究结果表明本文提出的方法具有很好的分类和鉴别作用,为蛾种类的鉴别提供了一种可行的方法.To study the feasibility of artificial neural network （ANN） on insect categorization it was brought forth that principal component analysis （PCA） and mathematical modelingwere combined to improve ANN in this paper, and six kinds of moths which belonged to Noctuidae, Lepidopter were used as samples to confirm the method. On the first step,13 math-morphological characters＇ （MMC） data was extracted form right front wings of 6 kinds of moths which using the character extracting software Bugshapel.0 in this paper. On the second step, PCA was applied to make new comprehensive variables from MMCs variables of moth wrings. The next step was to create BP neural network classifier with the combination of PCA. The results of PCA indicated that the first 5 principal components accumulated contribution rates to 85.52%, which meant all information got by character variables was included. On the base of PCA, a three-layer BP neural network classifier with 5 input nodes, 12 hidden nodes and 1 output node was set up. Totally 120 sets of character data from 20 samples each kind of moth were used to train and simulate the classifier. The other 60 sets of data were applied on verification. The results showed that relation coefficient R between simulated output values and destination ones was 0.977 and classification right rate was high up to 93.3%. Compared with the classifier built with BP neural network only and no PCA, the BP neural network classifier based on PCA had better performance and more accurate classification ability. The study results indicated that the method put forward in this paper has good effect on classification and discrimination and therefore a feasible way for moth species identification was provided.

关 键 词：蛾翅(夜蛾科) 数学形态特征 主成分分析 BP神经网络 分类鉴定 

分 类 号：Q96[生物学—昆虫学]