基于GA-ANN融合算法的棉田杂草特征降维及分类识别  

Feature Dimension Reduction and Category Identification of Weeds in Cotton Field Based on GA-ANN Complex Algorithm

作　　者：王淑芬 杨玲香 WANG Shufen;YANG Lingxiang(College of Sciences,Shihezi University,Shihezi 832003,China)

机构地区：[1]石河子大学理学院,新疆石河子832003

出　　处：《河南农业科学》2018年第2期148-154,160,共8页Journal of Henan Agricultural Sciences

基　　金：国家自然科学基金(21267020);石河子大学应用基础青年项目(2014ZRKXYQ07)

摘　　要：准确识别杂草是智能化除草技术所要解决的关键问题,为了实现棉花田间杂草的特征降维及分类识别,提出了遗传算法(GA)融合人工神经网络(ANN)的算法。试验中共采集棉花和杂草样本195个,提取棉花和杂草的形状特征、4个方向灰度共生矩阵纹理特征、HSV空间颜色特征等21个参数。将21个特征参数按照一定顺序组成码串作为遗传个体,融合神经网络模式识别算法,以实现特征参数的有效降维。对利用降维后的优良特征参数组合、全部特征参数以及主成分分析方法(PCA)降维识别的准确率进行了对比,结果表明:利用融合算法降维得到的不同特征组合,可将特征参数维数保持在8~13维,有效降低了特征参数空间的复杂度;融合算法平均分类准确率稳定在98%左右,明显优于PCA分析法。对降维后的优良特征参数组合进行自组织特征映射网络训练(SOM),可视化拓扑结构图表明,降维后的优良特征组合对各个类别的影响呈现出独立性、可区分性的显著特点,宽长比、H三阶矩特征与棉花样本的分类准确率呈强相关性,H一阶矩、S三阶矩对苘麻、龙葵草、灰菜、田旋花样本的分类影响显著,而对棉花样本的分类准确率影响较弱。The key of intelligentized weeding is to identify weeds accurately.In this paper,genetic algorithm-artificial neural network(GA-ANN)complex algorithm was proposed for feature dimensional reduction,classification and identification of weeds.A total of 195 cotton and weed samples were collected and 21 parameters,related to the shape feature,the texture feature,and the HSV spacial color feature,were extracted.The 21 parameters were coded as genetic individuals,and then the dimension of the parameters was reduced by the complex algorithm.The classification accuracies of weeds were calculated using the feature parameters reduced by GA-ANN complex algorithm,all the feature parameters and the parameters reduced by PCA,respectively,and compared.The results showed that the classification accuracy(about 98%)of the complex algorithm was significantly better than that of PCA.At the same time,the number of the feature parameters were reduced to 8-13 and the complexity of the feature parameters was reduced effectively.The reduced feature parameters were network-trained by self-organizing map(SOM),and the visual topology map showed that the effect of the reduced feature parameters on the classification demonstrated independence and distinguishability.The correlation analysis results showed that there was a strong correlation between the ratio of width to length,H third-order and the classification accuracy of cotton.There was a significant influence of H first-order,S third-order on the classification of Abutilon,Solanum nigrum,Chenopodium album,and Convolvulus arvensis,while no significant influence on cotton.

关 键 词：棉花 杂草 遗传算法 人工神经网络 自组织特征映射网络 降维分析 分类识别 

分 类 号：S451.0[农业科学—植物保护]