摘要
研究了一种快速、精确、无损检测桑叶农药残留的方法。以不含农药残留的桑叶、含有敌敌畏残留的桑叶、含有毒死蜱残留的桑叶、含有乙酰甲胺磷残留的桑叶、含有乐果残留的桑叶和含有辛硫磷残留的桑叶为实验对象,利用高光谱成像仪获取390-1 050 nm范围内的桑叶高光谱图像。利用ENVI软件确定叶片的感兴趣区域,并采用连续投影算法(SPA)优选出10个特征波长(452.51、469.88、517.28、539.85、578.92、643.72、727.24、758.34、785.67、819.67 nm)。利用基于径向基内核(RBF)的支持向量机(SVM)和10折交叉验证的方法建立桑叶农残检测模型,并讨论了3种参数寻优算法(网格搜索、遗传算法和粒子群算法)对模型性能的影响,发现采用网格搜索的SVM模型的性能最优,其交叉验证正确率为63.89%,预测正确率为78.33%。为了进一步提升模型的分类性能,将自适应提升算法(Adaboost)引入到SVM建模方法,基于特征波长下的光谱数据,对桑叶是否含有农药残留及农药残留品种进行分类建模。结果表明,Ada-SVM模型的预测准确率达到97.78%,较传统SVM模型的准确率提高了19.45个百分点。可见,利用高光谱图像技术结合Ada-SVM算法能够较准确地鉴别桑叶农药残留。
A non-destructive testing method was studied to rapidly and accurately detect pesticide residues on mulberry leaves. Six groups of mulberry leaves were chosen as experimental samples,which contained pesticide residues of dichlorvos,chlorpyrifos,acephate,dimethoate and phoxim as the first to fifth groups, respectively, and the sixth group without pesticide residues was taken as control.Hyperspectral images of samples in 390 - 1 050 nm were acquired by hyperspectral imaging devices. The region of interest from hyperspectral image was selected,and ten characteristic wavelengths,which were452. 51,469. 88,517. 28,539. 85,578. 92,643. 72,727. 24,758. 34,785. 67 and 819. 67 nm,were selected by the successive projections algorithm( SPA). Based on RBF kernel function of SVM and 10 fold cross-validation methods, the detection models of pesticide residues on mulberry leaves were established. The impacts of three parameter optimization algorithms( grid search,genetic algorithm and particle swarm optimization) on the model performance were discussed. The results showed that performance of SVM model by using grid search was the optimal one,and its cross-validation accuracywas 63. 89% and forecast accuracy was 78. 33%. In order to further enhance the classification performance of the model,the adaptive algorithm( Adaboost) was introduced into the SVM model,and Ada-SVM algorithm was used to build classification model,which can detect pesticide residues on mulberry leaves and identify the kinds of pesticide residues. The results showed that the prediction accuracy of Ada-SVM model reached 97. 78%,which was increased by 19. 45% compared with the original SVM model. Therefore,hyperspectral imaging technology combined with Ada-SVM algorithm can accurately identify the pesticide residues on mulberry leaves.
出处
《农业机械学报》
EI
CAS
CSCD
北大核心
2015年第6期251-256,共6页
Transactions of the Chinese Society for Agricultural Machinery
基金
国家自然科学基金资助项目(31471413)
江苏高校优势学科建设工程资助项目(苏政办发2011 6号)
江苏大学现代农业装备与技术重点实验室开放基金资助项目(NZ201306)
中国博士后科学基金资助项目(2014M561594)
江苏省博士后科研资助计划资助项目(1401175C)
关键词
桑叶
农药残留
支持向量机
高光谱图像
特征波长
Mulberry leaves
Pesticide residues
Support vector machine
Hyperspectral imaging
Characteristic wavelength