柑桔叶片可溶性糖近红外检测非线性模型研究  

Study on NIR detection non-linear model of soluble sugar in citrus leaves

作　　者：刘燕德[1] 肖怀春[1] 韩如冰[1] 孙旭东[1] 朱丹宁[1] 曾体伟 李泽敏[1] 

机构地区：[1]华东交通大学机电工程学院,江西南昌330013

出　　处：《广东农业科学》2016年第11期43-49,共7页Guangdong Agricultural Sciences

基　　金：国家"863"计划项目(SS2012AA101306);江西省科技支撑计划项目(20121BBF60054);南方山地果园智能化管理技术与装备2011协同创新中心(赣教高字[2014]60号);江西省优势科技创新团队(20153BCB24002)

摘　　要：为了监督柑桔叶片是否缺乏营养元素,对叶片可溶性糖进行分析。采用近红外光谱技术结合误差反馈神经网络(BPNN)和最小二乘支持向量机(LS-SVM)建立定量剖析非线性模型,运用主成分分析(PCA)进行数据压缩、无信息变量消除算法(UVE)和连续投影算法(SPA)进行有效波段筛选的方法来优化模型的输入变量,提高了模型检测精度。同时,利用Savitzke-Golay平滑(S-G)、多元散色校正(MSC)、导数和基线校正(Baseline)等预处理方法进行数据变换,来确定最佳建模方法。结果表明:波长筛选能优化模型,并提高运算速度,其中PCA优化效果最为明显,可溶性糖的相关系数Rp达到最大为0.91,均方根误差RMSEP最小为4.82,显著提高了模型的检测精度和稳健性,经过优化的输入变量所建模型,能够满足定量检测的要求,具有一定的可行性。In order to supervise the nutrional elements of citrus leaves, the soluble sugars in the leaves of citrus were analyzed. Combined with back propagation neural network （ BPNN ） and least squares support vector machine （ LS-SVM ）, quantitative analysis of the nonlinear model using near infrared spectroscopy was developed, at the same time, data were compressed using principal component analysis （ PCA ）, the effective wavelength bands were screened by Uninformative variable elimination （ UVE ） algorithm and Successive projections algorithm （ SPA ） . These methods were adopted to optimize the input variables of the model, which improved the detection accuracy. And spectra processing methods included Savitzke-Golay smoothing （ S-G ）, multiple scatter correction （ MSC ）, derivative and baseline correction （ Baseline ） and the combinations of these methods for data transformation, the best method for establishing models was determined. The MSC was adopted to eliminate baseline drift and amplify characteristic information, meanwhile amplify high frequency noise, which can be eliminated by 2th derivative. And smoothing was adopted to eliminate the interference noise and to make the spectrum smoother. It was concluded that the processing method was the best. The results showed that wavelength selection played an important role in optimization model, and improved the speed of computation. The effect of model optimization by the model PCA was most obvious and the maximum of correlation coefficient （ Rp ） of soluble sugar reached 0.91, the minimum of the root mean square error of prediction （ RMSEP ） reached 4.82. The results showed that the model accuracy and robustness were significantly improved, the prediction model could meet the requirement of quantitative detection after optimizing the input variables. Therefore, the prediction model has certain feasibility.

关 键 词：可溶性糖 近红外光谱技术 波段筛选 优化 预处理方法 

分 类 号：S682.264[农业科学—观赏园艺]