柑桔叶片黄龙病光谱特征选择及检测模型  

Spectral feature selection and discriminant model building for citrus leaf Huanglongbing

作　　者：刘燕德[1] 肖怀春[1] 孙旭东[1] 朱丹宁[1] 韩如冰[1] 叶灵玉 王均刚[1] 马奎荣 

机构地区：[1]华东交通大学机电工程学院,南昌330013

出　　处：《农业工程学报》2018年第3期180-187,共8页Transactions of the Chinese Society of Agricultural Engineering

基　　金：国家自然科学基金(61640417;31760344);"十二五"国家863计划课题(SS2012AA101306);江西省优势科技创新团队建设计划项目(20153BCB24002;南方山地果园智能化管理技术与装备协同创新中心(赣教高字[2014]60号);江西省研究生创新资金项目(YC2015-S238)

摘　　要：为探索高光谱技术诊断黄龙病及分类的可行性,通过变量筛选方法组合为高维数据实用化提供参考。采集柑桔叶片高光谱图像并进行普通(polymerase chain reaction,PCR)鉴别分为轻度、中度、重度、缺锌和正常5类样品。用无信息变量消除算法(uninformative variable elimination,UVE)剔除无关信息,组合遗传算法(genetic algorithm,GA)和连续投影算法(successive projections algorithm,SPA)筛选变量,对数据进行降维。结合极限学习机(extreme learning machine,ELM)和最小二乘支持向量机(least squares support vector machine,LS-SVM)构建柑桔黄龙病判别模型。对预测样品进行诊断分类,来评价模型判别能力。经对比发现,UVE组合SPA筛选变量后的LS-SVM模型效果最好,该模型以Link_kernel函数为核函数,惩罚因子(γ)最小为1.07,误判率最低为0。用全谱作输入变量时LS-SVM模型复杂程度最高且预测能力最差,误判率最高为11.9%,可能是包含无用信息和冗余信息变量造成的。研究显示,UVE组合SPA筛选变量,结合LS-SVM对柑桔黄龙病诊断并分类具有一定可行性,为高维度数据实用化提供一定参考价值。Citrus greening is a devastating disease of citrus fruit trees, and at present, it is potential for greening diagnosis by hyperspectral imaging technique. The purpose of this paper is to explore the feasibility of diagnosis and classification of greening using hyperspectral technique, and provide the reference for practical application of high-dimensional data. The hyperspectral images of citrus leaves were collected and divided into 5 types： slight greening, moderate greening, serious greening, nutrient deficiency and normal by common PCR（polymerase chain reaction）. The samples of normal, nutrient deficiency, slight, moderate and serious greening show bright band in turn and the bright band colors are getting brighter with the grade of the disease. The bright band of nutrient deficiency samples by PCR is more vague than the greening samples, which may be related to the lack of nutrient elements. A total of 169 samples are divided into the calibration and prediction set for calibrating the models and accessing their performance respectively according to the proportion of 3：1. The calibration set includes 25 slight citrus greening samples, 21 moderate citrus greening samples, 26 serious citrus greening samples, 26 nutrient deficiency samples and 29 normal samples. The prediction set includes 6 slight citrus greening samples, 13 moderate citrus greening samples, 6 serious citrus greening samples, 10 nutrient deficiency samples and 7 normal samples. From the representative pictures of the 5 kinds of leaves, it can be seen intuitively that the leaves of greening present similar symptoms with the nutrient deficiency leaves, which are obviously different from the normal leaves. But it is difficult to distinguish between the leaves with slight, moderate, serious greening and nutrient deficiency. The average spectrum of hyperspectral images of leaves is extracted in the region of interest. The results show that strong reflection peak of chlorophyll is located at 550 nm, and greening hinders plant photosynthesis,

关 键 词：作物 病害 模型 高光谱成像技术 连续投影算法 最小二乘支持向量机 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]