近红外光谱的玉米种子穗腐病特征提取与判别模型研究

玉米种子穗腐病是危害玉米产量的主要病害之一。利用近红外光谱开展了玉米种子穗腐病判别模型研究。246粒玉米种子由吉林省农业科学院海南育种基地提供，其中96粒玉米种子为穗腐病染病样本，其他150粒玉米种子为同种玉米正常样本。利用MATRIX-Ⅰ型傅里叶近红外光谱仪采集了样本800～2 500 nm范围的近红外光谱信息，并对样本近红外光谱数据利用多元散射校正（MSC）进行预处理。结合玉米内部有机物质的近红外光谱的敏感波段和样本近红外光谱吸收峰挑选了4个优选区间，并采用相关系数法(CA)、连续投影算法（SPA）和竞争性自适应重加权算法（CARS）三种不同原理的特征波长提取算法分别提取了4（1 362，1 760，2 143和2 311 nm）、5（1 227，1 310，1 382，1 450和1 728 nm）和10（1 232，1 233，1 257，1 279，1 313，1 688，1 703，1 705，2 302和2 323 nm）个特征波长。以提取得到的特征波长作为玉米种子穗腐病判别模型输入变量，用0-1（染病-正常）表示样本染病状况作为输出真实值建立支持向量机（SVM）模型，使用网格搜索法结合十折交叉验证法对模型参数进行优化。结果表明，CA-SVM，SPA-SVM和CARS-SVM三种判别模型中训练集和测试集建模准确率均在90%以上。该研究成果为玉米种子病害诊断装置提供了模型基础，且针对优选区间进行特征波长选择的方式也可以为建立其他种子病害判别模型提供参考。

Characteristic Extraction Method and Discriminant Model of Ear Rot of Maize Seed Base on NIR Spectra

Ear rot of corn seeds is one of the main diseases that harm the yield of corn. A discriminant model of ear rot of corn seeds was studied by near-infrared spectroscopy. The study samples were provided by the Hainan Breeding Base of Jilin Academy of Agricultural Sciences. 246 corn seeds were selected as the research objects, 96 of which were infected with ear rot, and the other 150 were normal samples of the same kind of corn. A Matrix-Ⅰ Fourier NIR spectrometer was used to collect the NIR spectra of the samples in the range of 800~2 500 nm, and the NIR spectra were preprocessed by Multiplicative Scatter Correction (MSC). Four optimal regions were selected combined with the sensitive band of NIR spectrum of organic matter in maize and the absorption peak of the NIR spectrum of samples. Correlation analysis (CA), successive projections algorithm, SPA) and Competitive Adaptive Reweighted Sampling (Competitive Adaptive Reweighted Sampling, Cars), 4 (1 362, 1 760, 2 143 and 2 311 nm), 5 (1 227, 1 310, 1 382, 1 450 nm) were extracted by three characteristic wavelength extraction algorithms with different principles, respectively 1 728 nm) and 10 (1 232, 1 233, 1 257, 1 279, 1 313, 1 688, 1 703, 1 705, 2 302 and 2 323 nm).The characteristic wavelengths extracted were used as input variables of the corn seed ear rot identification model. The disease status of samples was represented by 0-1 (infected normal) as the output true value to establish the support vector machine (SVM) model. The model parameters were optimized by the grid search method and the 10-fold cross-validation method. The results show that the modeling accuracy of the training and test set in three discriminant models, CA-SVM, SPA-SVM and CARS-SVM, is above 90%. The research results in this paper provide a model basis for the maize seed disease diagnosis device. The method of selecting characteristic wavelengths for the optimal region can also provide a reference for establishing other seed disease discrimination models.