玉米的单倍体鉴别技术对玉米单倍体育种技术非常重要。 近红外光谱分析技术以其操作简便, 可在线分析监测, 速度快, 无损, 测试成本低等特点在农业领域备受关注, 应用广泛。 实验首先通过美国JDSU公司的近红外光谱仪采集由国家玉米改良中心提供的玉米种子单倍体、 多倍体的近红外光谱数据, 然后对获得的原始数据做平滑(smoothing)、 一阶导(first derivative, FD)和矢量归一化(vector normalization, VN)预处理以消除其噪声影响, 再采用核函数为高斯核函数(Gaussian kernel function)的核主成分分析(kernel principal components analysis, KPCA)的方法将玉米种子的近红外光谱数据映射到高维空间中, 并对映射后的数据做非线性特征提取, 然后应用支持向量机(support vector machines, SVM)对提取的玉米种子单倍体、 多倍体光谱数据的非线性特征建立分类模型, 最后输入测试数据进行玉米单倍体、 多倍体的分类鉴别测试, 预测玉米种子是否是单倍体。 设计了两组对比试验, 其正确识别率的平均值分别达到95%和9357%。 在该实验中, 基于KPCA的玉米单倍体识别算法的性能表现较好、 识别率较高。 通过两组对比实验, 证明了玉米种子的近红外光谱数据更适于先将其映射于高维空间中进行特征提取, 再对提取的特征进行建模、 分类分析。 该实验为玉米单倍体识别技术提供了新的思路和方法。

The method of maize haploid identification plays a significant role in advancing Maize Haploid Breeding Technology. Given to its advantages of cost-effectiveness, high performance and being easy to operate, near-infrared spectroscopy (NIRS) has drawn great attention in the field of agricultural research. At the beginning of the experiment, NIRS data of both haploid and polyploidy maize seeds that are provided by National Maize Improvement Center of China are collected via US JDSU’s near-infrared spectrometer. After pre-processing that the original data were processed by smoothing, first derivative, and vector normalization in order to eliminate the influence of noise, the NIRS data is subsequently mapped in a high-dimensional space, where nonlinear feature extraction take place through Kernel Principal Components Analysis (KPCA) whose kernel function is Gaussian kernel. Then Vector Machines Support (SVM) was used to establish the classification model for it. Finally, based on the result of classified identification test, the experiment makes a prediction of whether maize seeds are haploid. In particular, this paper designs two sets of comparative experiments with average recognition rate being 95% and 9357%. The result indicates that the method based on KPCA is both feasible and valid. The above experiment proves that the process of “high-dimensional spatial mapping—nonlinear feature extraction—modeling classification analysis” is more suitable for studying maize seeds data collected via NIRS. Therefore, this paper may provide some new idea and method for Maize Haploid Identification technology.