应用太赫兹时域光谱反射成像技术结合广义二维相关光谱法探索玉米种子活力敏感太赫兹波段, 并结合支持向量机建立快速无损判别种子活力的分析模型。 实验以中地77玉米种子为例, 采用人工老化方式（40 ℃, 100%相对湿度）将种子样本分批老化0, 1, 2, 3, 4天制备不同活力的种子样本, 并按照GB/T 3543.4—1995进行种子发芽实验; 同时采用Terapluse 4000太赫兹时域光谱仪及反射成像附件采集上述不同老化程度种子样本的太赫兹光谱图像。 由于玉米种子的胚乳和种胚的成分差异显著, 为探究种子不同组织在老化过程中与活力的相关性, 本实验首先采用双高斯滤波器对THz图像中的像素点光谱消噪、 峰峰值差分重构图像增强以及阈值分割等预处理无损提取玉米种子不同组织太赫兹吸光度谱。 然后以老化天数为扰动量, 针对上述提取的样本胚乳和种胚光谱分别作广义二维相关分析, 根据实验中同步谱和异步谱中自动峰与交叉峰位置初步解析, 可得到与种子活力关系密切的THz波段主要集中在75和36 cm-1区域, 同时75和36 cm-1处的光谱信息存在强烈的协同变化且变化方向一致。 种子活力与老化天数密切相关, 因此根据老化天数分别建立了基于胚乳和种胚吸光度谱的五分类支持向量机模型用于种子活力定性判别, 但是其判别准确率仅为59.34%和71.28%, 表明该模型无法精细划分种子五个活力等级; 实验进一步根据GB4401.1—2008以玉米种子发芽率85%为阈值划分活力高低等级, 建立二分类种子活力判别模型, 可得胚乳和种胚测试集识别准确率分别可达88.61%和91.73%, 模型性能显著提升, 增强了THz技术用于种子活力无损粗筛的可行性。 实验结果表明: 太赫兹反射成像技术以其丰富的指纹谱、 低能安全以及图谱合一等特性, 有望成为单粒种子活力快速无损测定领域一项崭新、 有力的补充技术。

Sensitive terahertz bands related corn seed vigor were explored using terahertz time domain spectral reflection imaging technique combined with generalized two-dimensional correlation spectroscopy, and simultaneously the qualitative model to judge seed vigor nondestructively was established based on support vector machine and terahertz absorbance spectra. Take Zhongdi 77 (Corn variety) for example in this experiment. Firstly, there are 5 batches of different vigor seeds made by artificial aging (40 ℃, 100% relative humidity) for 0, 1, 2, 3, 4 days. 5 batches of seeds were conducted germination experiments according to GB/T 3543.4—1995 to obtain germination rate. Terahertz spectral images of seed samples were collected by Terapluse 4 000 terahertz time domain system with reflection imaging module. Because the composition of endosperm and embryo of corn seed is significantly different, it is meaningful to explore the correlation between the different tissues (endosperm and embryo) and the vigor in the aging process separately. A series of pretreatments was done to extract the terahertz absorbance spectra of different tissue of corn seed, such as denoising using double Gaussian filtering, image enhancement based on the peak-to-peak differential reconstruction and threshold segmentation. Take the aging days as the disturbance amount, and the generalized two-dimensional correlation spectrum analysis was carried out to the above-mentioned extracted spectra of endosperm and embryo. According to the preliminary analysis of the automatic peak and the cross peak in the synchronous and asynchronous spectra, the terahertz band closely related to seed vigor was mainly concentrated in the 75 and 36 cm-1 regions, and the spectral information at 75 and 36 cm-1 had great synergistic change and the change was consistent. Different aging days correspond to different vitality according to the germination experiment. Therefore, five-class support vector machines were built respectively based on endosperm and embryo absorbance spectra to identify seed vigor, and the recognition rate was only 59.34% and 71.28%. The results indicated that the model cannot precisely divide seed vigor to five levels. According to GB4401.1—2008, 85% germination rate was set as the threshold to divide seed vigor to two levels, then binary classifier based on support vector machines were built to distinguish seed vigor. The recognition accuracy of the endosperm and embryo test sets was 88.61% and 91.73% respectively. The recognition rate had improved significantly and the model can basically be used for rapid coarse screening of seed vigor. The experiment result showed terahertz reflection imaging with its rich fingerprint spectrum characteristics and lower energy is expected to be a new and powerful complementary technology to identify seed vigor rapidly and nondestructively.