高丹草中粗蛋白质以及碳水化合物的含量丰富, 适合青贮处理。 优质的高丹草种子是发展畜牧业十分重要的前提, 发芽率是检验种子质量最常规的指标之一, 播前种子发芽率检测与筛选十分必要。 现阶段采用发芽试验法进行种子发芽率的检测, 周期长、 成本高。 基于此, 提出利用近红外光谱对高丹草种子进行发芽率的快速、 无损检测。 选择适量的高丹草种子样品, 采集近红外漫反射光谱, 进行一阶导和二阶导预处理以及对比分析$R^{2}_{c},R^{2}_{p}$, RESEC和RMSEP。 采用支持向量机(SVM)建模, 使用MATLAB中调用的LIBSVM软件包来实现SVM训练和检测过程, 以检测不同发芽率的高丹草种子。 对来自不同省份的100组高丹草种子先剔除种子内的杂物、 破损以及不能满足试验条件的种子后, 用人工气候培养箱进行种子发芽试验, 获得100组种子样本的发芽率, 其发芽率分布在41%~64%的范围。 采用美国Unity Scientific 2600XT近红外光谱仪对样本进行光谱扫描。 随机分成校正集70份和检验集30份。 分别采用一阶导和二阶导进行了高丹草种子光谱的预处理, 将预处理之后的数据采用支持向量机的方法建模, 并对其参数进行了分析和讨论。 结果表明, 近红外光谱预测模型训练集相关系数($R^{2}_{c}$)和测试集相关系数($R^{2}_{p}$)分别为0.94和0.92, 校正均方根误差(RMSEC)、 预测均方根误差(RMSEP)分别为0.21和0.25, 两个产地的高丹草种子数据采用一阶导预处理时模型最优。 支持向量机的方法建模采用Rbf核函数, 当支持向量机惩罚因子 c=2 896.309 4和核函数 g=0.5时, 测试集种子发芽率的检测准确率为96.666 7%(29/30)。 该模型预测种子发芽率是可行的, 可以作为初步检测高丹草种子发芽率快速无损检测的手段之一, 能够有效的促进种子生产。

Sorghum-Sudan Grass is rich in crude protein and carbohydrate, suitable for silage treatment. High-quality seeds are a prerequisite for animal husbandry development, and germination rate is one of the most conventional indicators to test the seed quality. Therefore, testing and screening the germination rate of the seeds prior to sowing is essential. The germination test method is currently used to detect seed germination rate, which has a long cycle and high cost. In this study, a rapid and non-destructive method based on NIR was proposed to detect the germination rate of sorghum-sudangrass seeds. The near-infrared diffuse reflectance spectra of the seed samples were collected with 1-Der and 2-Der processing. Moreover, comparative analysis of the parameter values obtained for $R^{2}_{c},R^{2}_{p}$, RESEC and RMSEP was also performed. The support vector machine (SVM) was used for modeling, and the LIBSVM software package in Matlab was used to realize the SVM training and detection process to detect the seeds of sorghum-sudangrass seed with different germination rates. Using the Unity scientific 2600 XT Near-infrared spectrometer, 100 groups of sorghum-sudangrass seeds from different provinces were selected as samples. Before the experiment, the broken seeds and seeds that did not germinate were removed, and the germination test was carried out in the incubator. The germination rate of 100 samples was obtained, and the germination rate ranged from 41% to 64%. The seed samples were spectroscopically scanned and were randomly divided into calibration set (70 samples) and test set (30 samples). In this paper, the 1-Der and 2-Der method was used to preprocess the spectrum of sorghum-sudangrass seeds. SVM modeled the preprocessed data, and its parameters were analyzed. The results showed that the correlation coefficients of the training set ($R^{2}_{c}$) and test set ($R^{2}_{p}$) were 0.94 and 0.92 respectively, and the root mean square error of correction (RMSEC) and root mean square error of prediction (RMSEP) were 0.21 and 0.25 respectively, which reflected that the model was the best when the 1-Der was used to preprocess the seed data. When c=2 896.309 4, g=0.5, the detection accuracy of the test set was 96.666 7% (29/30) by using Rbf core functions of SVM modeling. These results suggested that the model was feasible to predict the seed germination rate, and could be used as one of the rapid and non-destructive detection methods for the preliminary detection of seed germination rate of sorghum-sudangrass could effectively promote the seed production.