基于THz-ATR光谱的玉米种子水分定量模型优化方法研究

应用太赫兹时域光谱技术结合区间偏最小二乘法筛选玉米种子水分THz特征波段，并采用支持向量机构建基于特征谱区的抗非线性干扰的种子水分快速定量分析模型。实验以郑单958玉米种子为例，制备含水量范围9.58%~12.71%的种子粉末样本40组(每组取样3份)，采用衰减全反射(ATR)附件扫描得到120份样本太赫兹时域光谱，根据SPXY(光谱-理化值共生距离算法)法划分得到训练集样本90份，测试集样本30份。种子水分对太赫兹波具有强烈吸收，首先采用基于偏最小二乘线性回归的移动区间（mwPLS）、独立区间(iPLS)、后向区间(biPLS)和联合区间(siPLS)方法筛选最优特征谱区组合；鉴于环境水分、种子其他成分及系统噪声对种子水分太赫兹光谱存在不可避免的非线性干扰，在上述光谱特征区间进一步采用基于RBF核函数的支持向量机和网格搜索法构建得到预测性能最优的种子水分快速定量分析非线性模型，训练集均方根误差为0.021 2，预测集均方根误差为0.069 7，相对分析误差为12.345 7，相较于传统偏最小二乘线性回归模型，模型性能得到提升。种子水分含量是影响种子贮藏安全和种子活力的重要因素，实验结果表明：太赫兹时域光谱结合化学计量学可以有效筛选种子水分特征吸收谱区，建立抗干扰、高精度的种子水分快速定量分析模型，有望成为未来种子质量快速测定领域一项极具应用潜力的补充技术。

Study on the Optimization Method of Maize Seed Moisture Quantification Model Based on THz-ATR Spectroscopy

Characteristic Terahertz(THz) bands of maize seed moisture were screened using the Terahertz time-domain spectroscopy technique combined with the interval partial least squares method. The support vector machine was used to construct a rapid quantitative analysis model of seed moisture based onthe characteristic spectral region against nonlinear interference. Take Zhengdan 958(Corn variety), for example, in this experiment, 40 sets of seed powder samples (3 samples from each set) with moisture content ranging from 9.58% to 12.71% were prepared. Terahertz time-domain spectra of 120 samples were collected by Terapluse 4 000 terahertz time-domain system with Attenuated Total Reflection (ATR) module. According to the SPXY method, 90 training set samples and 30 test set samples were obtained. Given the strong absorption of terahertz waves by seed moisture, the moving interval (mwPLS), independent interval (iPLS), backward interval (biPLS) and synergy interval (siPLS) methods based on partial least squares linear regression were firstly used to screen the optimal combination of the characteristic spectral regions. In view of the inevitable nonlinear interference of environmental moisture, other seed components and systematic noise on the terahertz spectrum of seed moisture, a nonlinear model for rapid quantitative analysis of seed moisture with optimal prediction performance was further constructed using support vector machine and grid search method based on RBF kernel function on the above spectral feature intervals. The optimal SVR model was obtained with a lower root mean square error of the training set (RMSEC) of 0.021 2, a lower root mean square error of the prediction (RMSEP) of 0.069 7 and a higher residual predictive deviation (RPD) of 12.345 7.The model performance was significantly improved compared with the traditional partial least squares linear regression model. Seed moisture content is an important factor in seed storage safety and seed vigour.The experimental results show that THz time-domain spectroscopy combined with the chemometric method can effectively be used to screen the characteristic absorption spectral region of seed moisture and establish an interference-resistant and high-precision model for rapid quantitative analysis of seed moisture, which is expected to be a Promising complementary technology in the field of rapid seed quality determination.