连续偏振光谱的稻种发芽率检测方法研究

针对目前传统稻种发芽率检测方法周期长、精度低的问题，提出新颖的基于连续偏振光谱技术实现稻种发芽率快速、无损检测的方法。以不同老化天数稻种为检测目标，10 min为检测时间点，使用起偏器将光纤准直光源调制成线偏振光垂直入射稻种浸出液，而后以5°为间隔旋转检偏器，并通过光纤光谱仪检测透射的光谱，对检测的偏振光谱通过归一化预处理后，根据不同发芽率稻种检测时偏振角及波长的贡献给出特征偏振角和特征波长，特征偏振角为0°，5°和25°，特征波长为576，620和788 nm，将获取的连续偏振光谱以特征偏振角和特征波长处的透射率为输入，构建稻种发芽率检测模型。分别比较运用偏最小二乘法回归(partial least squares regression，PLSR)、BP神经网络(back propagation neural network，BPNN)、径向基神经网络(radial basis function neural network，RBFNN)三种建模方法建立稻种发芽率检测模型。分别用老化天数为0，2，4，6 d的稻种，在不同的偏振角共测量1 520组实验数据，其中912组数据作为校正集，608组数据作为预测集，建模结果表明三种模型预测精度较高，其中RBFNN模型预测精度最高，其相关系数r为0.976，均方误差RMSE为0.785，平均相对误差MRE为0.85%。表明利用连续偏振光谱技术通过多维度光谱信息能够有效实现稻种发芽率的快速、准确检测。

Study on the Prediction of Germination Rate of Rice Seeds with Continuous Polarization Spectroscopy

With respect to the problem of long period and low precision in using traditional methods to predict rice seeds germina‐tion rate ,a novel method based on continuous polarization spectroscopy was proposed to achieve rapid and nondestructive predic‐tion .The paper set different aging rice seeds as prediction targets and ten minutes as prediction time ,using polarizer to modulate optical fiber collimating light source to linearly polarized light which issuing into rice seeds extract vertically before rotating the analyser every 5 degrees .The transmission spectrum was predicted through the optical fiber spectrometer .After normalization pretreatment to the polarization spectrum ,the article gave the characteristics of polarization angel and wavelength by 0 degree ,5 degrees ,25 degrees ,620 ,788 and 576 nm according to the contribution of polarization angel and wavelength when predicting different germination rate rice seeds and inputted obtained continuous polarization spectrum by wavelength ,polarization angel , transmissivity to construct rice seeds germination rate prediction model using three modeling methods to build rice seeds germina‐tion rate prediction model in comparison ,including Partial Least Squares Regression (PLSR) ,Back Propagation Neural Network (BPNN) and Radial Basis Function Neural Network (RBFNN) .1 520 sets of experimental data were measured in total at differ‐ent polarization angels through using rice seeds with different aging days (0 ,2 ,4 ,6) respectively ,setting 912 sets of data as calibration set and 608 sets of data as predicion set .The modeling results show that RBF model’s prediction accuracy is the high‐est .Its correlation coefficient is 0.976 ;the mean square is 0.785 ;and the average relative error is 0.85% .The research results show that the continuous polarization spectroscopy technique through multidimension spectral information can achieve rapid and accurate prediction of rice seeds germination rate .