基于随机森林的单粒玉米种子水分近红外快速定量检测

Rapid and Non-destructive Determination of Moisture Content of Single Maize Seed by Near Infrared Spectroscopy Based on Random Forest

玉米精量播种技术发展对单粒种子质量检测提出更高的检测要求,本研究重点探索了近红外光谱结合化学计量学方法建立单粒玉米种子水分检测模型的可行性。收集并测定了110份玉米样本的含水量,应用傅里叶变换红外光谱仪及单粒测样附件扫描得到样本集近红外光谱,利用SPXY法以3∶1比例划分训练集和测试集。采用多种光谱预处理方法消除单粒种子采集光谱时由于颗粒形态等引起的噪声干扰,再采用主成分分析(PCA)、去噪自动编码器(DAE)进行降维和特征提取,建立基于随机森林(RF)的单粒玉米种子含水量预测模型。实验结果表明,相对其他预处理方法而言,多元散射校正处理后建立的单粒种子水分模型性能较好,其训练集的R为0.9862,RMSEC为0.1414;测试集的R为0.9689,RMSEP为0.4457。DAE相较于PCA光谱特征提取效果更好,其训练集的R为0.9885,RMSEC为0.17531;测试集的R为0.9824,RMSEP为0.4206。研究结果表明,经过光谱预处理并结合光谱降维消噪后,基于RF的模型可以有效降低单粒玉米种子近红外光谱采集时引入的非线性干扰,有助于提升单粒玉米种子水分近红外快速无损检测实际应用可行性。

The development of maize precision sowing technology puts forward higher detection requirements for single seed quality detection.The present paper focused on the feasibility of establishing single seed moisture detection model by near infrared spectroscopy combined with chemometrics.The moisture content of 110 corn samples was collected and measured.The near infrared spectrum of the sample set was obtained by using Fourier transform infrared spectrometer and single grain test accessories.The training set and test set were divided by SPXY method with a ratio of 3∶1.A variety of spectral preprocessing methods were used to eliminate the noise interference caused by the particle shape when collecting the spectrum of single seed.Principal component analysis(PCA)and de-noising automatic encoder(DAE)were used for dimensionality reduction and feature extraction,and the prediction model of single seed moisture content based on random forest(RF)was established.The experimental results indicated that,compared with other pretreatment methods,the performance of single seed moisture model established by multiple scattering correction was better,the R of training set was 0.9862,rmsee was 0.1414;the R of test set was 0.9689,RMSEP was 0.4457.Compared with PCA,the R of training set was 0.9885,rmsee was 0.17531;the R of test set was 0.9824,RMSEP was 0.4206.The results indicated that the model based on RF can effectively reduce the non-linear interference in the near-infrared spectrum acquisition of single corn seed after spectral preprocessing combined with spectral dimensionality reduction and denoising,which was helpful to improve the practical application feasibility of rapid near-infrared nondestructive detection of single corn seed moisture.