基于竞争性自适应重加权采样和二维卷积神经网络的小麦水分含量的定量测定  

Determination of Moisture Content in Wheat Using Competitive Adaptive Reweighted Sampling and Two-Dimensional Convolutional Neural Networks

作　　者：曾小松 宦克为[1] 曹献文 金明杭 ZENG Xiaosong;HUAN Kewei;CAO Xianwen;JIN Minghang(School of Physics,Changchun University of Science and Technology,Changchun 130022)

机构地区：[1]长春理工大学物理学院,长春130022

出　　处：《长春理工大学学报（自然科学版）》2024年第6期36-44,共9页Journal of Changchun University of Science and Technology(Natural Science Edition)

基　　金：吉林省科技发展计划项目(20240404046ZP)。

摘　　要：本研究利用近红外光谱(NIRS)系统获取了小麦的光谱,然后提出了一种基于竞争性自适应重加权(CARS)与格拉姆角差场(GADF)的二维卷积神经网络(CARS-GADF-2DCNN)模型。CARS-GADF-2DCNN利用CARS方法选择全光谱中波长的最佳集合,随后利用GADF将选择结果编码为二维图像,最后使用二维卷积神经网络学习图像特征,完成了小麦水分的定量分析,并将该模型的预测结果与其他模型进行了比较。结果表明,与1DCNN、GADF-2DCNN、VCPA-GADF-2DCNN和IRIV-GADF-2DCNN相比,CARS-GADF-2DCNN的预测精度分别提高了68.8%、45.6%、20.2%和17.5%。综上,CARS-GADF-2DCNN解决了NIRS建模时预测准确度低和过拟合的问题。本研究为小麦的水分含量测定提供了一种准确快速的方法。A near infrared spectroscopy(NIRS)system was utilized to obtain wheat spectra,and then a two-dimensional convolutional neural network model(CARS-GADF-2DCNN)based on competitive adaptive reweighted sampling(CARS)and gramian angular difference field(GADF)was developed in this study.The CARS-GADF-2DCNN model employed CARS to identify characteristic wavelengths from the NIR spectra,converted the NIR spectra into two-dimensional images using GADF,and finally employed the 2DCNN to learn image features for quantitative analysis of wheat moisture content.The predictive performance of this model was evaluated in comparison with other models.The results demonstrated that CARS-GADF-2DCNN model improved the prediction accuracy for wheat moisture content by 68.8%,45.6%,20.2%,and 17.5%compared to 1DCNN,GADF-2DCNN,VCPA-GADF-2DCNN,and IRIV-GADF-2DCNN,respectively.In summary,the CARS-GADF-2DCNN alleviated the issues of low prediction accuracy and overfitting in NIRS modeling.This study provides an accurate and rapid method for determining wheat moisture content.

关 键 词：近红外光谱分析 小麦水分 变量选择 格拉姆角差场 卷积神经网络 

分 类 号：O657.33[理学—分析化学]