基于CNN和特征选择回归方法的小麦蛋白质含量测定

Determination of Protein Content in Wheat Based on CNN and Feature Selection Regression Methods

小麦蛋白质含量一般采用凯式定氮法测定,该方法操作流程复杂,分析时间长,无法实现样品的批量检测。小麦的近红外光谱特征与蛋白质含量间存在线性和非线性的映射关系,本研究提出一种具有线性和非线性映射能力的算法:基于卷积神经网络和特征选择回归的组合算法。使用多种预处理方法提高Wheat kernels近红外光谱数据信噪比,将一维的预处理光谱数据折叠成二维矩阵,使用二维卷积神经网络模型对小麦蛋白质含量进行预测,并提取中间层部分神经元输出的特征信息与预处理光谱数据集数据构成集成数据集,在集成数据集上使用套索、最小最大凹罚回归(MCP)和光滑切片绝对偏差回归(SCAD)方法构建小麦蛋白质含量测定模型,与多元线性回归、偏最小二乘回归等模型进行对比分析。实验结果表明:卷积神经网络的引入使得模型具有非线性映射能力,改善了小麦蛋白质含量预测模型性能。

Wheat protein content is generally determined by Kieldahl method,having a complicated operation process and long analysis time and cannot achieve batch testing of samples.There are linear and nonlinear mapping relationships between the near-infrared spectral features of wheat and protein content,and in this research,an algorithm with linear and nonlinear mapping capabilities was proposed,i.e.a combined algorithm based on convolutional neural network and feature selection regression.Multiple preprocessing methods were used to improve the signal-to-noise ratio of wheat kernels near-infrared spectral data,the one-dimensional preprocessed spectral data were collapsed into a two-dimensional matrix,a two-dimensional convolutional neural network model was used to predict the protein content of wheat,and the feature information output from the neurons in the middle part of the middle layer was extracted to form an integrated dataset with the data from the preprocessed spectral dataset.On the integrated dataset,the lasso,Minimax Concave Penalty regression(MCP)and Smoothly Clipped Absolute Deviation regression(SCAD)methods were used to construct a model for determining the protein content of wheat,compared and analyzed with the models of Multiple Linear Regression(MLR)and Partial Least Squares Regression(PLSR).The experimental results indicated that the introduction of convolutional neural network made the model with nonlinear mapping ability and improved the performance of wheat protein content prediction model.