硬度是评价小麦品质的一个重要质量参数, 对小麦的分类、 最终用途以及小麦籽粒组成的研究都非常重要。 为实现小麦硬度的快速、 准确检测, 在详细分析小麦籽粒成分对红外光吸收特性的基础上, 研究建立径向基函数(RBF)神经网络模型实现对未知样品硬度的准确检测, 并着重分析了不同光谱信号预处理方法对模型预测精度的影响。 从各小麦主产区收集111个小麦样品, 扫描样品获得可见-近红外光谱, 采用马氏距离判断并剔除异常光谱; 利用优化后的SPXY划分样品集合, 得到校正集84个样品, 预测集24个样品; 利用连续投影算法(SPA)从262个光谱波点中提取47个特征光谱; 分别使用一阶导数、 二阶导数和标准正态变量变换(SNV)及其不同组合对光谱进行预处理, 验证不同预处理方法之间的相互影响, 寻找最优的预处理方法组合。 校正集预处理后的特征光谱数据作为RBF模型的输入, 采用硬度指数法测定的对应样品硬度作为输出建立模型。 预测结果显示当采用SNV和SPA处理光谱数据时模型的效果达到最优, 评价指标判别系数(R2)、 预测标准差(SEP)和相对分析误差(RPD)可分别达到0.90, 3.02和3.11, 表明基于可见-近红外光谱的RBF神经网络模型能够准确地预测小麦的硬度, 与传统检测方法相比具有方便、 快速、 无损等优点, 为小麦硬度的检测提供一条更为便捷与实用的方法。

Grain hardness is an important quality parameter of wheat which has great influence on the classification, usage and composition research of wheat. To achieve rapid and accurate detection of wheat hardness, radial basis function (RBF) neural network model was built to predict the hardness of unknown samples on the basis of analyzing the absorptive characteristics of the composition of wheat grain in infrared, besides, the effects of different spectral pretreatment methods on the predictive accuracy of models were emphatically analyzed. 111 wheat samples were collected from major wheat-producing areas in China; then, spectral data were obtained by scanning samples. Mahalanobis distance method was used to identify and eliminated abnormal spectra. The optimized method of sample set partitioning based on joint X-Y distance (SPXY) was used to divide sample set with the number of calibration set samples being 84 and prediction set samples being 24. Successive projections algorithm (SPA) was employed to extract 47 spectral features from 262. SPA, first derivatives, second derivatives, standard normal variety (SNV) and their combinations were applied to preprocess spectral data, and the interplay of different prediction methods was analyzed to find the optimal prediction combination. Radial basis function (RBF) was built with preprocessed spectral data of calibration set being as inputs and the corresponding hardness data determined via hardness index (HI) method being as outputs. Results showed that the model got the best prediction accuracy when using the combination of SNV and SPA to preprocess spectral data, with the discriminant coefficient (R2), standard error of prediction (SEP) and ratio of performance to standard deviate (RPD) being 0.844, 3.983 and 2.529, respectively, which indicated that the RBF neural network model built based on visible-near infrared spectroscopy (Vis-NIR) could accurately predict wheat hardness, having the advantages of easy, fast and nondestructive compared with the traditional method. It provides a more convenient and practical method for estimating wheat hardness.