基于改进ABR模型的小麦籽粒类胡萝卜素含量近红外检测方法

Near Infrared Detection of Carotenoids in Wheat Grain Based on Improved ABR Model

摘要:

小麦籽粒类胡萝卜素含量是衡量小麦营养价值和育种品质的关键指标。为实现小麦籽粒类胡萝卜素含量快速无损检测，设计了一种小麦籽粒近红外光谱快速采集装置，可实现待测样品等量分样、快速多次采集近红外光谱数据，提高光谱采集效率。以180个小麦籽粒作为研究对象，获取900~1700nm范围内的近红外光谱数据，采用Savitzky-Golay平滑（Savitzky-Golay，SG）、标准正态变量变换（Standard normal variate transformation，SNV）、多元散射校正（Multiplicative scatter correction，MSC）、趋势校正（Trend correction，TC）、Savitzky-Golay平滑+趋势校正（SG+TC）、Savitzky-Golay平滑+多元散射校正（SG+MSC）、Savitzky-Golay平滑+标准正态变量变换（SG+SNV）、Savitzky-Golay平滑+一阶导数（SG+1D）、Savitzky-Golay平滑+二阶导数（SG+2D）、Savitzky-Golay平滑+三阶导数（SG+3D）10种预处理方法，Relief算法、遗传算法（Genetic algorithm，GA）、方差阈值算法（Variance threshold，VT）、连续投影算法（Successive projections algorithm，SPA）4种特征选择算法建立偏最小二乘回归（Partial least squares regression，PLSR）、支持向量机回归（Support vector regression，SVR）、自适应提升回归（Adaptive boosting regression，ABR）3种机器学习模型，对小麦籽粒类胡萝卜素含量进行预测。为进一步提高模型预测精度，引入纵横交叉（CSO）和自适应收敛因子（ACF）2种改进策略对ABR模型进行优化。结果表明，使用SNV预处理、Relief特征选择算法和CSO-ACF改进策略建立ABR模型预测效果最佳，校正集决定系数R2C为0.90，均方根误差（RMSEC）为0.29μg/g，预测集决定系数为R2P为0.90，均方根误差（RMSEP）为0.32μg/g，RPD为3.16。因此，该装置与模型算法相结合可实现小麦类胡萝卜素含量快速、无损预测。

Abstract:

Carotenoid content in wheat grain is the key index to measure the nutritional value and breeding quality of wheat. In order to realize the rapid and nondestructive detection of carotenoid content in wheat grain, a near-infrared spectrum rapid acquisition device for wheat grain was designed, which can realize the equal sample size of the sample to be measured, and quickly collect near-infrared spectrum data for many times, so as to improve the spectral acquisition efficiency. Taking 180 wheat grains as the research object, the near-infrared spectral data in the range of 900~1700nm were obtained. Savitzky-Golay (SG), standard normal variate transformation (SNV), multiplicative scatter correction (MSC), trend correction (TC), SG+TC, SG+MSC, SG+SNV, Savitzky-Golay+first derivative (SG+1D), Savitzky-Golay+second derivative (SG+2D), Savitzky-Golay+third derivative (SG+3D) ten pretreatment methods, four feature selection algorithms, relief algorithm (Relief), genetic algorithm (GA), variance threshold (VT) and successive projections algorithm (SPA), were used to establish three mathematical models of partial least squares regression (PLSR), support vector regression (SVR) and adaptive boosting regression (ABR) to predict carotenoid content in wheat seeds. The results showed that the prediction effect of ABR model based on SNV preprocessing, relief feature selection algorithm and CSO-ACF improved strategy was the best. The determination coefficient R2C of correction set was 0.90, the root mean square error (RMSEC) was 0.29μg/g, the determination coefficient R2P of prediction set was 0.90, the root mean square error (RMSEP) was 0.32μg/g, and the RPD was 3.16. Therefore, the device combined with the model algorithm can achieve rapid and nondestructive prediction of carotenoid content in wheat.