基于近红外光谱大豆蛋白质、脂肪快速无损检测模型的优化构建

Rapid Nondestructive Test of Soybean Protein and Fat by Near Infrared Spectroscopy Combined with Different Model Methods

为实现大豆蛋白质、脂肪含量的快速无损检测,采集350~2 500 nm光谱范围内的大豆近红外光谱。运用经典Kennard-Stone算法选取建模样本及验证样本,对近红外原始光谱进行卷积平滑(savitzky and golay, SG)+一阶微分、变量标准化(standard normal variate, SNV)+去趋势算法(de-trending,DT)、正交信号校正(orthogonal signal correction,OSC)处理;然后通过竞争性自适应重加权采样方法(competitive adaptive reweighted sampling,CARS)筛选出特征波长,比较偏最小二乘法(partial least squares,PLS)、BP神经网络法所建模型,最终获得对于大豆蛋白质、脂肪含量的快速、无损检测的最佳模型。结果表明:(1)经CARS特征波段挑选后,波长的变量个数由1 981个减少为100个以下,变量压缩率大于94.95%;(2)CARS波段选择能够提高建模精度,基于挑选的特征波段所建立模型的决定系数均>0.9;(3)OSC+CARS+PLS与OSC+CARS+BP该类数据处理组合方式在一定程度上能够实现大豆蛋白质、脂肪的快速、无损检测。优化构建的该模型能够精准快速无损的检测大豆蛋白质、脂肪含量,对大豆品质评估以及作物改良具有重要意义。

In order to realize the rapid nondestructive testing of soybean protein and fat content, the near infrared spectrum of soybean was collected in the range of 350-2 500 nm. The classical kennard-stone algorithm was used to select modeling samples and verification samples, SG + first-order differential, SNV+ DT, OSC processing. Then, the characteristic wavelength was selected by the competitive adaptive weighted sampling method(CARS), the detection model of soybean protein, fat content established by partial least squares(PLS) and BP neural network method was compared. Finally, the detection of soybean protein and fat content was achieved. It was demonstrated that:(1) After the CARS method was preferred, the number of wavelength variables was reduced from 1 981 to less than 100, and the variable compression ratio was greater than 94.95%.(2) The CARS band selection improved the modeling accuracy, and the determination coefficient of the model based on the selected feature bands was higher than 0.9.(3) The combination of OSC+CARS+PLS and OSC+CARS+BP could achieve rapidly and non-destructive detection of soy protein and fat to a certain extent. The optimized model can accurately, rapidly and nondestructively detect the accurate estimation of soybean protein and fat content, which is of great significance for soybean quality evaluation and crop improvement.