高光谱数据中不仅含有关键性信息还存在一些干扰信息和无效信息, 带有干扰信息和无效信息的数据建立模型会降低效率和模型精度。 从全波段数据中提取特征波长是提高关系模型精度的有效方法。 有序预测选择(OPS)是一种依据信息向量选择有效波长变量的特征波长提取算法, 在特征波长变量筛选方面表现了较好地性能。 但由于建立模型时, 没有去除重要性较低的变量, 导致过多的无效变量参与到模型中, 降低了模型的准确率。 论文以羊肉高光谱数据作为研究对象, 提出了一种改进的特征波长变量选择方法, 基于信息向量和指数衰减函数的有序预测选择方法(AW-OPS)对羊肉新鲜度进行检测, 该算法通过光谱数据和理化值数据之间的关系来计算信息向量并对波长变量进行排序, 采用指数衰减函数(EDF)通过多次迭代去除一些信息向量绝对值比较低的波长变量, 最后在已获取的有效波长变量中逐渐增加波长点建立多元回归模型, 选取交叉均方根误差(RMSECV)最小值的波长变量子集为特征波长变量。 实验时, 将OPS法和AW-OPS法在选取特征波长变量后, 分别构建羊肉TVB-N的偏最小二乘(PLS)关系模型, 同时与全光谱波段PLS模型的效果相比较。 结果表明: OPS算法运行程序平均用时为175.9 s, 优选出370个特征波长变量, OPS-PLS模型相关系数(RP)平均为0.963 1, 均方根误差(RMSEP)平均为0.727; 而改进的有序预测选择法(AW-OPS)运行程序平均用时为57.6 s, 优选出275特征波长变量, AW-OPS-PLS模型平均提升到0.973 1, RMSEP平均降低为0.572 8; 全光谱波长数目为1 414个波长变量, 其PLS模型的平均为0.920 8, RMSEP平均为1.048 3。 AW-OPS-PLS模型相较于OPS-PLS模型测试精度提高了21.2%, 相较于全光谱-PLS模型, 测试精度提高了45%, 证明AW-OPS是一种有效特征波长变量筛选方法, 提高了OPS模型精度和程序运行效率, 降低了模型复杂度。

Hyperspectral data contain not only critical information but also some interference information and invalid information, and using these data to build the model will reduce the reliability and accuracy of the relational model. Extracting feature wavelengths from full-band data is an effective way to improve the accuracy of prediction models. Ordered Predictive Selection (OPS) is a feature wavelength extraction algorithm that selects effective wavelength variables based on the information vector, and has shown good performance in feature wavelength variable screening. However, the model was built without removing the less important variables, resulting in too many invalid variables being involved in the model and reducing the model’s accuracy. The paper proposes an improved feature wavelength variable selection method based on an information vector and exponential decay function of ordered predictive selection method (AW-OPS) for lamb freshness detection, using lamb hyperspectral data as the research object. The algorithm calculates the information vector and ranks the wavelength variables by the relationship between the spectral data and the physicochemical value data. The exponential decay function (EDF) is used to remove some wavelength variables with relatively low absolute values of information vectors by multiple iterations. Finally, a multiple regression model was established by gradually adding wavelength points to the obtained effective wavelength variables, and the subset of wavelength variables with the lowest value of root mean square error (RMSECV) was selected as the characteristic wavelength variables. For the experiments, the partial least squares (PLS) relational models of lamb TVB-N were constructed by the OPS -and AW-OPS methods after selecting the characteristic wavelengths, respectively, and compared with the effects of FULL-PLS models. The results showed that the OPS algorithm took an average of 175.9 s to run the program, preferentially selected 370 characteristic wavelength variables, with an average OPS-PLS model correlation coefficient(RP)of 0.963 1 and an average root mean square error(RMSEP)of 0.727. while the improved ordered prediction selection method(AW-OPS)runs the program in an average time of 57.6 s, preferentially selects 275 characteristic wavelength variables, and the AW-OPS-PLS model RP improves to 0.973 1 on average, and RMSEP reduces to 0.572 8 on average. The number of full-spectrum wavelengths was 1 414 wavelength variables, and the average RP of its PLS model was 0.920 8, and the average RMSEP was 1.048 3. The AW-OPS-PLS model improved the test accuracy by 21.2% compared to the OPS-PLS model and 45% compared to the full-spectrum-PLS model, proving that the improved AW-OPS is an effective feature wavelength variable screening method that improves the accuracy of the OPS model and the efficiency of the program operation and reduces the complexity of the model.