基于偏最小二乘留一交叉验证法的近红外光谱建模样品选择方法的研究

目的提出一种新的挑选定标集的方法-偏最小二乘留一交叉验证法。方法以玉米为例,通过对玉米中水分含量的实际建模与外部验证,根据主成分数、相关系数、预测均方根差以及相对分析误差(ratio of performance to standard deviate,RPD)等因素,综合比较4种定标集挑选方法的优缺点。结果偏最小二乘留一交叉验证法结合样品和光谱性质,在保持原始样品覆盖范围的基础上,挑选出的定标集所建立的模型具有较低的模型复杂程度、较高的验证相关系数以及较高的RPD值。结论该方法既克服了随机挑选法存在的样品代表性不足的风险,同时也避免了含量梯度法和计算机识别法只考虑样品或者光谱的单一性质的不足,同时该方法具有操作简单、易于推广等优点,为食品安全检测提供了一种新的筛选样品的方法。

Study on the sample selection methods for building calibration model by near infrared spectroscopy based on partial least squares-leave one out-cross validation

Because of lots of samples were needed to build NIR models of agricultural products. Selection of representative samples for training can directly influence the representative accurateness of the model and reduce the workload of building models. Some methods to select samples for calibration were introduced, such as random method, content grads method, and Computer recognition?(K-S) method. A new method called partial least-squares leave one out-cross validation (PLS-LOO-CV) will be introduced. According to the values of R, RMSEP, and RPD, the above 4 methods have been compared with the example of the content of water in the coin. Results show that the method bring forward has good ability to choose calibration samples using the sample data and the spectrum data together. The conclusion was that this method could avoid the risk of lacking Representative cased by random method as well as the consideration of sample data or spectrum data only of all the samples using content grads method or computer recognition method, and at the same time the PLS-LOO-CV could keep the concentration range of origin set; the calibration set selected by the method could reduce the complexity of the model and the model has higher value of R, lowest value of RMSEP and higher value of RPD. In addition, the method is simple and easy to be used.