秸秆热值近红外光谱模型的外部验证结果间的统计比较分析

Statistical Comparison of Independent Validation Results for Near Infrared Spectroscopy Models Predicting Calorific Value of Straw

从全国24个省(市)收集到222个秸秆样品,包括172个稻秸样品和50个麦秸样品。采用近红外光谱技术,结合主成分回归、偏最小二乘回归和改进的偏最小二乘回归建立了秸秆热值的定量分析校正模型。近红外光谱模型的建立与优化过程中使用了不同的散射校正方法和光谱导数处理来帮助改善模型精度。对得到的54个模型采用统计学的方法分析外部验证的结果,通过比较外部验证的系统偏差(Bias)和Bias校正的预测标准差(SEP(C)),考察了不同光谱预处理和回归方法对秸秆热值的近红外模型预测性能的影响。结果表明:近红外光谱技术能够快速、准确地分析秸秆的热值,模型的SEP(C)在134～178J.g-1之间;对外部验证结果的统计分析,能够有效地选择较好的建模方法,确定较优模型。

Two hundred and twenty-two straw samples, consisting of 170 rice straw samples and 50 wheat straw samples, were collected from 24 provinces of China. Near infrared spectroscopy （NIRS）was applied to build quantitative models for calorific value of straw combining the use of principal component regression （PCR）, partial least square regression （PLS）and modified partial least square regression （MPLS）. Different scatter correction methods and derivative treatments were adopted to help improve the accuracy of NIRS models. A total of 54 NIRS models were obtained and independent validations were conducted using the same validation set of samples. A statistical comparison of independent validation results was then introduced to evaluate whether the models perform significantly. Bias and bias corrected standard error of prediction （SEP（C））, which are the mean and the standard deviation of the prediction residuals respectively, were compared by the proposed statistical procedures. It was concluded that near infrared spectroscopy was able to predict the calorific value of straw samples rapidly and accurately, with resuiting SEP（C）s between 134 and 178 J · g^-1 ; statistical comparison of biases and SEP（C）s was a reasonable and efficient way to compare spectral pre-processing methods, and select NIRS models predicting calorific value of straw.