基于组合偏最小二乘的特征波段优选方法在氨、碱化处理玉米秸秆粗蛋白检测中的研究

Study on Detection of Crude Protein in Ammonified and Alkalized Corn Straw by Spectrum Characteristic Band Selection Method Based on Synergy Interval Partial Least Squares

该文构建了玉米秸秆粗蛋白定量分析模型,并对光谱特征波段选取方法进行探讨及验证。首先对107个样本进行预处理,剔除两个异常样本后采用DB2小波缺省阈值4层分解方式进行光谱重构,预处理后粗蛋白模型交互验证决定系数R2CV从0.7889提高至0.9208,采用间隔偏最小二乘(IPLS)及其改进型方法后向区间间隔偏最小二乘(BIPLS)、组合间隔偏最小二乘(SIPLS)进行特征波段选取,并对比主成分分析、竞争性自适应重加权采样法、相关系数法、遗传算法、移动窗口最小二乘等结果,发现基于IPLS及其改进型BIPLS、SIPLS均可有效、准确定位特征波段区间,其中采用SIPLS 30波段间隔在10128~10398 cm^-1与11196~11462 cm^-1时具有最优模型,验证集相关系数(rp)为0.9784,验正集决定系数(R2P)为0.9572,验正集均方误差根(RMSEP)为0.2211,相比于其他波段选取方法表现出较好的实时准确性,该方法可为玉米秸秆氨碱化最优条件判定提供重要的数据支撑。

The quantitative analysis model for crude protein in corn straw was constructed in this paper,and the selection method for spectrum characteristic band was discussed and verified.Firstly,107 samples were preprocessed,then a DB2 wavelet transform method with default threshold 4-level decomposition was used to reconstruct the spectra after removing two abnormal samples.The determination coefficient of cross validation(R2CV)for crude protein model was increased from 0.7889 to 0.9208 after pretreatment.Interval partial least squares(IPLS)and its improved methods,i.e.backward interval partial least squares(BIPLS)and synergy interval partial least squares(SIPLS)were adopted to select the characteristic bands,that IPLS and it's improved approach BIPLS,SIPLS could locate the characteristic bands more effectively and accurately compared with principal component analysis(PCA),competitive adaptive reweighted sampling(CARS),correlation coefficient(CC),genetic algorithm(GA),moving windows partial least squares methods(MWPLS).When SIPLS was using 30 band interval,the optimal model validation results were obtained in the band ranges of 10128-10398 cm^-1 and 11196-11462 cm^-1,with the correlation coefficient of validation set is 0.9784,the determination coefficient R-square of prediction set is 0.9572 and the root mean square error of prediction set of 0.2211.IPLS method exhibited a better real-time accuracy,and it and has a certain practicability in data support for the determination of ammoniation and alkalization of corn straw.