关键词: 特征光谱, 小麦, 籽粒蛋白质, 无信息变量剔除, 连续投影算法, 模型构建

Abstract: 【Objective】The objective of the study was to select the characteristic spectrum of the whole grain of wheat grain protein and set up an optimization model for the rapid and non-damage detection of protein content of the whole grain of wheat grain protein, so as to provide a basis for designing field portable wheat grain protein content determination of spectrometer.【Method】 The experiment was carried out in 2012 and 2013 with eight winter wheat varieties with obvious difference in protein content as materials. Six treatments including three nitrogen levels and two irrigation levels were designed, and rich sample types of 176 wheat grain spectral data was collected. The original reflection spectrum obtained by ASD FieldSpec Pro optical spectrum instrument were transformed as absorption spectra. Then, through the S-G smoothing, the multiplicative scatter correction and baseline correction processing, the spectra were used to create model with cross validation of partial least squares regression, uninformative variables elimination (UVE) method, successive projections algorithm (SPA), multiple linear regression (MLR) provision and stepwise multiple linear regression (SMLR) and their combination, respectively. 【Result】The results show that, the unconcerned information with wheat grain protein could be eliminated by uninformative variables elimination (UVE) method, the original spectrum wavelengths were compressed from 1621 to 717, which realized the first screening without getting rid of protein information. After that, different screening methods were used to correct the characteristics spectrum model. In this study, firstly, the grain protein content irrelevant information variables were removed by UVE. Then, using the SPA eliminated the effects of collinearity in the band spectrum matrix. Last, using SMLR contribution to the whole grain of wheat grain protein prediction model, the 15 big characteristic bands were screened out. The root mean square prediction error (RMSEP) and R² are 0.5898 and 0.9410, respectively. 【Conclusion】To implement rapid determination of the whole grain of wheat grain protein under field conditions, the whole grain of wheat grain spectrum matrix can be effectively compressed using UVE, SPA and SMLR methods. The constructed forecasting model based on the screening protein content characteristic spectra can realize nondestructive and rapid determination of the whole grain of wheat grain protein content. The forecasting model is accuracy, reliable and cost effective, which lay a solid foundation for the design of field portable integrated wheat grain protein meter band selection and development.

Key words: characteristic spectrum, wheat, kernel protein, uninformative variables elimination, successive projections algorithm, model formation