东北寒地水稻茎秆纤维素含量近红外光谱反演

在水稻抗倒伏育种中,水稻茎秆纤维素含量作为重要的作物性状表现型数据,用传统方法获取时受人力成本和时间成本的约束,采集群体大小有限.利用高光谱技术能够实现对作物性状信息的快速、无损检测.为探究水稻茎秆纤维素含量近红外光谱反演模型,以田间小区试验的方式,采集水稻灌浆期至成熟期茎秆基部倒2、3节作为实验样本,并在实验室内使用...

Inversion Method for Cellulose Content of Rice Stem in Northeast Cold Region Based on Near Infrared Spectroscopy

In lodging resistance breeding of rice, the cellulose content of rice stem, as an important phenotypic data of crop traits, is constrained by human and time costs, which makes the size of the collecting population limited. The rapid and non-destructive detection of crop traits information can be achieved by using hyperspectral technology. In lodging resistance breeding of rice, rice stem cellulose content is one of the important character information. In order to explore the near-infrared spectral inversion model of cellulose content in rice stem, the bottom 2 and 3 segments of rice stem base formfilling stage to maturity stage were collected as experimental samples by field plot experiment, and the stem near-infrared reflectance spectrum data were measured by NIRQuest512 hyper-spectrometer in the laboratory. Standard normal variate (SNV), continuous wavelet transform (CWT), and the combination of the two methods (SNV-CWT) were used to pretreat the original near-infrared reflectance spectrum. Through comparative analysis, it was found that the original spectrum was optimized when it is firstly processed by SNV and then decomposed by CWT at 6 scales, and then the spectral characteristic variables are screened by the synergy interval PLS (SiPLS) method and iteratively retaining informative variables (IRIV) method for the characteristic spectral curve obtained by the optimal pretreatment (SNV-CWT). 64 and 16 characteristic variables were extracted, respectively. To optimize the model and improve its accuracy, the IRIV algorithm was used to conduct secondary screening of the characteristic variables selected by SiPLS, and 6 characteristic variables were obtained with the characteristic wavelengths of 1 200, 1 207, 1 325, 1 470, 1 482 and 1 492 nm. Finally, the support vector machine regression (ε-support vector machine regression, εSVR) and the kernel-based extreme learning machine (KELM) prediction model were established based on the selected characteristic variables. The model parameters (penalty coefficient C, kernel function coefficient γ and insensitive parameter ε) use grey wolf optimizer (GWO), differential evolution grey wolf optimizer (DEGWO) and self-adaptive differential evolution grey wolf optimizer (SaDEGWO) adaptive proposed in this paper for optimal selection. The results show that the SaDEGWO optimized εSVR model constructed by the characteristic variables selected by the SiPLS-IRIV method after spectral pretreatment with SNV-CWT method has the highest accuracy. The model parameters C, γ, ε are 302.838 2, 0.087 7, 0.070 8, respectively, and the coefficient of determination (R2_p) of the test set is 0.880. The root-mean-square error (RMSE_P) of the test set is 15.22 mg·g-1, residual predictive deviation (RPD) is 2.91. It indicates that the model has the good predictive ability, and this method can provide a reference for the prediction of cellulose content in rice stems.