水稻蛋白质含量NIR模型适配范围的研究

Study on Adaptability of NIR Models of Protein Content in Rice

目的比较不同类型样品建立水稻蛋白质近红外模型的效果和适配范围。方法通过对178份来自“II-32B/岳早籼6号”的重组自交系和496份水稻品种的近红外反射光谱的比较分析,选择其中59个株系和76份品种作为建模样品,采用偏最小二乘法建立基于品种、重组自交系和混合样品的3个蛋白质含量回归模型。结果经模型内部交叉验证和对模型外部重组自交系和品种样品的验证结果的比较分析,发现基于分离群体的模型因蛋白质含量范围较窄,样品来源较单一,适应范围仅局限于本群体内样品蛋白质含量预测,而品种和混合模型对群体和品种样品都表现出良好的适应能力,交叉验证决定系数大于0.90,外部验证决定系数大于0.89,本试验可为近红外建模的样本集选择提供良好的指导意义。结论不同类型样品对建模效果有显著影响,品种模型和混合模型的适配范围显著大于群体模型,研究结果不能支持用背景变异较小的样品建立较高精度回归模型的设想。

[Objective] This research was done in order to compare the effect and adaptability of NIR models in different calibration samples. [Method] With NIR spectra of 178 RILs from cross of＂Ⅱ-32B/Yue-Zao-Xian No.6＂ and 496 rice varieties, 59 RILs and 76 cultivars were selected to develop three PLS regression models. These wre based on RILs, cultivars and mixture samples, respectively. [Result] By comparing cross validation results with calibration sets and predictions for both RILs and varieties, we confirmed limitations of RILs based model due to narrow ranges of protein content and lack of diversity. A more reliable prediction method could be established only when the model was being applied to samples from RILs. However, both models that were based on representative varieties and mixture samples exhibited much better adaptability than samples from RILs or varieties with higher determination coefficients in cross validation （r^2 〉0.90） and testing set validation （r^2 〉0.89）, respectively. This gave suggestions on calibration sample selection. [Conclusion] Modeling results varied greatly in different types of calibration samples. The adaptability for calibration models in varieties and their mixtures were much broader than that of RILs. Using samples of smaller genetic variance, we were unable to create a regression model with more accuracy.