The effect of selecting calibration sample sets on the NIR predictive model for 1000- grain weight of paddy was investigated. NIR models were developed by using partial least square regression in the wavelength region from 600 nm to 1100 nm under the conditions of different calibration sets with various paddy quantities, different ratios of the calibration set to the validation set, and different methods for selecting the calibration sets. The developed NIR models were evaluated according to determination coefficients for cress- validation (Rv2) and for prediction (Rp2), and standard errors for cross- validation (SECV) and for prediction (SEP). The results showed that the quantity of paddy sample, the ratio of the calibration set to the validation set and the method for selecting calibration set all had significant influences upon the NIR model for 1000 - grain weight of paddy. The 7:3 was the optimal ratio of the calibration set to the validation set for the development of NIR model. The NIR model developed based on the calibration set which was selected with K - S algorithm had better predictive ability for 1000 - grain weight of paddy than that developed based on the calibration sets which were selected with the gradient and the random methods.%为研究样本集选择方法对稻谷千粒重NIR模型的影响,分别采用不同数量样品,不同定标集、验证集比例以及不同定标集选择方法,选出建模的定标集,在600～1100nm的波长区间,用偏最小二乘法建立稻谷千粒重的近红外光谱预测模型,根据内部交叉验证决定系数(Rv2)、外部验证决定系数(Rp2)、内部交叉验证误差(SECV)和预测误差(SEP)比较模型的预测能力.结果显示,样品数量、定标集和验证集比例以及定标集选择方法均对稻谷千粒重的NIR模型有明显影响.采用合适数量的样品可以得到较佳的NIR模型,以7∶3的比例分割定标集与验证集,得到的稻谷千粒重NIR模型具有相对高的预测能力,而与含量梯度法和随机抽取法相比,采用K-S算法进行定标集选择,可以得到预测精度更高的NIR模型.
展开▼
