用遗传区间偏最小二乘法建立苹果糖度近红外光谱模型

为了简化苹果糖度预测模型和提高模型的精度，用遗传区间偏最小二乘法(GA-iPLS)建立苹果近红外光谱预测模型。应用结果表明，整个光谱划分为40个子区间， GA-iPLS选择其中的第4，6，8，11，18号共5个子区间联合建立苹果糖度模型。遗传区间偏最小二乘法所建的模型，其校正时的相关系数r_c和交互验证均方根误差RMSECV分别为0.962和0.334 6，预测时的相关系数r_p和预测均方根误差RMSEP分别为0.932和0.384 2。与全光谱模型相比，该方法建立的模型不论对校正集还是预测集，模型的预测能力都提高了许多，且模型得到了很大的简化：其实际采用的波数点个数比全光谱模型采用的波数点个数大大减少，主因子数也比全光谱少，由此建立的模型更加简洁、数据运算量也更少。

Near Infrared Determination of Sugar Content in Apples Based on GA-iPLS

To improve and simplify the prediction model of sugar content, genetic algorithm interval partial least square (GA-iPLS) methods, the evolution of iPLS described by Lars NΦrgaard, were proposed and used to establish the calibration models of sugar content against apple spectra. The apple spectra data were divided into 40 intervals, among which 5 subsets, i.e. No. 4, 6, 8, 11 and 18, containing 362 data points were selected by GA-iPLS. The optimum GA-iPLS calibration model was obtained with the correlation coefficient (r_c) of 0.962, the root mean square error of cross-validation (RMSECV) of 0.334 6 and the root mean square error of prediction (RMSEP) of 0.384 6. Compared with the whole spectra data model, the data points and the factors in the GA-iPLS were decreased significantly. Consequently, the running time of the PLS model build by GA-iPLS was shorter than that of the whole spectra data model. Furthermore, the GA-iPLS model could not only improve precision, but also simplify the model.