中文摘要：

还原糖含量是评价马铃薯全粉品质的重要指标之一,该文研究基于近红外光谱技术结合最小二乘支持向量机（least squares support vector machine,LSSVM）算法的马铃薯全粉还原糖含量非线性数学模型。采用移动窗口偏最小二乘法（moving windows partial least square,MWPLS）和连续投影算法（successive projections algorithm,SPA）组合方法筛选出20个特征变量,作为LSSVM的输入向量。优化径向基函数（radial basis function,RBF）的惩罚因子和核参数,训练LSSVM校正模型。经比较,LSSVM校正模型预测结果最优,预测相关系数为0.984,预测标准差为0.223%,相对分析误差（standard deviation ratio,SDR）为5.62。结果表明：近红外光谱结合LSSVM算法提高了马铃薯全粉还原糖含量的预测精度。

英文摘要：

Reducing sugar content is one of the important indicators for evaluating the quality of potato granules. Near-infrared （NIR） spectroscopy has been attempted to determine reducing sugar content in potato granules using near-infrared （NIR） spectroscopy combined with least squares support vector machine （LSSVM） algorithm. NIR spectra were recorded in the wavenumber range of 10 000-4 000 cm-1 at a 4 cm-1 interval. The 110 samples were divided into calibration and prediction sets in terms of their respective actual value for avoiding bias in subset division. One of every four samples was divided into the prediction set according the range of actual value in calibration set covering the range in the prediction set. The calibration set contained 83 samples, and the remaining 27 samples constituted the prediction set. Three different variable selection methods, namely the moving windows partial least square （MWPLS）, MWPLS-genetic algorithm （MWPLS-GA）, and MWPLS-successive projection algorithm （MWPLS-SPA）, were performed comparatively to choose spectral variables associated with reducing sugar content distributions. The partial least square （PLS） models were developed with these selection spectral variables with the number of PLS components optimized according to root mean square error of cross validation （RMSECV） in the calibration set. The results derived by variable selection techniques were then compared with the performance of PLS models with new samples in the prediction set. The PLS calibration model exhibited a higher correlation coefficient of prediction （Rp） of 0.976, lower standard error of prediction （SEP） of 0.273%, and ratio of SEP and standard deviation （SDR） of 4.593, which was built using 20 spectral variables selected by the MWPLS-SPA method. Nonlinear models of the least squares support vector machine （LSSVM） were developed using different spectral variables selected by MWPLS, MWPLS-GA, and MWPLS-SPA. The main parameters of penalty factor （y） and nucle