基于抽取和连续投影算法的可见近红外光谱变量筛选

大多数短波CCD硅检测器为2 048或3 648像元,相邻波长间隔小,预处理算法对其适用性差.本文在600.09～980.47 nm光谱范围内,采用等间隔抽取方法重构光谱矩阵.经不同光谱预处理后,分别采用遗传算法(GA)和连续投影算法(SPA),筛选偏最小二乘法(PLS)建模变量.采用留一法交叉验证评价模型的预测能力,...

Selection of visible-NIR variables based on extraction and successive projections algorithm

The pixels were 2 048 or 3 648 for the most Si charge coupled device dector. The interval between the adjacent wavelengths was few. The pretreatment could not deal with the spectra well. Spectral matrix was reconstructed by equal interval extraction in the wavelength range of 600.09-980.47nm. The variables for developing partial least squares (PLS) models were chosen by genetic algorithm (GA) and successive projections algorithm (SPA) from the pretreatment spectra. The models' predictive ability was evaluated by leave-one-out cross validation. By comparison, the best results were obtained by the SPA-PLS models. The standard errors of cross validation (SECV) were 0.661 degrees Brix, 0.067% and 2.91 mg x (100 g)(-1) for soluble solids, total adicity and vitamin C, respectively. The results suggested that the predictive ability can be improved by equal interval extraction method and SPA for determinating the quality of Nanfeng mandarin fruits.