IPLS-SPA波长选择方法在近红外秸秆生物量中的应用研究

在近红外光谱分析模型中全谱信息通常含有大量冗余信息，会导致模型解析时间延长、 加大模型解析难度，因此如何快速有效地选取特征波长至关重要。采用基于间隔偏最小二乘(interval partial least squares, IPLS)结合连续投影算法(successive projections algorithm, SPA)对小麦秸秆发酵过程微生物生物量进行特征波长选择，共制备85个样本，采用氨基葡萄糖法测定微生物生物量，选择68个样本作为校正集，17个样本作为验证集。首先对全谱区520个波长点根据间隔点大小10, 20, 30, 40进行分段建模，选取出4 450～4 925和9 194～9 993 cm-1两个波段范围作为特征波段，将选取出的特征波段再进行连续投影算法及遗传算法(genetic algorithm, GA)特征波长点选取，并进行综合分析对比。实验结果表明采用IPLS-SPA算法选择4 450～4 925和9 194～9 993 cm-1的组合波段具有最佳建模效果，相比于全谱建模其参与建模的波长点由520个减少到10个，模型验证集决定系数(R-Square, R2)从0.884 9提升至0.945 28，验证集均方误差根(root mean square error prediction, RMSEP)从11.104 9降至8.203 3，GA遗传算法虽取得了更优的模型精度，但其实验结果并不稳定且随机性较强，而IPLS结合SPA方法能够稳定而准确的(地)选择特征波长信息，提高模型运算速度并降低模型拟合难度，可以作为一种新的波段选择参考方法。结果表明采用近红外光谱分析方法对秸秆发酵生物量进行快速检测是可行的。

Research of Straw Biomass Based on NIR by Wavelength Selection of IPLS-SPA

The whole spectrum usually contains a lot of redundant information in the near-infrared spectroscopy model, the presence of redundant information will increase the model resolution time and increase the difficulty of parsing model, Therefore, how to select the characteristic wavelength quickly and effectly is very crucial. In this paper, we combined the algorithm based on SPA (successive projections algorithm ) with IPLS (interval partial least squares ) to selec the characteristic wavelength in the fermentation of wheat straw microbial biomass, A total of 85 samples prepared by measuring microbial biomass using glucosamine method, 68 samples are chosen as calibration set and 17 samples are chosen as verification set. First, the whole spectral region 520 points are segmented modeling according to the interval wavelength point size 10, 20, 30, 40 and 4 450～4 925 cm-1, 9 194～9 993 cm-1 two-band range are selected as the characteristic wavelength band, then pick out the new feature wavelength points by Successive Projections Algorithm band and Genetic Algorithm (GA), comprehensive analysis and comparison the result of model. The experimental results show that the using of IPLS-SPA algorithm to select the combination band 4 450～4 925 cm-1 & 9 194～9 993 cm-1 has the best modeling effect, compared with the modeling of whole spectrum, the wavelength points decrease from 520 to 10, the correction coefficient of determination R2 rised from 0.884 9 to 0.945 28, root mean square error (RMSE) dropped from 11.104 9 to 8.203 3, although the genetic algorithm model achieved the better accuracy, but the results are instable and have a strong randomness , while IPLS combined SPA method can select characteristic wavelength information stability and accurately, which can improve the model calculation speed and reduce the fitting difficulty of the model, it can be used as a new reference method for band selection. The results show that using near infrared spectroscopy method for straw biomass rapid detection is feasible.