【作者】 陈华舟；

【导师】 潘涛；

【作者基本信息】 上海大学 ， 应用数学， 2011， 博士

【摘要】 近红外(NIR)光谱是一种直接定量分析技术,由于不需要化学反应,可以实时分析样品,使得它在应用上有很大的优势,同时也在方法上需要克服很大的困难,因为对于复杂体系而言,近红外光谱包含有各种噪音,必须利用有效的化学计量学方法去消除,这其中有很多具有挑战性的数学问题,如定标样品分析、光谱预处理模式、光谱波段的优选等等。土壤是农业可持续发展最重要的组成部分,土壤中的营养成分(有机质、总氮)的含量是衡量土壤肥力的一个重要指标。土壤成分的无试剂简便快速测定方法是现代农业急需的关键技术之一。由于土壤是含有多组分的复杂体系,土壤近红外光谱的高精度分析模型的研究具有重要意义,本文将以此为目标,研究若干核心的化学计量学方法。首先研究基于Savitzky-Golay(SG)平滑的光谱预处理方法;其次研究基于移动窗口偏最小二乘(MWPLS)的连续光谱波段的优选方法,和基于等间隔移动窗口多元线性回归(ECMWMLR)的离散光谱波长组合的优选方法,并结合SG平滑做进一步的模型优化;为了降低模型复杂性和设计专用仪器的参考,本文还提出一种基于最优组合波长的光谱降维方法,并利用实验证实了它的有效性;此外,为了得到稳定、可靠的模型,本文所有模型都是基于多个定标集和预测集的划分得到的,并且建立了一种合理的定标集和预测集的划分方法。本文建立计算机算法平台,集成上述NIR光谱分析的化学计量学方法,分别建立土壤有机质、总氮的NIR光谱分析模型,并进行模型检验。土壤有机质的最优MWPLS模型:波段为1692-1880 nm,PLS因子数为14,预测均方根偏差(RMSEP),预测相关系数(R_P)分别为0.275 (%),0.870;最优ECMWMLR模型:起点波长为1786 nm,点数为9,间隔为20,RMSEP,R_P分别为0.265 (%),0.871。土壤总氮的最优MWPLS模型:波段为1600-2198 nm,PLS因子数为11,RMSEP,R_P分别为0.0145 (%),0.886;最优ECMWMLR模型:起点波长为1716 nm,点数为9,间隔为31,RMSEP,R_P分别为0.0141 (%),0.891。结果表明,本文建立的优化模型,其效果明显优于传统的全谱PLS模型和SG-PLS模型,而且模型更为简单、稳定,为近红外光谱应用于土壤分析建立了高精度实用模型,所得到的光谱波段和光谱波长组合也为专用近红外仪器设计提供了重要参考。所建立的方法框架和计算机算法平台还可以应用到其他复杂体系的近红外光谱分析中。更多还原

【Abstract】 Near-infrared (NIR) spectroscopy is a direct quantitative analysis technique. It can analyze samples in real time without chemical reactions, which makes it a great advantage on the application, but also there are great difficulties in methods to overcome. As for complex systems, the near-infrared spectrum includes a variety of noises, chemometrics methods must be used to eliminate these noises, of which there are many challenging mathematical problems, such as the calibration samples analysis, spectral preprocessing modes, spectral waveband selection optimization and so on.Soil is the most important sustainable development of agriculture component. The nutritional content of the soil (organic matter, total nitrogen) is an important indicator to measure soil fertility. Simple and rapid reagent-free determination method for soil content is the critical need for modern agricultural technologies. As soil is a complex system with multi-component, the study of high precision models of near-infrared spectroscopy analysis for soil is much significant, taking this as the objective, we research a number of core chemometrics methods in this paper.Firstly, we study the spectral preprocessing methods based on Savitzky-Golay (SG) smoothing; secondly, research the optimization methods for continuous spectral waveband on moving window partial least squares (MWPLS), explore the optimization method for discrete spectral wavelength combination founded on equidistant combination moving window multiple linear regression (ECMWMLR), and then further optimize models joined with SG smoothing; thirdly, for reducing model complexity and providing the reference of designing special instruments, we propose a spectral dimension reduction method based on the optimal combinational wavelength, and experimentally confirmed its effectiveness. In addition, to get stable and reliable results, all optimal models in this paper are obtained by multiple divisions of calibration set and prediction set, and a rational dividing method is proposed.We build up a computer algorithm platform, for the integration of chemometrics methods for NIR spectroscopy analysis, and respectively establish NIR analysis models for soil organic matter and total nitrogen, and further examine the models. For organic matter, its optimal MWPLS model shows, the waveband is 1692-1880 nm, PLS factor is 14, root mean square error of prediction (RMSEP) and correlation coefficient of prediction (RP) are 0.275 (%) and 0.870, respectively; while in its optimal ECMWMLR model, the beginning wavelength is 1786 nm, the number of adopted wavelengths is 9, the gap of adopted wavelengths is 20, RMSEP and RP are 0.265 (%) and 0.871, respectively. For nitrogen, its optimal MWPLS model indicates, the waveband is 1600-2198 nm, PLS factor is 11, RMSEP and RP are 0.0145 (%) and 0.886, respectively; while in its optimal ECMWMLR model, the beginning wavelength is 1716 nm, the number of adopted wavelengths is 9, the gap of adopted wavelengths is 31, RMSEP and RP are 0.0141 (%) and 0.891, respectively. Results prove that the prediction effects of these optimal models are obviously better than that of the traditional analysis models, such as PLS and SG-PLS models on the whole spectral collecting region, and these optimal model is more simple and stable, providing high precision practical model for NIR spectroscopy applying to soil analysis, the spectral wavebands and the spectral wavelength combinations provide important references for designing specific NIR instrument. The methodological framework and the computer algorithm platform here can also be used for the NIR spectroscopy analysis of other complex systems.更多还原