Research Of Three New Wavelength Selection Methods In Near Infrared Spectroscopy Quantitative Analysis Area

Near infrared spectroscopy has become more and more important in agricultural products quality analysis area,since it is a fast and nondestructive detection technique.However,there are usually a lot of uninformative or even noise wavelengths in the NIR spectrum.Thus wavelength selection has become an important step in the optimization of near infrared spectroscopy analysis model and there are already as many as dozens of wavelength selection algorithms.In this dissertation,the rationale differences between existing wavelength selection algorithms were systematically analyzed,and commonly used wavelength selection algorithms were also divided into five categories according to their rationale and features,which includes algorithms based on parameters of partial least squares(PLS)model,intelligent optimization algorithms,successive projections strategy,model population analysis strategy,and spectral intervals respectively.In order to overcome common shortcomings such as poor stability and reliability of existing wavelength selection algorithms,several following researches have been implemented based on the spectral features of NIR technique and several new ideas including ensemble modeling,model population analysis and combination strategy.Meanwhile,three benchmarking sample sets were also used for performance validation of new algorithms including corn,soil and tablet samples.(1)A new wavelength selection algorithm named as moving window smoothing ensemble competitive adaptive reweighted sampling(MWS-ECARS)was proposed based on ensemble strategy and CARS.In this algorithm,moving window is applied to smoothing the accumulated selected frequencies of each wavelength by CARS after a number of repeated runs,which can not only overcome poor stability of CARS,but also optimize the width of selected wavelength intervals through adjusting the width of moving window and threshold setting.(2)A new wavelength interval selection algorithm named as interval combination optimization(ICO)was proposed under the framework of model population analysis(MPA).In this method,the optimal interval combination can be searched iteratively in a soft shrinkage manner firstly,and then the widths of selected intervals can be optimized by local search automatically.Results indicate that ICO not only inherits merits of soft shrinkage optimization,but also owns faster convergence speed and fewer tune parameters.In addition,it was also proved that the selection of interval rather than individual wavelengths can indeed reduce the risk of overfitting,as well as computational burden of MPA.What's more,WBS was proved to be a more efficient sampling method than WBMS for implementing MPA strategy,since it can overcome disadvantages of WBMS by introducing appropriate scales of random components into the sampling step of MPA.(3)Successive Projections Algorithm(SPA)was applied to refine the wavelengths selected by MWS-ECARS and ICO respectively based on combination strategy.Results indicate that SPA can guarantee that the predictive performance of finally selected wavelengths will not be decreased dramatically,while reducing the number of wavelengths selected by these two rough selection algorithms.However,the higher the complexity of one sample set is,the lower the refining intensity of SPA will be.(4)The influence of different spectral pretreatment methods on the distribution and predictive performance of wavelengths selected by ICO algorithm was investigated by experiments.Results indicate that different spectral pretreatment methods really have relative strong influence on both the distribution of wavelengths selected by ICO algorithm and the predictive performance of simplified model.