基于偏最小二乘回归的藻类荧光光谱特征波长选取  

Feature Wavelength Selection of Phytoplankton Fluorescence Spectra Based on Partial Least Squares

作　　者：余晓娅[1] 张玉钧[1] 殷高方[1] 肖雪[1] 赵南京[1] 段静波[1] 石朝毅[1] 方丽[1] 

机构地区：[1]中国科学院安徽光学精密机械研究所环境光学与技术重点实验室,安徽合肥230031

出　　处：《光学学报》2014年第9期294-299,共6页Acta Optica Sinica

基　　金：国家自然科学基金(61378041);国家863计划(2014AA06A509;2013AA065502);安徽省杰出青年科学基金(1108085J19);安徽省自然科学基金(1408085MD72);中国科学院仪器设备功能开发技术创新项目(yg2012071)

摘　　要：针对藻类荧光光谱解析中常见的信息冗余和光谱相关性问题,基于偏最小二乘(PLS)的方法,提出了区间蒙特卡罗偏最小二乘(IMC-PLS)方法,有效地解决了特征波长的选取问题。根据特征色素荧光峰位置预选出特征区域,综合利用了此特征区域内单个波段的信息和不同的随机波段组合对于模型的贡献,基于荧光光谱的三线性特点,联合了发射波长和激发波长的信息。研究结果表明,与无信息变量消除算法(UVE)相比,IMC-PLS反演4种藻类浓度得到的平均相对标准偏差分别降低了0%、34.3%、55.9%、30.5%,选择出的特征波长数和运算时间分别减少了80.1%、81.3%,IMC-PLS方法有效地解决了实时监测问题,也为离散三维荧光光谱仪器的研制提供了理论支持。For spectral information redundancy and correlation in phytoplankton spectral analysis, interval Monte Carlo partial least squares （IMC-PLS） which effectively solves the problem of feature wavelength selection is presented based on partial least squares （PLS）. Feature region is preselected according to the location of the pigment fluorescence peaks, the internal informations of a single band and the contributions of different random band combinations to the model are plenarily used. Based on three-linear feature of fluorescence spectra, emission wavelength band and excitation wavelength band are considered as a unit. The result shows that comparing with the uninformative variable eliminotion （UVE）, feature wavelength points and computation time obtained by IMC-PLS decrease by 80.1% and 81.3% and average relative tolerances （ARTs） by inversion of four algae concentrations decrease by 0%, 34. 3%, 55. 9%, 30. 5%. IMC-PLS algorithm effectively solves the problem of real-time monitoring, and provides theoretical support for the development of a discrete three-dimensional fluorescence spectrometer meanwhile.

关 键 词：光谱学 特征波长 区间蒙特卡罗偏最小二乘回归 无信息变量消除 荧光光谱 藻类 

分 类 号：O433.4[机械工程—光学工程]