首页 | 本学科首页   官方微博 | 高级检索  
     检索      

基于特征波段的高光谱技术检测水体中毒死蜱浓度的实验研究
引用本文:马瑞峻,张亚丽,陈瑜,张亚莉,邱志,萧金庆.基于特征波段的高光谱技术检测水体中毒死蜱浓度的实验研究[J].光谱学与光谱分析,2019,39(3):923-930.
作者姓名:马瑞峻  张亚丽  陈瑜  张亚莉  邱志  萧金庆
作者单位:华南农业大学工程学院,广东 广州,510642;华南农业大学工程学院,广东 广州,510642;华南农业大学工程学院,广东 广州,510642;华南农业大学工程学院,广东 广州,510642;华南农业大学工程学院,广东 广州,510642;华南农业大学工程学院,广东 广州,510642
基金项目:国家重点研发计划(2016YFD0800901)和国家自然科学基金项目(51309103)资助
摘    要:为了探究反射光谱检测水体中毒死蜱农药的可行性,使用由ASD公司的FieldSpecPro地物波谱仪构成的高光谱采集系统在室内、室外环境获取两种不同浓度区间的毒死蜱样品的光谱数据。基于偏最小二乘(PLS)和主成分分析(PCA)算法分别对毒死蜱样品光谱数据建立全波段定量模型,结果两种模型的预测能力均较高。通过相关性分析(CA)计算相关系数来选择毒死蜱样品光谱的特征波长,其中浓度区间为5~75 mg·L-1的室内、室外实验光谱的特征波长为388,1 080,1 276 nm和356,1 322,1 693 nm,浓度区间为0.1~100 mg·L-1的室内外实验样品光谱的特征波长为367,1 070,1 276,1 708 nm和383,1 081,1 250,1 663 nm。结合PLS算法建立样品特征波长光谱数据的定量模型,结果与全波段模型相比,浓度区间为5~75 mg·L-1的室内外实验光谱PLS特征波长模型的校正集决定系数R2C分别提高至0.987 5和0.999 2,预测集决定系数R2P分别提高至0.989 4和0.994 4,校正集均方根误差RMSEC分别降低为2.841和0.714,预测集均方根误差RMSEP分别降低为1.715和1.244;浓度区间为0.1~100 mg·L-1的室内外实验光谱特征波长PLS模型的校正集决定系数R2C分别提高至0.998 3和0.998 8,预测集决定系数R2P分别提高至0.998 4和0.999 0,校正集均方根误差RMSEC分别降低为1.383和1.186,预测集均方根误差RMSEP分别降低为1.510和1.229,验证集标准差与预测均方根误差的比值(RPD)有所增加,尤其是针对浓度区间为0.1~100 mg·L-1的实验,RPD值显著增加至21.7,说明基于特征波长建立的毒死蜱样品定量模型具有较高精度的预测能力,但是通过不同浓度区间范围的对比实验发现,ASD地物光谱仪对低浓度的毒死蜱溶液预测的相对误差偏大,存在客观上的检测下限。为了保证不同试验条件下的毒死蜱农药的特征波长都得到分析,增强模型使用的普适性与鲁棒性,根据特征波长选择出4个波段,即351~393,1 065~1 086,1 245~1 281和1 658~1 713 nm作为特征波段。特征波段模型的波长变量个数共38个,相比于全波段模型的432个波长变量,模型变量精简了91.2%,其中浓度区间为5~75 mg·L-1的室内外实验光谱PLS特征波段模型的R2C分别为0.993 7和0.987 8,R2P分别为0.979 8和0.998 2,RMSEC分别为1.690和2.516,RMSEP分别为1.987和0.659;浓度区间为0.1~100 mg·L-1的室内外实验光谱特征波段PLS模型的R2C分别为0.9882和0.9807,R2P分别为0.9391和0.9936,RMSEC分别为3.345和3.942,RMSEP分别为8.996和2.663,且四种实验情况下的模型RPD值均大于2.5,满足定量分析条件。因此采用高光谱采集系统对室内和室外环境中毒死蜱农药的快速检测具有一定的可行性,此研究结果对有机磷农药等面源污染物快速检测有实际的应用价值,可为农田水体有机磷农药快速检测仪器的开发提供理论基础。

关 键 词:高光谱  毒死蜱  偏最小二乘法  相关性分析法  定量模型  特征波长  特征波段
收稿时间:2018-06-02

Experimental Study on Detection of Chlorpyrifos Concentration in Water by Hyperspectral Technique Based on Characteristic Band
MA Rui-jun,ZHANG Ya-li,CHEN Yu,ZHANG Ya-li,QIU Zhi,XIAO Jin-qing.Experimental Study on Detection of Chlorpyrifos Concentration in Water by Hyperspectral Technique Based on Characteristic Band[J].Spectroscopy and Spectral Analysis,2019,39(3):923-930.
Authors:MA Rui-jun  ZHANG Ya-li  CHEN Yu  ZHANG Ya-li  QIU Zhi  XIAO Jin-qing
Institution:College of Engineering, South China Agricultural University, Guangzhou 510642, China
Abstract:In order to investigate the feasibility of reflectance spectroscopy for the detection of chlorpyrifos pesticides in water, indoor and outdoor spectral data of chlorpyrifos samples in two different concentrations were obtained using a hyperspectral acquisition system composed of ASD’s FieldSpecPro Spectrometer. The partial least squares (PLS) and principal component analysis (PCA) algorithms were used to establish quantitative models for spectral data of chlorpyrifos samples. The results showed that the predictable ability of the model is significantly reliable. Correlation analysis (CA) was used to calculate the correlation coefficient to select the characteristic wavelength of the spectrum of chlorpyrifos samples. The characteristic wavelengths of indoor and outdoor experimental spectra with concentration ranges of 5~75 mg·L-1 were 388, 1 080, 1 276 and 356, 1 322, 1 693 nm, respectively. And the characteristic wavelengths were 367, 1 070, 1 276, 1 708, and 383, 1 081, 1 250, 1 663 nm in the range of 0.1~100 mg·L-1 experiments. The PLS algorithm was used to establish a quantitative model of the sample characteristic wavelength spectral data. Compared with the full-band model, the calibration set determination coefficient (R2C) of the PLS characteristic wavelength model with concentration range of 5~75 mg·L-1 was increased to 0.987 5 and 0.999 2 in the indoor and outdoor experiment, respectively. And the prediction set determination coefficient (R2P) was increased to 0.989 4 and 0.994 4, respectively. The root mean square error of the calibration set (RMSEC) was reduced to 2.841 and 0.714, respectively. The root mean square error of the prediction set (RMSEP) was reduced to 1.715 and 1.244, respectively. The R2C of the characteristic wavelength PLS model with concentration range of 0.1~100 mg·L-1 in the indoor and outdoor experiment was increased to 0.998 3 and 0.998 8, respectively. The R2P was increased to 0.998 4 and 0.999 0, respectively, and the RMSEC of the correction set was reduced to 1.383 and 1.186, respectively, and the RMSEP of the prediction set was reduced to 1.510 and 1.229, respectively. The ratio of standard deviation of the validation set to standard error of prediction (RPD) were increased, especially for experiments with a concentration range of 0.1~100 mg·L-1. The RPD value increased to 21.7 significantly, indicating that the quantitative model based on the characteristic wavelength has higher accuracy of prediction ability. However, comparative experiments with different concentration ranges show that the relative error of the low-concentration chlorpyrifos solution prediction by the ASD spectrograph is large and there is an objective detection limit. In order to ensure that the characteristic wavelengths of chlorpyrifos pesticides under different experimental conditions are analyzed and the universality and robustness of the model are enhanced, four bands are selected according to the characteristic wavelengths, that is, 351~393, 1 065~1 086, 1 245~1 281 and 1 658~1 713 nm used as characteristic bands. The characteristic band model has a total of 38 wavelength variables. Compared with the 432 wavelength variables of the full-band model, the model variable was reduced by 91.2%. The R2C of indoor and outdoor experimental PLS models with concentration range of 5~75 mg·L-1 were 0.993 7 and 0.987 8, and R2P were 0.979 8 and 0.998 2, and RMSEC were 1.69 and 2.516, and RMSEP were 1.987 and 0.659, respectively. The R2C values of the experimental PLS model with concentration range of 0.1~100 mg·L-1 were 0.988 2 and 0.980 7 for the indoor and outdoor experiments, and the R2P were 0.939 1 and 0.993 6, and the RMSEC were 3.345 and 3.942, and the RMSEP were 8.996 and 2.663, respectively. All of the model RPD values were more than 2.5 and met the quantitative analysis conditions. Therefore, the hyperspectral system of the paper for the rapid detection of chlorpyrifos pesticides in indoor and outdoor environments has a certain feasibility. The results of this study have practical application value for the rapid detection of non-point source pollutants such as organic phosphorus pesticides, which can provide a theoretical basis for the development of an instrument for the rapid detection of organophosphorus pesticides in farmland water.
Keywords:Hyper-spectrum  Chlorpyrifos  Partial least squares  Correlation analysis  Quantitative model  Characteristic wavelength  Characteristic band  
本文献已被 CNKI 万方数据 等数据库收录!
点击此处可从《光谱学与光谱分析》浏览原始摘要信息
点击此处可从《光谱学与光谱分析》下载免费的PDF全文
设为首页 | 免责声明 | 关于勤云 | 加入收藏

Copyright©北京勤云科技发展有限公司  京ICP备09084417号