基于BP神经网络的地下水离子化合物含量高光谱反演方法研究

The Hyperspectral Inversion Method of Main Ionic Compounds Content in Groundwater based on BP Neural Network

地下水水质污染日益加重,监测地下水离子化合物含量有利于地下水的动态管理与精准防治。离子化合物光谱响应信号微弱且反演机理尚不明晰,现有研究多数对离子化合物进行简单的定性分析,较少采用数理统计方法综合估算其含量。基于离子化合物光谱机理和高光谱数据冗余的特性,通过测量实验室配比的不同浓度钠、钾、钙3种离子化合物标准液的可见—近红外反射光谱(400~1 000 nm),探究水体3种离子化合物的光谱响应机理、最佳预处理方式及特征波段优选算法,并基于最优特征波段构建BP神经网络模型以定量反演离子化合物含量。研究发现:(1)3种离子化合物整体反射率在波长400~1 000 nm处与含量成反比,与离子的电荷数和半径成正比;(2)基于主成分分析较连续投影法提取的特征谱段,构建的多元线性回归模型能够较好地反演水体离子化合物含量;(3)KCl最优反演模型的预处理方式为SG滤波,CaCl_(2)和NaCl最优反演模型的预处理为SG滤波后进行反射率归一化;(4)相较于传统线性反演模型,PCA-BPNN非线性模型取得了最优的反演结果,其中钾离子化合物含量反演结果最优,其训练集R^(2)和RMSE分别达到0.996 4和248.77,测试集R^(2)和RMSE分别达到0.998 8和156.89,研究可为地下水离子化物遥感反演提供重要的理论与技术支撑。

Groundwater quality is becoming increasingly polluted and monitoring the content of groundwater ion-ic compounds is beneficial for dynamic groundwater management and accurate prevention.Little is known about the weak spectral response and inversion mechanisms of ionic compounds,and most existing studies have per-formed simple qualitative analyses of ionic compounds,with less use of mathematical and statistical methods for comprehensive estimation of their content.Based on the spectral mechanism of ionic compounds and the redun-dant nature of hyperspectral data,the spectral response mechanism of three ionic compounds in water,the opti-mal pre-processing method and the algorithm of feature band selection were investigated by measuring the visi-ble-near infrared reflectance spectra(400~1000 nm)of three ionic compound standard solutions with different concentrations of sodium,potassium and calcium in the laboratory.And based on the characteristic spectral bands,a BP neural network model is constructed to quantitatively invert the ionic compound content.It was found that(1)The overall reflectance of the three ionic compounds is inversely proportional to the content at wavelengths from 400 to 1000 nm and proportional to the charge number and radius of the ions;(2)Compared with the continuous projection method,the multiple linear regression model constructed based on the character-istic spectral bands extracted by principal component analysis can better infer the content of ionic compounds in water bodies;(3)The preprocessing of the KCl optimal inversion model by SG filtering and the preprocessing of the CaCl_(2) and NaCl optimal inversion models by SG filtering followed by reflectance normalization;(4)Compared with the traditional linear inversion model,the PCA-BPNN nonlinear model achieves the best inversion results,among which the inversion results of potassium ion compound content are the best,with the R^(2) and RMSE of the training set reaching 0.9964 and 248.77,respectively,the R^(2) and RMSE of the test set reaching 0.9988 and 156.89,respectively.This study can provide important theoretical and technical support for groundwater ionization inversion.