基于随机变异—Kennard—Stone和偏最小二乘法的土壤重金属镉含量反演——以雄安新区西南部为例  

Retrieval of soil heavy metal Cadmium content based on Random Mutation,Kennard—Stone and partial least squares method:A case study of southwest of Xiong’an New District

作　　者：黄照强 倪斌 HUANG Zhaoqiang;NI Bin(Institute of Mineral Resources,China Metallurgical Geology Bureau,Beijing,101300)

机构地区：[1]中国冶金地质总局矿产资源研究院,北京101300

出　　处：《地质论评》2021年第5期1521-1532,共12页Geological Review

基　　金：中国冶金地质总局科技创新项目(编号:CMGB202001);国家重点研发计划项目(编号:2016YFC0600210);国家自然科学基金资助项目(编号:41272366)的成果。

摘　　要：土壤重金属污染对人类身体健康造成了严重威胁,作为快捷、高效、无损监测分析土壤重金属含量的方法之一,高光谱遥感反演土壤重金属含量的方法正逐步得到发展。本文以雄安新区西南部作为研究区,针对潜在生态风险最大且活性最强并易被植物吸收的重金属镉,开展高光谱反演研究。将采集来的426件土壤样品除杂、风干、过筛在实验室测得重金属含量,并用SVC便携式地面光谱仪测量样品350~2500 nm范围光谱。采用Savitzky—Golay卷积平滑方法进行光谱降噪平滑处理。由于粒径大小而不是化学成分差异可能会导致基线效应和漂移现象,为了增强光谱差异和光谱曲线形状,将数据进行标准正态变量变换(SNV)、一阶微分(FD)、二阶微分(SD)、多元散射校正(MSC)等数学变换,并分析变换后光谱与Cd含量的相关性。本文提出一种集成随机变异法—Kennard—Stone法—偏最小二乘回归的方法。针对变换后的光谱集和Cd含量集,第一步采用随机变异法—Kennard—Stone法将样本集分为70%训练集和30%验证集,使样本数随性质分布均匀并覆盖整个样本空间;第二步用偏最小二乘回归法结合交叉验证建立回归模型,用确定系数R2、均方根误差RMSE、偏差比值RPD、误差范围比值RER等参数开展模型评价,如果没有达到预期效果,则回到第一步,迭代反复选择,直至达到最优效果。结果表明,适用的最优反演重金属镉含量的技术方法是采用FD变换后,不断迭代集成RM—KS样本选择和PLS偏最小二乘法回归建立的模型,其验证综合效果最好,建模主成分数为11个,确定系数R2达到0.909,RMSE为0.604,RPD为2.696,RER达为15.516。成果可为类似区域快速、无损的土壤重金属Cd含量反演提供技术支撑。Soil heavy metal pollution poses a serious threat to human health. As one of the fast, efficient and non-destructive methods for monitoring and analyzing soil heavy metal concentration data, hyperspectral remote sensing method is gradually developing. In this paper, the hyperspectral inversion study in the southwest of Xiong’an New District was carried out for cadmium, which has the greatest potential ecological risk and the strongest activity and is easy to be absorbed by plants. 426 soil samples were collected for impurity removal, air drying and sieving. The heavy metal content was measured in the laboratory, and the spectrum of 350 ~ 2500 nm was measured by SVC portable ground spectrometer. Savitzky—Golay convolution smoothing method is used for spectral denoising and smoothing. Because the baseline effect and drift phenomenon may be caused by the difference of particle size rather than chemical composition, in order to enhance the spectral difference and the shape of spectral curve, the data were transformed by several mathematical transformations such as Standard Normal Variable Transformation(SNV), First-Order Differential(FD), Second-Order Differential(SD), Multiplicative Scattering Correction(MSC). And the correlations between the transformed spectrum and Cd content were analyzed. In this paper, a method of integrating Random Mutation(RM) — Kennard—Stone(KS) —Partial Least Squares Regression(PLSR) is proposed. For the transformed samples spectrum and Cd content set, the first step is to divide the samples set into 70% training set and 30% verification set by using RM — KS method, so that the number of samples is evenly distributed with the properties and covers the whole sample space;In the second step, the PLSR method combined with cross validation is used to establish the regression model, and the parameters such as determination coefficient R2, root mean square error(RMSE), ratio of percent deviation(RPD), ratio of error range(RER) are used to carry out the model evaluation. If the expecte

关 键 词：土壤重金属污染 光谱变换 RM—KS样本集选择 偏最小二乘回归 高光谱 

分 类 号：X833[环境科学与工程—环境工程]