光谱学与光谱分析, 2019, 39 (5): 1543, 网络出版: 2019-05-13  

土壤铅污染光谱的HHT鉴别及BC-PLSR铅含量预测模型

HHT Identification and BC-PLSR Prediction Model of Soil Lead Pollution Spectrum
付萍杰 1,2,*杨可明 1,2程龙 1,2王敏 1,2,3
作者单位
1 中国矿业大学(北京)煤炭资源与安全开采国家重点实验室, 北京 100083
2 中国矿业大学(北京)地球科学与测绘工程学院, 北京 100083
3 华北理工大学, 河北 唐山 063210
摘要
土壤重金属污染问题一直备受关注, 利用高光谱遥感对其进行研究取得了大量的成果, 主要集中在利用土壤光谱的导数变换、 连续统去除等常规方法预测土壤重金属含量上。 土壤光谱数据与非线性非平稳的机电信号、 医学信号等具有一定的相似性。 通过希尔伯特黄变换(Hilbert-Huang transform, HHT), 对土壤铅(Pb)污染光谱进行频率域分析, 实现土壤Pb污染光谱的HHT鉴别, 并建立土壤Pb含量预测模型。 首先, 进行土壤Pb污染实验, 采集土壤Pb污染样品的光谱、 含水率及有机质含量; 其次, 通过土壤Pb污染样品光谱的HHT时频分析和第二个本征模函数(intrinsic mode function, IMF)分量(IMF2)瞬时频率的二阶导数识别土壤Pb污染的特征波段; 最后, 选择合适的频率域参数、 土壤光谱一阶导数、 土壤有机质含量及土壤含水率作为参数, 利用箱形图、 聚类分析、 偏最小二乘法建立土壤Pb含量预测模型。 研究结果表明: 土壤Pb污染的HHT时频分析图可以鉴别土壤Pb污染光谱, 未受污染的土壤光谱HHT时频分析图在波段序列为250~430之间没有异常信号, Pb污染土壤的光谱HHT时频分析图在波段序列为250~430之间存在多个异常信号, 并且随着浓度的升高, 异常信号分布范围越来越广, 当污染浓度达到800 μg·g-1时, 土壤样品的光谱信号在波段序列为270处、 频率为0.3 Hz之前出现了较强的异常信号; 土壤Pb污染光谱经验模态分解(empirical mode decomposition, EMD)处理后, 得到的未受污染的土壤光谱IMF2的瞬时频率的二阶导数的突变非常微弱, 而Pb污染的土壤光谱IMF2的瞬时频率的二阶导数存在明显的突变点, 根据突变点及土壤Pb污染光谱的IMF2的瞬时频率的二阶导数识别的土壤Pb污染光谱的特征波段区间为2 150~2 300 nm; 利用不同浓度Pb污染下土壤光谱Hilbert能量谱峰值、 EMD能量熵、 一阶导数、 有机质和含水率, 通过箱形图去除了六组异常样品, 然后利用聚类分析的方法将去除异常样品后的土壤Pb污染样品分为两类, 最后将Hilbert能量谱峰值、 EMD能量熵、 2 134 nm波段一阶导数、 790 nm波段一阶导数、 1 276 nm波段一阶导数、 2 482 nm波段一阶导数、 有机质和含水率作为参数建立两类数据的BC-PLSR(boxplot cluster-partial least squares regression)模型预测土壤中Pb含量, 经验证模型精度较高, 相关系数分别为0.88和0.99。
Abstract
The problem of soil heavy metal pollution has always attracted attention. Therefore, many results have been achieved in this field by the research on the use of hyperspectral remote sensing, mainly focusing on predicting heavy metal content in soil using conventional methods such as derivative of soil spectra and continuous continuum removal. The soil spectral data showed tremendous similarity with non-linear non-stationary mechatronic signals, medical signals, etc. In this study, HHT was used to analyze the soil’s lead (Pb) pollution experimental spectra in the frequency domain. The purpose of the HHT analysis was to achieve the HHT identification of soil’s Pb pollution spectra, and establish the model for predicting Pb content in soil. Firstly, the soil Pb pollution experiment was conducted to collect the spectrum, water content, and organic matter content of soil Pb-contaminated samples. Secondly, the HHT time-frequency analysis and the second derivative of the instantaneous frequency of the second intrinsic mode function (IMF2) component of the Pb pollution spectra of soil samples were used to identify the characteristic bands of soil Pb contamination. Finally, the prediction model of soil Pb content that the parameters were appropriate frequency domain parameters, soil first-order derivative, soil organic matter content, and soil water content was established using boxplot, cluster analysis, and partial least squares. The results showed that HHT time-frequency analysis charts of soil Pb-contaminated could identify soil Pb contamination spectra. There was no abnormal signal in the band sequence between 250 and 430 from HHT time-frequency analysis plots of uncontaminated soil spectrum. There were many abnormal signals in the band between 250 and 430 from soil spectral HHT time-frequency analysis plots under Pb contamination, and with the increase of the concentration, the abnormal signal distribution range became wider and wider. When the pollution concentration reached 800 μg·g-1, a strong abnormal signal was obtained in the band sequence of 270 and the frequency before 0.3 Hz. The mutation of second derivative of IMF2 instantaneous frequency of uncontaminated soil spectrum was very weak after the EMD, while there were obvious mutation points of Pb-contaminated soil spectrum. The characteristic wavelength band of soil Pb-contaminated soil spectrum was 2 150~2 300 nm according to the mutation points and second derivative of IMF2 instantaneous frequency of Pb-contaminated soil spectrum. Six groups of abnormal samples were removed from boxplot based on Hilbert energy spectrum peaks, EMD energy entropy, first derivative, organic matter and water content under different Pb concentrations. Then the Pb-contaminated soil samplings were divided into two categories by cluster analysis. Finally, Hilbert energy spectrum peak, EMD energy entropy, spectral first derivative of 2 134, 790, 1 276 and 2 482 band, organic matter and water content were used as parameters. The BC-PLSR (Boxplot Cluster-Partial Least Squares Regression) models for the data of two categories were established to predict Pb content in soil. The accuracy of the validated model was high, and the correlation coefficients were 0.88 and 0.99, respectively.

付萍杰, 杨可明, 程龙, 王敏. 土壤铅污染光谱的HHT鉴别及BC-PLSR铅含量预测模型[J]. 光谱学与光谱分析, 2019, 39(5): 1543. FU Ping-jie, YANG Ke-ming, CHENG Long, WANG Min. HHT Identification and BC-PLSR Prediction Model of Soil Lead Pollution Spectrum[J]. Spectroscopy and Spectral Analysis, 2019, 39(5): 1543.

关于本站 Cookie 的使用提示

中国光学期刊网使用基于 cookie 的技术来更好地为您提供各项服务,点击此处了解我们的隐私策略。 如您需继续使用本网站,请您授权我们使用本地 cookie 来保存部分信息。
全站搜索
您最值得信赖的光电行业旗舰网络服务平台!