| 基于机器学习的铅锌铜矿区土壤重金属含量高光谱反演分析 |
| 摘要点击 250 全文点击 10 投稿时间：2025-02-26 修订日期：2025-07-05 |
| 查看HTML全文
查看全文 查看/发表评论 下载PDF阅读器 |
| 中文关键词 机器学习 土壤 重金属 特征波段 高光谱反演 |
| 英文关键词 machine learning soil heavy metals characteristic bands hyperspectral inversion |
| DOI 10.13227/j.hjkx.202502230 |
| 作者 | 单位 | E-mail | | 古甜甜 | 成都理工大学地球与行星科学学院, 地球勘探与信息技术教育部重点实验室, 成都 610059 | gutiantian@stu.cdut.edu.cn | | 刘严松 | 成都理工大学地球与行星科学学院, 地球勘探与信息技术教育部重点实验室, 成都 610059
四川三合空间科技有限公司, 成都 610094 | liuyansong2012@cdut.edu.cn | | 胡薇 | 成都理工大学地球与行星科学学院, 地球勘探与信息技术教育部重点实验室, 成都 610059 | | | 郑博骞 | 成都理工大学地球与行星科学学院, 地球勘探与信息技术教育部重点实验室, 成都 610059 | | | 唐方强 | 成都理工大学地球与行星科学学院, 地球勘探与信息技术教育部重点实验室, 成都 610059 | | | 任凤玲 | 成都理工大学地球与行星科学学院, 地球勘探与信息技术教育部重点实验室, 成都 610059 | |
|
| 中文摘要 |
| 针对土壤中重金属含量低，弱土壤光谱信息难以通过直接分析重金属的特征光谱来估算重金属含量问题，提出结合一阶微分（FD）、二阶微分（SD）、倒数对数（AT）等多种光谱变换方法和随机森林（RF）、支持向量机（SVM）、多元逐步线性回归（SMLR）、人工神经网络（ANN）这4种机器学习算法的土壤重金属含量高光谱反演模型组合方案. 首先采集土壤样本的XRF检测数据和反射光谱并利用FD、SD和AT等数学方法进行光谱变换，然后系统分析土壤重金属含量与变换后光谱特征的相关性，据此筛选出具有代表性的特征波段，最后基于特征波段对比RF、SVM、SMLR和ANN这4种建模方法的反演效果并通过精度评估确定各元素最优模型选择. 结果表明，预处理的5种光谱变化有效地提高了图像的信噪比，减少了光谱之间的间隔和差异，除去特征波段为0的变换方式（AT和ASD），其他光谱变化表现出的效果依次为：SD>FD>AFD. 4种机器学习模型均具有一定的预测能力，且模型效果决定系数R2几乎大于0.6. 适用于该矿区预测效果最佳模型依次为：SMLR>SVM>ANN>RF. Cu和Mn元素的最佳反演模型均为SD-SMLR模型，Fe元素的最佳反演模型为SD-SVM模型，Zn元素的最佳反演模型为AFD-SMLR模型. 研究结果可为铅锌铜矿区大范围测量重金属元素含量提供技术支持，并为环境监测提供了新的发展路径. |
| 英文摘要 |
| In view of the low content of heavy metals in soil， it is difficult to estimate the heavy metal content by directly analyzing the characteristic spectrum of heavy metals in weak soil spectral information. A combination scheme of the hyperspectral inversion model for soil heavy metal content was proposed， which combined multiple spectral transformation methods， such as first-order differential （FD）， second-order differential （SD）， and reciprocal logarithm （AT）， and four machine learning algorithms including random forest （RF）， support vector machine （SVM）， multiple stepwise linear regression （SMLR）， and artificial neural network （ANN）. Firstly， the XRF detection data and reflectance spectra of the soil samples were collected and transformed by FD， SD， AT， and other mathematical methods. Then， the correlation between the soil heavy metal content and the transformed spectral characteristics was systematically analyzed， and the representative characteristic bands were screened out. Finally， the inversion effects of the four modeling methods， RF， SVM， SMLR， and ANN， were compared based on the feature bands， and the optimal model selection of each element was determined by accuracy evaluation. The results show that： The five spectral changes in preprocessing effectively improved the signal-to-noise ratio of the image and reduced the intervals and differences between spectra， and the effects of the spectral changes， excluding the transformation method with a characteristic band of 0 （AT and ASD）， were in the order of SD > FD > AFD. All four machine learning models had certain predictive capabilities， and the determination coefficients R2 of the model effects were almost greater than 0.6. The best prediction models for this mining area in order were SMLR > SVM > ANN > RF. The best inversion models for both the Cu and Mn elements were SD-SMLR models， the best inversion model for the Fe element was the SD-SVM model， and the best inversion model for the Zn element was the AFD-SMLR model. The research results provide technical support for large-scale measurement of heavy metal element content in lead-zinc-copper mining areas and offer a new development path for environmental monitoring. |
