EGI OpenIR
基于高光谱的植物地球化学信息识别研究
Alternative TitleStudy on Plant Geochemical Information Recognition Based on Hyperspectral Measurements
崔世超
Subtype博士
Thesis Advisor周可法 ; 丁汝福
2020-06-30
Degree Grantor中国科学院大学
Place of Conferral北京
Degree Discipline工学博士
Keyword干旱荒漠区 白茎绢蒿 遥感植物地球化学 地球化学异常 隐伏矿产 Arid desert area Seriphidium terrae-albae Remote sensing phytogeochemistry Geochemical anomaly Concealed mineral
Abstract目前,随着露头矿、 地表矿及浅层矿的日益减少,急需一些新的找矿途径,而占世界陆地面积 2/3 的广大植被覆盖区仍有发现新的隐伏矿体的可能,是一个重要的深部矿产资源找矿方向。然而,由于植物的阻碍和干扰导致传统的地质勘查工作在植被覆盖区难以有效地开展, 急需一些新的找矿思路和找矿方法。近些年,随着遥感技术特别是高光谱技术的快速发展,遥感植物地球化学法在植物生长茂密的高植被覆盖区隐伏矿产勘查中得到了越来越广泛的应用,并取得了良好的应用效果, 但在植物生长相对稀疏的干旱荒漠区应用相对较少。事实上,干旱荒漠区在我国分布广泛且矿产资源也极其丰富,尤其是新疆地区的荒漠化面积达到 107.12 万平方公里,是我国干旱荒漠分布最大的区域。而与此同时, 新疆成矿条件好、矿产资源丰富、种类齐全,具有较大的开采潜力。为此,本研究选择广泛分布于新疆干旱荒漠区的一种小半灌木 — 白茎绢蒿( Seriphidiumterrae-albae) 为研究对象, 通过野外实地观测与采样,结合实验室盆栽实验, 研究不同程度的金属胁迫下白茎绢蒿的吸收聚集特征以及叶绿素、含水量和反射光谱的变化规律, 确定植物胁迫光谱和金属含量之间的定量关系, 研发一套利用高光谱技术快速、准确地识别植物中地球化学异常信息的方法, 为将来利用遥感植物地球化学法在干旱荒漠区开展隐伏矿产勘查工作奠定理论基础和技术支撑。 本文得到的主要研究成果如下:(1)白茎绢蒿几乎可以无障碍地吸收土壤中的 Au、 Zn、 Cu 以及 Ni 四种元素而不会出现死亡现象, 且植物中的这四种元素含量与土壤中对应元素含量具有较强的相关性(R2 分别达到 0.9524、 0.9568、 0.936 和 0.9177)。相比于生长在正常土壤环境中的植物,生长在高 Cu 和 Ni 含量的土壤中的白茎绢蒿体内聚集的金属元素明显增高,可以较为清晰地凸显出元素含量异常。在不同程度的 Cu和 Ni 胁迫下,白茎绢蒿的叶绿素含量、含水量以及反射光谱均会发生显著且有规律地变化。使用比值植被指数( 747742R R)可以较好地估算出白茎绢蒿 Cu 含量,构建的反演模型的决定系数 R2 达到 0.88,而对于 Ni 反演效果不佳,构建的植被指数反演模型的 R2 仅为 0.48。以上结果表明,当使用遥感植物地球化学法进行隐伏 Cu 矿的勘查时,白茎绢蒿可以作为有效的采样介质。(2)基于野外实测数据研究了波段宽度、光谱变换形式以及建模方法对白茎绢蒿 Cu 含量反演的影响。研究结果表明: 相比于随机森林(RF)、极限学习机(ELM)以及偏最小二乘支持向量机(LS-SVM)这三种建模方法,偏最小二乘(PLS)是最优建模方法,使用 PLS 构建的 Cu 含量反演模型不仅具有较高的精度且物理意义也比较明确; 相比于建模方法,光谱变换形式对 Cu 含量反演的影响更大,对原始光谱进行一阶以及二阶求导以后可以显著地增加 Cu 含量反演精度,其中把原始光谱平方根的二阶导数( R)''作为自变量输入到 PLS 中构建的Cu 含量反演模型的精度最高,其留一交叉验证得到的预测值和实测值之间的决定系数 R2 达到了 0.5447; 对比分析逐步回归分析法、遗传算法(GA)以及竞争加权重采样法( CARS)这三种波段筛选方法,将使用相关系数法从变换光谱( R)''中筛选出的特征波段输入到 PLS 中构建的模型的反演精度最高,留一交叉验证得到的预测值和实测值之间的决定系数 R2 达到了 0.5589,略高于基于全波段建模的 0.5447,而模型使用的波段个数却从 340 减少为 102 个,这极大增加了模型的稳定性,该模型是反演白茎绢蒿 Cu 含量的最优模型。(3) 相比于逐步回归分析、 GA 以及相关系数这三种波段筛选方法, 将使用 CARS 法从白茎绢蒿变换光谱 'lnR1( ) 中筛选出的 56 个特征波段输入到 PLS 中构建的土壤中 Au 含量反演模型的精度最高, 其留一交叉验证得到的预测值和实测值之间的 R2 达到了 0.8016,不仅具有较高的精度,模型的稳定性以及泛化能力也较高,该模型是反演土壤中 Au 含量的最优模型。(4)提出了一种新的指数 SCR(λ517-632、 697-737 nm),其代表植物 517-632 nm 和697-737 nm 范围内的反射光谱经过包络线去除以后得到的曲线包围 X 轴的面积,可以同时识别出在白茎绢蒿体内聚集而形成的 Au、 Ag 以及 Cu 含量异常。该指数的最大优点在于其可以识别组合元素含量异常,可以在一定程度上克服过往研究中利用单一元素异常信息进行矿产资源预测的不确定性和稳定性较差的缺点。(5)本研究发现,与使用单一时间段的植物光谱数据相比,使用多个时间段的植物光谱数据构建的土壤中 Cu 元素丰度识别模型的精度更高, 并利用从白茎绢蒿 6 月和 7 月的反射光谱中筛选出的 10 种植被指数和极限学习机分类模型(ELM) 构建了一个土壤中 Cu 元素丰度的识别模型,模型的识别精度达到了89.02%。该模型把光谱和时间数据有机地结合在一起,可以提供更加丰富的环境胁迫信息。 利用该模型可以快速圈定出土壤中 Cu 含量异常区域,为找矿靶区的圈定提供技术支持。(6) 基于 Matlab 语言和 Matlab GUI 功能研发了一个人机交互性较好、 可视化强的高光谱植物地球化学信息识别系统。该系统共包含了光谱预处理、光谱变换、特征波段提取、 金属含量反演、 金属元素丰度识别、最优波段组合的筛选、等值线图的绘制、遥感图像处理以及组合元素含量异常提取 9 大模块。 每个模块中又包含 2-3 个功能,总计共有 25 个功能。该系统把本研究中构建的 4 个植物地球化学模型有机地结合在了一起,利用该系统和我们课题组基于动力三角翼和HySpex 高光谱传感器搭载的“超低空探测平台”可以快速地进行植物地球化学信息的提取,进而为隐伏矿产的勘查提供技术支持。
Other AbstractCurrently, new prospecting focuses are urgently needed due to the decrease ofoutcrops, surface and shallow ore deposits. However, it is still possible to find newconcealed ore bodies in the vast vegetated area comprising 2/3 of global land area,which is an important prospecting focus for deep mineral resources. However, it isdifficult to effectively carry out traditional geological exploration in the vegetatedarea because plants obstruct and interfere with prospecting efforts. Therefore, newprospecting ideas and methods are urgently needed. Recently, the rapid developmentof remote sensing technology, especially hyperspectral technology, has resulted in theremote sensing phytogeochemical method becoming more widely used for exploringhidden minerals in densely vegetated areas. It has many applications that have beenapplied effectively, but few of these are in arid desert areas with sparse plant growth.Arid desert areas are widely distributed in China and have extremely abundantmineral resources, especially its Xinjiang Uyghur Autonomous Region, which has1.0712 million square kilometers of desertified areas and its largest distribution ofarid desert areas. Xinjiang has good mineralization, abundant mineral resources, andgreat mining potential. Therefore, this study uses the small semi-shrub Seriphidiumterrae-albae that is widely distributed in Xinjiang as the research object and uses fieldobservations and pot laboratory experiments to discuss its characteristics ofabsorption, aggregation and changes in chlorophyll, water content and reflectancespectrum under different levels of metal stress. The quantitative relationship betweenstress spectrum and metal content of plants was also constructed. The purpose of thispaper is to determine a set of methods to use hyperspectral technology to quickly andaccurately identify geochemical anomaly information in plants and to lay thetheoretical foundation and provide technical support for exploring concealed mineralresources in arid desert areas by using the remote sensing phytogeochemical method.The main research results are as follows:(1) Seriphidium terrae-albae can easily absorb Au, Zn, Cu, and Ni in soil, and itscontent of these four elements has a strong correlation with the corresponding elementcontent in the soil (R2 reaches 0.9524, 0.9568, 0.936 and 0.9177 respectively).Compared with plants growing in normal soil environments, the accumulation ofmetal in Seriphidium terrae-albae grown in soil with high Cu and Ni content issignificantly higher, which can clearly highlight abnormal metal content. Thechlorophyll content, water content, and reflection spectrum of Seriphidiumterrae-albae change significantly under different degrees of Cu and Ni stress. Theratio vegetation index (R747⁄R742) can be used to effectively estimate Cu content ofSeriphidium terrae-albae and the coefficient of determination (R2) of the inversionmodel constructed reached 0.88. However, the effect of Ni inversion was not good, asthe R2 of the inversion model of vegetation index constructed reached only 0.48. Asynthesis of the above results found that when using remote sensing plantgeochemical methods to conduct surveys of concealed Cu deposits, Seriphidiumterrae-albae can be used as an effective sampling medium.(2) In this study, based on field measurement data, the effects of bandwidth,spectral transformation form, and modeling methods on the inversion accuracy of Cucontent in Seriphidium terrae-albae were discussed. The research results show thatwhen comparing the three modeling methods (random forest (RF), extreme learningmachine (ELM), and partial least squares support vector machine (LS-SVM)), partialleast squares (PLS) is the best modeling method and the inversion model of Cucontent constructed using PLS not only has a high level of precision but also clearphysical meaning. Compared with the modeling methods, the spectral transformationshave a greater impact on the inversion accuracy of Cu content and the originalspectrum that has undergone the first and second derivations can dramatically increasethe inversion accuracy of Cu content. Among them, the Cu content inversion modelconstructed by inputting the second-order derivative of the square root of the originalspectrum (√R)′′as the independent variable to PLS has the highest accuracy. Thedetermination coefficient R2 between the predicted value and the measured value obtained by leave-one-out cross-validation reached 0.5447. It was also found thatcompared with stepwise regression analysis, genetic algorithm (GA) and competitiveweighted resampling (CARS), the highest accuracy can be obtained when the featurebands are selected from the transformed spectrum (√R)′′ using the correlationcoefficient method and inputted to PLS method to construct the Cu content inversionmodel. The determination coefficient R2 between the predicted value and themeasured value obtained by leave-one-out cross-validation reached 0.5589, whichwas slightly higher than 0.5447 based on full band modeling, while the number ofbands used in the model was reduced from 340 to 102, which greatly increased thestability of the model. This model is the best one for inversion of Cu content inSeriphidium terrae-albae.(3) Compared with stepwise regression analysis, GA and correlation coefficientmethod, the highest accuracy can be obtained when the 56 characteristic bands areselected from the transformed spectrum )'lnR1( of Seriphidium terrae-albae usingthe CARS method and inputted to PLS method to construct the inversion model of Aucontent in soil. The R2 between the predicted value and the measured value obtainedby leave-one-out cross-validation reached 0.8016, indicating that this model is notonly highly accurate, but also has high stability and generalizability. This model is thebest one for the inversion of Au content in soil.(4) This study proposed a new index SCR(λ517-632, 697-737 nm) that represents the areaaround the X-axis of the curve obtained from reflection spectrum in the 517-632 nmand 697-737 nm range of the plant after envelope removal, which can simultaneouslyidentify the abnormal Au, Ag, and Cu content formed by the aggregation ofSeriphidium terrae-albae. The biggest advantage of this newly constructed index is itsability to identify the abnormal content of combined elements, and can to some extentovercome the shortcomings of previous studies in their uncertainty and poor stabilityof mineral resource prediction by using abnormal information of a single element.(5) This study found that compared with plant spectral data in a single timeperiod, the identification model for Cu element abundance in soil constructed using plant spectral data in multiple time periods has higher accuracy. Meanwhile, arecognition model of Cu element abundance in soil was constructed using 10vegetation indices selected from the reflectance spectra in June and July along withthe extreme learning machine classification model, and the model had a recognitionaccuracy that reached 89.02%. This model can organically combine the spectrum andtime data, and can also provide more abundant environmental stress information. Thiscan be used to quickly delineate the abnormal area of Cu content in soil, which canprovide technical support for delineating a prospecting target area.(6) Based on Matlab and Matlab GUI, this research developed ahyperspectral-based plant geochemical information recognition system with goodhuman-computer interaction and strong visualization. The recognition system consistsof 9 modules: spectral preprocessing, spectral transformation, feature band extraction,construction of inversion model of metal content, identification of metal elementabundance, optimal band combination screening, isoline drawing, remote sensingimage processing and identification of the abnormal content of combined elements.Each module contains 2 to 3 functions, for a total of 25 functions. The systemorganically combines the four plant geochemical models built in this study. Using thissystem and the ultra-low altitude exploration platform based on the dynamic deltawing and HySpex hyperspectral sensor developed by the research group; it is possibleto quickly extract phytogeochemical information and provide technical support forexploring concealed minerals.
Subject Area地球探测与信息技术
Language中文
Document Type学位论文
Identifierhttp://ir.xjlas.org/handle/365004/15411
Collection中国科学院新疆生态与地理研究所
研究系统
Affiliation中国科学院新疆生态与地理研究所
First Author Affilication中国科学院新疆生态与地理研究所
Recommended Citation
GB/T 7714
崔世超. 基于高光谱的植物地球化学信息识别研究[D]. 北京. 中国科学院大学,2020.
Files in This Item:
There are no files associated with this item.
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[崔世超]'s Articles
Baidu academic
Similar articles in Baidu academic
[崔世超]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[崔世超]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.