基于高分辨率遥感影像的渭河水质遥感监测研究

【作者】 蒋赛；

【导师】 汪西莉；

【作者基本信息】 陕西师范大学 ， 计算机软件与理论， 2009， 硕士

【摘要】 随着遥感技术的发展和遥感影像尺度的精细化,利用遥感技术进行水环境质量监测的研究越来越多,对水质遥感监测的要求也从定性监测向更为精确的定量监测转变。由于海洋和湖泊等大面积水体中主要物质的光学模型比较成熟,针对这些物质的光学特性研究较多,所以目前遥感水质定量监测的研究对象多集中在海洋、内陆湖泊或水体面积较大的江河,而采用SPOT遥感数据对渭河水质进行定量遥感研究目前还没有。渭河流经陕西中部关中地区,沿途流经西安、宝鸡、咸阳、渭南等大中城市,该地区工业集中、人口密集、农业发达,是陕西省政治、经济、文化、金融及信息中心,而渭河是该地区主要河流和地表水源,因此对渭河进行水质遥感监测意义重大。基于此,本文以渭河陕西段水域为研究对象,采用法国SPOT-5遥感影像对该水域水质进行定量遥感研究。在文章结构上,本文在系统介绍了水质定量遥感监测的研究情况及方法流程的基础上,针对流程中各个步骤选择合适的方法进行实验并对各实验结果进行了较为详细的分析;在研究方法上,本文重点研究了遥感影像的大气辐射校正方法和基于参数优化的支持向量回归（Support Vector Regression,以下简称SVR）在渭河水质定量遥感中的应用。对于遥感影像的大气辐射校正方法,本文分析了目前遥感影像的大气辐射校正方法,结合自身的遥感影像及实验条件选择了合适的校正方法对SPOT-5遥感影像进行大气辐射校正;对于基于参数优化的SVR在渭河水质定量遥感中的应用,本文在分析了传统多元回归的缺点后引入了统计学习理论（Statistical Learning Theory,以下简称SLT）下的支持向量机（Support VectorMachine,以下简称SVM）及其在回归领域的推广SVR,并将SVR应用于渭河水质遥感反演。由于SVR模型及模型参数（惩罚系数C、核函数参数σ²和不敏感损失函数宽度ε）的选择对模型的精度存在较大的影响,且参数优选目前缺乏理论上的指导。因此本文通过实验分析选择径向基函数作为SVR核函数,利用交叉验证（Cross Validation,以下简称CV）估计模型推广误差并使用遗传算法（Genetic Algorithm,以下简称GA）优选SVR模型参数,最后使用构建好的SVR模型对部分渭河陕西段水质变量进行反演。本论文从以下几个方面进行了研究和探讨:（1）分析和筛选渭河陕西段水质实地监测数据,得到符合条件且具有代表性的四类水质变量高锰酸盐指数（CODmn）、氨氮（NH₃-N）、溶解氧（DO）、和化学需氧量（COD）。对SPOT-5遥感影像的预处理做了重点分析,特别是针对大气辐射校正,在分析目前大气辐射校正研究现状的基础上,综合使用七种大气校正方法对本文遥感影像进行大气辐射校正。遥感影像的几何校正使用ERDAS软件实现。（2）阐明相关性分析在水质遥感监测中的作用和意义。针对本文的具体情况,分别对各水质变量之间、遥感数据各波段之间、水质变量与遥感数据单波段、水质变量与遥感数据波段组合进行相关性分析,在此基础上使用传统统计多元回归方法构建各水质变量多元回归模型,并对模型做了检验和分析。（3）介绍了SLT下的SVR理论,分析了SVR的特征,包括核函数的构造和模型参数的优选方法。结合本文数据自身的小样本特点以及模型参数优选的复杂性,采用CV估计模型推广误差并使用GA优选SVR模型参数。将基于GA优选参数的SVR回归模型用于渭河陕西段水域水质变量的遥感水质反演,最终构建各水质变量的遥感反演模型。通过与传统统计多元回归得到的结果分析比较,使用基于GA优选参数的SVR能够在较高的精度上对各水质变量进行预测。更多还原

【Abstract】 With the development of remote sensing technology and the fine-scale of remote sensing images,remote sensing technology is used more and more in water quality research.And the requirement of remote sensing of water quality monitoring is changed from qualitative to quantitative monitoring with more exactitude.Because the optical models of the main material in large-scale water,such as oceans,lakes etc.,are relatively mature,and the research of these materal is wide,currently the studies of the quantitative remote sensing of water quality mostly make in the area of oceans,inland lakes and large rivers.But few people do quantitative remote sensing research about water quality of Weihe River based on SPOT data.Weihe River flows across the cental of Guanzhong in Shaanxi Province,including Xi’an,Baoji,Xianyang,where are the cental of political, economic,cultural,financial and information in Shaanxi Province,because of its industrial concentration,population density,agricultural development.Besides those,Weihe River is the Main River and surface water in the area.Thus it is significant to make remote sensing monitoring of water quality of Weihe River.According to the reasons above,quantitative remote sensing study of the water quality using the French SPOT-5 remote sensing image is done,based on the Weihe River in Shaanxi Province.As to the structure of this paper,we firstly systematically introduce the relative researches and methods of remote sensing monitoring of water quality.Secondly,we find perfect methods in various processes and analyze these experimental results in detail.In the paper,we focus on atmospheric radiation correction methods of remote sensing image and the applications of Support Vector Regression （SVR） in the Weihe River water quality in quantitative remote sensing based on parameter optimization.For the method of correction of atmospheric radiation,the current methods are analyzed firstly,and then we choose an appropriate method to finish SPOT-5 remote sensing atmospheric radiation image correction based on its own remote-sensing images and the experimental conditions.For the application of Support Vector Regression based on parameter optimization in the Weihe River water quality in quantitative remote sensing,after analyzing the shortcomings of the Traditional Multiple Regression,we introduce the Support Vector Machine （SVM） and its promotion,which called SVR,in the field of regression based on Statistical Learning Theory（SLT）.Finally the SVR is applied to the inversion of Weihe River water quality remote sensing.The selection of SVM model and its parameters（such as penalty coefficient C,kernel function and the parameter of the kernel functionσ²,and the parameter of insensitive loss functionε） can greatly affect the accuracy of the model and there is absence of the guidance in theory.Thus we choose the Radial Basis Function as kernel function of SVR according to experimental analysis,use Cross-Validation（CV） to estimate the promote error and use Genetic Algorithm（GA） to optimize the parameters of SVR model.In the end,the SVR model we have built is used to inverse some water quality variables of Wei River in Shaanxi Province.The main works in this thesis are as follows:（1） The water quality field data of Weihe River in Shaanxi Province is analyzed and chose to get the eligible and representative water quality variables,including Permanganate Index（CODmn）, ammonia（NH₃-N）,Dissolved Oxygen（DO）,and Chemical Oxygen Demand（COD）.Besides that, we concentrate on analyzing the preprocessing of SPOT-5 remote sensing images,especially the correction of atmospheric radiation.Based on analyzing current research of it,we comprehensively use seven methods to finish the correction of atmospheric radiation in remote sensing images of this paper.And ERDAS software is used to realize the Geometric Correction of remote sensing images.（2） The role and significance of correlation analysis in water quality monitoring is clarified.We respectively analyze the relevance among the various water quality variables,among the wave bands of remote sensing data,between water quality variables and single-band remote sensing data, between the water quality variables and the wave bands combination of remote sensing data,based on which the multiple regression models of various water quality variables are built by using the method of traditional statistical multivariate regression.And finally the model is checked and analyzed.（3） The SVR theory which is based on the SLT is introduced and its chatecters are analyzed, including the construction of kernel function and the optimization method of model parameters.We use CV to estimate the promote error and use GA to optimize the parameters of SVR model.The SVR regression model which is based on GA optimization parameters is used to do the remote sensing of water quality retrieval of Weihe River in Shaanxi Province and then various remote sensing retrieval models of water quality variables are built.Compared with the Traditional Multiple Regression,SVR model which is based on GA optimization parameters can predict various water quality variables in higher accuracy.更多还原