【作者】 傅文杰；

【导师】 戴塔根； 洪金益；

【作者基本信息】 中南大学 ， 国土资源信息工程， 2006， 博士

【摘要】 西部高寒山区是我国极具找矿潜力的重要成矿区域。以遥感技术为先导，结合地质、地球物理、地球化学勘查技术，在西部高寒山区进行矿产资源勘查是快速、有效的勘查方法。矿化蚀变信息提取是遥感矿产勘查中最关键的技术，但是由于矿化蚀变信息在遥感图像上是一种弱信息，使用传统的信息提取方法效果往往不尽人意。因此，不断研究有效的遥感蚀变信息提取新技术新方法，提高遥感找矿的可信度和效益，具有非常重要的现实意义。 支持向量机(Support vector machine, SVM)是机器学习界的研究热点，并在很多领域都得到了成功的应用。光谱相似尺度(Spectralsimilarity scale, SSS)是一种能同时度量光谱向量的大小和方向差异的新方法，能提高对光谱向量间相似性描述的准确性。 本文基于国家“十五”科技攻关计划项目(2003BA612A-04)中的子课题“SVM遥感数据矿化信息提取技术研究”，对遥感矿化蚀变信息提取中的两种新方法—SVM和SSS进行了研究。 研究工作中首先介绍了遥感矿化蚀变信息提取及SVM的研究现状，接着讨论了SVM和SSS理论基础及基本概念和要解决的关键技术问题，然后选择几个典型矿化区段作为试验区，对这两种新方法在遥感矿化蚀变提取信息中的应用进行了系统的研究，主要的研究内容概括如下： 利用生成的遥感模拟影像对SVM应用中的几个关键问题进行了的实验分析。实验中分别研究了不同的核函数类型、核参数、错误惩罚因子、多类算法及训练样本数目对SVM分类器性能的影响，实验结果表明核参数和惩罚因子比核函数对分类精度影响更大。 探讨了模型选择方法中较实用的网格搜索及交叉验证法的基本原理和方法，对网格搜索法确定支持向量机核函数参数和惩罚因子的过程作了详细描述，确定了SVM在遥感矿蚀变信息应用中模型参数的选择方法。 将SSS应用于遥感矿化蚀变信息提取中。通过对提取结果的验证及与传统提取方法对比，表明该方法有利于减少异物同谱现象的影响，提高遥感蚀变信息提取的精度。 将SVM方法应用于遥感矿化蚀变信息提取，通过野外实地验证和更多还原

【Abstract】 The western high-cold mountainous districts are the most important mineral-forming areas, with extremely high mineral exploration potential. Using Remote Sensing technique and combining with geology, geophysics and geochemistry technology, mineral exploration in high-cold mountainous areas is fast and efficient way. Altered rock’s information extraction is the most important technique in Remote Sensing mineral exploration. However altered rock’s information is weak in Remote Sensing image, sometimes the effect with traditional information extraction approachs is not very good. So we can see that it is considerably significant to explore new approach of mineral resources searching and evaluating in accordance with west’s natural condition.Support Vector Machine(SVM) is a hotspot in machine learning field and has been successfully applied in different fields. Spectral Similarity Scale(SSS) is a new approach which can simultaneously measure the magnitude and direction of spectral vectors, and it can improve the veracity of similarity description among spectral vectors.Based on the subordinate subject of the brainstorm state projects of the "Tenth Five-year Plan", two new approachs-SVM and SSS in Remote Sensing altered rock’s information extraction are discussed.Firstly, the paper introduces the present researching situation of Remote Sensing altered rock’s information extraction. Secondly, it discusses the SVM basic theory, concept and key technical problems by using SVM and SSS, and then the paper systematically researches the new approachs in the application of Remote Sensing information extraction selecting several typical mineralization districts as testing areas. The following research work has been done:Experimental attestation is made to several key problems in the application of SVM using inborn Remote Sensing simulative images. Author researches how different kernel functions, kernel parameters, penalty parameters, multi-classifying algorithm and the number of training examples affect the performance of SVM classifier. The result更多还原