【作者】 俞乐；

【导师】 徐世浙； 张登荣； Eun-Jung Holden； Alok Porwal； Michael C.Dentith；

【作者基本信息】 浙江大学 ， 地球探测与信息技术， 2010， 博士

【摘要】 星载和机载遥感数能有效用于岩性信息提取与制图,特别是在植被覆盖稀少的干旱、半干旱地区。随着传感器技术的快速发展,高空间分辨率、光谱分辨率和时间分辨率的遥感数据源类型不断增加。由于不同岩石单元在不同类型数据源上有不同的反应,综合多源遥感数据(包括航空地球物理数据)的各自优势,能克服单一数据源的不完全性和不确定性,增强岩性信息以利于准确、可靠的岩性制图。但是采用多源遥感数据进行自动岩性信息提取还存在较多的技术障碍。首先,面对海量增加的遥感数据,以人工作业方式为主的常规遥感图像配准技术难以满足快速预处理的要求,这将制约岩性等专题信息的提取效率。如何将不同分辨率、不同波段、不同时相等的多源遥感数据进行快速自动配准是当前遥感图像处理的难点。其次,植被覆盖是遥感岩性信息提取的一个常见障碍。常用的图像增强处理,只能凸显地表信息；而少数能对植被进行抑制的增强方法,也存在对实地光谱数据或特定地质现象(如蚀变)的依赖。亟需提出一种更加通用的,对原始数据要求更少的植被抑制技术,以满足覆盖区岩性增强的要求。再次,采用传统的遥感影像分类方法对具有高维特性的多源遥感数据进行岩性信息自动提取时,会陷入“维数灾难”,即分类精度会随着维度的增加而降低。因此,需要探索和试验能克服这一现象的分类器用于岩性自动提取。本文针对当前多源遥感数据岩性信息自动提取中的上述关键技术问题,以光学遥感和航磁数据为例,采用计算机视觉和机器学习领域的方法,开展岩性信息自动提取研究,取得了以下主要成果：(1)综合采用SIFT算法,Harris角点检测,小波金字塔方法,交叉匹配法以及TIN纠正等构建了适合于多源遥感数据的全自动快速配准流程。通过对不同类型高分辨率和中等分辨率遥感影像的试验,表明了本文方法的有效性。(2)提出了“掩模-强迫不变”植被抑制方法,通过对杭州地区的Landsat ETM+数据的试验表明本文提出的方法能显著抑制原始影像中的植被影响。(3)提出了采用SVM模型综合ASTER、DEM和航磁数据进行自动岩性分类。试验表明,SVM的岩性分类精度以及分类结果与岩性图的相似度均高于最小马氏距离法、最大似然法和神经网络(包括BPNN和PNN)等分类方法。另外通过对美国内华达州Cuprite矿区AVIRIS高光谱数据岩性分类的试验也表明了SVM模型的有效性。(4)采用了SVM-RFE对ASTER-DEM-航磁数据岩性分类进行特征选择。试验表明,该技术能在不显著降低原始分类精度的前提下,将原始特征维度从36个降低到19个,并且提供了适于岩性信息提取的知识。本文的主要创新成果和新进展有：(1)建立了基于SVM的多源遥感岩性信息提取模型,为遥感地质应用中的高维数据处理提供了新的技术手段；并采用了SVM-RFE方法对参与岩性分类的多源数据特征进行知识挖掘；(2)在遥感数据预处理和信息增强方面,提出基于点特征的多源遥感影像快速自动配准方法和“掩模-强迫不变”植被覆盖区地质信息增强技术。更多还原

【Abstract】 Remote sensing data from space-and air-borne sensors can be effectively used for lithological classification and mapping, especially in areas of high outcrop density in arid regions. With the rapid advances in sensor systems, remote sensing data are now available with high spatial, spectral, temporal, and radiometric resolution. Lithological units generally have distinct responses in different remote sensing datasets, hence integration of data from multiple remote sensors, including air-borne geophysical sensors, can potentially provide enhanced information about lithologies and facilitate more accurate and reliable lithological mapping. However, the integration of multi-source remote sensing datasets to extract lithological information is not straight forward, and requires rigorous pre-and post-processing. Firstly, co-registration of different remote sensing images is a pre-requisite before applying integration and classification algorithms. However, conventional manual co-registration is cumbersome especially with large volumes of remote sensing data. Automated co-registration, on the other hand, requires addressing complex problems such as varying illuminations and resolutions of the images, different perspectives and local deformations within images, etc. Secondly, vegetation cover is an impediment to lithological mapping from remote sensing data, and, in order to enhance the underlying geological information in such terrains, it is desirable to suppress the reflectance component of vegetation. Thirdly, supervised classification algorithms are conventionally used for lithological classifications, but classifiers such as maximal likelihood, neural networks etc. are not capable of learning complex patterns in high dimensional feature spaces, which is a critical characteristic of multi-source datasets.This thesis aims to provide new solutions to the above issues by integrating methods from computer-vision and machine-learning domains for automatic lithological information extraction from multi-source remote sensed datasets (take optical remotely sensed datasets and aeromagnetic dataset for example). Major contributions of this thesis are as follows.(1) A fully automatic and fast non-rigid image registration technique for multi-source imagery is developed. It incorporates the scale invariant feature transform (SIFT) method, the Harris corner detector, wavelet pyramid, cross-matching strategy and triangulated irregular networks (TINs). Experiments with multi-source high and moderate resolution remote sensing images demonstrate the efficiency and the accuracy of the proposed technique.(2) A "masking-forced invariance" algorithm is proposed for the suppression of the vegetation reflectance component in a densely vegetated study area. An evaluation based on comparison with the geological map shows that the forced invariance technique results in significant enhancement of the lithological information in the processed image.(3) Support vector machine (SVM) algorithm is applied to integrate Advanced Space-borne Thermal Emission and Reflection Radiometer (ASTER) imagery with ASTER-derived digital elevation mode (DEM) and aeromagnetic data to implement automated lithological mapping of the study area. Experiments with other supervised classification techniques such as maximum likelihood classifier, minimum distance classifier and neural network classifer (including BPNN and PNN) show that the SVM provides higher accuracy both in terms of classification of independent test samples as well as similarity with the available bed-rock geological map. Another experiment on AVIRIS hyperspectral data in Cuprite, Nevada, USA also shows the effectiveness of using SVM for lithological classification.(4) A feature selection algorithm based on support vector machine called SVM recursive feature elimination or SVM-RFE is used to find the importance of different input features in SVM-based lithological classification. The experimental results for the above lithological classification case study indicate that the feature subset generated by the SVM-RFE is acceptable in terms of classification accuracy. The accuracy remains almost the same when only 19 features out of 36 are reserved. SVM-RFE not only is proved to be an effective tool to shrink high dimensionality dataset, but also is a useful tool to rank the importance (knowledge) of different inputs.In summary, the main innovations of this thesis are:(1) a SVM-based automatic lithological classification method for multi-source remote sensing datasets, namely, ASTER, DEM and aeromagnetic; then a SVM-RFE based data mining method to shrink data dimensionality and discover knowledge of lithology classification from input datasets. (2) in remote sensing image pre-processing and enhancement procedures, a fast and fully automatic registration technique based on point features for multi-source remote-sensing image; and a "masking-forced invariance" approach for suppression of vegetation in multispectral remote sensing images.更多还原