【作者】 王旭红；

【导师】 周明全；

【作者基本信息】 西北大学 ， 计算机软件与理论， 2005， 博士

【摘要】 随着传感器技术的发展,遥感影像的数量以飞快地速度增长。人们收集和存储影像的能力已经大大超过分析和从影像上获取信息的能力。这一切促使我们发展图像挖掘技术,它需各学科如图像处理、数据库、信息提取、机器学习和软件设计等同仁共同付出努力。图像挖掘旨在发现隐藏在数据库中含蓄的不明确的知识、影像数据的关系或其它模式,是数据挖掘的一个重要分支。 遥感图像数据挖掘(remote sensing image mining,(ReSIM))技术不仅是图像挖掘技术在遥感领域的应用,也是空间数据挖掘技术的一个重要拓展分支。它既要应用图像挖掘的一般性的理论和技术,又要结合遥感数据和空间数据的特殊性如独特的空间位置信息、复杂的空间关系和空间尺度,是空间数据挖掘与图像挖掘交叉的研究学科。其中,分类和预测方法是遥感图像分析和信息挖掘的重要研究内容,也是研究的重点。 本文围绕遥感影像信息自动化与智能化的获取和利用这一线索,对遥感影像数据挖掘理论和技术进行了研究,主要的研究成果和创新点如下: (1)研究了功能驱动和信息驱动两种图像挖掘模式,提出了信息驱动的遥感影像挖掘原型系统结构图和流程图,并指出系统应具备的功能和必要的工具。 (2)实现了最常用的两种分类器一监督分类(bayes)和非监督分类(Isodata)算法,并提出了bayes算法改良方法;实现了灰度共生矩阵纹理表示法;研究了图像中对象的空间结构和空间关系。 (3)在研究几种数据挖掘理论如模糊分类法、证据理论、人工神经网络(BP算法和SOM网络)、支持向量机、关联规则、决策树算法基础上,提出了基于这几种理论的遥感影像挖掘方法。 (4)上述的数据挖掘方法都是针对像元的图像分析,所能够得到的信息是极其有限的,不能够反映像邻域间的上下文信息(contextual information)。提出了“面向对象”图像挖掘方法,给出了该方法的流程和算法,实现了该流程和算法。 (5)在研究知识——颜色、纹理和边界等语义特征、混合光谱特征、高维数据的特征、GIS数据、地学专家知识等表示方法的基础上,提出了GIS数据辅助遥感图像数据挖掘的两种途径——以逻辑波段形式直接参与分类和融于空间数据库中系统化应用,并给出应用模型或系统框架结构;指出挖掘查询语言应为类似SQL的适用于地理信息挖掘的输入请求语言——GMQL(Geo-Mining Query Language);提出了知识库的表示方法;实现了基于规则化知识库遥感图像挖掘方法;实现了基于数据降维的高维数据特征提取算法。 (6)在分析Web环境下数据挖掘现状的基础上,提出了Web环境下图像挖掘系统框架图;进一步提出了Web环境下遥感影像数据挖掘系统框架图。更多还原

【Abstract】 The volume of remote sensed imagery continues to grow at an enormous rate due to the advances in sensor technology, and our capability for collecting and storing images has greatly outpaced our ability to analyze and retrieve information from the images. This motivates us to develop image mining techniques, which is very much an interdisciplinary endeavor drawing upon expertise in image processing, databases, information retrieval, machine learning, and software design. Image mining is armed to find the implicit and ambiguous informations, relationship of image data or other patterns in database.Remote sensing image data mining isn’t only the applications of image mining in the field of remote sensing, but also a important embranchment of the spacial data mining.It uses the ecumenical theory and technology of image mining as well as combines the particularity of remote sensing image and spacial data, such as the particular spacial locations, the complicated spatial relations and the spatial scale. It is a important crossed research subject between the spacial data mining and image mining. Finally, the method of classification and forecast is an important research field in the remote sensing image analysing and mining, also, is a reaearch focus of this dissertation.Encircling a key of remote sensing image information automatically acquired and used, mainly studies the theory and technology in remote sensing image data mining. The main reaearch results and creative proposals are following :(1) Studying two frameworks for image mining---an information-driven frameworkand a funtion-driven framework, presents a framework figure of a prototype system for remote sensing image mining and a flow chart of it. Also, points out necessaryfunctions and tools for the system.(2) Achieves the algorithm of two frequently used classification implement—a supervised and non.supervised, bayes and isodata algorithm. Also improves on the bayes algorithm. Achieves texture feature extraction by using gray-level co-occurrence matrices. Studys the special structure and relations of image objects.(3) On the basis of studying of a few of data mining theories, including fuzzy classification, evidence theory, artificial nerve network (BP and SOM), support vector machines, association rules and decision-tree algorithm, presents the method of remote sening mining based on the theorys.(4) The above-mentioned methods of data mining almost aims at the image analyzing of pixels. The information acquiring from pixels is greadtly limited, which don’t show the contextual information of neighboring pixels. In the dissertation, presents the " Object Oriented "image mining concept for the first time. Also, designs the flow and algorithm of the method and realizes it.(5) In research of the information of expression, including semantic feature, such as color, texture, border, as so on, spectral mixture feature, high- dimension data feature extraction,Gis data and expert information,presents two approach by using Gis data for image mining—straightway classifying in form of logistic waveband and systematizing applicat ion in the spatial database. Designs the using model or system framework. Points out mining Language would be the same with geographical information mining, resembling SQL input request Language—GMQL (Geo-Mining Query Language). Presents a expression method of a knowledge.data base. Realizes the remote sensing image data mining based on rule. Realizes high, dimension data feature extraction algorithm by using debasing dimension of data.(6) After analyzing the actuality of web data mining, brings forward a framework of the web image mining system. On second thoughts, puts forward a framework of the web Remote sensing image mining system.更多还原