【作者】 杜根远；

【导师】 苗放；

【作者基本信息】 成都理工大学 ， 地球探测与信息技术， 2011， 博士

【摘要】 随着全球立体对地观测系统的逐步形成和完善,空间数据的数量、大小、复杂性和传输速度都在飞速增长,全球、海量、多源是其显著特征,其中,遥感图像数据是应用最为广泛的一种空间数据。目前,遥感应用的水平滞后于空间遥感技术的发展,从而造成空间数据资源的巨大浪费,其应用价值得不到充分利用,形成了空间数据的生产和传输能力远远大于空间数据解析能力的局面。研究海量遥感图像数据的高效组织与快捷应用、快速检索有效空间信息、提高遥感图像分析识别的精度,是目前遥感应用中亟待解决的问题,具有十分重要的科学意义和应用价值。解决这一问题的关键是发展有效的空间数据管理和内容检索方法,这也是近年来海量遥感图像检索所面临的瓶颈之一。目前,图像内容检索技术取得了一些研究进展,但是针对遥感图像内容检索的研究却相对缓慢,无论是理论体系还是应用系统,都还不成熟,遥感图像具有尺度大、主题不明确、多时相、语义丰富等特点,普通图像中的研究成果不能直接应用于遥感图像内容检索中去。对于一个完备的遥感图像内容检索系统,其数据组织、存储与管理、特征描述及提取、相似性度量、网络服务模式、系统架构设计及实现等研究工作面临着许多困难与不足,研究所涉及的各项关键技术势在必行。本论文针对海量遥感图像内容检索所涉及的关键技术,提出了一些创新性思路和方法,并分别从理论和技术的角度对其价值和实用性予以分析和验证。论文的主要创新性研究成果和贡献如下:(1)提出了一种结合进化聚类和模糊C均值聚类的遥感图像分割方法,并提出了一种基于改进模糊C均值聚类的遥感图像序列分割方法结合进化聚类和模糊C均值聚类算法,提出了一种遥感图像分割方法(Evolving Clustering-Fuzzy C-Means, EC-FCM)。利用ECM解决模糊C均值聚类算法的初始化中心选择问题,再利用FCM算法对获得的聚类中心进行优化,完成模糊聚类划分,通过去模糊化转换为确定性分类,实现聚类分割。在上述方法基础上,提出了一种基于改进FCM聚类的遥感图像序列分割方法(Sequence Segmentation Method, SSM)。颜色空间选用相关性更低的HSI空间,采用更适合遥感图像的基于标准协方差矩阵的Mahalanobis距离,利用进化聚类解决FCM算法的初始化中心选择问题,并根据策略对图像进行序列分割。理论分析及编程实验结果表明,上述方法同FCM算法比较能以较少的迭代次数收敛到全局最优解,有较好的分割效果,能够有效地提高遥感图像阈值分割的精度和效率,可用于遥感图像分类和基于内容的遥感图像检索系统中。(2)提出了一种可用于内容检索的基于粒计算的图像区域相似性度量方法基于粒计算理论,提出了一种可用于内容检索的图像区域相似性度量方法(Image Region Similarity Measure, IRSM),将图像特征信息表转化为有序矩阵形式,通过对有序矩阵进行研究,引入图像特征粒、?阶粒库的概念,从不同的粒度层次分析图像特征的重要性,保持了图像特征信息表中区域间的序关系,并基于粒计算理论给出图像特征的权值,实现了一种图像区域相似性度量方法。实例表明,该相似性度量方法能客观、有效的度量图像区域间相似程度,为粒计算理论在遥感图像内容检索中的研究提供了一种新的思路和方法。(3)提出了一种在G/S模式下的空间剖分数据存储调度服务模型结合客户端聚合服务的空间信息网络服务G/S模式和地球剖分组织理论,提出了一种在G/S模式下的空间剖分数据存储调度服务模型,给出了空间剖分数据网络服务体系的架构、数据访问流程,设计了剖分数据存储调度服务模型的地址编码结构及地址解析过程,形成了一种有效的“数据分散存储,客户端信息汇聚”的空间剖分数据组织管理、按需整合、快捷调度机制。经原型测试对上述思路进行了部分验证,具有访问速度快、数据更新容易、对大数据适应的特点,能有效解决海量遥感图像数据的组织效率瓶颈和快捷应用难题,对于开发遥感图像内容检索系统具有一定的理论意义和应用价值。(4)设计了一种适合内容检索的分布式遥感图像数据库,建立了一个遥感图像内容检索原型系统基于Oracle Spatial和GeoRaster,设计了一种适合内容检索的分布式遥感图像数据库;利用Visual C++语言和OCCI(Oracle C++ Call Interface)接口,设计并实现了遥感图像内容检索原型系统(Content-based Remote Sensing Image Retrieval, CBRSIR),提供了一定的查询检索功能,作为本论文研究的算法测试平台及实例证明。运行结果表明,网络资源占用较小,系统效率受数据库所在服务器内存大小影响,在保障检索成功率的基础上,检索性能有一定的提升。更多还原

【Abstract】 With the gradual development and perfection of the global stereo earth observation system, the amount, size, complexity and transmission rate of spatial data which possess the characteristic of globalization, mass and multisource are growing rapidly. The remote sensing image data is the most widely used kind of spatial data. At present, the application of remote sensing technology lags behind its development and results in a tremendous waste of the spatial data resources. This forms the situation that the production and transmission capacity of spatial data is much greater than the analysis capacity of spatial data. So the problems, having significant theoretic and application values, need to be solved urgently are the effective organization and quick application of large-scale remote sensing image data, the quick search of effective space information and the promoting of remote sensing image analysis and recognition accuracy.The key to solve these problems is finding an effective way to manage the spatial data and to retrieve the content of spatial data, which is also the bottleneck of large-scale remote sensing image retrieval in recently years. At present, the study on the technology of content-based image retrieval has made some progress, but on the content-based remote sensing image retrieval the progress is relatively slow, no matter on theoretical system or application system. The research achievement of general image can not be used directly in content-based remote sensing image retrieval because that the remote sensing images are large scale, vague, multidate and semantically rich. The study on the organization, storage, management, description and extraction of data, similarity measure, the network service mode and the design and realization of system architecture are facing many difficulties and shortages for a complete remote sensing image content retrieval system.In this thesis, we put forward some innovative ideas and methods related to the key technology of large-scale content-based remote sensing image retrieval and further verify their value and practicality from theoretical and practical aspect respectively. The innovative achievements and contributions of this thesis are as follows.(1) This thesis puts forward a new method, which combines ECM with FCM, to segment the remote sensing image. Based on this method, we propose a new method for remote sensing image sequence segmentation on the basis of the modified FCM.Combining ECM with FCM, this thesis puts forward a new remote sensing image segmentation method evolving clustering-fuzzy C-means (EC-FCM). ECM is used to choose the initialized center of the fuzzy C-means clustering algorithmic, and then optimize this cluster center by using the FCM to accomplish the division of fuzzy clustering. Finally the genetic clustering can be realized by changing the fuzzy clustering into the certain category through the defuzzification.On the basis of the proposed theory, this thesis further puts forward a new method of the remote sensing image sequence segmentation based on the modified FCM (SSM). This method adopts the low relativity HSI space, the Mahalanobis distances which is more suitable for the remote sensing image. According to this method evolutionary clustering is used to choose the initialized center of the FCM algorithmic, further the image is segmented according to the strategies.Both the theoretical analysis and the results of experiment show that the proposed method, compared with FCM algorithmic, can converge to the global optimum solution with few iterative times, can effectively improve the precision and efficiency of remote sensing image threshold segmentation, and can be applied in the classification of remote sensing image and content-based remote sensing image retrieval system.(2) Based on the granular computing, this thesis proposes a new method of image region similarity measure (IRSM) which can be used in content retrieval. This thesis proposes a new method of image region similarity measure (IRSM) which can be used in content retrieval on the ground of the granular computing. On the basis of the granular computing theory we convert the characteristics information of the image into the ordered matrix. Then the conception of feature granular, ? -order granular base are introduced based on the study of ordered matrix. The importance of the image features are analyzed from the different level of granularity so as to keep the order relation among the regions in the image feature information list. Further the weight of the image feature is given.An example shows that using this method we can measure the image region similarity objectively. Further this method provides a new way to use granular computing theory in the study of the content-based remote sensing image retrieval.(3) This thesis proposes a new spatial subdivision data storage and scheduling service model in the G/S mode.Combined with the service mode of client aggregation services and the global subdivision theory, this thesis proposes a new spatial subdivision data storage and scheduling service model in the G/S mode, provides the framework, the data access process of the spatial subdivision data network service system, designs the address coding stricture and the address resolution process of the spatial subdivision data storage and scheduling service model, shapes the mechanism of management, integration and scheduling of spatial subdivision data. The proposed method is partly verified through the prototype testing, and the verification results indicate that this prototype has a high data access speed, can update easily, and is especially suitable for the large-scale remote sensing image data, beside this method can effectively solve the bottleneck of organizational efficiency and the quick application of mass remote sensing image data. We can infer that it has theoretical and practical value in the development of content-based remote sensing image retrieval system.(4) A distributed remote sensing image database which is suit for the content-based retrieval is designed in this thesis, at the same time a content-based remote sensing image retrieval prototype system is built.Based on the Oracle Spatialand and GeoRaster, a distributed remote sensing image database which is suit for the content-based retrieval is designed in this thesis. A CBRSIR prototype system is designed and realized by using VC++ language and Oracle C++ Call Interface to provide certain retrieval functions as the testing environment and actual example of the study.The running results indicate that the efficiency of the system depends on the size of the server memory, the system consumes less network bandwidth, and the retrieval performance is definitely improved.更多还原