任务驱动的遥感影像检索案例推理方法研究

发布时间：2018-03-21 12:30

本文选题：任务驱动　切入点：遥感影像检索　出处：《武汉大学》2014年博士论文　论文类型：学位论文

【摘要】：遥感信息作为保障国家安全与国民经济建设的一种重要战略性资源,在农业、减灾等众多问题的宏观决策方面发挥着不可替代的作用。随着对地观测技术的不断发展,遥感数据源将不断丰富,面临的关键问题之一是如何面向各类用户的不同需求,提供有效的遥感影像检索手段,实现遥感数据的快速获取、处理及高效服务。目前遥感影像的获取主要依赖于专业查询订购服务或者空间信息门户,其中都要求用户提交的查询中包含对专业遥感信息不同详细程度的定性或定量描述,不具备特定应用领域的遥感信息语义查询能力。因此,本文提出任务驱动的遥感影像检索概念,旨在通过遥感影像应用任务智能检索所需的遥感影像,简化其获取方式,提高其服务水平。然而,遥感影像应用任务与遥感影像之间的关系由于时空地理环境的复杂性,较难抽象为一种通用的规则或模型。因此,本文引入基于案例的推理技术,将遥感影像应用任务与遥感影像之间的复杂时空关系隐藏在案例中,并通过类比推理来利用这种隐藏的关联,实现以遥感影像应用任务为驱动智能检索出所需的遥感影像。具体研究工作及成果如下： (1)提出了任务驱动的遥感影像检索案例推理方法。该方法将难以抽象为规则的遥感影像应用任务、空间、时间与遥感影像之间的关系使用案例来表示,通过类比推理来利用这种关系。用户提交查询后通过相似性检索获得已有相似遥感影像应用案例,并对已有案例进行适当调整使之满足用户查询,最终返回用户所需的遥感影像。该方法较传统基于规则、本体的方法,其知识源更加丰富,获取难度更低,且更加容易表达特殊例外的情况。 (2)建立了以时间、空间、遥感影像应用任务以及遥感影像为主体的遥感影像应用案例语义表达模型。模型中的时间、空间、遥感影像应用任务以及遥感影像等元素都使用本体定义和描述,并对每个元素进行了特征建模,建立了其中的时空关系。模型包含概念模型和描述模型两个层次,概念模型描述概念对象的本质属性和关系,描述模型则是从自然语言层次描述对象的属性和关系。描述模型可通过语义推理实现向概念模型的转化,概念模型和描述模型的结合既保证案例具有足够的表达能力和推理能力,又搭建了从自然语言文本向计算机可理解文字转化的桥梁。 (3)建立了基于深层语义的遥感影像应用案例检索模型,包括相似性度量模型和组织检索模型。建立了遥感影像应用案例中时间、空间、遥感影像应用任务要素的局部相似性度量方法,以及这些元素构成的案例整体相似性度量方法。在时间相似性上,基于时间结构和遥感影像应用任务特点建立了时间语义相似性度量模型。在空间相似性上,将空间对象实例之间的空间关系划分为三类,并对应给出了基于空间关系联系强度、基于分类要素特征向量、基于图谱的相似性计算方法。在遥感影像应用任务相似性上给出了基于属性的相似性度量模型。建立了顾及空间特征的案例检索网络,扩展了虚拟空间、时间索引节点,空间索引节点使用双层R树实现,时间索引节点使用倒排结构,且提出虚拟案例的概念改进了原始扩展激活算法,使得其top-k检索性能与案例数量基本无关。该检索网络还支持不同用户背景下案例元素动态权重的检索。 (4)建立了基于知识的遥感影像应用案例调整模型。该模型一方面利用已有的时间、空间、传感器等本体领域知识作为调整知识,另一方面以案例库自身作为训练集,从中挖掘差异调整规则和频繁规则作为案例调整知识。其中重点对差异化调整规则的表达、学习算法进行了阐述,提出了基于概念要素细分的遥感影像应用案例差异化内容表达,以及顾及时空特征、语义特征的泛化准则和基于规则网的泛化算法；最后给出了面向对象的和查询驱动的调整知识应用方式。 (5)建立了任务驱动的遥感影像检索原型系统iGeoPortal,并在此基础上开展了验证实验,旨在验证本文提出的检索方法的性能及有效性。同时给出了基于信息抽取的遥感影像查询及案例的自然语言处理方法,建立了时空语义推理模型实现描述模型向概念模型的转化,提出了以查询为主体的地名定位方法和基于求交的模糊地名空间范围快速计算方法。初步试验表明,本文提出的任务驱动的遥感影像检索CBR方法,及其中的关键技术是有效和可行的。
[Abstract]:The remote sensing information as an important strategic resource security, national security and national economic construction in agriculture, plays an irreplaceable role in macro decision of disaster reduction and many other issues. With the development of earth observation technology, remote sensing data source will be continuously enriched, one of the key issues facing is how different needs for various types of the user, provide effective remote sensing image retrieval method, to achieve quick access to remote sensing data processing and efficient service. At present, the acquisition of remote sensing image mainly depends on the professional or query subscription service portal, which requires users to submit a query to include professional remote sensing information qualitatively or quantitatively different detailed description and query of remote sensing the semantic information does not have a specific application ability.
Therefore, the remote sensing image retrieval task driven concept, aims to retrieval of remote sensing image required by the application of remote sensing image to simplify the task of intelligent access, improve the service level. However, due to the complexity of air environment when the relationship between the application of remote sensing images and remote sensing images of the task, difficult to abstract a general a rule or model. Therefore, this paper introduces the case-based reasoning technique, the complex spatial and temporal relationships between the application of remote sensing image task and remote sensing images hidden in the case, and by using the analogy to hide, to realize the application of remote sensing image task driven intelligent retrieval of remote sensing images is needed. The specific research work and the results are as follows:
(1) proposed the case-based reasoning method of remote sensing image retrieval. The method of task driving will be difficult to abstract rules of the application of remote sensing image task space, the relationship between the time and the use of remote sensing image case represented by analogical reasoning to exploit this relationship. The user submits the query by similarity retrieval has obtained similar images application of the case, and the case has been adjusted to satisfy the user query, remote sensing image eventually return required by the user. The method is based on the traditional rules, the method of ontology, the knowledge source is more abundant, the difficulty of obtaining lower, and more easily express special exceptions.
(2) established by time, space, expression model of remote sensing image semantic application case of the application of remote sensing images and remote sensing image as the main task. The model of time, space, the application of remote sensing images and remote sensing image task elements using ontology definition and description, and each of the elements of the feature modeling, established the relationship between time and space the model contains. The conceptual model and description model of two levels, the concept model to describe the concept of object attributes and relationship description model is the description of object attributes and relations from the level of natural language. The description model can realize the transformation to the conceptual model by combining semantic reasoning, conceptual model and description model not only guarantee the case with sufficient expression and reasoning ability, but also to build a bridge to understand text conversion from natural language to computer.
(3) the deep semantic retrieval model is established based on the case of the application of remote sensing images, including similarity measure model and organization model is established. Retrieval time, the application of remote sensing images in the case of space, application of remote sensing image task elements local similarity measure method, and these elements constitute the whole case similarity measure method. In the similar time on the time structure and the application of remote sensing images based on the characteristics of the establishment of task time semantic similarity measure model. The spatial similarity, the space between an instance of the object spatial relations are divided into three categories, and the corresponding spatial relations is presented based on the contact strength calculation based on the classification of feature vector elements, similarity based on mapping method. In the application of remote sensing image similarity task gives the measurement model based on similarity attribute was established. Considering spatial characteristics of case retrieval network expansion Virtual space, time index node, node using double R tree spatial index, time index node using an inverted structure, and put forward the concept of virtual case improved the original extended activation algorithm, the retrieval performance of Top-k and the number of cases is independent. The network also supports the retrieval of different users under the background of case elements of dynamic weight retrieval.
(4) to establish the application of remote sensing image case adjustment model based on knowledge. On the one hand, the model uses the existing time and space, such as the adjustment of sensor knowledge domain ontology knowledge, on the other hand the case base itself as the training set, mining frequent difference adjustment rules and rules from as the case to adjust the knowledge which focus on the expression. On the differential adjustment rules, learning algorithm is discussed, put forward the concept of the application of remote sensing image segmentation element case expression based on the difference in content, and take into account the temporal and spatial features, semantic features and generalization criteria based on generalization algorithm rules; finally, the object oriented and query driven adjustment are given. The application of knowledge
(5) the establishment of remote sensing image retrieval task driven iGeoPortal prototype system, and carried out on the basis of experiment, to verify the proposed retrieval performance and effectiveness. At the same time gives the Natural Language Processing method of remote sensing image query and case information extraction based on established spatio-temporal semantic reasoning model description model to achieve transformation the conceptual model, put forward methods to query the placename location as the main body and the fast calculation of space intersection fuzzy placename method based on remote sensing images. Experiment results show that the proposed task driven retrieval method of CBR, and the key technology is effective and feasible.

【学位授予单位】：武汉大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：P237

【参考文献】