【作者】 李光强；

【导师】 朱建军； 贺安生； 邓敏；

【作者基本信息】 中南大学 ， 地图制图学与地理信息工程， 2009， 博士

【摘要】 时空异常很可能是一类当前未知的、潜在有用的重要信息,代表地理现象或地理过程的特殊性。时空异常探测已在环境保护、地质灾害监测、地球物理勘探和化探数据分析、公众安全与卫生等领域受到很大关注。本文围绕时空异常探测理论与方法展开研究,提出了空间、时空异常探测的若干方法。主要研究内容包括:(1)在基础理论与方法方面,首先系统地回顾了现有空间、时空异常探测的研究成果,分析归纳了现有方法存在的问题;然后,总结了时空数据的特征及分类方法,进而分析了时空异常的特征,探讨了时空异常的分类方法;最后,讨论了时空异常探测的框架,并将异常探测分为异常探测和异常可靠性评估两个阶段。(2)针对基于统计学的异常探测方法存在的问题,提出了基于邻近域的空间、时空异常统计判别法,顾及了空间、时空的基本特性。该方法通过为每个对象建立空间、时空邻近域,并在邻近域内使用“k倍标准差”准则判别对象专题属性的异常性。进而,发展了有约束的空间异常探测方法。(3)针对现有空间聚类方法大都只考虑空间距离、忽略专题属性相似性的问题,本文提出了基于双重距离的空间聚类方法(DDBSC),将所有空间邻近且专题属性相似的空间对象聚为一类,并在聚类结果中探测空间异常。为适应空间局部密度差异的特性,提出了密度自适应的空间聚类方法(ADBSC),并与专题属性概念格方法结合,进而探测空间异常。(4)为了使用聚类方法探测时空异常,本文提出了基于专题属性概念格的时空聚类方法(CLBSTC)。CLBSTC综合运用ADBSC聚类和概念格构造方法,首先将时空上邻近的、专题属性概念格相同的对象聚为一类,然后在聚类结果中探测未归属任何时空簇的时空异常。(5)在运用智能计算探测空间、时空异常方面,本文将BP神经网络引入空间、时空异常探测过程,探讨了相应的BP神经网络的拓扑结构、学习样本的设计、学习规则等内容;然后将网络输出结果与原始数据的距离定义为异常度,进而探测空间、时空异常。通过多组实验表明,在输入向量中加入空间、时空聚类编号和异常数据项相关的专题属性项时,BP神经网络输出误差最小,探测的空间、时空异常最为稳定。(6)由于空间、时空数据本身和计算过程都不可避免地带有不确定性,因此需要对探测结果进行可靠性评估。本文将异常可靠性评估分为异常过滤和异常评价两个步骤,提出了基于关联规则的时空异常过滤方法,从候选异常集合中剔除所有符合关联规则的数据。为了定量评价时空异常的可靠性,本文在关联规则挖掘表的基础上,提出了异常支持度和置信度的概念,用于描述异常的可靠性。为了有效获取空间、时空关联规则,本文亦提出了基于Voronoi图的空间关联规则挖掘方法和基于事件影响域的时空关联规则挖掘方法。最后,总结了本文的研究成果,并展望了本文后续研究工作,主要集中在:(1)使用模糊集、决策树等理论,进一步研究时空异常的不确定性评价方法,(2)综合使用三维可视化和图表集成显示技术,发展时空异常的可视化方法。更多还原

【Abstract】 Spatio-temporal outliers (STOs for short) may contain some kind of potential and unknown knowledge about geographical phenomena. The detection of spatio-temporal outlier (STO for short) is very significant and necessary for better understanding spatio-temporal data, discovering the spatial relationships among spatio-temporal entities. Currently, many approaches have been proposed for the detection of STOs, and have been applied to many fields, such as weather, forest fires, geological disasters, environmental protection, public safety, and so on. This thesis focuses on the development of theories and methods about the detection of STOs, and all are summarized as follows:(1) After overview of existing research results about the detection of STOs, the characteristics of spatio-temporal data (STD for short) are summed up and the STD classification method is presented. And then, the characteristics and classification method of the STOs are studied. The framework of the STOs detection is explored, which includes the STOs candidates detecting step and the STOs evaluating step.(2) To solve the problems in the traditional statistical method for the detection of outliers, the method is expanded up to spatio-temporal domain and the nearest-neighbors and statistical-criteria based spatial outliers (SOs for short) detection (NNSCBSOD for short) is developed. The NNSCBSOD employs the k-times-standard-deviation rule in the each nearest neighbor to discriminate the outlying-ness of the spatial object.(3) For using cluster-analysis to detect STOs, since the existing spatial and spatio-temporal clustering methods only considers the spatial or spatio-temporal distance, while ignoring the thematic attributes, the dual-distance based spatial clustering methods (DDBSC for short) is proposed, which form all the adjacent and attributes similar spatial objects into a spatial cluster. Then, in the clustering results, all the isolating objects emerge and compose the SOs set. In view of the existing spatial clustering methods not adapting the uneven distribution of spatial data, an adaptive density-change based spatial cluster (ADBSC for short) method is proposed, too. The ADBSC and the concept-lattice approach are combined and utilized to detect SOs.(4) In order to use Clustering method to detect STOs, a concept-lattices based spatio-temporal Cluster (CLBSTC for short) is raised in this dissertation. The CLBSTC synthetically uses the ADBSC and concept-lattices approach to discovery the spatio-temporal clusters, and then, form all the spatio-temporal adjacent objects within the same concept lattice into a spatio-temporal cluster (STC for short). In the clustering results, all objects, not belonging to any STC, are STOs.(5) To employ the intelligent computing technology to detect STOs, this dissertation introduces the back propagation (BP for short) neural network into STO-detection procession, and STO detection neural network (STODNN for short) is described. Then, the constructions, learning samples design, learning rules about STODNN are discussed. After that, a STO measure is put forward via using the distance between the network output and the original data. Depending on the STO measure, the STOs are detected. Many experiments showed that the BP neural network output errors are smallest when using the input vector including the STC number and related attributes items, and STOs detection results are most stable.(6) Because of uncertainty in the STDs and the calculating process, it is required to evaluate the reliability STOs. In this dissertation, the reliability evaluation process is divided into two steps: STO filter and evaluation. Furthermore, the association-rules (ARs for short) based STO filtering methods (ARBF for short) is proposed. The ARBF prunes all the STOs candidates, which consistent with the ARs. In order to evaluate the reliability of the STOs, basing on the mining table for ARs, the outlying support and the outlying confidence are defined to measure the reliability of the STOs. In the end, Voronoi-diagram based spatial ARs mining method (VDSAR for short) and event-coverage based spatio-temporal ARs mining method (ECSTAR for short) are proposed so as to discovery space and spatio-temporal ARs.Finally, after concluding all the research work in this dissertation, some deficiencies and the following work are discussed, which include (1) employing fuzzy sets, decision trees, etc. to further study evaluation the uncertainty methods for the STOs, and (2) integrating the 3-dimensional visualization and chart technologies to visualize the STOs.更多还原