【作者】 李三平；

【导师】 李德玉； 葛咏；

【作者基本信息】 山西大学 ， 计算机软件与理论， 2006， 硕士

【摘要】 客观世界本身内在的不确定性和人类对客观世界认知的局限性,是导致科学对客观真实的描述存在大量不确定性的主要原因。不确定性问题以其普遍性和现实性,成为当前研究领域中的一个难点和热点。 遥感的研究对象是空间实体,描述空间实体的空间数据中,包括与时间有关的位置数据和属性数据,都存在着大量不确定性,有的甚至严重影响产品的可靠性。数据质量是数据的核心,而数据质量的关键则是对不确定性的评估。 如何全面、准确地度量和可视化表达遥感信息处理中不确定性的程度和空间分布是遥感信息不确定性研究的关键问题之一。传统的度量方法,例如误差矩阵和Kappa系数,直接在训练数据上估计分类模型的性能。而事实上这样做并不合适,不能将以训练样本集为基础的误差矩阵当作总分类精度的量测尺度,我们需要估计模型对于“样本外数据”的性能。因此我们提出将粗糙集理论作为度量遥感信息属性不确定性问题的应用框架,对样本区数据和分类影像中的像元、目标和影像整体等不同尺度上的空间对象进行属性不确定性度量,并以我国黄河三角洲地区的Landsat5TM遥感影像进行不确定性度量的实例分析。 由于混合像元而造成的影像不确定性属于遥感技术固有的不确定性特征,产生混合像元的原因则主要是遥感影像数据的栅格编码格式与传感器的空间分辨能力和空间尺度等因素。混合像元的存在限制了传统像元级遥感分类的精度,而这些分类方法中又都很少考虑整个影像内相邻像元或者该像元与周围像元间的关联关系,即空间数据的空间依赖或空间自相关特性。论文根据遥感影像的空间自相关特性,提出混合像元内部亚像元空间分布的模拟算法,利用像元内部基本组分的面积比信息,在亚像元尺度上模拟混合像元内部的地物空间分布,实现了进入像元内部,在亚像元级别上确定地物边界的目的。同时利用人工影像数据和合成影像数据对模拟算法的性能和效率在不同配置的硬件平台上进行了测试。 遥感信息中存在大量不确定性,如果只是仅仅对其进行不确定性程度上的数学度量,而忽略其在空间域和时间域上的分布特征,则很难完整准确地描述和理解遥感信息中的不确定性。从视觉的角度表现不确定性信息是不确定性建模的重要组成部分,不确定性可视化技术能够辅助用户探查原始数据的不确定性以及不确定性的大小、分布、空间结构和趋势,使用户简单明确地意识到数据不确定性的存在,以更多还原

【Abstract】 The inherent uncertainties in the real world and the limitation of human cognition are primary factors of the abundance of uncertainty existing in the judgment of sciences to objectivities. Due to its universality and reality, the uncertainty problem is becoming a hotspot and a difficult issue in the present research fields.The research target in remote sensing is the spatial entity. An amount of uncertainties exist in the spatial data describing the spatial entity, including temporal-relative location data and attribute data, some of which may seriously distort the reliability of products. Data quality is the core of data, and the key of data quality is evaluating uncertainty.One of the key points in research of uncertainties in remote sensing is measuring and visualizing the degree and spatial distribution of uncertainty completely and accurately in the processing of remotely sensed image. In the classical fashions, e.g. error matrix and Kappa coefficient, the performance of the classification models is estimated directly on the training data. Whereas it is actually not appropriate. The error matrix based on the training data set can not be regarded as the measurement of overall accuracy of classification models, and these models’ performance need to be evaluated on "out-of-sample-data" - data that have not been used in constructing the models. In this paper, we apply the rough sets theory as the application framework of measuring the attribute uncertainties in remote sensing information, and several measures are proposed for assessing the attribute uncertainties in sample area data and different spatial objects based on the scale of pixel, landcover class and the whole image in classified remotely sensed imagery. These measurements could measure effectively attribute uncertainties and facilitate to trace the propagation of error and uncertainties in classified remotely sensed data. Subsequently, the remotely sensed imagery of Landsat 5 TM located at the Yellow River Delta is utilized as a case study of uncertainty analysis.Remotely sensed images often contain a combination of both pure and mixed pixels. The presence of mixed pixels is primarily due to the raster encoding format of remotesensing data, characteristic of the sensor’s response to radiation, spatial resolution of the sensor, and spatial scale. The uncertainty in the imagery resulted from the aforementioned reflects the inherent characteristic of uncertainty kept in the technique of remote sensing. The existence of mixed pixels restricts the accuracy of classical classifiers of remotely sensed information based on the level of pixel, and these classifiers rarely consider the relationship of spatial conjunction between the neighboring pixels in the whole image, in other words, the spatial dependency of spatial data or spatial autocorrelation. In this paper, a novel simulating algorithm for identifying the spatial distribution of subpixels in a mixed pixel is proposed based on the assumption of spatial autocorrelation. Cooperating with the information on the proportions of every endmember component presented within a pixel derived from a soft classification, the algorithm simulates the spatial distribution of landcovers within the mixed pixel, and implements crossing through the pixel and identifying the boundary between the field objects at the subpixel scale. Subsequently, the performance and efficiency of the simulating algorithm is tested on different configuration of hardware platforms with the artificial imagery and synthetic imagery respectively.Large amount of uncertainties exist in the remote sensing information. If only mathematics measurement on the uncertainty degree of the remote sensing information is utilized, while the distribution characteristics in the temporal and spatial domains is ignored, the uncertainties in remote sensing information will be difficult to be completely and accurately described and understood. Representing uncertainty information from the point of view of visual sensation is an important section of modelling uncertainty. The uncertainty visualizing technology will help the users to explore uncertainty of the raw data and the size, distribution, special structure and tendency of the uncertainty. Then, the existence of uncertainty in data and the effects of uncertainty on the final decision would be simply and exactly felt by the users and products of remote sensing can be better understood and employed. In this paper, methods and technologies of uncertainty visualization are discussed and applied by the compartmentalization of static and dynamic visualized variable and feature visualization. The technique of the parallel coordinates is applied with emphasis in the feature space of remote sensing information to visualize the imagery’s character of spectral and uncertainty, and then some visualizing examples of uncertainty measurements in the classified remotely sensed imagery are provided.At last, the design and the flow of data handling of the software UnVis for analyzing uncertainty in the remote sensing information are described, and the performance of the two finished functional modules of attribute uncertainty measuring and subpixel spatial distribution simulating is tested.更多还原