【作者】 刘金梅；

【导师】 王国宇；

【作者基本信息】 中国海洋大学 ， 计算机应用技术， 2014， 博士

【摘要】 不断进步的遥感技术丰富了可供使用的数据资料。不同类型传感器获取的数据特点不同，多源遥感数据融合技术是综合利用各种数据提取全面地物信息的有效手段。目前，大多数遥感影像融合研究集中在像素级融合方面，特征级和决策级融合研究相对较少。本文提出了特征级和决策级融合的新方法，并将提出的方法应用于城市“热岛效应”研究和景观格局分析领域。本文主要工作如下：一、研究多特征融合方法。特征融合的常用方法是串行融合，串行融合只是实现了几何意义上的合并，没有真正有效利用特征之间的相互关系。文中介绍了两种多特征融合方法。本文首次将多视角谱嵌入算法用于遥感影像特征融合。多视角谱嵌入理论利用矩阵等价变换的原理，将不同特征空间的数据变换到同一表示空间，实现不同空间数据的融合。文章提出了基于特征接近度的融合算法。定义吸引力表示特征矢量与原型模式的空间接近程度。用特征矢量与各类型的原型模式之间的吸引力组成接近度矢量，作为融合特征。将这两种算法用于多光谱和高光谱遥感影像地物分类，实验证明，这两种特征融合算法对样本分类正确率超过90%，而且都能有效降低特征矢量维数。二、研究决策级融合方法。针对多分类器动态集成研究存在的一些问题，如算法的时间复杂度较高，集成过程中需要采用额外的动态机制，动态集成分类的稳定性差，提出了基于判决矩阵的决策融合算法。受行为知识空间融合算法启发，通过训练样本的先验知识提取分类器选择依据。训练样本输入到基本分类器中，得到基本分类器对输入类型的判断。在判决矩阵中记录基本分类器对输入样本类型的判断以及能够对输入正确分类的分类器序号。当输入待分类数据时，根据基本分类器的输出查找判决矩阵，得到样本可以信赖的分类器子集，按照服从多数的原则确定样本类型。分类器的动态选择机制记录在矩阵中，不存在人工参数设定问题，集成结果读取简单。在实验数据和基本分类器相同的情况下，用基于判决矩阵的融合算法和行为知识空间融合算法进行决策融合，基于判决矩阵的融合算法取得了更好的分类效果。三、以青岛市“热岛效应”研究和青岛市景观格局分析为例，使用本文中提出的方法处理实际问题。在青岛市“热岛效应”研究中发现，人口聚集区域的温度比崂山山区的地面温度高6°左右。研究结果表明，人类的聚集改变了周边环境，人类在活动过程中释放的热能加剧了城市“热岛效应”。景观格局指标显示，青岛市居民区斑块破碎度大，物质流通方便；绿地零星分布在居民区，对缓解城市“热岛效应”作用有限；林地斑块较多，海岸线长，自然条件优越。通过景观格局分析发现，青岛城市化进程对环境影响显著。未来城市发展规划制定应注重环境保护，增强城市宜居指数。更多还原

【Abstract】 The progress on remote sensing technology enriches available data. Data acquiredby various types of sensors are different. The technology of multi-source remotesensing data fusion is an effective means for the comprehensive utilization of variousdata.At present, mostly remote sensing imagery fusion research concentrates on thefusion methods of pixel level. Research on feature level and decision level isrelatively little. This paper proposes new fusion methods on feature level and decisionlevel, which are applied in the research of urban "heat island effect" and landscapeanalysis. The main work is as follows.(I) Multi-feature fusion method is studied. Serial fusion is a common used methodfor feature fusion, which is a combination in geometric meaning and can noteffectively use the correlation between features. In this paper, two algorithms aredescribed to fuse multi-feature. The multi-view spectral embedding is introducedto remote sensing imagery fusion for the first time. Multi-view spectral embeddingtheory is based on the principle of matrix equivalent to transform data fromdifferent feature spaces into the same representation space. Moreover, thepaper presents a novel fusion algorithm based on feature proximity. Attraction isdefined to represent the relationship of feature vector and prototype models. Theproximity vector formed by attraction between prototype models and feature vector isadopted to represent fused feature. The two algorithms are used in multi-spectral andhyper-spectral remote sensing image classification. Experimental results show thatthe two fusion algorithms classify the samples at the correct rate of over90%and they reduce the dimension of feature space significantly.(II) Decision fusion method is studied. To solve some problems in dynamicintegration of multiple classifiers such as high time complexity, additional dynamicmechanisms, instability of dynamic integration, a decision fusion algorithm based on judgment matrix is proposed. Inspired by the behavior knowledge space fusion, theselection of classifier is based on priori knowledge of the training samples. A trainingsample is inputted into the basic classifiers, and its corresponding type is judged bythe basic classifiers. Record the classification results and the classifier numbers thatcan correctly identify the input type in judgment matrix. In the process of dataclassification, the judgment matrix is accessed according to the output of basicclassifiers to find the reliable classifiers. The type of test sample is obtained inaccordance with the majority. The dynamic selection mechanism is recorded in thematrix so the problem of setting artificial parameters does not exist. It is easy toacquire the integration results. The fusion algorithms based on decision matrix andbehavior knowledge space are performed on the same basic classifiers. Experimentalresults show that the fusion algorithm based on decision matrix can achieve betterclassification results.(III) The proposed methods are used to analysis the "heat island effect" andlandscape pattern of Qingdao. In the research of "heat island effect", the groundtemperature in densely populated area is6degrees warmer than that in mountainousarea of Laoshan. The results show that the aggregation of human changes thesurrounding environment and energy release in the process of human activitiesaggravates the “heat island effect”. The landscape pattern indexes show thatresidential area patches are fragmentized, which makes material circulationconvenient. The green lands are scattered in residential areas so their ability to relieve"heat island effect" is limited. There are many forest patches and a long coastline,which makes Qingdao own nice natural conditions. The urbanization has a significanteffect on environment. Therefore, the future development plan should pay moreattention to environmental protection and enhance the city livable index.更多还原