导　　师： 刘艳芳

学科专业： 120405

授予学位： 博士

作　　者： ;

机构地区： 武汉大学

摘　　要： 面向土地利用的高分辨率遥感影像分类是一个传统而复杂的课题。其困难存在于两个方面：一是分类过程和结果如何更好地面向GIS专题应用；二是高分辨率遥感影像中存在大量的“同物异谱”和“异物同谱”现象,增加了分类的难度。通过研究人类的目视解泽过程和特点,我们发现解决这两个问题的关键在于获得有效的专题知识,特别是空间知识来辅助分类。因而,全文的研究围绕着三个问题展开：需要什么样的土地利用空间知识,如何挖掘和表达土地利用空间知识、如何应用这些土地利用空间知识实现分类。 构建分类的模型框架是研究的首要内容,它为解决上述三个核心问题提供了总体策略和指导。其基本思想是模拟人类的目视解译行为：首先,对土地利用空间知识的概念和来源进行了总结,从空间知识的尺度效应原理入手,讨论了本研究所需要的知识类型、重点研究它们之间的内部关联机制,从而建立起土地利用空间知识的认知金字塔；第二,引入更加灵活的算法从GIS数据及遥感影像数据中挖掘深层次的知识,并讨论如何利用GIS、RS与空间数据挖掘技术之间的集成模式来实现这些算法；第三,分析了蚁群智能优化算法在模拟人工目视解译方面的优势和原理,并初步探讨了如何将获取的土地利用空间知识融入到其中。最终,形成了基于数据挖掘技术的土地利用空间知识提取、土地利用空间知识渐进式表达、融合空间知识的蚁群优化算法分类三个核心的模型内容。 从GIS数据库和RS影像中挖掘有用的土地利用空间知识是本研究的核心内容和基础内容。在这一过程中,本文根据人类在遥感分类过程中的认知习惯,采取渐进的方式,综合运用了多种空间数据挖掘技术,提取的土地利用空间知识包括：分区知识、样本分布知识、样本光谱知识、地物形态知识、多尺度纹理知识、关联知识：1/)由于土地利用空间知识是具有一定时空属性的,因此,我们首先以地域分异规律为指导,通过叠置法空间分区技术,获取土地利用规律相对一致的区域,从而保证从中提取的土地利用空间知识具有典型性和代表性；2/)其次,为了提高样本选择的效率和自动化程度,利用GIS数据辅助提取样本知识,研究中发现,传统的提取方法采用的是离散像元特征统计法,选取的样本受噪声影响很大,且样本空间分布分散,不易于理解。对此,本文改进并设计了基于空间数据聚类法的采样算法。主要思路是：在GIS图像分割的基础上,采用基于分割块的频数统计方法,其优点在于：充分利用了历史数据库中各种图斑的位置知识,获得的是融合了空间信息的光谱分布知识,使得该算法对于光谱噪声的敏感性降低,保证了训练样本选取的质量,并且能够获得空间连续的训练样本,选取结果更加稳定、直观；3/)纹理是一种重要的图像空间知识,而纹理尺度是影响该知识有效性的主要因素。但是传统的经验法和枚举法都不能快速有效的选择尺度参数。针对这一问题,提出了一种基于空间几何知识构建多尺度纹理特征的算法从纹理基元的角度出发,分析了纹理尺度与地物形态的关系,并得出结论：分类对象的形态特征才是决定纹理尺度的最重要因素。从历史数据库中获取稳定的地物形态知识,可以帮助我们确定最佳的纹理尺度参数。并且,针对不同的土地刊用类别,建立起有效的纹理描述指标,可以提高类别的可分性,使类别总数增加,更加符合土地利用要求。基于上述认识,本文首先研究了基于MER算子描述图斑的空间几何形态的方法,其次设计频率直方图统计算法,为每个类别选取出最佳的纹理尺度；将其与枚举法进行对比分析,结果表明分类对象的形态特征与纹理尺度因子之间具有很高的相关性,所选尺度参数合理,且算法效率较枚举法有大幅度的提高；4/)引入Apriori空间关联规则挖掘算法提取土地利用空间关联规则。作为一种成熟的数据挖掘技术,Apriori算法可以自动化、智能化地处理数据,使得提取的土地利用关联规则具有连续性和自适应性,弥补了传统研究中人工归纳方法的不足；5/)基于MER算法和LDM算法提取土地利用图斑的方向知识。为了在蚁群分类算法中为某类蚂蚁提供搜索像元的基本方向,本研究对各种地理要素在形态上表现出来的方向特征进行统计,统计对象是各类别的土地利用图斑,主要思路是：先利用MER算法提取单个面状地物的主方向,再根据LDM算法统计出一组面状要素的主方向；6/)最后,针对遥感分类需求,将挖掘出来的各种知识的进行统一表达,并归纳入库。 融合空间知识的蚁群优化分类算法是实现遥感分类的关键环节。研究的主要内容是蚁群智能优化算法、遥感影像分类、以及土地利用空间知识三者的有机集成方式。首先从蚁群系统的时空环境、分类问题描述、求解过程三个基本问题入手,阐述了蚁群系统与遥感影像分类问题耦合的基本思想。在此基础上,采用人工蚁群作为土地利用空间知识的信息载体,反之,利用空间知识为蚁群分类算法提供各种参数,由此建立起融合空间知识的人工蚁群分类模型,并设计了相关的算子。改进和创新的内容归纳如下：1/)利用样本分布知识辅助设计初始信息素；2/)在邻域像元搜索时,为了提高算法的稳定性和运行速度,设计了一些禁忌搜索算子实现改良；3/)在设计状态转移算子时,综合利用社会信息和蚂蚁个体对搜索方向的偏好决定状态转移概率,并且对方向权重提出了新的计算方法；4/)对于类别匹配算子,在度量特征向量之间距离的基础上,进一步根据蚁群算法的邻域搜索行为,利用图像中的空间信息,即相邻地物间的关联规则,对解的质量进行评价。主要思路是设计一个隶属度函数来描述像元与各类别样本的相似性,以及相邻地物间的空间关联规则；5/)在信息素更新阶段,引入了奖惩策略,可以动态增强优质解的信息素浓度,抑制劣质解的信息素浓度,帮助算法尽快收敛。该算法充分发挥了基于知识分类以及基于人工智能分类方法的优势。 以海南省昌江县的高分辨率遥感影像土地利用分类为例,对本文提出的方法进行验证及分析。在分类总体框架的指导下,利用GIS、RS软件、二次开发技术、以及matlab开发技术相集成,实现了土地利用空间知识的挖掘、表达、分类算法,实验结果表明：土地利用空间知识在遥感影像专题分类中具有重要作用,本文所提出的基于土地利用空间知识挖掘的遥感影像专题分类方法具有以下优势：一方面,拓展了GIS数据的应用范围和技术层次,提高了GIS数据库使用效率；另一方面,相较于传统的分类方法,其过程更符合人类的认识及思维过程,使分类类别更加丰富,而由于空间知识地有效利用,减少了由“同物异谱”和“同谱异物”带来的分类误差,改进的各种算子都对提高算法性能起到了一定的作用,更适合用于遥感影像土地利用专题分类。 It is a traditional and complex subject for land-use high-resolution remote sensing image classification There are two main difficulties for this issue:First, how to make the classification results more better for GIS applications; Second, there are a lot of phenomenon of "same object different images" and "same image different objects" in high-resolution RS images, which have made it more difficult. The key to solving the two problems has been search out by studying the human visual interpretation process in image classification:that is using thematic knowledge, especially spatial knowledge in classification process. Thus, the paper is all around these three issues:what kinds of land-use spatial knowledge are need, how to mine such spatial knowledge, and how to apply them for land use image classification. The primary task is establishing the classification model and framework, which can provides an overall strategy for the three core issues above. We take the simulation of human visual interpretation behavior as guidance. First, the concept of land use spatial knowledge and their scale effect were summarized, in order to make sure the types of land-use spatial knowledge needed in this study. What is more important is to analysis the internal correlation mechanisms of these knowledge, for establishing a cognitive pyramid; Second, which algorithms will be chosen to knowledge mining, as well as how to compute realize these algorithms has been discussed; Third, the advantages of ant colony optimization algorithm in the field of simulation of artificial visual interpretation have been analyzed, and it will be presented as image classifier with land use spatial knowledge. The core and basis content in this paper is mining useful land-use spatial knowledge from the GIS database and the RS Images. According to the process of human cognitive habit in image classification, a more gradual and comprehensive various spatial knowledge mining technology have been used by:1/) realizing space districting technology using overlay analysis method to update the region of land use with relatively consistent rule, from where can extract typical and representative land use spatial knowledge; 2/) using GIS auxiliary data to automatically extract sample knowledge. Comparing to the traditional sampling method using the discrete points statistical, which is sensitive to noises and difficult to understand, this paper designed a method based on spatial data clustering algorithm, which; 3/) presenting a algorithm of object geometric knowledge-based multi-scale texture feature extraction. Texture is an important image special knowledge, whereas scales are the major factor to affect its validity. But the traditional method, such as empirical analysis and enumeration analysis, cannot obtain the scale parameter quickly and effectively. This paper analyzed the relationship between texture scale and features shape, and concluded:the texture scale parameter is strong correlation with feature geometry characteristic, and it can be determined based on stable features geometry characteristic from a historical database. Furthermore, the multi-resolution texture descriptors can improve the separability among different land use categories, so that the total number of classification categories can be increased and more in line with the land use requirements. Based on the above, this study first describing the spatial geometry characteristic by MER algorithm, followed extracting texture scale parameters by relevant histogram statistical algorithm; 4/) introducing Apriori algorithm for spatial association rule mining of land use. As a mature technology in spatial data mining, Apriori algorithm is automatic and intelligent, so it can make land use association knowledge continuity and adaptability, which fill up a deficiency in traditional study by using artificial induction method; 5/) proposing an method for descritping the directional chractistic of each category by MER algorithm and LDM algorithm. It includes two steps:extracting the main direction of each feature polygon by MER algorithm, then, statisting the main direction of each category by LDM algorithm, which can be used as original direction in ant colony algorithm; 6/) finally, a classification-oriented knowledge database is established with unified expression of all knowledge above. Spatial knowledge based ant colony optimization classification algorithm is the key step of image classification. The main content is how to realize the organic integration mode among ant colony intelligent optimization algorithm, remote sensing image classification, and land-use spatial knowledge. First of all, the coupling concept between ant system and image classification is described from three aspects:the time-space environment of the ant colony system, classification problem description, and the process of solving. Then, ant colony have been designed as the carriers of land use spatial knowledge, whereas land use spatial knowledge generate some parameters for ant colony classification algorithm, based on it, algorithm model is established, as well as the relevant operators are designed. The improvement and innovation are summarized as follows:1/) using the sample distribution knowledge-aided into the of design of initial pheromone of ant colony; 2/) designing a set of tabu search operator for neighborhood pixel search, in order to improve the algorithm's stability and speed; 3/) using social information and individual information of direction preference to decision the state transition probability integrally; 4/) putting forward a new direction weight operators for state transition operator; 5/) defining a membership function to measure the solution similarity on the basis of distance function for category matching operator; 6/) introducing the incentive strategy to enhance the pheromone intensity of high-quality solutions and restrain the pheromone intensity of low-quality solutions for pheromone update operator, to help the algorithm converge as soon as possible. Example Analysis is the final but useful step. Changjiang County of Hainan Province has been taken as experimental region, to realize the high-resolution RS image classification of land use using the proposed method in this paper. With the guidance of classification framework, various effective land use spatial knowledge were mining and expressing, then successfully used in ant colony classification algorithm. In addition, some contrast analyzed experiments have been designed. Experiments show that:land use spatial knowledge plays an important role in RS image classification, the image classification method proposed has the following advantages:On the one hand, the application range of GIS data has been extended, and the utilize efficiency of GIS database has been improve; On the other hand, comparing to traditional classification method, its process is more in line with the intelligent behavior of human understanding and recognition, which can achieve more richer categorizes, moreover, as spatial knowledge were used, the classification errors which brought by the phenomenon of "same object different images" and "same image different objects" have efficiently reduced, so the classification result is more reasonable, and more suitable for RS image thematic classification of land use.

关 键 词： 高分辨率遥感影像 土地利用专题分类 土地利用空间知识挖掘 遥感认知的时空尺度 土地利用空间分区 基于空间数据聚类法的样本知识提取 图斑几何特征描述子 纹理尺度参数选择算法 地物方向特征描述子 算法土地利用空间关联规则挖掘 融合土地利用空间知识的蚁群智能优化分类算法

分 类 号： [TP751]

领　　域： [自动化与计算机技术] [自动化与计算机技术]