【作者】 孙家波；

【导师】 郧文聚； 朱德海； 杨建宇；

【作者基本信息】 中国农业大学 ， 农业工程， 2014， 博士

【摘要】 耕地保护一直以来都是土地资源管理的核心,耕地数量及其分布信息的获取是实现这一目标的前提,遥感技术能够实现从广域和局域的空间尺度客观地获取耕地信息。高空间分辨率(下文简称“高分辨率”)遥感影像目前已经广泛地应用于大比例尺的土地利用调查与遥感监测业务,但是由于自身存在的高度细节化、数据量大、类内差异大等原因,导致影像自动解译的难度增加,使得实际业务工作中仍然以人工目视解译为主,缺乏自动化程度较高的流程化的工作方式。面向对象的影像分析技术适合于高分辨率影像的信息提取,但是受到分割方法选择、特征选择、阈值确定等因素影响,限制了该技术在实际业务中的应用。本文以高分辨率影像城市郊区的耕地提取为主线,结合基于知识遥感影像分类的知识获取、表达、推理、积累的一般过程,通过提高面向对象影像分类的自动化,以分类规则集的形式实现耕地的自动提取。本文的主要研究和工作涉及以下几个方面：(1)建立了具有一定通用性的高分辨率影像信息提取知识框架。从信息提取的角度,结合遥感影像目视解译的基本过程及应用到的地学知识,从解译影像、解译区域、解译对象以及地学辅助数据四个方面,系统地提出了一个较为全面的知识框架,为耕地、建设用地等典型地类提取规则的构建提供特征库。(2)改进SEaTH (Seperabality and Threshold)算法实现了耕地分层提取策略的自动获取。针对目前耕地提取存在的自动化程度低、普适性差等缺点,通过改进SEaTH算法,在实现分类特征自动选择和特征阈值自动确定的基础上,分析耕地与非耕地类别提取的难易程度以及区分过程的相似性,并基于分层分类的思想,依据从易到难、从少到多、优先使用非纹理特征等原则,实现耕地提取规则集的自动构建。(3)提出了耕地提取规则集在同一区域、不同时相影像条件下的推广和调整方案。结合规则集从源影像应用到目标影像时分类特征、特征阂值、分类顺序等的变化,在保证耕地提取精度满足要求的前提下,提出了规则集推广和调整的方案,以实现规则集的有效复用。(4)将规则引擎技术引入到遥感影像信息提取过程实现了基于专家系统的耕地自动提取。利用开源的NxBRE规则引擎工具和RuleML语法规范,通过建立知识库文件,研发规则定制与管理、界面展示、知识推理等功能模块,将耕地提取规则集的管理和应用与代码的实现过程分离,实现耕地提取规则的表达、存储、更新、执行和可视化,推动耕地自动提取相关研究成果的实际应用。更多还原

【Abstract】 Cultivated land protection has always been the core and focus of land resource management, the cultivated land quantity and distribution of information acquisition is the precondition of achieving this goal. Remote sensing technology is applied to get cultivated land information objectively from a wide-area and local-area spatial scale. Very high resolution (VHR) satellite imagery is used widely for large scale land use survey and monitoring. Due to its highly detailed information, large data volume, big differences within the same classes, VHR satellite imagery is difficult to be interpreted automatically and is always managed in the means of manual visual interpretation in the actual operation, lacking a high degree of automation of process-oriented way of working. Object-oriented image analysis (OBIA) technology is suitable for VHR satellite imagery information extraction. However, OBIA technology has not been used in the actual business, which is caused by the constraints in the process of segmentation and classification, for example, the selection of segmentation algorithm, the classification features selection, and the determination of theresholds. Taking cultivated land extraction on the outskirts of cities of VHR satallited imagery as the main line and combining with the knowledge-based classification fundamental, such as knowledge acquision, knowledge representation, knowledge reasoning, and knowledge accumulation, this paper achieves a rapid and real-time extraction of cultivated land in the form of classification rule set on the basis of enhancing OBIA automation level. The main research and work involves the following aspects:(1) An universal knowledge framework for VHR imagery information extraction is established. From the perspective of information extraction, a detailed knowledge framework is built to supply the classification rule set for cultivated land, construction land, and other land cover types with feature library from four aspects of interpretation imagery, interpretation area, interpretation objects and geographical secondary data, which is combined with the basic process of visual interpretation and the use of geographical knowledge.(2) Hierarchical extraction strategy for cultivated land is accessed automatically on the basis of the improvement of the SEaTH (Seperability and Thersholds) algorithm. The SEaTH algorithm is improved to solve the problems of cultivated land extraction, such as the low degree of automation and poor universality. Firstly, automatic feature selection and thresholds determination is reliazed; and then cultivated land extraction rule set is constructed hierarchically and automatically by analyzing the similarity of distinction process and the degree of distinguishment between cultivated and non-cultivated land, in terms of several principles about information extraction, for example, from easy to difficult, from less to more, priority in the use of non-texture features.(3) A promotion and adjustment scheme is presented for cultivated land extraction rule set under the condition of multi-temporal images in the same area. Considering possible changes of classification features, rule thresholds, and extraction orders when applying the cultivated land extraction rule set from the source image to the target image, a promotion and adjustment scheme is presented to achieve an effective repeated use of rule set on the premise of a high classification accuracy.(4) The rules engine technology is introduced into the remote sensing image information extraction process to achieve the rapid extraction of the cultivated land based on the expert system. An open source rule engine tool NxBRE and its knowledge representation specification RuleML are utilized to estabilish the knowledge base files. The custom and management of rules, interface display, and knowledge resoning function modules are developed to isolate the applyment of cultivated land extraction and the implementation of code, and to ealize the expression, storage, update, execution, and visualization of rule set, promoting the practical application of relevant research results.更多还原