基于Agent的遥感影像分类方法及其应用研究

发布时间：2019-03-23 20:05

【摘要】：利用影像进行分类是遥感信息提取的重要方法之一,分类后处理是改善初始分类图像质量的必要手段。通过对国内外现有遥感影像分类研究进行综述与分析,并结合实际生产任务的需要,发现当前研究存在不足:①绝大多数的研究偏重于分类方法的改进及应用,对分类后处理却鲜有涉及;②部分学者提出的分类后处理方法仅针对分类图像本身,而影像包含的信息量未能得到充分利用;③人工分类后处理严重依赖经验,且费时费力,现有商业软件的自动化处理工具尽管在一定程度上改善初始分类图像质量,但存在过度聚类的现象。计算Agent是存在于动态环境中的一种抽象模型,在数字图像处理中表现出高度的智能性。本文尝试将Agent理论和模型引入遥感影像分类研究,在梳理Agent相关概念和理论的基础上,构建用于遥感影像分类后处理工作的Agent和多Agent系统。针对初始分类图像的特点以及遥感影像上可挖掘的地物增强信息,构建由分类型、决策型和综合调节型Agent三者协同工作的多Agent系统,通过Agent对分类图像环境和遥感影像环境的感知、推理和信息利用,能够对初始分类图像上常见缺陷进行自动调整。之后利用IDL语言开发了Agent分类后处理工具的核心功能模块,便于以工作流的方式实现基于Agent的分类后处理任务。研究中以北京市作为实例,对预处理后的北京市Landsat 8 OLI影像,采用最大似然法、神经网络法和光谱角法实施监督分类,同时从影像上提取NDVI、亮度、绿度等9类特征信息,在Agent分类后处理工具的支持下,对初始分类图像进行自动化的调整,从精度统计和目视解译两方面验证了本文提出方法的有效性,并与ENVI内置工具处理结果进行了对比。研究结论有:①本文提出的基于Agent分类后处理工作模式可以实现遥感影像分类后处理任务的自动化,能有效抑制“椒盐噪声”等问题,最高可将总体分类精度提高5.5%;②Agent分类后处理工具综合利用初始分类图像和遥感影像,避免单纯基于滤波处理方式导致的过度聚类问题,且在辅助资料缺失时,本文方法仍然适用;③分类图像初始精度相对较低情况下,Agent分类后处理方法对分类精度提升效果优于初始分类图像精度较高的情况;④以IDL开发核心功能模块,便于与ENVI软件集成或开发独立系统,利于推广应用。
[Abstract]:Image classification is one of the most important methods for remote sensing information extraction, and post-classification is a necessary method to improve the quality of initial classification image. Through summarizing and analyzing the existing remote sensing image classification research at home and abroad, and combining with the needs of actual production tasks, it is found that there are some deficiencies in the current research: 1 the vast majority of research focuses on the improvement and application of classification methods. The post-processing of classification is rarely involved; (2) the classification post-processing method proposed by some scholars only aims at the classification image itself, but the information contained in the image has not been fully utilized; (3) the post-processing of manual classification depends heavily on experience and is time-consuming and laborious. Although the automatic processing tools of existing commercial software improve the quality of initial classification images to a certain extent, there is an over-clustering phenomenon. Computing Agent is an abstract model that exists in the dynamic environment. It shows a high degree of intelligence in digital image processing. This paper attempts to introduce the Agent theory and model into the research of remote sensing image classification. On the basis of combing the related concepts and theories of Agent, a Agent and multi-Agent system for the post-processing of remote sensing image classification is constructed. According to the characteristics of the initial classification image and the mineable feature enhancement information on the remote sensing image, a multi-Agent system is constructed, which consists of classification, decision-making and integrated adjustment-based Agent. Through the perception, reasoning and information utilization of the classified image environment and remote sensing image environment by Agent, the common defects in the initial classification image can be automatically adjusted. Then the core function module of Agent classification post-processing tool is developed with IDL language, which is convenient to realize the post-processing task based on Agent in the way of workflow. Taking Beijing as an example, the maximum likelihood method, neural network method and spectral angle method were used to monitor and classify the pre-processed Landsat 8 OLI images in Beijing. At the same time, 9 kinds of feature information, such as NDVI, brightness and green degree, were extracted from the images. With the support of the post-processing tool of Agent, the automatic adjustment of the initial classification image is carried out. The validity of the proposed method is verified from the two aspects of precision statistics and visual interpretation, and the results are compared with those of the ENVI built-in tool. The conclusions are as follows: (1) the post-processing mode based on Agent can realize the automation of the post-processing task of remote sensing image classification, and can effectively suppress the "pepper and salt noise" and so on. The overall classification accuracy can be increased by 5.5% at the highest level. The 2Agent classification post-processing tool uses the initial classification image and the remote sensing image synthetically, avoids the over-clustering problem caused by the simple filtering processing, and when the auxiliary data is missing, the method in this paper is still applicable. (3) when the initial accuracy of the classification image is relatively low, the Agent classification post-processing method is better than the initial classification image in improving the classification accuracy; (4) develop the core function module with IDL, it is convenient to integrate with ENVI software or develop independent system, which is helpful to popularize and apply.
【学位授予单位】：中国地质大学(北京)
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP751

【参考文献】