【作者】 付磊；

【导师】 王金亮；

【作者基本信息】 云南师范大学 ， 地图学与地理信息系统， 2008， 硕士

【摘要】 近年来,从事全球环境变化研究的科学家逐步认识到土地利用/覆盖变化研究在全球环境变化研究中的重要性和必要性,并逐渐加强了这方面的研究。由于不同领域遥感图像的应用对遥感图像处理提出了不同的要求,所以图像处理中重要的环节——图像分类也就显得尤为重要。探索一种高效准确的分类方法成为遥感图像分类研究的热点。本研究将粗糙集作为神经网络的预处理单元,利用粗糙集消除冗余特征,减少神经网络的输入节点,降低了网络规模,加快了训练速度。粗糙集神经网络利用粗糙集原理进行知识的表达、推理和简化,利用神经网络的并行特点完成网络学习运算,能更有效地处理不确定、不精确及冗余的数据。本文在前人研究的基础上,把粗糙集理论、遗传算法、神经网络等理论和方法有机结合,力图从数学形态学的角度而不是传统的统计学角度实现土地利用/覆盖信息的分类。其具体研究工作可以归纳为以下几个方面：(1)根据粗糙集理论基本原理与算法,结合粗糙集在图像处理中的研究现状,找到其在遥感图像分类中的理论支持。(2)根据遥感数据的特点,提出粗糙集理论在处理遥感数据不确定性中的优势,分析几种粗糙集框架下的遥感信息的知识发现和表达方法,为研究制定合理有效的技术实现手段。(3)在遥感图像的预处理中,采用了正射校正和辐射校正相结合的办法,有效减小数据的空间域和频率域的误差。(4)在粗糙集属性简约中的辐射量度决策信息离散化中,采用了遗传算法的方式,弥补了粗糙集核约简存在的不足。结果表明,粗糙集简约后的决策信息放入RBF神经网络中进行运算,输出结果与BP网络运算结果进行对比,在运算时间和测试精度上均优于BP网络。虽然基于粗糙集理论的RBF神经网络遥感图像分类模型可以解决土地利用/覆盖分类中的一些问题,但仍然需要地形因子等地理信息、地学知识、专家知识的辅助决策支持,才能够有效提高分类识别类型及精度。更多还原

【Abstract】 Nowadays, the scientists engaged in global environmental research change gradually realized that land use/land cover change research in global environmental change research is importance and necessity, and gradually stepped up this kind of research. As different areas of the application of remote sensing images of remote sensing image processing to a different request, so the remote sensing image classify is particularly important. To explore an efficient and accurate classify become a hot spots of research in this area.This research will be rough set as a neural network pretreatment unit, using rough sets to eliminate redundant features to reduce the importation of neural network nodes, reducing the size of network, speed up the training pace. Rough sets neural networks using rough set the principle of knowledge, reasoning and simplification, the use of neural network parallel computing features complete net study, to more effectively deal with uncertainty, imprecise and redundant data.In this paper, on the basis of previous studies, the rough set theory, genetic algorithms, neural networks, and other theory and methods organic combination, in an attempt to mathematical morphology from the perspective rather than the traditional statistical point of view to achieve land use/land cover information classification. Their specific research work can be summarized in the following areas:(1) According to rough set of basic principles and algorithms, the combination of rough set in image processing of the status, found out the the theoretical to support the remote sensing image classify.(2) According to the characteristic of the remote sensing data, found out the rough set theory dealing with the uncertainty of remote sensing data in the advantage, analysed the several framework of the rough set to remote sensing information and knowledge mining, ways to study and formulate reasonable and effective means of technology.(3) At the remote sensing image preprocessing, using the ortho-correction and a combination of radiation calibration methods, to effectively reduce the data space domain and frequency domain error. (4) At the rough set of attributes simple to measure the radiation dispersion of information for decision-making, using a genetic algorithm, make up a rough set of nuclear reduction of the deficiencies.The results showed that:after the rough set of simple information for decision-making add to RBF neural network for operation, compared with the results, computing time and accuracy tests on the network are better than BP. Although based on the rough set theory RBF neural network of the remote sensing image classify model can solve some of the problems about the land use/land cover classify, but still need to topographical factors such as geographic information, and knowledge and expertise to support the decision-making, in order to effectively improve the classify type and accuracy.更多还原