【作者】 刘钦龙；

【导师】 焦斌亮；

【作者基本信息】 燕山大学 ， 电路与系统， 2009， 硕士

【摘要】 遥感图像分类一直是遥感研究领域的重要内容,如何解决多类别图像的分类识别并满足一定的精度是遥感图像研究中的一个关键问题,具有十分重要的意义。传统的遥感图像自动分类识别主要采用决策理论(或统计)方法进行分类,但是,由于遥感图像本身的空间分辨率低,以及“同物异谱”、“异物同谱”现象的存在,往往出现较多的错分、漏分情况,导致分类精度不高。人工神经网络具有高度并行处理能力、自适应能力、非线性映射能力、泛化能力,使得人工神经网络在遥感图像分类应用研究中提供了新的方法。传统的遥感图像分类法主要分为监督分类法和非监督分类法,本文首先简要概括了几种常用的监督分类和非监督分类法。介绍了人工神经网络的结构原理、学习规则和在遥感图像分类中典型的神经网络模型,详细的分析了BP神经网络模型的遥感图像分类,并优化了基本的BP算法。重点讨论了自组织神经网络在遥感图像分类中的应用,针对自组织神经网络算法的学习率收敛慢做了改进,采用幂函数收敛。本文以Matlab为平台,构建BP神经网络和改进的自组织神经网络,对北京市海淀区遥感影像进行实验,并对实验结果进行了分析、比较。结果表明,改进的自组织神经网络分类方法分类精度远高于传统的遥感图像分类法和BP神经网络分类法分类精度,采用改进自组织神经网络模型的遥感图像分类法是有效的。更多还原

【Abstract】 Remote sensing(RS)image classification is always a pivotal part of remote sensing study. How to improve the accuracy of RS interpretation is a urgent problem in RS application. The automatic classification of traditional RS image is principal used in decision-making theory (or statistical) methods to classify. However, the existence of the phenomenon about RS images of low spatial resolution is appeared as well as "synonyms spectrum", "foreign body spectrum with", more often at fault, leakage points, which induce the low classification accuracy .The artificial neural network (ANN) has high degree of parallel processing, adaptive capacity, non-linear mapping capabilities, generalization ability. So artificial neural networks in remote RS in classifications applied research provides a new approach.Traditional RS classification is divided into supervised classification and unsupervised classification. At first, the paper briefly introduced several commonly used classification of supervised and unsupervised methods. Then we introduce the ANN structure and learning rules , as well as typical neural network models in remote sensing image classifications. The BP neural network model of remote sensing image classification is analysised detailed, and the basic algorithm for BP is optimized.Lastly,self-organizing feature neural network(SOFM) in remote sensing image classification is emphasized .In this paper,the power function convergence is used to solve the slow rate of SOFM learning algorithm convergence.Based on MATLAB as a platform, we build BP neural network and improved SOFM neural network, making Haidian District, Beijing remote sensing imaging as experiments object. We analysis and comparison the experiments results. The experiment shows that the improvement of SOFM neural network classification accuracy is higher than the traditional classification of RS image and BP neural network classification accuracy. So the improved SOFM neural network model of RS image classification is valid.更多还原