【作者】 于君娜；

【导师】 赵春晖；

【作者基本信息】 哈尔滨工程大学 ， 信号与信息处理， 2007， 硕士

【摘要】 光谱分辨率的提高是光学遥感不断发展的趋势。高光谱分辨率(简称为高光谱)遥感技术是过去二十年中人类在对地观测方面所取得的重大技术突破之一，是当前遥感的前沿技术。由于高光谱所特有的高光谱分辨率的性质，其潜在的可应用性受到人们的广泛关注。高光谱图像分类的研究是高光谱遥感应用的主要内容之一。人工神经网络模型是近年来发展起来的综合数据分类方法之一。神经网络高光谱模式分类方法有以下优点：1)不要求同类地物在特征空间满足正态分布，无需对原始类别做概率分布假设，不存在求解概率分布参数的问题，是一种无参分类器：2)可以将多种数据，如纹理信息、地形信息和光谱信息等方便而有效地融合到分类中来。由于神经网络具有以上的优点，研究基于神经网络的高光谱遥感图像分类方法是十分有必要的。本论文的目的就是对常见的神经网络如BP(误差反向传播)神经网络、RBF(径向基函数)神经网络的结构、训练算法进行研究，并结合其它的理论知识，寻找有效的高光谱遥感图像分类方法。论文主要完成以下工作：1．基于误差梯度下降的标准BP算法对大多数的实际应用而言都太慢了，针对这些不足，人们已经提出了许多改进标准BP算法的方案。论文介绍了几种改进BP算法的基本原理，并将它们用到实际高光谱遥感图像分类中，对各种算法的性能作了归纳和对比。2．BP网络初始权值的选取缺乏理论性的指导，成为BP神经网络应用的一个弊端，而初始权值选取不当会使BP神经网络的性能受到严重的影响。为此，论文利用BP神经网络和决策融合理论相结合的办法来提高高光谱遥感图像的分类精度。3．在RBF神经网络结构中，影响分类精度的主要参数为：隐层节点个数、径向基函数的宽度及中心位置、隐层到输出层之间的权值以及输出层节点的阈值。这五个参数如果选取不当，将会使分类效果受到严重影响。论文将RBF神经网络用于高光谱遥感图像分类中，针对高光谱遥感图像特点，结合具体影像实例，提出了一种设计以上五个参数的新方法。4．在进行高光谱遥感图像分类处理时，经常会遇到“同物异谱”的情况。论文将基于目标分解的神经网络用到高光谱遥感图像的分类中来。基于目标分解的神经网络分类方法由三部分组成，即目标分解、神经网络分类和“亚类”归并。它对于具有“同物异谱”特性的地物，首先把它分解为几种不同的“亚类地物”，使得分解后的每一种“亚类地物”自身的光谱特征是一致的，其分布呈现单峰正态分布。然后把分解后的几种“亚类地物”当成不同的地物类别送入神经网络去训练。最后在网络输出时加上一个逻辑运算，使得“亚类地物”重新归并到原来的类别中去，以此来改善高光谱图像的分类效果。5．遥感图像中，由于传感器的空间分辨率以及地面的复杂多样性而含有大量的混合像元。如何有效地解译混合像元是高光谱遥感应用的关键问题之一。论文采用一种方差纯化样本的方法来提取“端元”，并用RBF网络对整幅遥感图像进行混合像元分解。更多还原

【Abstract】 The enhancement of spectral resolution is the trend in development of the optical remote sensing. High spectral resolution (hyperspectral for short) remote sensing is one of the significant technological breakthroughs for the observation to ground in the past 20 years, and is the advancing front technology of the current remote sensing. The hyperspectral’s latent application receives widespread attention because of its high spectral resolution.The research of hyperspectral imagery classification is one of the main contents of the hyperspectral remote sensing application. The neural network is an integrated data classification method which is developed in recent years. The neural network has some merits when it is used to the hyperspectral imagery classification. First, it doesn’t require the same culture distributing as normal, and doesn’t need to do the probability distributional assumption to the primitive classification, not existing the problem of solving the probability distribution parameter. Second, it can make kinds of data, such as texture information and topographical information and spectral information, fused conveniently and effectively for classification.Because of the above merits, the research on the hyperspectral imagery classification by the neural network is very necessary. The purpose of this thesis is to carry on research to the structure of common neural network such as BP network, RBF network, training algorithm, and combine other theory knowledge to look for the valid method of hyperspectral imagery classification. This thesis mainly finishes the following work:First, the standard BP algorithm based on the error gradient descent is too slow for most practical application, to these insufficient, people have already proposed a lot of improvement schemes of standard BP algorithm. The thesis introduces the basic principles of several kinds of improvement BP algorithms, and sums up and compares the performance of various algorithms when using them to the hyperspectral imagery classification.Second, the choice of initial weights of BP network lacking theory guidance become a drawback of BP network application, and it will make the performance of BP network receive serious influence that initial weights are chosen improperly. For this reason, this thesis combines BP network with the decision fusion theory to improve the accuracy of the hyperspectral imagery classification.Third, the main parameters that affect the classification accuracy of the RBF network are: the number of the hidden neurons, the widths and centers position of kernel function of the hidden neurons, the weights between hidden-layer and output-layer and the biases of the output neurons. This paper proposes a new method to design the five parameters according to the characteristic of the hyperspectral image.Forth, the situation of commaterial with different spectrum often appears in the hyperspectral imagery classification. The thesis uses the neural network based on target decomposition to the classification. The classification method of neural network based on target decomposition is made up of three parts, the target decomposition, the neural network classifier and the subclass merged. It divides the same class with different spectrum into several kinds of subclasses. The spectrum of each subclass is identical and unimodal normal distribution. Then the subclasses decomposed are regarded as different ground objects and sent into the neural network to train. Finally, a logic operation is added when the network exports to make the subclasses merge into the original classification again, so as to improve the accuracy of hyperspectral imagery classification.Fifth, the hyperspectral image contains a large amount of mixing pixels because of the spatial resolution and complicated variety on the ground. The problem of resolving the mixing pixel effectively is an outstanding question of the hyperspectral application. This paper uses a method of variance purifying samples to extract endmember, and uses the RBF network to unmix the whole hyperspectral image.更多还原