【作者】 周爽；

【导师】 苏宝库；

【作者基本信息】 哈尔滨工业大学 ， 控制科学与工程学科， 2010， 博士

【摘要】 高光谱遥感数据提供的丰富地表信息使得高光谱遥感的应用越来越广泛,如何充分地利用这大量的信息,以及如何在如此大量的信息中提取有用信息,并使这些信息能够为我们的应用服务,是摆在研究者面前的重要课题。降维和分类是高光谱图像处理中的两个关键技术。无论是高光谱图像的降维,还是高光谱图像的分类,其根本都是要从大量的高光谱数据中提取出能够满足人们需要的特定或要求的信息,而这些信息的提取综合起来都可以说是一个信息的组合优化处理过程。蚁群算法作为一种群智能仿生优化新技术,其突出特点是自组织性、鲁棒性、并行性,非常适合于求解非确定性的离散组合优化问题。本文在分析高光谱图像的光谱分辨率特性、空间相关性、谱间相关性、数据维、信息量等特性的基础上,重点分析了蚁群算法在高光谱图像处理中的应用,证明了蚁群算法完全适用于高光谱图像的降维和分类处理。高光谱图像的降维包括波段选择和特征提取两种方式。本文提出一种基于蚁群算法的高光谱图像波段选择方法,以解决最优波段选择算法复杂度高,计算量大的问题。将每个波段看作是蚂蚁觅食时所经过的节点,选取不同的评价函数作为蚂蚁选择路径的依据,应用蚁群算法选择出多次搜索中最优的一组波段组合。蚂蚁寻找最优路径的过程,就是最优波段组合的形成过程。通过蚁群算法从众多的光谱波段中挑选出能够反映地物光谱空间分布的特征波段,形成降维的波段子空间达到了降低高光谱图像数据维的目的。接着,利用高光谱图像高度相关的波段成组出现的特点,提出一种基于蚁群算法的高光谱图像子空间分解方法,在子空间中采用特征变换的方式降低高光谱特征空间的维数。该方法同样将波段作为蚂蚁觅食时经过的节点,蚂蚁依据波段之间的相关性来决定路径的选择,蚂蚁经过优化搜索之后将高维高光谱数据空间分解为几个较低维的数据子空间,再采用传统的主成分分析方法在子空间中提取有效特征,进而实现对高光谱图像的降维。高光谱图像的分类方法包括监督分类和非监督分类两种。本文提出一种基于蚁群算法的高光谱图像监督分类方法。该方法先依据图像的信息熵将高光谱图像各个波段的单幅图像中的灰度属性分段离散化,然后将这些离散化的灰度属性作为条件项并集合到一起形成一个备选条件项数据集。在训练样本中,将高光谱图像各个波段数据经过离散化后形成的条件项当作蚂蚁的候选路径节点,用条件项的信息熵作为蚂蚁路径转移的启发函数。经过蚂蚁的迭代搜索,每只蚂蚁都构造出一条分类规则,通过信息素浓度的调整,将蚂蚁们构造的规则中质量较好的规则保留下来,而质量不好的规则则在搜索过程中逐渐被淘汰。在所有训练用的地物类别都被归类之后,最终形成用于分类的分类规则。最后,提出一种基于蚂蚁化学识别系统的蚁群聚类算法。该聚类算法依据蚂蚁之间的相似程度来决定蚂蚁的类别归属,最终相似程度高的蚂蚁可以聚集成一类。在该算法中,遥感图像中的每个像素都被看作是一只蚂蚁,该蚂蚁所携带的信息除其所代表的像素点的各波段光谱信息外,还包括所属类别的标号,类别属性等信息。本文依据基于蚁群算法的高光谱图像波段选择方法获得的选择结果,从中提取若干个特征波段作为数据源,采用上述聚类方法进行实验,并与传统的k均值算法比较实验结果。为了客观地评价聚类结果的优劣,本文综合考虑了聚类算法本身的聚类性能,即类内距,类间距,以及聚类图像与标准图像的相关度,提出一种综合上述参数的聚类图像客观评价指标,并应用该指标对本文提出的聚类算法和传统的k均值算法所获得的图像作出了客观评价。更多还原

【Abstract】 The applications of hyperspectral remote sensing data have become broader and broader for its providing the abundant surface information. It is an important issue for the researchers how to extract useful information from the observed data, how to fully utilize such large amount of information, and how to make them become availability to our applications. Dimensionality reduction and classification are two key techniques in the hyperspectral image processing. Whether in hyperspectral image dimensionality reduction or in classification, both of them are the extraction of required information from a large number of hyperspectral data. And the integration of the extracted information is a process of combinatorial optimization. Ant colony optimization algorithm (ACO) is a new type of bionic optimization algorithm developed in recent years. Its main characteristics are self-organization, robustness, parallelism, which are very suitable for solving non-deterministic discrete combinatorial optimization problem. Based on the hyperspectral image characteristics, such as spectral resolution, spatial correlation, spectrum correlation, data dimensionality, information capacity, etc, the applications of ant colony algorithm to the processing are researched, and it is proven that the ant colony algorithm is suitable for hyperspectral image dimensionality reduction and classification in this dissertation.Dimensionality reduction of hyperspectral images includes band selection and feature extraction. In this dissertation, an ant colony algorithm based band seclection method of hyperspectral image is proposed to solve the complication of optimal algorithm and the computation load. In the new method, each band is regard as a node that an ant passes through during its foraging, the different evaluation function is chosen as a measure that ants select the path, and the optimal combination of bands is searched and obtained by using ant colony algorithm. The process of ants searching optimal path is the process of forming the optimal band combination. By the ant colony algorithm, several signature bands reflecting the material spectral distribution are selected from the whole spectral bands, and the band subspaces of reduced dimensionality are formed, in which the purpose for the dimensionality reduction of hyperspectral data is reached. Then, according to the characteristics that highly correlated specral bands are presented in group, a method of subspace decomposition to hyperspectral image is proposed based on ant colony algorithm. The new method adopts the feature transform to reduce the dimensionality of hyperspectral feature space. Also every band is regarded as a node that an ant passes through during its foraging. The ant judges its path in terms of the correlation between the bands. High-dimensional hyperspectral data space is decomposed into several lower-dimensional data sub-space by ant optimal searching process. Then by using the principal component analysis to extract the effective features of subspace, the dimensionality reduction of hyperspectral images is achieved.There are two kinds of classification method of hyperspectral image, i.e. supervising classification and unsupervising classification. A hyperspectral image classification method based on ant colony algorithm is proposed in this dissertation. Firstly, in terms of the information entropy of image, the gray-scale of every band in hyperspectral image is dispersed in partition. Then these condition items of discrete gray-scale are combined and a data set of condition items is formed. In the training samples, the condition terms formed by dispersing the hyperspectral image data is regarded as the candidate nodes of ants. The information entropy of condition items is regarded as heuristic function of ant transferring path. After an iteration searching of ant, every ant can construct a classification rule, and the best rule is reserved while the worse rule is gradually eliminated by adjusting the concentration of pheromone. After all training samples is classified, the classification rule is formed.Finally, an ant colony clustering algorithm is proposed based on ant chemical recognition system in this clustering algorithm. According to the similarity between ants, the ownership of ants is decided, and the ants being a high degree of similarity can assembled into a class. Each pixel in the remote sensing images is considered as an ant, in which besides the spectral information of every pixel in every band, the category labels, category attributes and so on are also included. In this dissertation, adopting the results acquired by the band selection of hyperspectral image based on ACO, the feature bands are extracted as simulation data, and the clustering method experiment is performed. The experiment results are compared with the results of traditional k-means algorithm. In order to objectively evaluate the performance of the clustering results, the clustering performance of the new algorithm is taken into account in the dissertation, which is intra-class distance, inter-class distance, and the correlation between clustering image and standard image. An objectively comprehensive evaluation measure that integrates the above parameters is proposed. The objective assessment to the new clustering algorithm and traditional k-means algorithm are acquired.更多还原