【作者】 陈丹；

【导师】 刘艳芳；

【作者基本信息】 武汉大学 ， 土地资源管理， 2010， 博士

【摘要】 遥感影像的分类方法的研究是遥感领域中的一项重要内容,它直接影响到遥感数据的应用范围和应用效果。随着遥感技术的发展,遥感影像在遥感平台和影像分辨率上都有了长足的发展,遥感分类的方法研究也向着多方法,多分类器相结合,取长补短进行了更多的新方法的应用研究。本文研究的是高分辨率遥感影像分类,通过对遥感影像的分类方法以及基于象元的遥感影像的分类方法的总结,并进一步分析不难发现,对于现阶段高分辨率的遥感影像的分类方法主流的方法集中在基于象元的分类方法,为提高遥感影像的分类精度也需要结合一些辅助的信息进行分类。通过对基于象元的遥感影像的分类方法的综述,可以发现许多基于象元的方法都存在黑箱操作的问题,例如人工神经网络法,支持向量机等,其分类结果难以解释,分类规则不明晰。遗传算法和免疫算法等多参数算法也都被应用于遥感影像分类中,其算法可行性得到了肯定,但是这类算法由于需要设置较多的参数以及确定的阈值,使得其算法的分类精度的稳定性还不够。本文选择改进决策树算法进行遥感影像分类避免了以上提到的黑箱问题,得到可解释的分类规则及分类结果。针对高分辨率遥感影像的决策树分类方法减小了混合像元对影像分结果的影响,并且决策树分类是一个可以处理离散数据的方法,非常适合引进其它辅助分类信息进行遥感影像分类。现阶段决策树的分类方法仍然存在许多不足,主要的可归纳为以下几点：(1)连续属性的处理。大多数决策树算法倾向于处理由离散型属性组成的决策表,而在决策树的实际应用中连续属性是不可避免会出现的。连续属性的离散化及其相应的阈值划分标准一直是归纳学习的一个难点。(2)局部非回溯的启发式很难获得全局最优的决策树。如Quinlan的Interaction Detection 3 (ID3)算法采用的信息增益就不是最优的启发式。(3)决策树算法通常假设训练样本数据是完全的理想数据,导致决策树系统在实际应用中容错性较差,样本中数据的不完全,不完整等噪声对决策树的分类性能影响较大。本文针对以上决策树分类算法的不足进行改进,并结合遥感影像分类特点和分类知识,构建一个面向影像分类的优化的决策树模型。本文的研究内容可归纳为以下几个方面：(1)现阶段研究中,决策树算法通常利用完全的理想数据作为训练样本,导致在实际应用中真实样本数据的不完整性等噪声影响了决策树算法的分类能力,决策树的分类能力对样本的选择具有很强的依赖性。本文首先针对该缺点,结合遥感影像分类本身的特点,提出建立特征优先级机制对节点的选择进行控制。该机制是基于影像知识的,通过影像分类的先验知识,建立起各类地物以及整体影像的特征优先级,并构建特征优先级与决策树节点选择之间的控制模式,完成基于影像知识的决策树准则对决策树方法的优化。(2)针对ID3算法采用的信息增益熵为局部非回溯的启发式的缺点,本文引入了模拟退火算法的思想,来重建决策树算法的启发式,利用信息增益熵和整体影像的破碎度因子相结合,得到模拟退火的目标函数,利用模拟退火有一定几率跳出局部最优的特点,完成决策树算法从局部最优中的跳出,避免过拟合现象的产生,完成基于全局影像的决策树准则对决策树方法的优化。(3)通过以上两个准则的建立,得到面向影像分类的决策树优化模型的构架,并对优化决策树模型的逻辑流进行了详细的设计。在进行优化的决策树节点选择之前,首先对样本数据进行一个模拟退火的纵向聚类。这个聚类比常见的模拟退火聚类增加了一个决策属性在进行聚类时的判断。也就是模拟退火的过程除了考虑样本的相关性系数,收敛程度还需要考虑聚类的样本其决策属性是否相同。纵向聚类过程,设计了一个小样本系数用于保证样本特征的完整性,同时小样本系数的设计也保证了样本的速度不至于下降过快,不会产生没有足够的样本参与分类的情况。样本的纵向聚类进一步减低了决策树建立的结果对样本数据的依赖性。通过选择合适的算法在纵向聚类完成后的样本上进行样本属性的聚类和离散化过程,得到决策树可处理的数据。在此基础上设计了决策树的剪枝过程,和算法性能的评价指标。(4)通过实验研究,比较优化的决策树方法和ID3方法的决策树构建结果以及QuickBird影像和SPOT影像的分类结果,证明了优化的决策树方法能有效地提高各尺度下影像的分类精度。虽然比较树的构建结果可知优化的决策树算法并不能在结构上对常规的决策树有明显的改善,甚至在一些尺度下优化的决策树算法的结果较之ID3算法更为复杂,但是其复杂的程度在实质上不影响计算的效率,但是在利用其构建的树的规则进行影像分类时,能有效地提高影像分类精度。通过比较最优尺度下优化的决策树算法与神经网络算法和最大似然法对实验区域的影像分类精度,进一步证实了优化的决策树算法的有用性。实验中出现的优化的决策树算法对占区域比例小的地物类别分了精度不高的缺点,为后期的研究明确了方向。如何利用方法改进和结合,将该方法用于中低分辨率的遥感影像分类也是今后的研究方向之一。更多还原

【Abstract】 Classification methods for remote sensing image an important part of the research in remote sensing field, which directly affects the using range and results of the application of remote sensing (RS) data. With the development of the remote sensing technology, the remote sensing science has come a long way on the development of the platforms and the resolution. The classification methods of RS images also have been developed towards multi-method and multi-classifier combination. The classification methods have complemented each other for further studys.This purpose of this paper is to research the classification of remote sensing imanges.It is not hard to find that recently the classification algorithms for high-resolution remote sensing images mainly focus on pixel-based classification ones and in order to improve the accuracy of classification of remote sensing image, some auxiliary information is also need to be integrated,according to the analysis of the classification algorithms of remote sensing imanges and pixel-based classification algorithms of remote sensing image.By the synthesis of pixel-based classification algorithms of remote sensing image, we can find many algorithms exist the problem of "the black-box operation",such as artificial neural networks, support vector machines,therefore the classification results are difficult to be interpreted and classification rules arenot clear.Genetic algorithm and immune algorithm which need many parameters are used in the classification of remote sensing image and the feasibility of these algorithms were affirmed.However, these algorithms need to set a lot of parameters and accurate thresholds, which makes the unstability of the classification accuracy. In this paper, the improved decision tree algorithm is adopted to research the classification of the remote sensing image,which avoid the problem of "black box" as mentioned above, and the classification rules and classification results that coud be clearly explained are acquired.The improved decision tree algorithm for the classification of high-resolution remote reduces the effect of mixed pixels to the classification results, and the classification algorithm of decision tree can deal with discrete data, so it is very suitable to be used into the classification of remote sensing image by introducing other assistant classification information.There are still many deficiencies in the classification algorithm of decision tree at the present stage.The main ones are summarized as follows:(1) The treatment of continuous attributes. Most decision tree algorithms tend to handle the decision tables which are composed by discrete attributes,but in the practical application of the algorithm of decision tree, continuous attributes will arise inevitablely. Therefore the discretization of continuous attributes and their corresponding threshold criteria for the classification is a difficulty for the inductive learning at all times.(2) It is difficult to get the global optimal decision tree with the local non-backtracking heuristic. For example, the Quinlan’s Interaction Detection 3 (ID3) algorithm[30] which uses the information gain is not optimal heuristic.(3) A lot of decision tree algorithms often assume that training data is complete, which results in poor fault tolerance in the application of the decision tree system and the noise of incomplete data samples has the great impact on the classification algorithm of decision tree.For the deficiencies of the above classification algorithm of decision tree, this paper makes some improvement to the decision tree algorithm and builds a RS image classification-oriented optimized model of decision tree by combining the classification characteristics of remote sensing image and the knowledge of classification. The content of this study can be summarized as follows:(1) In the present study, the decision tree algorithms usually use complete samples as training data, leading to the bad classification rusults with the noise of incomplete data samples in practical applications,because the effect of classification algorithm of decision tree strongly depends on the samples selected.In order to improve the shortcomings.Firstly, a characteristics priority mechanism for the control of the selection of nodes is proposed by combining the own characteristics of the classification of remote sensing images. The mechanism is based on the knowledge of images. The priority of characteristics between various types of surface features and the overall image are built, and the control mode between the priority of characteristics and the node selection in the decision tree are designed with the prior knowledge of classification of RS images, then the decision tree is optimized by the rules based on knowledge of RS images.(2) For the shortcomings of ID3 algorithm wich uses information gain entropy as the local non-backtracking heuristic, the idea of simulated annealing algorithm is introduced to reconstruct the decision tree’s heuristic, and the objective function of simulated annealing is obtained by the combination of information gain entropy and the overall image of the fragmentation phase factors. Simulated annealing algorithm has the advantage of jumping out from local optimum at some certain chance, which can help the jumping from local optimum in the decision tree algorithm, avoid the over fitting phenomenon, and optimize the decision tree algorithm based on rules of global RS images.(3)By the establishment of the above two criteria, the model of RS images oriented-optimized decision tree is built, and the logical flow of decision tree is designed in detail.The simulated annealing of the vertical cluster is operated on the samples before node selection of optimized decision tree. Compared with usual simulated annealing clustering,this clustering adds a judgment of the attribute of decision into the clustering process.In other words,simulated annealing process takes the correlation coefficient of the sample into account, the convergence speed of clustering also needs to consider whether the decision-making properties of samples are same. In the vertical clustering process, a factor of small sample is designed to ensure the integrity of the sample, which at the same time guaranteed the rate of sample will not get down too fast and there are enough samples participated in the classification. The vertical clustering of samples further reduces the establishment result of the decision tree’s dependence on samples.By operating the clustering and discretization process on the samples in the vertical clustering using the appropriate clustering algorithm, the data that could be handled by decision tree is obtained. Then the pruning process of the decision tree and the evaluation indexes of algorithm’s performance are desiged.(4)By comparing the construction and classification results of QuickBird and SPOT images of decision tree between optimized decision tree and the ID3 algorithms, we can get the following conclusions:The optimized decision tree algorithm can effectively improve the classification accuracy of RS images with different scales. Although by comparing the construction result of tree of the optimized decision tree and the ID3 algorithms, the optimized decision tree could not make significant improvement to the structure, and even the results of the optimize decision tree algorithm is more complex than the ID3 algorithm in dealing with RS images with some scales,but in fact the complexity of the optimize decision tree algorithm does not affect the efficency of calculation,and by using the rules of the tree constructed of the optimize decision tree algorithm, the accuracy of classification can be improved efficiently.By comparing the accuracy of image classification in the decision tree algorithm,neural network algorithms and maximum likelihood experiment in the optimal scale, the usefullness of the optimized decision tree algorithm are confirmed.In the experiment, there is the shortcoming that the accuracy of classification of small surface features with minor proportion in the total region is not high, so this also shows a clear direction for the further research. How to improve and integrate the method to match the classification of the low-resolution remote sensing image is one of directions of the future research.更多还原