【作者】 徐芳；

【导师】 郑肇葆；

【作者基本信息】 武汉大学 ， 摄影测量与遥感， 2004， 博士

【摘要】 影像解译是目前生产中急需，但尚未完全解决的摄影测量与遥感技术热点之一，也是亟待解决的一个瓶颈问题。支持向量机是国际上机器学习领域新的研究热点，是Vapnik等人根据小样本统计理论—统计学习理论发展的一种新的通用学习算法，能够较好的解决小样本学习问题。因此本文将支持向量机引入到航空影像的分类与分割中，期望探索一种新的航空影像解译的有效途径，为实现航空影像的自动解译打下一定的基础。 本文的主要研究内容包括以下几个方面：航空影像纹理分类与影像分割的支持向量机方法，遗传模糊-C均值的支持向量机样本预选取方法，最小二乘支持向量机及其稀疏性在航空影像分割中的应用，支持向量机与其它方法用于航空影像分类与分割的优劣比较。 (1) 提出将支持向量机用于航空影像的纹理分类与影像分割中，在对多种线性不可分的特征进行分类时，用SVM方法得到了较好的分割与分类结果。 研究了支持向量机参数(核函数、惩罚因子C)和影像特征维数对航空影像分割与分类的影响。SVM中核函数的选择对航空影像纹理的正确分类没有太大的影响，选择不同核函数所对应的最高分类正确率相差不多；但不同的核函数对航空影像分割的影响较大；航空影像纹理分类和分割对常数C敏感；可采用cross-validation方法确定惩罚因子C的初值，调整此初值，达到最好的分割结果。在选择特征时，应尽量多选择特征以得到好的分割与分类结果。 鉴于航空影像的复杂性，决策函数的较小变化会使超平面附近的样本类别发生变化，因而提出了保留支持值α_i=C对应的样本，来保证分割的正确率。 提出了两级金字塔影像上的决策树支持向量机方法，解决航空影像中多类地物的分割问题。 (2) 支持向量机的研究热点之一是对其训练算法的研究，训练学习过程中需要计算和存储的数据大小与训练样本数的平方相关，因此随着样本数目的增多，所需要的内存也就增大。本文提出遗传模糊-C均值的样本预选取方法，保留了最优分类超平面附近的样本点，去除远处样本点，减小训练样本集，从而减少了内存的开销。 对不同的样本集，样本集减小比例略有不同，但样本集都是可以减小的，只要减少后的样本集进行SVM训练的迭代次数和SV个数变化不大，决策函数就变化不大，就可以通过减小样本集，减少内存的开销。同时，通过减小样本集，使SV所占比例提高，也使优化学习过程更有效的集中在SV的优化上。 (3) 支持向量机中惩罚因子C对分类与分割的精度有很大的影响，而C是由人确定的，与人的经验有关。最小二乘支持向量机避免了C值的选择问题，本文用最小二乘支持向量机分割航空影像，其结果比经典方法略差。 提出用LS-SVM的稀疏化处理方法分割航空影像，稀疏化后的分割结果与未作稀疏化处理的分割结果相差较小，因而，可根据最小二乘支持向量机的稀疏性简化决策函数，提高测试速度。 (4)神经网络是近年来广泛应用的一种方法，因其具有并行处理、自学习和高容错性，得到了众多学者的青睐。本文在使用相同的样本和特征的情况下，利用支持向量机和神经网络进行分类和分害lJ，结果表明支持向量机方法好于神经网络方法。原因在于神经网络完全依赖初始权值，而初始权值的确定还没有一个稳健的方法，支持向量机方法依赖于惩罚因子(常数)C值和核函数的选择，结果也不够稳健，其对C的依赖性略大于核函数，但在同一核函数条件下凭经验可在有限次数内找到最优C值。 模糊一C均值方法是一种常用的分割方法，但是，由于本文试验中的影像特征是线性不可分的，即使采用了监督方法，FCM也无法准确将每一像素正确归入它应在的类别。因而，无论是监督的FCM方法还是非监督的FCM方法，其对航空影像的分割都比支持向量机方法差。更多还原

【Abstract】 Image interpretation is a interest area and a bottle problem for photogrammetry and remote sensing. Support Vector Machines (SVM) is a hot research field in Machine Learning. SVM is a kind of novel machine learning method based on Statistical Learning Theory (SLT) proposed by Vapnik. SVM can better solve the learning problem of small sample. This paper proposes that the SVM is used in the classification and segmentation on aerial image. The purpose of the paper is to play foundation for researches on aerial image automatically interpretation.The main contents include the method of Support Vector Machines (SVM) on aerial image texture classification and image segmentation, the pre-selection sample method of Genetic Algorithm fuzzy - C mean, the application of Least Square Support Vector Machines(LS-SVM) and its sparseness, the compare of the SVM and the other methods on aerial image texture classification and image segmentation.(1) This paper proposes that the SVM is used in the aerial image texture classification and image segmentation. The results of SVM with mutli-nonlinear features are good .This paper researches the parameters (kernel, penalty parameter C) of SVM and the dimension of feature, which influence aerial image segmentation and classification. The selection of Kernel function less affects the correct results of aerial image texture classification. The best correct classification rate of different Kernel is similitude. But the effect of different Kernel is large for aerial image segmentation. The effect of C is large for aerial image segmentation and classification. The original value of C could use the method of cross-validation. The original value is adjusted to the best result. The more feature the more better results.Whereas the complexity of aerial image, the decision function occurs little difference which changes the classification of sample near super-plane. This paper proposes that hold the samples of ai = C for assure the segmentation correctness.This paper proposes the method of decision-tree SVM on two levels pyramid image, it could solve the segmentation problem of multi-classes objects on aerial image.(2) One of focus is the training method of SVM. The storage amount is interrelated with the number of samples in the learning process. This paper proposes the pre-selection sample method of Genetic Algorithm fuzzy - C mean. The result is that hold the samples nearing the supper plane, delete the samples far off the supper plane, decrease the training set and the storage.The reducing rate has little different for different of sample set. While the change of iteration number and SV number is little, the change of decision function is small, the memory is reduced through reducing sample set. At the same time, the proportion of SV has been increased, learning more validly focus on theSV’s optimization.(3) The influence of penalty parameter C for classification and segmentation is large. People decide the value of C. LS-SVM avoid the selection the value of C. The segmentation results of LS-SVM are a little bad than SVM. The briefness of decision function is reached by the sparseness of LS-SVM.This paper proposes that the sparseness of LS-SVM is used in the aerial image segmentation. The results between sparseness and non-sparseness has little difference. Thus the decision function could be briefed by the sparseness of LS-SVM. The test speed could be improved.(4) Artificial Neural Network (ANN) is applications widely near years. The paper use SVM and ANN based on the same samples and features to classify and segment. The results indicate that SVM is better than ANN. The reason is ANN complete depends on the original power value, which don’t have a robust decided method. People could find the best C of SVM in several times based on the experience.FCM is a general segmentation method. Because the features in the experiments aren’t linear separate, FCM isn’t successful even used supervised method, and the segmentation results of FCM are bad than SVM.更多还原