基于局部支持向量机的高分辨率遥感图像分类

发布时间：2018-08-31 11:33

【摘要】：随着图像采集传感技术的发展以及社会需求的不断提高,遥感图像呈现出高分辨率的发展趋势。分辨率越高的遥感图像所含信息内容越丰富,对各种应用领域的发展促进更大。因此,高分辨率的遥感图像成为地物目标识别的重要来源。遥感图像分辨率的日益增加也给遥感图像处理领域带来越来越大的挑战。随着遥感图像分辨率的不断提升,应用领域对其精度的需求不断增加,导致需要处理的数据量呈指数级的增长。因此,传统的人工目视解译的图像处理模式已完全不能满足现实需求,利用数据挖掘、机器学习等依靠计算机自动进行图像分类成为遥感图像处理领域的主流。然而,即使是利用计算机自动分类也因为数据量的爆炸性增长而导致处理时间长,分类精度达不到预期效果等现象；另一方面,高分辨率遥感图像具有更多的图像特征,除了包括传统图像具有的光谱信息以外,还有纹理特征、几何特征以及上下文信息等等。而传统的遥感图像处理主要依据单个特征进行分类,这种方式因对地物中存在的“同物异谱”以及“同谱异物”问题无法很好解决而导致分类结果差。支持向量机(SVM)模型是依靠训练样本集中的支持向量学习而来,具有稀疏性好,分类精度高以及运行速度快等优势,特别适合处理训练数据集小、非线性以及高维特征的分类问题。针对具有海量数据的高分辨率遥感图像分类,支持向量机具有较大的优势。然而,这种传统支持向量机没有全局一致性,为了进一步提高其分类精度,融合近邻分类器KNN和SVM的局部支持向量机算法被应用于遥感图像的分类问题中。局部支持向量机结合了两个分类器的优势,因此分类精度更高。根据高分辨率遥感图像分类问题的特点,为了进一步提高对高分辨率遥感图像的分类效果,降低其计算的复杂度,文章在对现有研究成果的梳理下,对局部支持向量机在高分辨率遥感图像分类中的应用进行了深入研究,主要的研究工作如下：1、提出了基于不确定性的改进KNNSVM的局部支持向量机—-BKNNSVM算法。详细分析了具有代表性的局部支持向量机KNNSVM算法存在的算法时间复杂度高的不足。对KNNSVM算法进行深入研究后,发现KNNSVM算法是通过严格的K近邻分类器(SKNN)和局部支持向量机(LSVM)两个分类器的融合来确定样本类标。分析KNNSVM算法的时间开销曲线,发现需要建立SVM分类器确定类标的样本越多,时间开销越大。考虑到对任何一个未标记样本其近邻分布都是满足二项式分布的,借助Beta分布对未标样本不确定性的计算,论文提出了BKNNSVM算法,该算法通过设定不确定性阈值及近邻K值两个参数值的大小,通过增加(减少)参与KNN分类器的分类样本数,从而减少(增加)需要建立SVM分类器的个数来调节BKNNSVM算法的时间复杂度。实验结果显示,设置合理的阈值及K值的BKNNSVM能够在保持KNNSVM算法精度的同时,明显地降低算法的时间复杂度。2、提出了基于距离的局部支持向量机算法DLSVM。通过对SVM错分样本在超平面附近的分布特点,发现超平面附近的样本分类错误率最高,分类精度最低,但这些样本利用局部支持向量机进行分类时,分类精度较高。为了寻找这些被错分的点,我们借助了主动学习的相关理论,认为离超平面越近,传统支持向量机对样本的错分率越高。因此,DLSVM首先计算未标记样本离超平面的距离,针对离超平面较近的样本建立局部支持向量机,提高其分类精度,而远距离的样本直接采用传统的SVM分类以减少分类时间。由于只对少数近距离样本采用局部支持向量机算法,DLSVM在传统支持向量机的基础上时间复杂度增加不多,但分类精度有显著提高,特别地,该算法的时间复杂度远小于KNNSVM算法。3、提出了多类分类问题的有向无环图局部支持向量机(DAGLSVM)分类方法。支持向量机分类器SVM是用来解决二分类问题的。因为其具有分类精度高、泛化能力强,因此也被推广运用于多分类问题。SVM解决多分类问题的主要方法有一对一,一对多、有向无环图DAGSVM及二叉树SVM等。考虑到如果将KNNSVM算法直接应用到一对一的多类分类问题中,建立局部支持向量机的数量将在二分类问题的基础上呈二次方倍增,当类数目较多时,时间开销将无法忍受。因此,为了降低时间复杂度,提出了基于有向无环图(DAG)的多类分类局部支持向量机算法——DAGLSVM。该算法为每一个未标记样本建立近邻训练样本集,在DAG拓扑结构中的每个节点,选择未标记样本类不确定性最小的两个类优先进行决策,以降低导致累积误差的风险。最后的实验结果显示,在合适的K值参数下,DAGLSVM算法在运行时间消耗增加不大的情况下,能够有效提高高分辨率遥感图像的分类精度。4、设计了遥感图像处理软件平台,并通过该平台完成文章中涉及的所有实验项目。该平台在台湾大学林智仁教授开发的开源代码一-LIBSVM的基础上,利用JAVA平台设计开发了遥感图像支持向量机分类软件。该软件以LIBSVM算法为核心,综合数据挖掘软件平台——WEKA平台,实现了KNN、KNNSVM、BKNNSVM、DLSVM、DAGSVM以及DAGLSVM等多种分类算法。完成了对遥感图像的显示,图像特征提取、分类、分类结果存储以及结果显示等功能。总之,由于局部支持向量机不仅能够提高传统支持向量机的精度,还能保持其泛化性好、支持小训练样本集的特点,因此,本文针对局部支持向量机算法提出的各种优化算法能够提高高分辨率遥感图像的分类性能,降低其运算的时间复杂度,有利于进一步提高遥感图像的处理速度,发挥其更大的社会经济效益。
[Abstract]:With the development of image acquisition and sensing technology and the continuous improvement of social demand, remote sensing images show a trend of high resolution. The higher the resolution of remote sensing images, the richer the information content, the greater the development of various applications. Therefore, high resolution remote sensing images become an important source of object recognition. The increasing resolution of remote sensing image also brings more and more challenges to the field of remote sensing image processing. With the continuous improvement of the resolution of remote sensing image, the demand for its accuracy in the application field is increasing, resulting in an exponential increase in the amount of data to be processed. Unable to meet the actual needs, the use of data mining, machine learning and other computer-based automatic image classification has become the mainstream in the field of remote sensing image processing. High-resolution remote sensing images have more image features, including texture features, geometric features and context information besides the spectral information of traditional images. However, traditional remote sensing image processing is mainly based on a single feature classification, which is due to the existence of "similarities and differences" and "similarities and differences" in terrain. Support Vector Machine (SVM) model is based on the support vector learning of training sample set. It has the advantages of good sparsity, high classification accuracy and fast running speed. It is especially suitable for dealing with the classification problem of small training data set, nonlinear and high-dimensional features. However, this traditional support vector machine does not have global consistency. In order to further improve the classification accuracy, the local support vector machine (LSVM) algorithm which combines the nearest neighbor classifier KNN and SVM is applied to the classification of remote sensing images. According to the characteristics of high-resolution remote sensing image classification problem, in order to further improve the classification effect of high-resolution remote sensing image and reduce the computational complexity, this paper combs the existing research results, on the basis of local support vector machine in high-resolution remote sensing image classification. The application of image classification is studied deeply. The main research work is as follows: 1. An improved KNNSVM-BKNNSVM algorithm based on uncertainty is proposed. The shortcomings of KNNSVM algorithm with high time complexity are analyzed in detail. After that, it is found that the KNNSVM algorithm determines the sample class scale by the fusion of the strict K-nearest neighbor classifier (SKNN) and local support vector machine (LSVM). By analyzing the time cost curve of the KNNSVM algorithm, it is found that the more samples need to be established to determine the class scale by SVM classifier, the greater the time cost. The nearest neighbor distribution satisfies the binomial distribution. The BKNNSVM algorithm is proposed by calculating the uncertainty of unmarked samples with Beta distribution. By setting the uncertainty threshold and the size of the nearest neighbor K value, the algorithm reduces (increases) the number of samples participating in the KNN classifier by increasing (decreases) the number of samples participating in the classification. The experimental results show that BKNNSVM with reasonable thresholds and K values can significantly reduce the time complexity while maintaining the accuracy of the KNNSVM algorithm. 2. A distance-based local support vector machine algorithm DLSVM is proposed. Near the hyperplane, the sample classification error rate is the highest and the classification accuracy is the lowest, but these samples are classified by local support vector machine with high classification accuracy. Therefore, DLSVM firstly calculates the distance between unlabeled samples and hyperplane, and establishes local support vector machine (LSVM) for the samples nearer to the hyperplane to improve the classification accuracy. However, the traditional SVM classification is directly used for the remote samples to reduce the classification time. In particular, the time complexity of this algorithm is much less than that of KNNSVM. 3. A directed acyclic graph local support vector machine (DAGLSVM) classification method for multi-class classification problems is proposed. Support vector machine classifier SVM is used to solve the two-class classification problem. Because of its high classification accuracy and strong generalization ability, SVM is also widely used in multi-class classification problems. The main methods of SVM to solve multi-class classification problems are one-to-one, one-to-many, directed acyclic graph DAGSVM and binary tree SVM. Considering that if KNNSVM algorithm is directly applied to one-to-one multi-class classification problems, the locality is established. The number of support vector machines will be quadratically multiplied on the basis of binary classification problem, and the time cost will be unbearable when the number of classes is large. Therefore, in order to reduce the time complexity, a multi-class classification local support vector machine algorithm based on directed acyclic graph (DAG) called DAGLSVM is proposed. Neighborhood training sample sets, each node in the DAG topology, choose two classes with the least uncertainty of unlabeled sample classes to make decisions in order to reduce the risk of cumulative errors. Finally, the experimental results show that the DAGLSVM algorithm can effectively improve the performance of the DAGLSVM algorithm under the appropriate K-value parameters without increasing the running time. The classification accuracy of high-resolution remote sensing image is 4.The software platform of remote sensing image processing is designed, and all the experimental projects involved in this paper are completed through this platform.Based on the open source-LIBSVM developed by Professor Lin Zhiren of Taiwan University, the classification software of remote sensing image is designed and developed by using JAVA platform. The software takes LIBSVM algorithm as the core and integrates data mining software platform-WEKA platform. It implements various classification algorithms such as KNN, KNNSVM, BKNNSVM, DLSVM, DAGSVM and DAGLSVM. It completes the functions of remote sensing image display, image feature extraction, classification, classification result storage and result display. It can not only improve the accuracy of traditional support vector machine, but also maintain its good generalization and support the characteristics of small training sample set. Therefore, various optimization algorithms proposed in this paper for local support vector machine algorithm can improve the classification performance of high-resolution remote sensing images, reduce the time complexity of its operation, which is conducive to further improving remote sensing. The speed of image processing will bring greater social and economic benefits.
【学位授予单位】：中国地质大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：P237

【相似文献】