基于机器学习的高光谱图像地物分类研究

发布时间：2018-01-31 07:14

本文关键词： 高光谱图像拉普拉斯支持向量机半监督学习　出处：《厦门大学》2014年硕士论文　论文类型：学位论文

【摘要】：基于高光谱图像的地物分类是遥感领域的研究热点。高光谱遥感数据最主要的特点是：将传统的图象维与光谱维信息融合为一体,在获取地表空间图象的同时,得到每个地物的连续光谱信息,从而实现依据地物光谱特征的地物成份信息反演与地物识别。高光谱内部存在着流形结构和内部数据空间信息等特点。高光谱标记样本少而昂贵,高光谱的维度高且数据量很大,使得对高光谱的数据处理出现问题。本文从高光谱数据特点入手,对高光谱地物分类进行分析。主要研究成果如下：首先,引入拉普拉斯支持向量机对高光谱数据进行地物分类。普拉斯算子能够将高维的高光谱数据映射到低维空间中进行非线性降维,发现内在的流形结构。因此,将拉普拉斯应用于高光谱分类能有效的利用高光谱的特点使地物分类达到更高的准确度。在解决方案中,初始预共轭梯度优化方法和对偶方法是两个不同的解法。预共轭梯度解法的拉普拉斯支持向量机解法在原始方法的对偶框架上加入了基于预测稳定性的梯度下降早期停止条件。本文将拉普拉斯预共轭梯度引入到高光谱图像的地物分类当中,提出了利用预共轭梯度解法来对高光谱数据进行地物分类,极大地减少了训练和分类时间以及复杂度。、其次,针对分类器参数优化方面,本文提出了模拟退火算法对各类分类器进行参数的优化。在拉普拉斯的三个参数即核参数、惩罚因子、拉普拉斯因子的设置方面,模拟退火算法对参数进行优化,与传统的网格搜索优化算法相比节省了大量的时间,特别是存在多个参数的情况下,较少了计算复杂度。本文通过利用Indian Pine数据集中的六类地物类别进行分类试验,同时对比了其他不同种类的分类器,显示了本论文所提出方法的优越性和分类的精确性。
[Abstract]:The classification of ground objects based on hyperspectral images is a hotspot in the field of remote sensing. The most important feature of hyperspectral remote sensing data is that the traditional image dimension and spectral dimension information are integrated to obtain the surface spatial image at the same time. Get the continuous spectral information for each object. In order to realize the retrieval and recognition of the composition information of the ground objects according to the spectral characteristics of the ground objects, the hyperspectral interior has the characteristics of manifold structure and internal data spatial information, and the hyperspectral labeling samples are small and expensive. The hyperspectral dimension is high and the amount of data is very large, which makes the processing of hyperspectral data appear problems. This paper starts with the characteristics of hyperspectral data, and analyzes the classification of hyperspectral features. The main research results are as follows: Firstly, Laplace support vector machine is introduced to classify hyperspectral data. Plath operator can map high-dimensional hyperspectral data to low-dimensional space for nonlinear dimensionality reduction. Therefore, the application of Laplace to hyperspectral classification can effectively make use of the characteristics of hyperspectral to achieve higher accuracy in the classification of ground objects. The initial preconjugate gradient optimization method and the dual method are two different solutions. The Laplace support vector machine method for the preconjugate gradient method adds gradient descent based on predictive stability to the dual frame of the original method. In this paper, Laplace preconjugate gradient is introduced into the classification of ground objects in hyperspectral images. A preconjugate gradient method is proposed to classify the ground objects of hyperspectral data, which greatly reduces the training and classification time and complexity. Secondly, for the optimization of classifier parameters, this paper proposes simulated annealing algorithm to optimize the parameters of all kinds of classifiers. In Laplacian, three parameters, kernel parameters, penalty factor. In the setting of Laplace factor, the simulated annealing algorithm optimizes the parameters, which saves a lot of time compared with the traditional mesh search optimization algorithm, especially in the case of multiple parameters. Less computational complexity. In this paper, by using the Indian Pine data set of the six categories of ground objects classification experiments, at the same time compared with other different types of classifiers. The advantages of the proposed method and the accuracy of classification are demonstrated.
【学位授予单位】：厦门大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP751;TP181

【参考文献】