面向大规模高光谱数据的半监督地物分类方法

发布时间：2018-06-28 21:18

本文选题：高光谱数据 + 半监督图　；参考：《西安电子科技大学》2015年硕士论文

【摘要】：高光谱遥感技术的快速发展,为地物的精确分类带来契机。对于高光谱数据而言,无监督分类方法不能获得很好的分类效果,而监督分类方法由于在获取标记样本时耗费人力物力,使得分类器泛化能力不强。因此,能够同时利用少量标记样本以及大量未标记样本的半监督方法成为研究的热点,其中基于图的半监督方法由于能通过图有效的反映数据之间的信息,逼近数据之间的流形结构而受到广泛关注。然而,高光谱的高特征维度和数据规模,使得传统基于图的算法承受巨大的计算和存储压力。本文为了解决该问题,提出了几种以锚点图正则为核心的面向大规模高光谱数据的半监督地物分类方法:(1)提出一种基于模糊锚点图正则的高光谱图像分类算法,选取数目远小于样本总数的锚点用于标签预测,能够有效减轻传统基于图的半监督算法中由于高光谱数据规模巨大而带来的计算和存储问题。首先,通过基于空间近邻加权值的核模糊C均值聚类选取锚点并得到隶属度矩阵,然后计算交叉相似度矩阵,并利用该矩阵构建模糊锚点图正则,通过优化求解目标函数,得到锚点的标签,最后根据线性关系计算未标记样本的标记。根据实验仿真,与一些经典的高光谱图像分类算法进行比较,可以看出本章算法与其它算法相比具有较好的分类效果。(2)提出一种基于局部模糊锚点图正则和松弛聚类的高光谱图像分类算法,在上一章算法的基础之上,利用交叉相似度矩阵对高光谱数据进行聚类与空间分割,将图像分割成为一系列的超像素,将每个超像素与周围的超像素联合起来组成超像素集,计算超像素集中样本之间的相似度,将其用于构建局部模糊锚点图正则,并引入松弛聚类假设,减小混合像元对分类精度带来的负面影响,之后对每个超像素集迭代计算锚点标签和样本属于各个类别的概率向量,使用迭代得到锚点的标签对超像素中样本的进行标签预测。在实际高光谱数据上进行仿真,可以看出本章算法具有明显的优势,分类效果更好,具有空间局部一致性。(3)提出一种基于模糊锚点图正则的高光谱分类算法分布式实现,首先采用随机采样和规则采样对数据进行分配,并发送给各个子系统,接着使用分布式基于空间近邻加权的核模糊C均值聚类方法选取锚点,针对不同的数据采样方式实现相应基于锚点图正则的分布式算法。通过实验可以看出,本方法能够被实现,说明并行方法的可行性。
[Abstract]:The rapid development of hyperspectral remote sensing technology brings opportunities for accurate classification of ground objects. For hyperspectral data, the unsupervised classification method can not achieve a good classification effect, but the supervised classification method does not have strong generalization ability because it consumes manpower and material resources to obtain labeled samples. Therefore, the semi-supervised method, which can use a small number of labeled samples and a large number of unlabeled samples at the same time, has become a hot research topic, and the graph-based semi-supervised method can effectively reflect the information between the data through the graph. Approximation of manifold structures between data has attracted widespread attention. However, hyperspectral hypercharacteristic dimension and data scale make the traditional graph-based algorithms bear huge computational and storage pressures. In order to solve this problem, several semi-supervised classification methods for large-scale hyperspectral data based on anchor map regularization are proposed. (1) A hyperspectral image classification algorithm based on fuzzy anchor map regularization is proposed. The selection of anchor points which are far less than the total number of samples for label prediction can effectively alleviate the computational and storage problems caused by the large scale of hyperspectral data in traditional graph-based semi-supervised algorithms. Firstly, the anchor points are selected by the kernel fuzzy C-means clustering based on the weighted value of spatial nearest neighbor, and the membership matrix is obtained. Then, the cross-similarity matrix is calculated, and the fuzzy anchor graph is constructed to be regular, and the objective function is optimized. Finally, the labels of the unlabeled samples are calculated according to the linear relationship. According to the experimental simulation, compared with some classical hyperspectral image classification algorithms, It can be seen that this algorithm has better classification effect than other algorithms. (2) A hyperspectral image classification algorithm based on local fuzzy anchor graph regularization and relaxation clustering is proposed, which is based on the previous chapter. Using cross-similarity matrix to cluster and segment hyperspectral data, the image is segmented into a series of super-pixels, and each super-pixel is combined with the surrounding super-pixel to form a super-pixel set. The similarity between samples in hyperpixel set is calculated and used to construct local fuzzy anchor graph regularization. The relaxed clustering hypothesis is introduced to reduce the negative effect of mixed pixels on classification accuracy. Then the anchor label and the probability vector of the sample belonging to each class are calculated iteratively for each super-pixel set, and the label of the anchor point is used to predict the label of the sample in the super-pixel. Simulation on the actual hyperspectral data shows that this algorithm has obvious advantages, better classification effect and spatial local consistency. (3) A distributed hyperspectral classification algorithm based on fuzzy anchor graph regularization is proposed. Firstly, random sampling and regular sampling are used to distribute the data and send them to each subsystem. Then the distributed kernel fuzzy C-means clustering method based on spatial nearest neighbor weighting is used to select anchor points. A distributed algorithm based on anchor graph regularization is implemented for different data sampling methods. The experiment shows that the method can be implemented and the feasibility of the parallel method is demonstrated.
【学位授予单位】：西安电子科技大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TP751

【相似文献】