高空间—高光谱分辨率的遥感图像城市场景分类识别研究

发布时间：2018-07-05 17:04

本文选题：高光谱 + 城市场景　；参考：《哈尔滨工业大学》2014年硕士论文

【摘要】：高光谱技术是近几十年来地球观测技术取得的最重大成就之一。高光谱图像光谱分辨率较高，能够提供非常丰富的光谱信息，因此在许多领域得到了广泛应用，但是由于空间分辨率较低，在用来对城市场景进行分析时受到很大限制。城市中场景的面积较小且分布密集，使用低空间分辨率的图像不能有效区分地物。随着高光谱传感器技术的发展，高光谱图像的空间分辨率有了较大提升，许多小区域场景能够由像素来描述，使得利用高空间-高光谱分辨率的遥感图像对城市场景进行分析成为可能。本文将利用具有较高空间分辨率的高光谱图像对城市场景进行分类识别研究。具体工作内容如下：首先，根据高光谱图像的数据特性进行光谱特征提取。在采用传统的局部Fisher判别分析方法和近邻保留嵌入方法的基础上，将二者结合起来，提出了一种半监督局部判别分析方法。本方法综合考虑了已知样本的可分性信息和未知样本的结构信息。基于该方法对城市场景的光谱特征进行提取，利用支持向量机和最大似然方法进行分类，通过与其他特征提取算法进行对比分析，验证了半监督局部判别分析方法提取特征的分类效果。然后，利用城市高光谱图像的高空间分辨率特点，提取空间特征，包括形态学特征、形状特征等。进而对空谱特征的联合方式进行了重点研究，主要采用以下三种方式：第一，直接对光谱和空间特征进行组合；第二，采用基于核函数的特征组合方式，将不同的特征在核变换空间进行组合；第三，采用多特征组合框架对特征进行组合，对不同特征进行降维，通过保留尽可能多的信息来实现组合。使用单种特征和各种空谱联合特征进行支持向量机分类实验，结果表明空谱联合特征会提升光谱或空间特征的分类精度。最后，针对城市场景分类识别中可能存在的训练样本不足问题，，结合主动学习方法进行了研究，并利用判别随机场模型对分类结果进行优化。在传统主动学习算法的基础上，利用已知样本信息，提出了一种确定候选样本集的方法。该方法最大的优势在于主动学习的过程中不需要人工标记选出的新样本。论文还针对归属类别概率的输出问题，研究了一种基于逻辑回归模型的多项式逻辑回归分类方法。通过使用多项式逻辑回归方法和支持向量机对提出的确定候选集的方法进行了验证，实验表明该方法不但节省了人力消耗，而且可以在小样本情况下有效提高分类精度。
[Abstract]:Hyperspectral technology is one of the most important achievements in Earth observation technology in recent decades. Hyperspectral images have high spectral resolution and can provide rich spectral information, so they are widely used in many fields. However, because of their low spatial resolution, they are limited in analyzing urban scenes. Because of the small area and dense distribution of the scene in the city, the low spatial resolution image can not effectively distinguish the ground objects. With the development of hyperspectral sensor technology, the spatial resolution of hyperspectral images has been greatly improved, many small regional scenes can be described by pixels. It is possible to use high spatial and hyperspectral resolution remote sensing images to analyze urban scenes. In this paper, hyperspectral images with high spatial resolution are used to classify and recognize urban scenes. The main work is as follows: firstly, spectral feature extraction is carried out according to the data characteristics of hyperspectral image. Based on the traditional local Fisher discriminant analysis method and the nearest neighbor retention embedding method, a semi-supervised partial discriminant analysis method is proposed. In this method, the separability information of known samples and the structural information of unknown samples are considered. Based on this method, the spectral features of urban scenes are extracted. Support vector machine (SVM) and maximum likelihood method are used to classify and compare with other feature extraction algorithms. The classification effect of the discriminant analysis method is verified. Then, the spatial features, including morphological features and shape features, are extracted by using the high spatial resolution of urban hyperspectral images. Then, the paper focuses on the joint method of space-spectrum features, mainly using the following three ways: first, the spectral and spatial features are combined directly; second, the kernel-based feature combination method is adopted. The different features are combined in kernel transform space. Thirdly, multi-feature combination framework is used to combine the features, reduce the dimension of different features, and achieve the combination by retaining as much information as possible. The support vector machine (SVM) classification experiment is carried out by using a single feature and a variety of space-spectrum joint features. The results show that the space-spectrum joint feature can improve the classification accuracy of spectral or spatial features. Finally, aiming at the problem of insufficient training samples in urban scene classification and recognition, combining with active learning method, the classification results are optimized by discriminant random field model. Based on the traditional active learning algorithm and using the known sample information, a method to determine the candidate sample set is proposed. The biggest advantage of this method is that it does not need to mark the new samples in the process of active learning. A polynomial logistic regression classification method based on logical regression model is also studied in this paper. The polynomial logic regression method and support vector machine are used to verify the proposed method to determine the candidate set. The experimental results show that the proposed method not only saves manpower consumption, but also can effectively improve the classification accuracy in the case of small samples.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP751

【参考文献】