基于多特征的面向对象高分辨率遥感图像分类

发布时间：2018-05-25 20:55

本文选题：图像分割 + 图像分类　；参考：《电子科技大学》2017年硕士论文

【摘要】：目前高分辨率遥感图像的应用呈现两个增长趋势,一个是应用领域的增加,一个是应用复杂度的增加。高分辨率遥感图像在城市土地使用情况统计、城市生态评估、灾害评估、农业灌溉等方面均有重要应用。遥感图像空间分辨率的增加,一方面使得图像中的地物细节更清晰,另一方面也增加了图像信息分析的难度。为了提高分类的正确率,本文结合了空间特征和颜色特征等多种特征。同时,为了处理高分辨率图像的大量数据并减少计算量,在图像分割和分类中应用了面向对象的思想。本文主要工作如下:1.将面向对象的思想引入遥感图像分割。面向对象的分析方法不仅具有良好的抗噪声性,且在降低计算量的同时能够保证分割结果的准确性。为限制计算复杂度,本文通过适当的分水岭变换得到图像的超像素表示。本文采用区域邻接图(region-adjacency graph,RAG)度量初始超像素块的相似性,将图像分割问题转化为图割问题。多组实验表明,使用基于超像素的分割方法所得分割结果几乎不存在过分割现象,并且分割结果的边界正确率得到较好保证。2.建立包含多种特征的特征集合,用于高分辨率遥感图像的分类。传统分类算法中,依靠光谱特征和纹理特征实现遥感图像的分类。然而高分辨率图像中大量增加的地物细节对特征提出了新的要求。为了有效描述图像的空间信息,增加形态学特征APs(morphological attribute profiles)。APs特征可以根据选择的属性类型生成不同的特征。与常规的基于预定义的结构元的形态滤波器相比,APs可以提供一个多层次的图像分析,从而得到更精确的空间信息。本文通过大量实验验证了APs特征用于高分辨率图像分类的有效性。颜色特征的引入,丰富了特征集合,增强了不同类别之间的区分度。本文实验表明,颜色特征的增加能够改善图像中阴影等地物的分类情况。由于不同特征在提取图像信息时各有侧重,因而如何选择合适的特征是图像分类的关键问题之一。本文研究了不同特征组合的分类结果,利用SVM分类算法实现面向对象的分类。分类以结合图论的基于超像素的分割算法所得分割结果为基础进行。统计6种特征组合的分类结果并分析,发现联合光谱特征和空间特征以及颜色特征的特征集合可以获得较为理想的分类结果。
[Abstract]:At present, the application of high resolution remote sensing image shows two increasing trends, one is the increase of application field and the other is the increase of application complexity. High-resolution remote sensing images have important applications in urban land use statistics, urban ecological assessment, disaster assessment, agricultural irrigation and so on. The increase of spatial resolution of remote sensing image makes the details of ground objects more clear on the one hand and makes the analysis of image information more difficult on the other hand. In order to improve the accuracy of classification, this paper combines spatial features and color features. At the same time, in order to deal with a large number of high-resolution image data and reduce the amount of computation, the object-oriented idea is applied in image segmentation and classification. The main work of this paper is as follows: 1. The object-oriented idea is introduced into remote sensing image segmentation. The object-oriented analysis method not only has good anti-noise performance, but also can ensure the accuracy of the segmentation results while reducing the computational complexity. In order to limit the computational complexity, the super-pixel representation of the image is obtained by appropriate watershed transformation. In this paper, region-adjacency graph rag is used to measure the similarity of initial superpixel blocks, and the image segmentation problem is transformed into graph cutting problem. Multi-group experiments show that there is almost no over-segmentation phenomenon in the segmentation results obtained by using the hyperpixel segmentation method, and the boundary accuracy of the segmentation results is better guaranteed. 2. A feature set consisting of multiple features is established for the classification of high resolution remote sensing images. In the traditional classification algorithm, remote sensing image classification is realized by spectral feature and texture feature. However, a large number of feature details in high resolution images require new features. In order to effectively describe the spatial information of the image, the addition of morphological features APs(morphological attribute profiles).APs features can generate different features according to the selected attribute types. Compared with the conventional morphological filter based on predefined structure elements, the APs can provide a multi-level image analysis to obtain more accurate spatial information. The effectiveness of APs feature in high resolution image classification is verified by a large number of experiments in this paper. The introduction of color features enriches feature sets and enhances the differentiation between different categories. Experiments show that the increase of color features can improve the classification of objects such as shadows in images. Because different features have different emphases in extracting image information, how to select suitable features is one of the key problems in image classification. In this paper, the classification results of different feature combinations are studied, and the object-oriented classification is realized by using SVM classification algorithm. The classification is based on the segmentation results obtained from the hyperpixel segmentation algorithm combined with graph theory. The classification results of six kinds of feature combinations are analyzed and it is found that the combination of spectral features and spatial features and the feature sets of color features can obtain more ideal classification results.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP751

【参考文献】