基于纹理分析的高分辨率影像面向对象分类研究

发布时间：2018-04-19 23:31

本文选题：面向对象分类 + 灰度共生矩阵　；参考：《北京师范大学》2014年硕士论文

【摘要】：随着卫星遥感数据空间分辨率的不断提高，使用传统的基于像元的遥感影像处理方法不仅无法充分利用高分辨率影像中的空间细节信息，还会因为“同物异谱”以及“同谱异物”的现象导致分类结果出现较多的漏分和误分，同时分类结果还会呈现严重的“椒盐噪声”，严重影响了分类精度。应运而生的面向对象分类方法，能有效的抑制上述问题，因此受到了众多学者的关注。纹理信息作为一种重要的影像空间特征信息，在遥感影像分类中有着广泛的应用。众多学者利用纹理信息辅助分类获得了较好的效果。但是目前的研究大多基于传统的面向像元分类，即便是基于面向对象分类的纹理研究也多采用了单一尺度的面向对象分类。并未就纹理信息在多尺度面向对象分类中对分类精度的影响进行深入研究。因此本文针对前人研究，使用唐山市丰南区的IKONOS数据构建了多尺度的面向对象分类体系，并以此研究了纹理信息的添加对于分类精度的影响，得到以下的研究成果：本文通过ESP(Estimation of Scale Parameters)工具和多次试验，确定了研究区域内主要地类最适宜的分割尺度和分割参数，建立了三级的多尺度分割层次（81,47,16），并依此建立了多级的分类体系，体现了多尺度分割在面向对象分类中的优势。在面向对象分类的基础上，提取了8种GLCM（灰度共生矩阵）纹理和3种LSS（局部空间统计）纹理，在SVM（Support Vector Machine）和NN（最邻近）两种分类器下，研究了不同纹理对于总体精度以及各类别PA（用户精度）和UA（制图精度）的影响。实验证明，该实验条件下添加单纹理信息能有效提高总体精度以及大部分类别的PA和UA，在SVM分类器下，纹理信息的添加对总体精度的影响较小，均在1%左右，而在NN分类器下，纹理信息的添加对总体精度的影响较大，Geary's C纹理拥有最佳的总体精度（79.06%），相比单纯使用多光谱的分类结果提升了4.6%的总体精度。选择了部分纹理来研究纹理尺度参数对于分类精度的影响，结果显示尺度参数的变化会对分类结果产生一定的影响，，但这种影响会因为面向对象分类本身的机制问题而削弱。提出了一种基于蚁群算法的最优纹理特征组合选择方法，能够在保证较高分类精度的情况下大幅缩减特征维数，可以在未分类的情况下，仅根据样本就可以得到最优的特征组合。得到了两种分类器在实验区内的最优纹理特征组合，并进行了验证。
[Abstract]:With the continuous improvement of spatial resolution of satellite remote sensing data, the traditional pixel based remote sensing image processing method can not make full use of the spatial details of high-resolution images. Because of the phenomenon of "isomorphism" and "isospectral foreign body", there will be more missing points and false scores in the classification results. At the same time, the classification results will also present serious "salt and pepper noise", which seriously affects the classification accuracy. The object-oriented classification method which arises at the historic moment can restrain the above problems effectively, so many scholars pay close attention to it. As an important spatial feature information, texture information is widely used in remote sensing image classification. Many scholars use texture information to assist classification to obtain better results. However, most of the current researches are based on traditional pixel oriented classification, and even the texture research based on object oriented classification is mostly based on single scale object oriented classification. The effect of texture information on classification accuracy in multi-scale object-oriented classification is not studied. Therefore, this paper constructs a multi-scale object-oriented classification system based on the IKONOS data of Fengnan District, Tangshan City, and studies the effect of texture information on classification accuracy. The following research results are obtained: By means of ESP(Estimation of Scale parameters tool and many experiments, this paper determines the most suitable segmentation scale and segmentation parameters for the main ground classes in the study area, and establishes a multi-scale multi-scale segmentation level of 81D 4716m, based on which a multilevel classification system is established. It shows the advantage of multi-scale segmentation in object-oriented classification. On the basis of object oriented classification, eight GLCM textures and three LSS (Local Spatial Statistics) textures are extracted, which are based on SVM(Support Vector Machine and NN (nearest neighbor) classifiers. The effects of different textures on the overall accuracy, as well as various types of PAs (user accuracy) and UA( cartographic accuracy) are studied. Experiments show that adding single texture information can effectively improve the overall precision and the PA and UAA of most categories under the experimental condition. In SVM classifier, the effect of adding texture information on the overall accuracy is less, which is about 1%, while in NN classifier, the effect of adding texture information on the overall accuracy is about 1%. The addition of texture information has a great influence on the overall precision. The GearyCtexture has the best overall precision 79.060.Compared with the classification results using multi-spectral method, the overall accuracy is increased by 4.6%. Some textures are selected to study the effect of texture scale parameters on classification accuracy. The result shows that the variation of scale parameters will have a certain impact on the classification results, but this effect will be weakened by the mechanism of object oriented classification. An optimal texture feature combination selection method based on ant colony algorithm is proposed, which can greatly reduce the feature dimension in the case of high classification accuracy, and can be used in the case of no classification. The optimal feature combination can be obtained only according to the sample. The optimal texture feature combinations of the two classifiers in the experimental region are obtained and verified.
【学位授予单位】：北京师范大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：P237

【参考文献】