多谱遥感影像特征提取及协同解译研究

发布时间：2018-06-29 07:50

本文选题：多谱 + 协同解译　；参考：《哈尔滨工业大学》2014年硕士论文

【摘要】：随着遥感对地观测技术的发展，多谱遥感数据越来越多地应用在军事目标检测、城市绿化监测、农业遥感等领域，如拥有近乎连续的光谱信息的高光谱影像、可提供更精细的形状纹理等空间信息的高空间分辨率遥感影像、可反映地物热辐射信息的红外数据、可获得地物数字高程模型的激光雷达数据等。然而对这些多平台数据的应用都面临着数据量更加巨大、信息更加复杂的问题，因此无论是针对高空间分辨率还是高光谱分辨率的遥感影像，都迫切地需要我们对其信息提取以及分类解译方法进行不断深入的研究。本论文主要针对多谱遥感影像数据中的高光谱数据及高空间分辨率遥感影像进行协同处理，利用光谱与空间的联合信息实现多谱遥感影像在特征层的协同解译。首先，针对高光谱影像，分析了其数据特点，利用高光谱影像具有的图谱合一、非线性和稀疏性的特点，研究了受限玻尔兹曼机算法原理，介绍了深层网络模型框架下的半监督的贪婪逐层学习方法，同时尝试了经典的主成分分析及非负矩阵分解的特征提取方法，对比了支持向量机与深度置信网络模型对高光谱影像地物分类解译的解译精度及训练、测试时间，为后续协同解译奠定理论研究基础。实验验证了深度置信网络对于具有复杂光谱信息的类别具有出色的分类效果。其次，为了解决随着高分辨率遥感影像分辨率越高，纹理信息越复杂，分类精度随之下降的问题，计算不同地物的最优分割尺度，采用分形网络演化算法对影像进行多尺度分割，而后分别采用基于样本对象和基于多层次体系结构的面向对象的信息提取方法对各类别地物进行信息提取分类，与基于像素的处理方法对比，验证了面向对象处理的有效性。最后，利用高分辨率遥感影像的信息提取结果，合并相邻的同类地物对象，对合并后的对象提取空间特征，，与高光谱影像的光谱特征进行特征层融合，将空谱联合特征输入深度置信网络模型，讨论了不同学习率、隐层单元数、网络层数、训练样本比例等网络参数对模型的影响，利用两种数据源各自的优势，提高了单一数据源的解译精度，实现多谱遥感影像特征层的高精度协同解译。
[Abstract]:With the development of remote sensing Earth observation technology, multispectral remote sensing data are more and more used in military target detection, urban greening monitoring, agricultural remote sensing and other fields, such as hyperspectral images with almost continuous spectral information. High spatial resolution remote sensing images which can provide more precise spatial information such as shape and texture can reflect the infrared data of the thermal radiation information of the ground objects and obtain the lidar data of the digital elevation model of the ground objects and so on. However, the applications of these multi-platform data are faced with the problems of more huge amount of data and more complex information, so it is not only for the remote sensing image with high spatial resolution or hyperspectral resolution, but also for the remote sensing image with high spatial resolution or hyperspectral resolution. It is urgent for us to study the methods of information extraction and classification. In this paper, the hyperspectral data and the spatial resolution remote sensing image of multi-spectral remote sensing image are processed in cooperation, and the cooperative interpretation of multi-spectral remote sensing image in the feature layer is realized by using the joint information of spectrum and space. Firstly, the characteristics of hyperspectral images are analyzed, and the algorithm principle of constrained Boltzmann machine is studied, which is based on the characteristics of unifying, nonlinear and sparsity of hyperspectral images. In this paper, the semi-supervised greedy learning method based on deep network model is introduced, and the classical principal component analysis (PCA) and non-negative matrix decomposition feature extraction method are tried. The accuracy, training and test time of classification interpretation of hyperspectral images based on support vector machine and depth confidence network model are compared, which lays a theoretical foundation for subsequent cooperative interpretation. The experimental results show that the depth confidence network has a good classification effect for the categories with complex spectral information. Secondly, in order to solve the problem that the higher the resolution of high resolution remote sensing image is, the more complex the texture information is and the lower the classification accuracy is, the optimal segmentation scale of different ground objects is calculated. The fractal network evolution algorithm is used to segment the image at multi-scale, and then the object oriented information extraction method based on the sample object and the multi-level architecture is used to extract and classify the information of each kind of ground objects. Compared with the pixel-based processing method, the effectiveness of object-oriented processing is verified. Finally, using the information extraction result of high-resolution remote sensing image, merging the adjacent objects of similar ground objects, extracting the spatial features of the combined objects, and fusion with the spectral features of hyperspectral images. The spatial spectrum joint feature is input into the depth confidence network model. The influence of the network parameters such as different learning rate, hidden layer number, network layer number and training sample ratio on the model is discussed. The advantages of the two kinds of data sources are utilized. The interpretation accuracy of single data source is improved, and the cooperative interpretation of multi-spectral remote sensing image feature layer is realized.
【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】