基于深度神经网络的遥感图像变化检测

发布时间：2018-03-05 10:16

本文选题：深度神经网络　切入点：遥感图像　出处：《南京理工大学》2017年硕士论文　论文类型：学位论文

【摘要】：变化检测是指对同一地区不同时相影像的地表变化信息进行监测,可以定性或定量地分析地表的变化特征和过程。检测出不同时相的遥感图像之间变化的区域是一项挑战性的技术,传统的方法中检测流程较繁琐,如需要配准、辐射校正等预处理过程,构造差异图,分析差异图等。深度神经网络的多层的非线性映射使其具备了拟合逼近任意函数形式的能力,因此能构造出复杂的分类面,高质量地完成模式分类识别任务。本文研究利用深度网络解决遥感图像变化检测问题,相对传统方法实现了流程的简化,变化检测精度的提升。本文主要包括以下三个方面的工作:(1)提出了一种基于深度神经网络的遥感图变化检测问题中的差异图构造方法。该方法使用了两个深度置信网分别接受两幅待检测变化的遥感图的输入,由两深度置信网的输出构造差异图,根据变化检测问题的需要,设计了双向BP调优算法用于对深度置信网进行训练,增加构造出的差异图中变化区域与不变区域的分离度。比传统的方法相比,提出的方法免去辐射校正预处理过程,差异图的质量显著提升。(2)设计了一种组合神经网络分类器,并推导给出了它的训练算法。两个独立的深度置信网连接Logistic Regression分类器构成组合神经网络,组合神经网络完成由两遥感图像像素值到变或不变类别的完整映射,将传统的方法中的繁琐流程如辐射校正,构造差异图,分析差异图等统一地纳入组合神经网络之中,方法具有显著的简洁性,且具有较高精度的变化检测结果。组合神经网络方法与一种已提出的用于遥感图像变化检测问题的单深度置信网模型相比,更适合于遥感图像变化检测问题,分类正确率更高,能得到噪声点更少的变化检测结果。(3)研究了使用深度置信网分析差异图方法。设计了两种差异图分析方法,深度置信网对差异图进行特征提取而后在特征空间中进行聚类,称为DBN-kmeans方法,对DBN进行有监督训练,实现对差异图中像元点进行变化或不变的二分类方法称为DBN分析法。与主成分聚类分析方法相比,在一些数据集上,基于深度置信网方法的结果能勾勒更正确的局部轮廓,变化检测结果的质量有所提升。
[Abstract]:Change detection refers to the monitoring of surface change information in different images of the same area. It is a challenging technique to detect the regions of variation between remote sensing images in different phases, and the traditional methods are more complicated, such as registration. Radiation correction and other preprocessing processes, such as constructing difference map, analyzing difference map, etc. The multi-layer nonlinear mapping of depth neural network has the ability to fit and approximate any function form, so it can construct complex classification surface. The task of pattern classification and recognition is accomplished with high quality. In this paper, the problem of remote sensing image change detection is solved by using depth network, and compared with the traditional method, the flow chart is simplified. This paper mainly includes the following three aspects: 1) this paper presents a method of constructing difference map in remote sensing image change detection problem based on depth neural network. Two depth sets are used in this method. The network receives input from two remote sensing images to be detected, According to the need of change detection problem, a bidirectional BP optimization algorithm is designed to train the depth confidence net by constructing the difference map from the output of the two depth confidence nets. Compared with the traditional method, the proposed method eliminates the process of radiation correction preprocessing, and the quality of the difference map is significantly improved. (2) A combined neural network classifier is designed. Its training algorithm is derived and given. Two independent depth confidence nets are connected with Logistic Regression classifiers to form a combined neural network. The combined neural network completes the complete mapping from pixel values of two remote sensing images to variable or invariant categories. The traditional methods, such as radiation correction, construction of difference maps and analysis of difference maps, are brought into the combinational neural network in a unified way. Compared with the proposed single depth confidence net model for remote sensing image change detection, the combined neural network method is more suitable for remote sensing image change detection, and the classification accuracy is higher. The method of using depth confidence net to analyze the difference map is studied. Two different map analysis methods are designed. The depth confidence net extracts the feature of the difference map and clusters it in the feature space. The two classification method called DBN-kmeans method, which can train DBN supervised and realize the change or invariance of pixel points in the difference graph, is called DBN analysis method. Compared with principal component clustering analysis method, on some data sets, The results based on the depth confidence net method can outline more accurate local contours and improve the quality of the change detection results.
【学位授予单位】：南京理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP751;TP183

【相似文献】