CASIA OpenIR  > 毕业生  > 博士学位论文
高分辨率遥感图像变化检测技术研究
徐元
学位类型工学博士
导师潘春洪
2016-05-27
学位授予单位中国科学院研究生院
学位授予地点北京
关键词变化检测 高分辨率遥感图像 多尺度分析 字典学习 字典学习 深度学习
其他摘要遥感图像变化检测是遥感图像处理中的热点研究问题,在自然环境监测、灾 情分析、城市规划管理、军事区域监视等领域具有十分重要的应用价值。目前遥 感图像变化检测技术正向着精细化应用的方向发展。针对高分辨率遥感图像的变 化检测可以获取更加丰富而细致的变化信息,因而受到广泛关注。然而,受高分 辨率图像自身特性的影响,将传统变化检测算法直接进行应用往往达不到高精度 的检测目标。因此有必要针对高分辨率遥感图像的变化检测技术进行深入细致的研究。本文以城市场景高分辨率遥感图像为研究对象,结合其同物异谱、同谱异物、 细节冗余等特点,分析了变化检测中存在的难点问题,并针对多尺度变化信息利 用、变化特征提取与学习、变化判断决策等高分辨率图像变化检测关键技术开展 深入研究。本文的贡献主要包含以下几个方面:提出一种基于鲁棒变化特征的多尺度传播半监督变化检测方法。具体地,论 文首先针对城市场景中常见的因视角变化造成的突出地物图像位置偏移问 题,引入了一种局部搜索策略,并在此基础上提取出可靠的变化特征;其次, 根据用户给出的少量标注样本来消除变化定义的不确定性,并进一步采用半 监督分类的方式获得用户兴趣变化的变化概率信息;最后,提出了一种自上 而下的多尺度变化信息传播算法,在充分利用上下文信息的同时降低了变化 定义对尺度的依赖性。所提出的方法不需要大量的标注信息,利用少量标注样本即可检测出用户感兴趣的变化。对比实验验证了所提出方法的有效性。提出一种基于提出一种基于稀疏变化描述子和鲁棒判别字典学习的区域级多尺度融合有监督变化检测方法。具体地,论文首先提出了一种可同时描述变化程度和变化类型的稀疏变化特征描述子,提高了变化类和非变化类的类间可区分性;其次,利用$k$NN分类器来获得粗分类;然后,基于区域级变化特征协同稀疏表示和鲁棒判别字典学习来获得精细分类,并在此基础上计算区域内像素是否发生变化的概率;最后,利用条件随机场通过图上的判别学习方式来融合所有尺度上的粗分类和精细分类,从而获得区域级变化检测结果。所提出方法通过在图上融合多尺度判别分类信息,提高了检测的精度。对比研究表明,本文所提出的方法能取得较优的变化检测结果。提出一种基于样本子集选择的变化检测算法。考虑到在实际的应用中,高分辨率图像尺寸普遍较大,往往需要针对应用场景,在检测速度和检测精度之间做出权衡。为此,提出了一种通过调整参数或改变参数学习方式,即可在样本集压缩率和分类精度之间做出权衡的样本子集选择算法。首先对每个样本赋予一个非负权重,通过优化稀疏正则kNN的分类精度得到样本权重。然后选择前k个具有最大权重的样本作为子集元素。与同类方法相比,所提出的方法在保证相当分类精度的同时,具有更好的样本集压缩率,因此提供了一种高效的大尺寸图像变化检测方法。实验验证了该方法在大尺寸图像变化检测中的有效性。提出了一种基于自编码机的无监督变化检测算法。其核心思想是采用深度学习模型来学习非线性的变化特征。具体地,论文构建了一种面向高分辨率变化检测的自编码机深度学习模型,提出了一种针对目标多时相图像的平凡变化的自编码机深度学习模型训练策略,利用自编码机生成了一种可消除平凡变化影响的模拟特征图;进一步通过比较真实特征与模拟特征的差异得出最终变化检测结果。与现有的人工变化特征提取方法相比,论文提供了一种基于深层神经网络的无监督变化特征学习范例。对比实验验证了所提方法的有效性。; Change detection of remote sensing images has already played an important role in many application fields, such as urban planning and management, natural environment monitoring, military surveillance, etc. Nowadays great progress has been made in remote sensing applications toward the direction of fine granularity. As high resolution remote sensing images can provide abundant detail change information, change detection technology for images of this category has been received wide attention. Nevertheless, it will not achieve the goal with direct employment of traditional change detection algorithm for the intrinsic characteristics of high resolution remote sensing images. As a result, it is necessary to develop appropriate change detection technologies for high resolution remote sensing images.

This dissertation takes high resolution remote sensing images in urban scene as objects of study and focus on several difficulties in their change detection, considering the phenomenon of different objects which have the same spectrum and the same objects which have different spectrum and the characteristics of detailed redundancy. Besides that, we explore a number of key technologies on aspects of error reduction in preprocessing (radiation correction, image registration), features extraction of remote sensing images, change feature description, usage of multi-scale change information, improvement of efficiency in change detection, etc. The contributions of this dissertation are list as follows:

A pixel-based multi-scale propagation semi-supervised change detection algorithm is proposed based on robust change features. With the strategy of local-area search, this method gets rid of the influence of visual angel to the projecting ground objects in urban scene and can extract the reliable change features. Additionally, it eliminates the uncertainty of change definition according to a small amount of samples selected by user and acquires the change probability which reflects user preferences with semi-supervised classification. A top-down multi-scale change information propagation based on back-propagation algorithm was developed, which takes context-information into account and reduces the dependence of dimension in change definition. The experimental results proved the effectiveness of this algorithm.

A multi-scale region-level fusion supervised change detection algorithm is proposed based on sparse change descriptor and robust discriminative dictionary learning. To improve the distinction between change feature categories, this algorithm introduced a sparse change descriptor which combines both the change degree component and change pattern component. We employed the region-level co-sparse representation to estimate the change probability of each pixel-pair according to rough classification with k-nearest neighbors and fine classification with robust discriminative dictionary learning. Comparative experiments indicate the validity of the proposed approach.

High resolution remote sensing images are usually of large size so that making a tradeoff between detection speed and detection accuracy is necessary in practical applications. Based on sample subset selection, we proposed a change detection algorithm which can change sample compression ratio and classification accuracy by changing the learning method of parameters or tuning them. With the capability of direct extension to a semi-supervised or unsupervised one, this algorithm can guarantee not only a comparative classification accuracy but a better sample compression ratio. The experimental results proved the effectiveness of this algorithm.

For the lacking of a suitable change descriptor for high resolution remote sensing images, we proposed an un-supervised change detection algorithm based on auto-encode model. This algorithm is designed to extract non-line change features of remote sensing images using deep learning model. Deriving from the difference between ordinary change and change of interest, we introduced a training strategy which can generate simulated feature map and eliminate the influence of ordinary change with auto-encode model. Experimental results indicate that our approach can achieve better performance of change detection, Compared with the state-of-the-art methods.
语种中文
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/11688
专题毕业生_博士学位论文
作者单位中国科学院自动化研究所
推荐引用方式
GB/T 7714
徐元. 高分辨率遥感图像变化检测技术研究[D]. 北京. 中国科学院研究生院,2016.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
明审thesis.pdf(7293KB)学位论文 暂不开放CC BY-NC-SA请求全文
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[徐元]的文章
百度学术
百度学术中相似的文章
[徐元]的文章
必应学术
必应学术中相似的文章
[徐元]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。