文章导读
总览 评价 张普照 ， 公茂果 * （ 西安电子科技大学智能感知与图像理解教育部重点实验室, 西安 710071 ； ） 摘要： 变化检测与分析是空时遥感影像联合解译领域中的一个重要研究课题。 随着遥感影像时间、空间和光谱分辨率的提高,仅仅检测变化与否已无法满足
张普照， 公茂果*
（
西安电子科技大学智能感知与图像理解教育部重点实验室, 西安 710071 ； ）
摘要：
变化检测与分析是空时遥感影像联合解译领域中的一个重要研究课题。 随着遥感影像时间、空间和光谱分辨率的提高,仅仅检测变化与否已无法满足人们对空时遥感影像理解的需求,如何针对变化区域进行精准、高效的分析是当前空时影像处理的迫切需求。而变化检测与分析任务的核心是如何有效地表示差异和评估差异程度,本文基于深度神经网络提出了一种原创的差异表示学习网络,该网络首先将两时影像块映射到特征空间获得特征表示, 以提取关键信息并抑制噪声。然后,两时特征表示经合并层的自动特征选择和后续网络的特征抽象,进而获得差异表示。差异表示经深度网络分类和K-means聚类可分别获得分类误差和聚类误差,这两个误差通过反向传播对网络参数进行调整,从而学到对过聚类标签和期待的聚类均友好的差异表示。对学到的差异表示进行K-means聚类,即可得到变化检测与分析的结果。实验结果证实了本文所提出的差异表示网络的有效性和优越性。
关键词：
变化检测;变化分析;差异表示;深度神经网络;表示学习
ZHANG Pu-Zhao， GONG Mao-Guo*
（
Key Laboratory of Intelligent Perception and Image Understanding of Ministry of Education of China, Xidian University, Xi'an 710071 ； ）
Abstract：
Change detection and analysis (CDA) is an important research topic in the field of joint interpretation of spatial-temporal remote sensing images. As the increase of remote sensing images in temporal-, spatial- and spectral- resolutions, it cannot meet the requirement on image understanding only to identify one pixel changes or not, and it is an urgent demand to make a more precise and effective analysis on the changed regions. The core of CDA is to effectively represent the difference and evaluate the difference degree. In this paper, we propose a novel difference representation learning network (DRLnet) based on deep neural networks. DRLnet firstly maps bi-temporal image blocks into a suitable feature space to extract key information and suppress noise. Then, after automatic feature selection by the merging layer, the subsequent layers can learn more abstract difference representation. By applying network forward pass and K-means clustering, we can obtain the corresponding classification errors and clutering errors, which can be used to tune network parameters in the back propagation of errors. After training, the learned difference representation will be friendly to both the over-clustering classification labels and the desired clustering, and the final change detection and analysis reulsts can be obtained by carrying out K-means clustering on the learned differene representation. The experimental results on multispectral images also have demonstrated the effectiveness and superiority of DRLnet.
Tag：
点此返回栏目查看更多>>>参考论文
