基于缩影的多时相遥感语义变化检测方法  

A method of detecting multitemporal semantic changes based on epitomes

作　　者：景维鹏[1] 王健[1] 张文钧 谷俊涛 陈广胜[1] JING Weipeng;WANG Jian;ZHANG Wenjun;GU Juntao;CHEN Guangsheng(College of Information and Computer Engineering,Northeast University of Forestry,Harbin 150040,China;Heilongjiang Research Center for Cyberspace,Harbin 150090,China)

机构地区：[1]东北林业大学信息与计算机工程学院,哈尔滨150040 [2]黑龙江省网络空间研究中心,哈尔滨150090

出　　处：《华中农业大学学报》2023年第3期123-132,共10页Journal of Huazhong Agricultural University

基　　金：国家自然科学基金项目(32171777);黑龙江省应用技术研究与开发计划项目(GA20A301)。

摘　　要：针对高分辨率遥感图像标签稀缺和标签技术增长缓慢限制了多时相语义变化检测发展的问题,提出了采用有噪声、低分辨率的弱标签替代高分辨率标签进行多时相语义变化检测的方法。首先,采用低分辨率卫星数据平滑高分辨率遥感图像输入的质量差异。其次,通过将缩影(epitomes)模型和标签超分辨率算法作为统计推理算法相结合的方法预估高分辨率遥感图像分类图,并拟合一个小型FCN网络对生成的遥感图像分类图进行后处理来改善其分类的效果。最后,通过对比不同时相土地覆盖分类图像之间的差异得出变化检测结果。结果表明,本研究提出的方法与其他多时相语义变化检测方法 FCN/all相比,平均交并比(mIoU)提高了8.9个百分点,能够有效检测土地覆盖分类变化。The detection of multitemporal semantic changes is often used to monitor changes in agricultural ecology and to track the development of agricultural land because it uses semantic information to analyze the specific types of changes.A method of detecting multitemporal semantic changes using weak labels with noise and low resolution instead of high-resolution labels was proposed to solve the problem that the scarcity of high-resolution remote sensing image labels and the slow growth of labeling technology limit the development of detecting multitemporal semantic changes.First,low resolution satellite data were used to smooth the quality differences of high-resolution remote sensing image inputs.Secondly,the high-resolution remote sensing image classification map was estimated by combining the epitomes model and the label super-resolution algorithm as a statistical inference algorithm,and a small FCN network was fitted to postprocess the remote sensing image classification map generated to improve its classification.Finally,the results of detecting change were obtained by comparing the differences between different simultaneous land cover classification images.The results showed that the proposed method improved the mean intersection over uion(mIoU) by 8.9 percentage points compared with other methods of detecting multitemporal semantic changes,and detected the changes of land cover classification effectively.

关 键 词：弱监督 标签超分辨率 缩影 土地覆盖变化图 语义分割 遥感影像 变化检测 

分 类 号：TP79[自动化与计算机技术—检测技术与自动化装置]