融合密集连接和自注意力机制的高分辨率遥感影像变化检测方法  

A high-resolution remote sensing images change detection method via the integration of dense connections and self-attention mechanisms

作　　者：庞世燕 郝京京 左志奇 兰晶晶 胡翔云[3,4] PANG Shiyan;HAO Jingjing;ZUO Zhiqi;LAN Jingjing;HU Xiangyun(Faculty of Artificial Intelligence in Education,Central China Normal University,Wuhan 430079,China;College of Informatics,Huazhong Agricultural University,Wuhan 430070,China;School of Remote Sensing and Information Engineering,Wuhan University,Wuhan 430079,China;Hubei Luojia Laboratory,Wuhan 430079,China)

机构地区：[1]华中师范大学人工智能教育学部,湖北武汉430079 [2]华中农业大学信息学院,湖北武汉430070 [3]武汉大学遥感信息工程学院,湖北武汉430079 [4]湖北珞珈实验室,湖北武汉430079

出　　处：《测绘学报》2024年第12期2244-2253,共10页Acta Geodaetica et Cartographica Sinica

基　　金：教育部人文社会科学项目(22YJC880058);中央高校基本科研业务费项目(CCNU22QN019)。

摘　　要：遥感影像变化检测是遥感影像分析中的一项重要任务,在城市动态监测、地理信息更新、自然灾害监测、违章建筑物查处、军事目标打击效果分析及国土资源调查等方面应用广泛。作为像素级别的预测任务,当前的变化检测有两个比较突出的问题:一是受限于自注意力计算的效率问题,遥感影像中的远程上下文信息利用不足;二是侧重于深度变化语义特征的提取,而忽略了包含高分辨率和细粒度特征的浅层信息的应用。针对第1个问题,本文采用Transformer对提取的两期深层影像特征进行上下文建模,提升深层变化特征的质量。为了兼顾Transformer的效率问题,本文将图像特征图转化为稀疏标记,以此来显著减少Transformer的令牌数量。针对第2个问题,本文采用密集跳跃连接的方式,以保留浅层变化特征中的高分辨率和细粒度信息。采用了3个公开的数据集进行试验验证,结果表明,与其他先进的变化检测方法相比,本文方法的IoU指标分别达到了85.44%、84.15%和94.61%,均优于其他对比方法。Remote sensing image change detection is an important task in remote sensing image analysis,which is widely used in urban dynamic monitoring,geographic information update,natural disaster monitoring,illegal building investigation,military target strike effect analysis,and land and resources survey.As a pixel-level prediction task,the current methods of change detection have two prominent problems:one is the computational efficiency of self-attention between arbitrary pixel pairs is low,and the long context information in remote sensing images is insufficiently utilized;the other is that the current methods focus on the extraction of deep change image features while shallow information containing high-resolution and fine-grained features are ignored.To address the first problem,Transformer is used to perform context modeling on the extracted bitemporal image features to improve the quality of the deepest change features.To take into account the efficiency of the Transformer,the proposed method converts the images into sparse tokens,thereby significantly reducing the number of tokens of the Transformer.For the second problem,the proposed method uses dense skip connections to retain high resolution in shallow change features.Three publicly available datasets were used for experiments.Extensive experiments show that compared with the state-of-the-art change detection methods,the IoU metric of the proposed method reached 85.44%,84.15%and 94.61%,respectively,which is better than other comparison methods.

关 键 词：遥感影像 变化检测 TRANSFORMER 密集连接 多尺度特征融合 

分 类 号：P237[天文地球—摄影测量与遥感]