基于尺度感知与空间选择层级交互的遥感影像变化检测  

Remote Sensing Image Change Detection Based on Scale-Aware and Spatial Selection Hierarchical Interaction

作　　者：邵攀 管宗胜 符潍奇 曾凡宇 程泽敏 石卫超 SHAO Pan;GUAN Zongsheng;FU Weiqi;ZENG Fanyu;CHENG Zemin;SHI Weichao(Hubei Key Laboratory of Intelligent Vision Monitoring for Hydroelectric Engineering,China Three Gorges University,Yichang 443002,China;College of Computer and Information Technology,China Three Gorges University,Yichang 443002,China;National ATR Key Laboratory,College of Electronic Science and Technology,National University of Defense Technology,Changsha 410073,China;School of Physics and Technology,Wuhan University,Wuhan 430072,China)

机构地区：[1]三峡大学水电工程智能视觉检测湖北省重点实验室,宜昌443002 [2]三峡大学计算机与信息学院,宜昌443002 [3]国防科技大学电子科学学院ATR重点实验室,长沙410073 [4]武汉大学物理科学与技术学院,武汉430072

出　　处：《航天返回与遥感》2024年第5期89-100,共12页Spacecraft Recovery & Remote Sensing

基　　金：国家自然科学基金项目(41901341);湖北省自然科学基金项目(2024AFB867)。

摘　　要：目前,深度学习遥感影像变化检测方法在处理尺度变化显著影像时效果仍不够理想,且多数方法在解码阶段缺乏不同层级特征之间的有效交互。针对上述问题,文章以经典U-net网络为基础,提出一种基于尺度感知与空间选择层级交互的高分辨率遥感影像变化检测方法。首先,通过分块并行不同大小的深度可分离卷积提取特征后引入通道注意力,设计一种尺度感知模块,以便有效提取不同形状尺度的变化对象;然后利用空间注意力交叉增强浅层特征与深层特征,提出一种空间选择层级交互模块,细化特征的表征能力;最后,基于两期遥感影像的差异图给出一种差异多尺度注意力模块,来突出变化信息,并抑制未变化信息。文章所提出的方法在WHU、Google、LEVIR和GVLM四个公开数据集上的精确率和召回率的调和平均数(F_(1)值)分别达到91.72%、85.17%、90.82%和88.03%,相比于现有的FC-EF、FC-Conc、IFN、SNUNet、BIT和MSCANet等6种对比变化检测网络,F_(1)值得到显著提升。Currently,remote sensing image change detection methods based on deep learning are still not effective enough to be satisfactory when dealing with images with significant scale changes,and most of the methods lack effective interactions between different layers of features in the decoding stage.Aiming at the above problems,the paper proposes a high-resolution remote sensing image change detection method based on scaleaware and spatial selection hierarchical interaction in view of the classical U-net network.Firstly,a scale-aware module is designed by introducing channel attention after extracting features through chunked parallel depthwise separable convolutions of different sizes,in order to efficiently extract changing objects with different shape scales.Then,by utilizing spatial attention cross-enhancement between shallow and deep features,a spatial selection hierarchical interaction module is presented to refine the representational capabilities of the features.Finally,based on the difference maps of the two remote sensing images,a difference multi-scale attention module is given to highlight the changed information and suppress the unchanged information.The method proposed in the paper achieves F_(1) scores(the harmonic mean of precision and recall)of 91.72%,85.17%,90.82%,and 88.03%on four public datasets:WHU,Google,LEVIR,and GVLM,respectively.Compared to existing six change detection networks such as FC-EF,FC-Conc,IFN,SNUNet,BIT,and MSCANet,the F_(1) score is significantly improved.

关 键 词：深度学习 遥感影像变化检测 尺度感知 空间选择层级交互 U-net网络 

分 类 号：TP75[自动化与计算机技术—检测技术与自动化装置]