联合时空差异注意力与层级细节增强的遥感影像变化检测  

Joint spatial-temporal differential attention and hierarchical detail enhancement for remote sensing image change detection

作　　者：管宗胜 邵攀 杨子超 程泽敏 余快 Guan Zongsheng;Shao Pan;Yang Zichao;Cheng Zemin;Yu Kuai(Hubei Key Laboratory of Intelligent Vision Based Monitoring for Hydroelectric Engineering,China Three Gorges University,Yichang Hubei 443002,China;College of Computer&Information Technology,China Three Gorges University,Yichang Hubei 443002,China;Advanced Copper Industry,Jiangxi University of Science&Technology,Yingtan Jiangxi 335000,China)

机构地区：[1]三峡大学水电工程智能视觉监测湖北省重点实验室,湖北宜昌443002 [2]三峡大学计算机与信息学院,湖北宜昌443002 [3]江西理工大学先进铜产业学院,江西鹰潭335000

出　　处：《计算机应用研究》2025年第3期937-943,共7页Application Research of Computers

基　　金：国家自然科学基金资助项目(41901341,42101469);湖北省自然科学基金资助项目(2024AFB867)。

摘　　要：目前,基于U-Net的深度学习遥感影像变化检测方法包含许多伪变化信息,且多数方法缺乏层级特征间的有效交互。针对上述问题,以经典U-Net为基础,提出一种联合时空差异注意力与层级细节增强的高分辨率遥感影像变化检测方法。首先,分别提取两期影像的单时相特征与级联特征,基于两期单时相特征的欧氏距离与差值特征,提出一种时空差异注意力模块,强化级联特征对变化区域的学习;然后,利用混合空间通道注意力交互相邻层级特征间的信息,构建一种层级细节增强模块,促进特征解码;最后,结合分块策略和空洞条形卷积,设计一种轻量级的多尺度边界细化模块,提取多尺度特征并缓解边界信息的丢失。在四个常用公开数据集上的实验结果表明,该方法相比于现有8种变化检测网络,取得了最好的评价指标。Currently,deep learning remote sensing image change detection methods based on U-Net contain many pseudo-change information,and most of them lack effective interaction between layer-level features.Aiming at the above problems,based on the classical U-Net network,this paper proposed a joint spatial-temporal differential attention and hierarchical detail enhancement method for high-resolution remote sensing image change detection.Firstly,this paper extracted the single time-phase features and cascade features of the two periods of images respectively,and based on the Euclidean distance and diffe-rence features of the single time-phase features of the two periods,it proposed a spatial-temporal differential attention module to strengthen the learning of cascade features to the changing region.Secondly,using hybrid spatial channel attention to interact information between adjacent hierarchical features to construct a hierarchical detail enhancement module that facilitated feature decoding.Finally,this paper designed a lightweight multi-scale boundary refinement module to extract multi-scale features and mitigate the loss of boundary information by combining the chunking strategy and atrous strip convolution.Experimental results on four commonly used public datasets show that the method achieves the best evaluation metrics compared to eight existing change detection networks.

关 键 词：深度学习 遥感影像变化检测 时空差异注意力 层级细节增强 U-Net 

分 类 号：TP751.1[自动化与计算机技术—检测技术与自动化装置]