时空差异增强与自适应特征融合的轻量级遥感影像变化检测网络  

A lightweight remote sensing images change detection network utilizing spatio-temporal difference enhancement and adaptive feature fusion

作　　者：龚良雄 李星华 程远明 赵兴友 谢仁平 王红根 GONG Liangxiong;LI Xinghua;CHENG Yuanming;ZHAO Xingyou;XIE Renping;WANG Honggen(Nanchang Institute of Surveying and Mapping Co.,Ltd.,Nanchang 330038,China;School of Remote Sensing and Information Engineering,Wuhan University,Wuhan 430079,China;Nanchang Urban Planning&Design Institute Group Co.,Ltd.,Nanchang 330038,China;School of Computer Science and Technology,Dongguan University of Technology,Dongguan 523808,China)

机构地区：[1]南昌市测绘勘察研究院有限公司,江西南昌330038 [2]武汉大学遥感信息工程学院,湖北武汉430079 [3]南昌市城市规划设计研究总院集团有限公司,江西南昌330038 [4]东莞理工学院计算机科学与技术学院,广东东莞523808

出　　处：《测绘学报》2025年第1期136-153,共18页Acta Geodaetica et Cartographica Sinica

基　　金：国家自然科学基金(42171302,62001113)。

摘　　要：针对现有遥感影像变化检测方法存在多时相差异特征利用不足、多尺度特征融合不足等问题,提出一种时空差异增强与自适应特征融合的轻量级遥感影像变化检测网络。本文设计了轻量级时空差异增强模块,采用语义变化感知和空间变化感知的双分支结构,组合利用语义自适应增强机制和混合注意力机制,增强双时相特征图的空谱差异。不同尺度特征图通过边缘细化残差模块进一步优化变化区域边缘。还改进了双向特征融合金字塔结构,采用可学习的权重参数来量化不同尺度特征的重要性,实现多尺度特征的有效融合。选取10种主流的变化检测方法,在WHU-CD、LEVIR-CD、SYSU-CD和SECOND数据集上进行模型对比试验,结果表明:SEAFNet相较于多种主流的变化检测方法,在定性分析、定量分析、网络复杂度与准确度平衡方面均取得了比较优异的表现。To address the limitations in existing change detection methods of remote sensing images,such as insufficient utilization of multi-temporal difference features and inadequate multi-scale feature fusion,a lightweight remote sensing images change detection network named SEAFNet is proposed,which integrates spatio-temporal difference enhancement with adaptive feature fusion.This paper designs the lightweight spatio-temporal difference enhancement module,which employs a dual-branch structure with semantic change perception and spatial change perception.This module combines a semantic adaptive enhancement mechanism and a mixed attention mechanism to enhance the space-spectrum differences in the bi-temporal feature maps.To further refine the edges of the change regions,different scale feature maps are optimized through an edge refinement residual module.The bi-directional feature fusion pyramid structure is also improved by using learnable weight parameters to quantify the importance of features at different scales,achieving effective multi-scale feature fusion.Comparative experiments with ten mainstream change detection methods on WHU-CD,LEVIR-CD,SYSU-CD and SECOND datasets demonstrate that SEAFNet outperforms these methods in qualitative and quantitative analysis,and the balance between network complexity and accuracy.

关 键 词：遥感影像 时空差异增强 注意力机制 自适应特征融合 变化检测 

分 类 号：P237[天文地球—摄影测量与遥感]