基于编码对齐生成对抗网络的异源遥感影像地表覆被变化检测方法  

Change Detection Method for Surface Cover of Heterogeneous Remote Sensing Image based on Code-Aligned Generative Adversarial Network

作　　者：张成才[1] 刘威[1] 杨峰 彭凯[1] 周雪丽 ZHANG Chengcai;LIU Wei;YANG Feng;PENG Kai;ZHOU Xueli(Yellow River Laboratory(Zhengzhou University),Zhengzhou 450001,China;Chushandian Reservoir Management Bureau,Henan Province,Xinyang 464000,China;Henan Map Institution,Henan Province,Zhengzhou 450003,China)

机构地区：[1]黄河实验室(郑州大学),河南郑州450001 [2]河南省出山店水库建设管理局,河南信阳464000 [3]河南省地图院,河南郑州450003

出　　处：《遥感技术与应用》2023年第5期1107-1117,共11页Remote Sensing Technology and Application

基　　金：河南省自然科学基金项目“联合主动微波遥感和光学遥感数据的大型灌区土壤水分反演方法研究”(222300420539);河南省水利科技攻关项目“河南省出山店水库流域覆被遥感监测方法”(GG201902);国家自然科学基金重点项目“基于大数据的城市洪涝灾害预报预警理论与方法研究”(51739009)资助。

摘　　要：相较于同源遥感影像地表覆被变化检测,异源影像能综合不同卫星传感器间数据特征和现势性等优势,更好满足应用需求。针对异源遥感影像变化检测中存在的光谱差异和特征空间不一致问题,研究提出编码对齐生成对抗网络实现异源影像的高精度变化检测。考虑到异源影像间通道和数据类型上存在差异,难保持重构前后影像空间结构的一致性,研究通过添加自编码器和构造编码对齐损失,约束编码器输出特征的空间结构变化,使重构前后影像空间结构一致,有效减少信息丢失;在跨域映射过程中为减少源域与目标域间影像的色彩差异,采用循环一致对抗生成网络在无成对影像情况下进行色彩迁移,实现两时相异源影像的相互跨域映射,生成能与原始影像直接对比的无色偏重构影像;利用设计的变化概率权重使网络在训练过程中自动选择样本,有效提取覆被变化信息。实验结果表明:该方法与CGAN、SCCN等方法相比能更充分提取影像特征,降低跨域映射函数的随机性;在4组公开数据集的检测精度分别达到0.93、0.96、0.97、0.88,精度最高;变化检测结果与参考图的一致性、检测差异图质量均最优。因此,该方法在异源遥感影像中能够进行高精度变化检测。Compared to the change detection of homologous remote sensing images,heterogeneous images can integrate the advantages of different satellite sensor data features and temporal relevance,better satisfying application requirements.To address the issues of spectral differences and inconsistent feature spaces in change detection of heterogeneous remote sensing images,this study proposes an aligned generative adversarial network for high-precision change detection of heterogeneous images.Considering the differences in channels and data types between heterogeneous images,it is difficult to maintain the consistency of spatial structures before and after reconstruction.The study incorporates autoencoders and constructs alignment loss to constrain the spatial structure changes of encoder output features,ensuring consistency in spatial structures between the reconstructed im‐ages and reducing information loss effectively.In the cross-domain mapping process,to minimize the color differences between source and target domain images,a cycle-consistent adversarial generative network is used for color transfer in the absence of paired images,enabling mutual cross-domain mapping between two temporally distinct heterogeneous images,generating color-preserving reconstructed images that can be directly compared with the original images.By utilizing designed change probability weights,the network automatically selects samples during the training process,effectively extracting land cover change information.Experimental results demonstrate that compared to methods such as CGAN and SCCN,the proposed method can more fully extract image features and reduce the randomness of cross-domain mapping functions.The detection accuracies on four publicly available datasets reach 0.93,0.96,0.97,and 0.88,with the highest accuracy achieved.The consistency between the change detection results and the reference maps,as well as the quality of the difference maps,is optimal.This method enables high-precision change detection in heterogeneous remote sen

关 键 词：异源遥感影像 变化检测 自适应 跨域映射 

分 类 号：P237[天文地球—摄影测量与遥感]