基于Cycle Siamese VGG16的遥感影像建筑物变化检测

Change Detection of Remote Sensing Images for Building Structures Based on Cycle Siamese VGG16

遥感影像建筑物变化检测在城市规划、环境监测、灾害评估等领域中发挥着关键作用。但现有方法忽略了不同时相的影像色彩差异带来的域间隙,使得模型难以拟合欧氏距离过大的变化特征。另外,现有常规解码方法无法在感受野内聚合上下文信息,不能准确识别建筑物变化检测结果的边缘。针对以上问题,文章从时间-色彩关联性角度提出一种建筑物变化检测方法。在数据层面,考虑前、后时相影像色调不一致现象,基于循环一致生成对抗网络迁移后时相风格,缩短影像域间隙。在特征拟合过程中,在特征级联后嵌入时空注意力模块,通过增强对建筑物的关注度,解决检测结果假阴性问题。基于建筑屋顶的纹理相似性,嵌入上下文增强模块,利用影像的深层上下文信息,避免出现建筑物孔洞现象;考虑建筑物边缘平滑性,使用感知重组模块对建筑物变化信息进行自适应感知,以提升建筑物边界位置准确性。实验结果表明,相对于其他方法,所提出的模型在建筑物变化检测任务上取得了最佳F1值。

Remote sensing image building change detection plays a critical role in urban planning,environmental monitoring,disaster assessment,natural resource management and other fields.However,existing methods ignore domain gaps caused by differences in image colors between different time periods,making it difficult for the model to fit change features with large Euclidean distances.In addition,current conventional decoding methods cannot aggregate contextual information within the receptive field and cannot accurately identify the edges of building change detection results.To address these issues,this paper proposes a building change detection method from the perspective of time-color correlation.At the data level,considering the phenomenon of inconsistent tone between pre-and post-time-phase images,based on the cycle-consistent adversarial networks,the phase style is transferred to shorten the image domain gap.In the feature fitting process,a spatial-temporal attention module is embedded after feature cascading to enhance attention to buildings and solve false negatives in detection results.Based on the roof texture similarity of buildings,a context enhancement module is embedded to use deep contextual information of images to avoid building holes.Considering the smoothness of building edges,a perception recombination module is used to adaptively perceive building change information and improve accuracy of building edge positions.Experimental results show that compared with other methods,the proposed model achieves the best F1 score in building change detection tasks.