结构先验感知的高分辨率遥感变化检测  

A VHR Remote Sensing Change Detection Model with Structure Prior Perception

作　　者：陈江伟 孟小亮 CHEN Jiangwei;MENG Xiaoliang(School of Remote Sensing and Information Engineering,Wuhan University,Wuhan 430000,China)

机构地区：[1]武汉大学遥感信息工程学院,武汉430000

出　　处：《航天返回与遥感》2025年第2期157-169,共13页Spacecraft Recovery & Remote Sensing

基　　金：2023年湖北省重大科技攻关项目(2023BAA025);国家自然科学基金(41971352)。

摘　　要：近年来,深度学习和计算机视觉领域的重大进展推动了遥感变化检测的发展。然而,现有的方法仍然依赖于单一的视觉模态,无法有效利用其他模态的信息,如高程图或深度图等结构先验。为了充分利用深度图等结构先验信息,文章提出了一种新型的具有结构先验感知能力的高分辨率遥感变化检测框架SPP-CD,该框架接受双时相光学遥感影像和对应的深度图作为输入,采用孪生编码器结构对输入数据进行特征提取,特征交互模块对双时相特征进行交互与融合,并使用特征解码器对融合后特征进行解码,最终输出精细的像素级变化检测图。基于SPP-CD框架设计了双路径融合多模态特征编码器,利用基于跨模态注意力的全局路径和基于卷积的局部路径使模型兼具远距离上下文建模、跨模态特征建模和精细特征提取能力。实验结果表明,在LEVIR-CD数据集上,与现有单模态基准方法相比,文章提出的方法在关键指标F1分数、交并比和总体准确率上分别达到92.36%、85.81%和99.21%,优于单模态基准方法,从而证明了集成空间结构先验信息能够有效提升变化检测性能,可以缓解现有方法主要依赖视觉信息的局限性。In recent years,significant progress in deep learning and computer vision has promoted the development of remote sensing change detection.However,the existing methods still rely on a single visual modality and cannot effectively utilize information from other modalities,such as structure priors such as elevation maps or depth maps.In order to make full use of structure prior information such as depth maps,this paper propose a novel very-high-resolution remote sensing change detection framework SPP-CD with structure prior perception ability.The framework accepts bi-temporal optical remote sensing images and corresponding depth maps as input,uses a siamese encoder structure to extract features from the input data,and then uses a feature interaction module to interact and fuse the bitemporal features.Finally,a feature decoder is used to decode the fused features and output a fine-grained pixel-level change detection map.A dual path fusion multimodal feature encoder is designed based on the SPP-CD framework,which utilizes a global path based on cross modal attention and a local path based on convolution to enable the model to combine long-distance context modeling,cross modal feature modeling,and fine feature extraction capabilities.Experimental results on the LEVIR-CD dataset show that compared with the existing single modal baseline methods,the method proposed by this paper achieved 92.36%,85.81%and 99.21%on the key indicators F1,IoU and OA respectively,surpassing the single-modal baselines,thus proving that integrating spatial structure prior information can effectively improve the change detection performance and alleviate the limitations of existing methods that mainly rely on visual information.

关 键 词：遥感影像 变化检测 多模态融合 先验感知 深度学习 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置] V19[自动化与计算机技术—控制科学与工程]