多尺度特征融合与空间优化的弱监督高分遥感建筑变化检测  

Weakly supervised building change detection integrating multi-scale feature fusion and spatial refinement for high resolution remote sensing images

作　　者：鄢薪 慎利[1,2] 潘俊杰 戴延帅 王继成 郑晓莉 李志林 YAN Xin;SHEN Li;PAN Junjie;DAI Yanshuai;WANG Jicheng;ZHENG Xiaoli;LI Zhi-lin(State-Province Joint Engineering Laboratory of Spatial Information Technology for High-Speed Railway Safety,Southwest Jiaotong University,Chengdu 611756,China;Faculty of Geosciences and Environmental Engineering,Southwest Jiaotong University,Chengdu 611756,China;Key Laboratory of Ministry of Education on Land Resources Evaluation and Monitoring in Southwest China,Sichuan Normal University,Chengdu 610066,China;Key Laboratory of Investigation,Monitoring,Protection and Utilization for Cultivated Land Resources,MNR,Sichuan Institute of Land Science and Technology(Sichuan Center of Satellite Application Technology),Chengdu 610045,China)

机构地区：[1]西南交通大学高速铁路运营安全空间信息技术国家地方联合工程实验室,四川成都611756 [2]西南交通大学地球科学与环境工程学院,四川成都611756 [3]四川师范大学西南土地资源评价与监测教育部重点实验室,四川成都610066 [4]四川省国土科学技术研究院(四川省卫星应用技术中心)耕地资源调查监测与保护利用重点实验室,四川成都610045

出　　处：《测绘学报》2024年第8期1586-1597,共12页Acta Geodaetica et Cartographica Sinica

基　　金：国家重点研发计划(2022YFB3904202);国家自然科学基金(42071386,41971330);四川省科技厅基本科研业务费项目(2023JDKY0017-3)。

摘　　要：针对建筑物变化检测中深度学习方法严重依赖大量高成本高难度的像素级标注样本进行模型训练的问题,本文提出一种基于图像级标注样本的高分辨率遥感建筑物弱监督变化检测方法MDF-LSR-Net。该方法首先提取双时相多尺度差异特征,并对多尺度差异特征进行渐进式融合,利用充分融合后的多层次多尺度差异特征来生成变化热力图;然后,利用低层融合差异特征的局部空间相似性来优化初始的变化热力图,进一步增强热力图中变化区域的完整性和准确性;最后,基于高质量的变化热力图训练最终的变化检测模型。在公开的建筑物变化检测数据集WHU和LEVIR上的多组试验结果表明,本文方法能够获取更加完整且准确的变化热力图,从而使得基于此训练的变化检测模型也取得更高的检测精度,其中最终的变化检测模型在WHU数据集上的IOU和F 1值分别可达65%和79%以上。To alleviate the heavy dependence of deep learning methods on large-scale high-cost pixel-level annotations,in this paper,we propose a novel weakly supervised method,named MDF-LSR-Net,for high-resolution remote sensing building change detection.Specifically,the proposed method first designs a multi-scale difference feature aggregation module to make better use of multi-scale difference features to generate change heatmaps.Then,by utilizing the local spatial consistency of the low-level fused difference features,MDF-LSR-Net presents a local spatial refinement module to enhance the integrity and accuracy of change regions in heatmaps.Finally,the change detection model is trained based on the high-quality change heatmaps.Experimental results on publicly available datasets,including WHU and LEVIR,demonstrate that our proposed method can obtain more integral and accurate change heatmaps,leading to significantly improved detection performance of the final change detection model.The final model has achieved over 65% points in IOU and over 79% points in F 1 on the WHU dataset.

关 键 词：高分辨率遥感影像 建筑物变化检测 深度学习 弱监督学习 多尺度特征融合 

分 类 号：P237[天文地球—摄影测量与遥感]