多级对比学习下的弱监督高分遥感影像城市固废堆场提取  

Multi-level contrastive learning for weakly supervised extraction of urban solid wastes dump from high-resolution remote sensing images

作　　者：王继成 郭安嵋 慎利[2,3] 蓝天 徐柱[2,3] 李志林 WANG Jicheng;GUO Anmei;SHEN Li;LAN Tian;XU Zhu;LI Zhilin(Key Laboratory of Ministry of Education on Land Resources Evaluation and Monitoring in Southwest China,Sichuan Normal University,Chengdu 610066,China;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)

机构地区：[1]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 [2]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 [3]西南交通大学地球科学与环境工程学院,四川成都611756

出　　处：《测绘学报》2024年第6期1212-1223,共12页Acta Geodaetica et Cartographica Sinica

基　　金：国家重点研发计划(2022YFB3904202,2022YFB3904205);国家自然科学基金(42394063,42101412);四川省科技计划资助(2023NSFSC19179);四川省科技厅基本科研业务费项目(2023JDKY0017-3)。

摘　　要：城市固体废物是城市化进程中的重要污染源,对城市生态环境和公共健康造成了巨大危害。高分影像固废堆场智能解译是实现自动排查,提升监测效率的核心和关键技术。基于深度学习的固废堆场自动提取方法严重依赖于获取成本高、制作难度大的高质量像素级标注。为此,本文提出使用更易获取的影像级标注,利用影像自监督学习实现像素级固废堆场提取。围绕固废堆场的影像特征,本文方法在尺度对比约束下综合像素、影像两个层次的对比学习方法,对固废堆场的类别激活图细化和完善,并基于此生成高质量的固废堆场伪像素级标注,用于训练固废堆场提取模型。试验结果表明,本文方法在固废堆场提取的F 1值和IoU分数方面分别达到了71.58%和55.74%,显著优于所有对比方法。这说明利用多级对比学习的弱监督方法能够获得更加完整且准确的类别激活图,从而取得更高的固废堆场提取精度。Urban solid waste is a major pollutant in the urbanization process that endangers the urban ecology and public health.Intelligent interpretation of high-resolution satellite imagery for solid waste dumps(SWD)is crucial for automated monitoring.However,deep learning-based automatic extraction methods for SWD heavily rely on costly and labor-intensive high-quality pixel-level annotations.This paper presents a weakly supervised method that only uses image-level annotations to perform pixel-level SWD extraction.The method leverages the image characteristics of SWD and applies contrastive learning at both pixels,image levels under constraints of scale contrast to improve the class activation maps(CAMs)of SWD.Based on the CAMs,the method generates high-quality pixel-level pseudo-labels that are used to train the SWD extraction model.The experiments on a self-created SWD dataset demonstrate that the proposed method achieves an F 1 score of 71.58%and an IoU score of 55.74%,which are significantly higher than the baseline methods.This shows that the multi-level contrastive learning-based weakly supervised method can produce more complete and accurate CAMs of SWD,leading to better extraction performance.

关 键 词：城市固废堆场 高分辨率遥感影像 对比学习 弱监督信息提取 类别激活图 

分 类 号：P258[天文地球—测绘科学与技术]