基于HRNetV2的城市内涝视频图像监测方法  

Urban Waterlogging Video Image Monitoring Method Based on HRNetV2

作　　者：陈笑娟 徐艺芙[1,3] 吕鑫 魏军 李婷 陈小雷[1,2,3] CHEN Xiao-juan;XU Yi-fu;LÜXin;WEI Jun;LI Ting;CHEN Xiao-lei(A China Meteorological Administration Xiong'an Atmospheric Boundary Layer Key Laboratory,Xiong'an NewArea 071800,Hebei Province,China;Key Laboratory of Meteorology and Ecological Environment of Hebei Province,Shijiazhuang 050021,Hebei Province,China;Meteorological Disaster Prevention and Environment Meteorological Center of Hebei Province,Shijiazhuang 050021,Hebei Province,China;School of Remote Sensing&Geomatics Engineering,Nanjing University of Information Science&Technology,Nanjing 210044,Jiangsu Province,China)

机构地区：[1]中国气象局雄安大气边界层重点开放实验室,河北雄安新区071800 [2]河北省气象与生态环境重点实验室,河北石家庄050021 [3]河北省气象灾害防御和环境气象中心,河北石家庄050021 [4]南京信息工程大学遥感与测绘工程学院,江苏南京210044

出　　处：《中国农村水利水电》2025年第4期164-168,175,共6页China Rural Water and Hydropower

基　　金：河北省科技厅重点研发计划项目22375421D。

摘　　要：近年来,城市内涝事件频发,给城市居民的生活和财产带来了巨大损失。传统监测手段受设备安装与维护成本高昂、人工效率低等问题局限,同时传统的卷积神经网络无法对分辨率不同的监控设施视频图像加以区分,难以全面、高效地应对内涝灾害,因此研究创新性地提出了一种基于HRNetV2的城市内涝监测方法。该方法充分利用了HRNet模型在目标检测、图像分类等人体姿态估计应用方面展现出优势性能,可在并行多个不同分辨率的卷积分支的同时共享卷积权重,减少模型参数量和计算量,提高模型训练效率。研究通过对监测采集和收集的社会化城市内涝积水图像构成的数据集进行训练,并采用精度和复杂度两项关键评价指标,将HRNetV2与Unet、PSPNet、ResUnet、DeeplabV3+四种主流模型的训练结果进行了全面对比。实验结果表明,HRNetV2在积水图像识别方面展现出了卓越的性能。其交并比、精确度、召回率以及F1分数分别高达92.19%、96.90%、95.76%和95.83%,均显著优于其他4种对比模型。与此同时,HRNetV2在复杂度方面也有着出色的表现,相较于其他模型,其计算复杂度大幅降低,更加适合在实际监测场景中应用。这一研究成果不仅为城市内涝监测提供了一种全新的技术手段,可以更加准确、高效地监测城市内涝情况,同时也可为城市规划、灾害管理等领域提供有价值的参考。In recent years,urban flooding events have occurred frequently,causing significant losses to the lives and property of urban residents.Traditional monitoring methods are limited by high installation and maintenance costs of equipment,as well as low labor efficiency.Additionally,traditional convolutional neural networks are unable to distinguish between video images from surveillance facilities with different resolutions,making it difficult to comprehensively and efficiently respond to flooding disasters.Therefore,this study innovatively proposes an urban flooding monitoring method based on HRNetV2.This method fully exploits the advantages of the HRNet model,which has demonstrated superior performance in applications such as object detection,image classification,and human pose estimation.It can share convolutional weights while operating in parallel on multiple convolutional branches with different resolutions,reducing the number of model parameters and computations,and improving model training efficiency.In this study,a dataset composed of urban flooding images collected by monitoring equipment and social sources was used for training.Two key evaluation indicators,accuracy and complexity,were adopted to comprehensively compare the training results of HRNetV2 with four mainstream models:Unet,PSPNet,ResUnet,and DeeplabV3+.The experimental results demonstrate that HRNetV2 exhibits exceptional performance in floodwater image recognition.Its Intersection over Union(IOU),accuracy,recall rate,and F1 score reached 92.19%,96.90%,95.76%,and 95.83%respectively,significantly outperforming the other four comparison models.Meanwhile,HRNetV2 also performs excellently in terms of complexity,with a significant reduction in computational complexity compared to other models,making it more suitable for practical monitoring scenarios.This research not only provides a novel technical means for urban flooding monitoring,enabling more accurate and efficient monitoring of urban flooding situations,but also offers valuable references for

关 键 词：城市内涝 智能监测 HRNetV2 多分辨率融合 深度可分离卷积 

分 类 号：P429[天文地球—大气科学及气象学]