摘要
针对山区阴影、植被茂密和SAR相干斑噪声易导致水体误提和水体边界细节保持困难等问题,本文综合利用多极化SAR数据,提出了一种融合多尺度雷达后向散射特征深度学习的SAR影像水体提取方法。该方法采用多分辨率残差卷积和多尺度特征恒等映射来更好地表征山区复杂环境下多极化SAR影像丰富的多尺度地物特征,以提高水体提取的精度和完整性。试验采用覆盖山西省某水库的Sentinel-1双极化数据,并选择经典SVM、MRF算法和U-Net网络模型与本文算法进行水体提取性能的定性定量对比分析。F1、Kappa系数、IoU和OA分别达到92.14%、91.39%、85.43%和98.62%,表明本文方法水体提取综合性能最优,而且能够较好地保持水体提取的边界细节。
To address the challenges of frequent water misclassification and difficulty in preserving water boundary details caused by mountain shadows,dense vegetation,and SAR coherent speckle noise in complex terrains,this study proposes a novel deep learning-based water extraction method for SAR imagery that integrates multi-scale radar backscattering features through multi-polarization SAR data.The developed approach employs multi-resolution residual convolution and multi-scale feature identity mapping to effectively characterize the rich multi-scale terrain features in complex mountainous environments using multi-polarization SAR imagery,thereby enhancing both accuracy and completeness in water body extraction.Experimental validation was conducted using Sentinel-1 dual-polarization data covering a reservoir in Shanxi Province,with comparative analysis against classical SVM,MRF algorithm,and U-Net model through qualitative and quantitative assessments.The experimental results showed that 92.14%F1,91.39%Kappa coefficient,85.43%IoU,and 98.62%OA.This indicates that the method proposed in this paper has the best comprehensive performance in water extraction and can effectively maintain the boundary details of water extraction.
作者
原志芳
李俊晓
康倩
YUAN Zhifang;LI Junxiao;KANG Qian(College of Civil and Transportation Engineering,Shenzhen University,Shenzhen 518000,China;Shanxi Jiuli Information Technology Co.,Ltd.,Taiyuan 030000,China)
出处
《测绘通报》
北大核心
2026年第2期1-6,共6页
Bulletin of Surveying and Mapping
基金
国家自然科学基金面上项目(42371431)
国家重点研发计划(2022YFB3903705)。
关键词
水体提取
多极化SAR
深度学习
多尺度
山体阴影
waterbody extraction
multi-polarization SAR
deep learning
multi-scale
mountain shadows
