SAR图像洪水淹没区语义分割方法研究  

Research on Semantic Segmentation of Flood Inundation Areas in SAR Images

作　　者：王杰 黄本胜[1] 陈亮雄[1] 杨静学[1] WANG Jie;HUANG Bensheng;CHEN Liangxiong;YANG Jingxue(Guangdong Research Institute of Water Resources and Hydropower,Guangzhou 510635,China;School of Civil Engineering,Sun Yat‐sen University,Zhuhai 519082,China)

机构地区：[1]广东省水利水电科学研究院,广东广州510635 [2]中山大学土木工程学院,广东珠海519082

出　　处：《测绘地理信息》2024年第6期125-130,共6页Journal of Geomatics

基　　金：广东省重点领域研发计划(2020B0101130018);广州市科技计划项目(202201011275);广东省水利科技创新项目(2022-02,2020-15)

摘　　要：研究了合成孔径雷达(synthetic aperture radar,SAR)图像水体与非水体的后向散射特性,围绕样本自动标注与增强训练这两个关键问题,利用阈值分割、水文约束与马尔科夫随机场设计了自动标注算法,并将特征增强网络与嵌入式样本增强相结合,提出了一种在有限样本条件下的SAR图像水体语义分割方法。以广东省“22·6”北江特大洪水为实验案例,采用了潖江蓄滞洪区的GF-3影像为实验数据。通过实验证明,本研究提出的方法能够有效识别洪水淹没范围,总体分类准确率达到了92.6%左右。This paper studies the backscattering characteristics of SAR images of water and non-water,and focuses on the two key issues of automatic annotation and enhanced training strategy.Threshold segmentation,hydrologic constraint and Markov random fields(MRF)are used in designing the automatic labeling algorithm using,with the integration of the feature enhancement network and embedded sample enhancement,resulting in a semantic segmentation method for SAR images with limited samples.In this study,the“22·6”Beijiang extreme flood is taken as an experimental case,and the GF-3 images of Pajiang River detention area are used as the experimental data.According to the experimental results,it is evident that the proposed method is capable of distinguishing water from non-water with 92.6%overall accuracy.

关 键 词：遥感 洪水淹没区 自动标注 增强训练 语义分割 

分 类 号：TV122[水利工程—水文学及水资源] TN957.52[电子电信—信号与信息处理]