注：国家自然科学基金项目“基于二维材料-金属的可调控构建敏感膜及其在水中多参数重金属传感器研究”（No. 61901042）；国家自然科学基金项目“中医脉诊指压传感器材料与器件研究”（No. 62071054）；北京市教委科研计划一般项目（No. KM202011232016）；传感器国家重点实验室开放课题（No. SKT1902）；传感器北京市重点实验室开放课题（No. 2019CGKF007）
作者：张铭飞^1,2,3，高国伟^1,2，胡敬芳^1,2,3，宋钰^1,2,3
单位：1. 北京信息科技大学 传感器北京市重点实验室，北京 100101；
2. 北京信息科技大学 现代测控技术教育部重点实验室，北京 100192；
3. 传感器联合国家重点实验室，中国科学院空天信息创新研究院，北京 100190
中图分类号：TP751                
文献标识码：A                
文章编号：1006-883X(2021)03-0009-06
收稿日期：2021-02-18

摘要：地表水的勘测对于海岸线变化、环境保护、防灾减灾、水质检测都有重要的意义，借助遥感图像可以快速、反复、精确地获取到地表水的时空分布特征。文章调研国内外学者在遥感图像水体识别方向的研究成果，简述基于遥感技术的水体识别方法。其中，阈值法通过对水体和背景地物的光谱曲线进行分析，选取适合的阈值进行图像分割，操作简单便利，存在信噪比低、易将水体与背景地物混淆的问题。决策树法和自动提取水体法解决了阈值法的显著缺点，然而很难在精确度上得到进一步的提升。近年来，随着深度学习的广泛应用，逐渐被用于遥感图像的水体提取，深度学习方法具有优秀的特征提取能力，在提取精度上有很大提升，然而深度学习过度依赖带有标签的样本数据，因此具有一定的局限性。对样本进行标记需要消耗大量的时间、人力，因此，无监督学习对于遥感图像的水体识别具有重要意义。
关键词：水体识别；SAR图像；阈值法；深度学习；神经网络

Summarize of Research on Water Recognition and Detection based on Remote Sensing Image
ZHANG Mingfei^{1, 2, 3}, GAO Guowei^{1, 2}, HU Jingfang^{1, 2, 3}, SONG Yu^{1, 2, 3}
1. Beijing Sensor Key Laboratory, Beijing Information Science & Technology University, Beijing 100101, China; 2. Key Laboratory of Modern Measurement and Control Technology, Beijing Information Science & Technology University, Beijing 100192, China; 3. State Key Laboratory of Transducer Technology, Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing 100190, China
Abstract: Surface water survey is of great significance to coastline changes, environmental protection, disaster prevention and mitigation, and water quality detection. With the help of remote sensing images, the temporal and spatial distribution characteristics of surface water can be quickly, repeatedly, and accurately obtained. The article investigates the research results of domestic and foreign scholars in the direction of remote sensing image water body recognition, and briefly describes the water body recognition method based on remote sensing technology. Among them, the threshold method analyzes the spectral curves of water bodies and background features, and selects appropriate thresholds for image segmentation. The operation is simple and convenient, and there is a problem of low signal-to-noise ratio and easy to confuse water bodies with background features. The decision tree method and the automatic water extraction method solve the significant shortcomings of the threshold method, but it is difficult to further improve the accuracy. In recent years, with the widespread application of deep learning, it has gradually been used for water extraction from remote sensing images. Deep learning methods have excellent feature extraction capabilities and have greatly improved the extraction accuracy. However, deep learning overly relies on labeled samples Data, so it has certain limitations. It takes a lot of time and manpower to label samples, so unsupervised learning is of great significance for water recognition in remote sensing images.
Key words: water body identification; sar image; threshold value method; deep learning method; neural network

阅读全文

备注：2021年 第27卷 第03期