Abstract: Water information extraction is an important prerequisite for water resources management and protection. Remote sensing technology is an important means of water extraction. Most of the exist. ing water extraction methods are applied to low and medium resolution remote sensing images, while there is little research on water extraction from high-resolution remote sensing images, especially GF-7 satellite images. Combined with the characteristics of high spatial resolution and huge pixel number of GF-7 data, an automatic water information extraction method based on semantic segmentation network is proposed in this paper. In this study, the digital surface model DSM is extracted from the data based on the stereo image of GF-7, and learning and training in the form of band combination are carried out, which effectively alleviates the water misclassification and missing in complex terrain or background. The network uses hole convolution and spatial pyramid pool structure, which effectively improves the ability of large-scale spatial feature extraction and semantic information learning at multiple spatial scales. Taking Quanhe river basin in Hubei province as the study area, water extraction experiments and comparative experiments of different methods are carried out. Due to the large amount of non-water information in the image, the average intersection ratio MIOU is introduced as the evaluation index. The experiment results show that this method can effectively realize the high-precision extraction of water information from GF-7 image, and the extraction accuracy is significantly improved compared with the traditional water index method, support vector machine and U-Net method. The water extraction accuracy can reach 95.3% and MIOU can reach 0.932.

Key words: water information extraction, satellite remote sensing, GF-7, semantic segmentation network

摘要： 水体信息提取是水资源管理与保护的重要前提。遥感技术是水体提取的一个重要手段，现有的水体提取方法大多应用于中低分辨率遥感影像，而对高分辨率遥感影像，尤其是GF-7卫星影像的水体提取研究较少。笔者针对GF-7数据空间分辨率高和像素级数庞大等特点，提出了一个基于语义分割网络的水体信息自动提取方法。在该研究中，基于GF-7的立体像对数据进行数字地表模型（DSM）的提取，并通过波段组合的形式进行学习训练，有效缓解了复杂地形或背景下的水体错分和漏分现象。网络使用空洞卷积和空间金字塔池化结构有效提高了模型大尺度空间特征提取和多空间尺度下的语义信息学习能力。以湖北省泉河流域为研究区，进行了水体提取实验及不同方法的对比实验。由于影像非水体信息较多，引入平均交并比MIOU 作为评价指标。实验结果表明，该方法可有效实现GF-7影像水体信息的高精度提取，与传统水体指数法、支持向量机及U-Net方法相比，提取精度显著提高，水体提取精度可达到95.3%，MIOU 高达0.932。

关键词: 水体信息提取, 卫星遥感, GF-7号, 语义分割网络