基于深度学习的遥感图像水体提取综述

A review of water body extraction from remote sensing images based on deep learning

对江河湖泊等水体目标的空间分布、时序变化进行及时、准确的检测和统计具有十分重要的意义和应用价值,已成为当前遥感地表观测研究的重要热点。传统水体提取方法依靠经验设计的指数模型进行水体阈值分割或分类,易受到植被、建筑物等地物的阴影以及水体自身泥沙含量、盐碱浓度等理化特性变化的影响,难以在不同时空尺度环境下保持鲁棒性。随着海量多源、多分辨率的遥感图像的快速获取,深度学习算法在水体提取方面的优势逐渐凸显并被国内外学者广泛关注。得益于深度神经网络模型强大的学习能力和灵活的卷积结构设计方案,研究人员陆续提出了各种模型和学习策略用以提高水体提取的鲁棒性和精度,但目前缺少对该类研究进展的全面综述和问题剖析。因此,文章对近年来国内外发表的相关研究成果进行总结,重点归纳不同算法在遥感图像水体提取方面的优缺点及存在的问题,并对基于深度学习的遥感图像水体提取方法研究的发展提出了建议和展望。

Timely and accurate detection and statistical analysis of the spatial distributions and time-series variations of water bodies like rivers and lakes holds critical significance and application value.It has become a significant interest in current remote sensing surface observation research.Conventional water body extraction methods rely on empirically designed index models for threshold-based segmentation or classification of water bodies.They are susceptible to shadows of surface features like vegetation and buildings,and physicochemical characteristics like sediment content and saline-alkali concentration in water bodies,thus failing to maintain robustness under different spatio-temporal scales.With the rapid acquisition of massive multi-source and multi-resolution remote sensing images,deep learning algorithms have gradually exhibited prominent advantages in water body extraction,garnering considerable attention both domestically and internationally.Thanks to the powerful learning abilities and flexible convolutional structure design schemes of deep neural network models,researchers have successively proposed various models and learning strategies to enhance the robustness and accuracy of water body extraction.However,there lacks a comprehensive review and problem analysis of research advances in this regard.Therefore,this study summarized the relevant research results published domestically and internationally in recent years,especially the advantages,limitations,and existing problems of different algorithms in the water body extraction from remote sensing images.Moreover,this study proposed suggestions and prospects for the advancement of deep learning-based methods for extracting water bodies from remote sensing images.