基于分层特征提取和多尺度特征融合的高分辨率遥感影像水体提取深度学习算法  

Deep Learning Algorithm for Water Body Extraction from High-resolution Remote Sensing Images Based on Hierarchical Feature Extraction and Multi-scale Feature Fusion

作　　者：盛晟 万芳琦 林康聆 胡朝阳[3] 陈华[1] SHENG Sheng;WAN Fangqi;LIN Kangling;HU Zhaoyang;CHEN Hua(State Key Laboratory of Water Resources Engineering and Management,Wuhan University,Wuhan 430072,China;Jiangxi Institute of Natural Resources Surveying,Mapping and Monitoring,Nanchang 330009,China;Fujian Provincial Investigation,Design&Research Institute of Water Conservancy&Hydropower,Fuzhou 350001,China)

机构地区：[1]武汉大学水资源工程与调度全国重点实验室,湖北武汉430072 [2]江西省自然资源测绘与监测院,江西南昌330009 [3]福建省水利水电勘测设计研究院,福建福州350001

出　　处：《人民珠江》2024年第2期45-52,共8页Pearl River

基　　金：国家重点研发计划项目(2022YFC3002701)。

摘　　要：高精度的水体提取有助于水资源监测和管理。目前基于遥感影像的水体提取方法缺乏对于边界质量的重视,造成边界划分不准确,细节保留度低的问题。为了提升遥感影像水体提取的边界与细节的精度,提出了一种基于多尺度特征融合的高分辨率遥感影像水体提取深度学习算法,包括分层特征提取模块与融合多尺度特征的堆叠连接解码器模块。分层特征提取模块中,引入了通道注意力结构,用于整合高分辨率遥感影像中水体的形状、纹理和色调信息,以便更好地理解水体的形状和边界。在融合多尺度特征的堆叠连接解码器模块中,进行了多层次语义信息的堆叠连接,并加强了特征提取,同时捕捉了广泛的背景信息和细微的细节信息,以实现更好的水体提取效果。在自行标注的数据集与公开数据集上的试验结果表明,模型的准确率达到了98.37%和91.23%,与现有的语义分割模型相比,提取的水体边缘更加完整,同时保留细节的能力更强。提出的模型提升了水体提取的精度和泛化能力,为高分辨率遥感影像水体提取提供了参考。Highly accurate water body extraction can be helpful for water resources monitoring and management.The current methods of water body extraction based on remote sensing images lack attention to boundary quality,resulting in inaccurate boundary delineation and low detail retention.To improve the boundary and detail accuracy of water body extraction for remote sensing images,this paper proposes a deep learning algorithm for water body extraction from high-resolution remote sensing images based on multi-scale feature fusion.The model includes a hierarchical feature extraction module and a stacked-connected decoder module that fuses multi-scale features.In the hierarchical feature extraction module,a channel attention structure is introduced for integrating shape,texture,and hue information of water bodies in high-resolution remote sensing images to better understand the shapes and boundaries of water bodies.In the stacked-connected decoder module that incorporates multi-scale features,the stacked connection of multi-level semantic information and enhanced feature extraction are performed.Meanwhile,broad background information and fine detail information are captured to achieve better water body extraction results.Experimental results on both self-annotated and publicly available datasets show that the model yields 98.37%and 91.23%accuracy,and extracts more complete edges of water bodies while retaining more details than existing semantic segmentation models.The proposed model improves the accuracy and generalization ability of water body extraction and provides references for water body extraction from high-resolution remote sensing images.

关 键 词：水体提取 高分辨率遥感影像 深度学习 多尺度特征融合 

分 类 号：P332[天文地球—水文科学]