基于改进的U-Net卷积神经网络的遥感影像水体信息提取方法  

Water body information extraction method for remote sensing images based on improved U-Net convolutional neural network

作　　者：宋子俊 董张玉[1,2,3] 张鹏飞 张远南 SONG Zijun;DONG Zhangyu;ZHANG Pengfei;ZHANG Yuannan(School of Computer Science and Information Engineering,Hefei University of Technology,Hefei 230601,China;Anhui Province Key Laboratory of Industry Safety and Emergency Technology,Hefei 230601,China;Intelligent Interconnected Systems Laboratory of Anhui Province,Hefei University of Technology,Hefei 230601,China)

机构地区：[1]合肥工业大学计算机与信息学院,安徽合肥230601 [2]工业安全与应急技术安徽省重点实验室,安徽合肥230601 [3]合肥工业大学智能互联系统安徽省实验室,安徽合肥230601

出　　处：《合肥工业大学学报（自然科学版）》2024年第4期488-495,515,共9页Journal of Hefei University of Technology：Natural Science

基　　金：安徽省重点研究与开发计划资助项目(202004a07020030);安徽省自然科学基金资助项目(2108085MF233);中央高校基本科研业务费专项资金资助项目(JZ2021HGTB0111)。

摘　　要：针对当前遥感影像水体信息提取存在细节水体提取能力较弱、重要特征损失较大的问题,文章提出一种基于改进的U-Net网络实现遥感影像水体信息提取的方法。该方法首先通过引入Resnet残差卷积模块深化传统U-Net网络架构提升特征挖掘能力,并引入Respath残差连接模块减少跳跃连接过程中的语义差距,同时引入PSConv多尺度卷积模块、Eca有效通道注意力机制模块,提高网络特征学习能力,构建PS-Eca-Multiresunet网络模型,弥补传统U-Net网络存在的细节特征提取能力较弱问题。选择“2020年第四届中科星图杯高分遥感图像解译软件大赛”数据集进行实验,结果表明,与传统U-Net网络模型相比,该方法水体提取的平均交并比提高了9.08,像素精度提升了7.4%。改进的网络提取结果能够有效避免阴影影响,提高对细节水体的提取精度,实现遥感影像水体信息的高精度提取。Aiming at the problems of weak ability to extract detailed water bodies and large loss of important features in current water body information extraction from remote sensing images,this paper proposes a method to extract water body information from remote sensing images using an improved U-Net network.The method firstly deepens the traditional U-Net network architecture by introducing the Resnet residual convolution module to improve the feature mining ability,and introduces the Respath residual connection module to reduce the semantic gap in the skip connection process,while introducing the PSConv multi-scale convolution module and Eca effective channel attention mechanism module to improve the network feature learning ability,and constructs the PS-Eca-Multiresunet network model to compensate for the shallow feature loss problem that exists in general networks.The dataset of 2020 GEOVIS Cup Gaofen Challenge on Automated High-Resolution Earth Observation Image Interpretation is selected for the experiment.The results show that the average intersection ratio of water extraction by this method is 9.08 higher than that of the traditional U-Net network model,and the precision of image elements is 7.4%higher than that of the traditional U-Net network model.The improved network extraction results can effectively avoid the influence of shadows,improve the extraction accuracy of detailed water bodies,and achieve high-precision extraction of water body information from remote sensing images.

关 键 词：水体提取 深度学习 多尺度卷积 有效通道注意力机制 Multiresunet网络 

分 类 号：TP751.1[自动化与计算机技术—检测技术与自动化装置] TP753[自动化与计算机技术—控制科学与工程]