改进U-Net模型的水体分割方法  

Improved U-Net Model for Water Body Segmentation

作　　者：蔡宏超 龚建华 张友松 王建茹 胡卫东 CAI Hongchao;GONG Jianhua;ZHANG Yousong;WANG Jianru;HU Weidong(School of Geology and Geomatics,Tianjin Chengjian University,Tianjin 300384,China;Zhejiang-CAS Application Center for Geoinformatics,Jiaxing,Zhejiang 314100,China;Aerospace Information Research Institute,Chinese Academy of Sciences,Beijing 100094,China;Zhejiang Marine and Fishery Law Enforcement Corps,Hangzhou 310007,China)

机构地区：[1]天津城建大学地质与测绘学院,天津300384 [2]浙江中科空间信息技术应用研发中心,浙江嘉兴314100 [3]中国科学院空天信息创新研究院,北京100094 [4]浙江省海洋与渔业执法总队,杭州310007

出　　处：《遥感信息》2024年第5期140-147,共8页Remote Sensing Information

基　　金：中国科学院空天信息创新研究院前沿科学与颠覆性技术研究先导基金(E0Z21101)。

摘　　要：针对遥感影像水体分割算法在应对细小水系繁多、水环境复杂的情况时分割准确率低,以及易混淆地物、错分率高等问题,设计了一种改进的U-Net模型。首先,设计并建立改进的U-Net模型,通过对原始图像增加上采样部分,使该模型形成上下对称的结构;同时,采用S形循环,通过增加神经网络的中间层层数,保留更多的图像特征。其次,对于改进模型的深度进行调整,即一次上下采样和两次上下采样,文章据此提出两种不同的网络结构,并对比分析二者的分割精度。实验证明,OSUNet-V2模型(两次上下采样)精度更高、效果更佳,相较于U-Net,精度提高1.28%,交并比提高2.19%,对地物的分辨能力及抗易混淆地物的干扰能力高于U-Net,可以为城市水体分布情况的快速获取提供数据参考和支撑。To address challenges faced by water body segmentation algorithms in remote sensing imagery,including low accuracy in handling numerous and intricate small water systems,complexities in water environments,and high misclassification rates of confusable land features,an enhanced U-N et al gorithm is designed.Firstly,an improved U-Net model is designed and established.This model forms a symmetric structure by adding an upsampling component to the original image and uses an S-shaped loop,while increasing the number of intermediate layers in the neural network to retain more image features.Secondly,the depth of the improved model is adjusted,including one-time and two-time up-and-down sampling.This article introduces two different network structures and compares the segmentation accuracy of both.Experimental results demonstrate that the OSUNet-V2 model(two-time upsampling)achieves higher accuracy and better performance.Compared to U-Net,this model shows a 1.58%increase in accuracy and a 3.68%improvement in intersection over union.The model exhibits a higher capability for discriminating land features and demonstrates superior resistance to interference from easily confusable terrain types compared to U-Net,making it a valuable resource for rapidly acquiring data to understand urban water body distribution.

关 键 词：深度学习 水体提取 高分辨率遥感影像 特征提取 U-Net 

分 类 号：TP79[自动化与计算机技术—检测技术与自动化装置]