融合Siam-U-Net++和注意力机制的水体变化检测——以三峡库区奉节县为例
Water Change Detection Integrating Siam-U-Net++ and Attention Mechanism——Taking Fengjie County in Three Gorges Reservoir Area as An Example
林娜 1郭江 1王斌 2周俊宇 1何静1
作者信息
- 1. 重庆交通大学 智慧城市学院,重庆 400074
- 2. 重庆市地理信息与遥感应用中心,重庆 401147
- 折叠
摘要
由于三峡库区的复杂地形特点和受到部分区域的阴影遮盖影响,因此水体变化检测容易出现漏检误检.针对这一情况,为了能够提高水体变化检测结果的精度,提出了一种融合Siam-U-Net++和scSE注意力机制的三峡库区水体变化检测算法.利用编码阶段共享权值的Siam-U-Net++为主干网络,在编码器网络单元上选择拥有残差结构的ResNet34网络,快速高效获取水体变化特征信息.在Siam-U-Net++的上采样之后引入结合通道注意力和空间注意力的scSE注意力机制模块.在自制的库区奉节县数据集上分别与LinkNet、U-Net++、DeeplabV3+网络模型以及NDWI进行应用测试,并引用不同注意力机制进行对比实验,其召回率、精确率和F1值分别为94.57%、90.75%、92.62%,均优于其他模型.实验结果表明该算法有效提高了在水体变化检测结果上的效果.
Abstract
Due to the complex terrain characteristics of the Three Gorges Reservoir area and the shadow cover-age in some areas,the detection of water body changes is prone to missed detection and false detection.In view of this situation,in order to improve the accuracy of water body change detection results,a water body change detection algorithm in the Three Gorges Reservoir area is proposed that integrates Siam-U-Net++ and scSE attention mechanism.The Siam-U-Net++ with shared weights in the encoding stage is used as the backbone network,and the ResNet34 network with residual structure is selected on the encoder network unit to quickly and efficiently obtain the characteristic information of water body changes.The scSE attention mechanism mod-ule combining channel attention and spatial attention is introduced after the upsampling of Siam-U-Net++.On the self-made data set,it is tested with LinkNet,U-Net++,DeeplabV3+ network models and NDWI,and different attention mechanisms are used for comparative experiments.The recall rate,accuracy rate and F1 value were 94.57%,90.75%and 92.62%respectively,which were better than other models.Experimental re-sults show that the algorithm can effectively improve the results of water body change detection.
关键词
水体变化检测/深度学习/三峡库区/Siam-U-Net++网络/ResNet34/注意力机制Key words
Water change detection/Deep learning/Three Gorges reservoir area/Siam-U-Net++ network/ResNet34/Attention mechanism引用本文复制引用
基金项目
重庆市自然科学基金面上项目(CSTB2023NSCQ-MSX0781)
国家重点研发计划项目(2021YFB2600600)
国家重点研发计划项目(2021YFB2600603)
出版年
2023
NETL
NSTL
万方数据