一种联合空间变换和置换注意力机制的近岸水产养殖区信息提取方法
An information extraction method from offshore aquaculture area based on spatial transformation and shuffle attention mechanism
巫统仁 1张显 2刘培 3文婷婷 1邹振学1
作者信息
- 1. 海南热带海洋学院崖州湾创新研究院,海南三亚 572025;海南省海洋与渔业科学院,海南海口 571126;河南理工大学测绘与国土信息工程学院,河南焦作 454003
- 2. 河南理工大学测绘与国土信息工程学院,河南焦作 454003
- 3. 海南热带海洋学院崖州湾创新研究院,海南三亚 572025;海南省海洋与渔业科学院,海南海口 571126;北华航天工业学院遥感信息工程学院,河北廊坊 065000
- 折叠
摘要
为解决水产养殖区遥感提取过程中近岸坑塘养殖目标和网箱养殖目标地物背景复杂,受房屋、植被、海水和船只等干扰提取精度较低等问题,提出了一种联合空间变换和置换注意力机制的近岸水产养殖区信息提取方法SA-STN-Net,选取海南省文昌市八门湾和万宁市坡头港为研究区域,先利用光谱特征和纹理特征构建水产养殖目标先验知识,然后在U-Net模型基础上联合空间变换网络(spatial transformer net-work,STN)和置换注意力机制(shuffle attention,SA),用于增强养殖目标空间特征、减少复杂地物的干扰并聚焦近岸水产养殖区域.结果表明:与原始U-Net模型相比,SA-STN-Net模型的总体提取精度和平均交并比提高了 3.3％和5.7％;与当前较为先进的A2fpn、Swin-Transformer和Dc-Swin等深度学习分割算法相比,SA-STN-Net模型具有更好的分割性能,F1分数分别提高了 6.7％、4.2％和7.2％.研究表明,本文提出的SA-STN-Net模型能适应近岸水产养殖目标地物背景复杂的情况,可对近岸养殖目标进行有效提取,本研究结果可为近岸规划与管理部门提供技术支持.
Abstract
In order to solve the problem of low accuracy in extracting aquaculture areas using remote sensing tech-nology due to complex background of onshore aquaculture and offshore cage culture areas and to easily disturbed by factors including houses,vegetation,seawater,and ships,a complex deep learning method that combines shuffle attention mechanisms and spatial transformation network was proposed,and tested in Bamen bay in Wenchang city and Potou Port in Wanning City.With the help of GF-2 high resolution remotely sensed data,the prior knowledge of aquaculture targets was constructed using spectral and texture features.Then,based on the U-Net model,the spatial transformation network(STN)and the shuffle attention(SA)mechanism are combined to enhance the spa-tial characteristics of the aquaculture area and to reduce the interference of complex backgrounds.The test results showed that the overall accuracy and mean intersection over union of SA-STN-Net model were enhanced by 3.3％and 5.7％compared with the preliminary U-Net model,respectively.Swin-Transformer,Dc-Swin,and F1 score of SA-STN-Net model were found to be increased by 6.7％,4.2％and 7.2％in the score compared with the most state-of-art deep learn model such as A2fpn,respectively.The findings demonstrate that the proposed SA-STN-Net model is adapted to the complex environmental background of offshore aquaculture,effectively extracts offshore aq-uaculture targets,and can provide technical support for offshore planning and management departments.
关键词
空间变换网络(STN)/置换注意力(SA)/深度学习/水产养殖区信息提取Key words
spatial transformer network(STN)/shuffle attention(SA)/deep learning/extraction of aquaculture area引用本文复制引用
基金项目
海南热带海洋学院崖州湾创新研究院开放课题重点项目(2022CXYKFKT03)
海南热带海洋学院崖州湾创新研究院组学生创新重点项目(2022CXYXSCXXM07)
海南省自然科学基金(423MS120)
海南省海洋与渔业科学院省本级课题(KYL-2024-06)
河北省自然科学基金面上项目(D2020409002)
出版年
2024
NETL
NSTL
维普
万方数据