基于RANet互导融合学习的遥感影像浒苔检测方法
Detection method of enteromorpha prolifera in remote sensing images based on RANet mutual guidance fusion learning
张蔡辉 1任鹏 2陈艳拢 3初佳兰 3郑斌2
作者信息
- 1. 中国石油大学(华东) 海洋与空间信息学院 山东 青岛 266580;国家海洋环境监测中心 辽宁 大连 116023
- 2. 中国石油大学(华东) 海洋与空间信息学院 山东 青岛 266580
- 3. 国家海洋环境监测中心 辽宁 大连 116023
- 折叠
摘要
针对遥感影像浒苔检测标注数据少的问题,本文提出一种基于残差注意力网络(residual attention network,RANet)互导融合学习的浒苔检测方法.首先,本文搭建了残差卷积模块联合注意力机制的RANet模型用于浒苔检测.其次,在双网络架构下利用两个RANet模型相互引导,挑选双模型融合后的高置信度伪标签,结合数据增强来逐步扩充训练集,从而对双模型迭代学习实现高精度的浒苔检测.实验结果表明,与阈值法、生成对抗网络(generative adversarial network,GAN)、经典分割模型(FCN、SegNet、UNet、PSPNet和DeepLabv3+)相比,基于RANet互导融合学习的浒苔检测方法具有更高的检测准确性.本研究构建的模型具备进行大规模浒苔监测的可行性,可为大规模浒苔暴发时的灾情监测提供技术支撑.
Abstract
Aiming at the problem of insufficient labeled data in remote sensing image detection of enteromorpha prolifera,this paper proposes a method for enteromorpha prolifera detection based on residual attention network(RANet)mutual guidance fusion learning.Firstly,this paper builds a RANet model with residual convolution module and attention mechanism for enteromorpha prolifera detection.Secondly,the two RANet models are used to guide each other under the dual network architecture,and the high confidence pseudo labels after the fusion of the double models are selected,and the training set is gradually expanded by combining data augmentation,so as to achieve high-precision enteromorpha prolifera detection by the iterative learning of the double models.The experimental results show that compared with threshold method,GAN,classical segmentation model(FCN,UNet,SegNet,PSPNet and DeepLabv3+),the detection method based on RANet mutual guidance fusion learning has higher detection accuracy.The model constructed in this study has feasibility for large-scale enteromorpha prolifera monitoring,which could provide technical support for the disaster monitoring of large-scale enteromorpha prolifera outbreak.
关键词
浒苔检测/互导融合学习/残差卷积/注意力机制/伪标签/数据增强Key words
detection of enteromorpha prolifera/mutual guidance fusion learning/residual convolution/attention mechanism/pseudo label/data augmentation引用本文复制引用
基金项目
国家自然科学基金项目(62071491)
国家自然科学基金项目(41706105)
出版年
2024
NETL
NSTL
万方数据