摘要
针对利用SAR图像实现冰间水道检测任务时,深度学习模型训练受限于大量人工标注数据获取困难的问题,结合传统图像分割方法和Deeplabv3+深度学习网络,提出了一种伪标签引导的无监督冰间水道检测方法。首先引入大津算法和分水岭算法对SAR图像进行粗分割;然后根据粗分割结果设计伪标签生成策略,并生成确定度高的伪标签图像;最后用该伪标签指导Deeplabv3+网络的训练,实现冰间水道检测。在波弗特海区域哨兵一号卫星1级地距多视产品制作的数据集上,实验结果表明该方法性能指标mIoU、F1-score、OA分别达到88.19%、93.62%和97.50%,能够准确识别海冰SAR图像中的冰间水道。
In the task of sea ice channel detection based on SAR sea ice image,the supervised learning method needs a large number of labeled data,but it is difficult to label manually.Combined with the traditional image segmentation method and the deep learning method of Deeplabv3+,an unsupervised sea ice channel detection method guided by pseudo label is proposed.Firstly,the OTSU algorithm and watershed algorithm are introduced to perform course segmentation on SAR images;then,a pseudo-label generation strategy is designed to generate pseudo-label images with high certainty based on the course segmentation results;finally,the pseudo labels are used to guide the training of Deeplabv3+network to achieve ice channel detection.The experiments use the dataset produced from the 1-level ground distance multi-view product of Sentinel-1 satellite in the Beaufort Sea Region.The results show that the proposed method achieves mloU 88.19%,F1-score 93.62%,and 0A 97.50%,providing accurate sea ice channel detection with SAR images.
作者
宋巍
张博涵
祝敏
张明华
葛梦滢
SONG Wei;ZHANG Bohan;ZHU Min;ZHANG Minghua;GE Mengying(College oflnformation,Shanghai Ocean University,Shanghai 201306,China;Engineering Training Centre,Shanghai University,Shanghai 200444,China)
出处
《海洋测绘》
北大核心
2025年第3期61-64,68,共5页
Hydrographic Surveying and Charting
基金
国家重点研发计划项目(2021YFC3101601)。
关键词
图像分割
无监督学习
传统图像分割算法
SAR影像
伪标签
image segmentation
unsupervised learning
traditional image segmentation algorithms
SAR images
pseudo labels
