基于SE-ConvLSTM的时空特征融合SAR图像海冰分类
Sea Ice Classification of SAR Images based on SE-ConvLSTM Spatial-temporal Feature Fusion
葛梦滢 1高稳 2祝敏 2郭伟其 3宋巍2
作者信息
- 1. 上海海洋大学 信息学院,上海 201306;上海大学 工程训练中心,上海 200444
- 2. 上海海洋大学 信息学院,上海 201306
- 3. 上海东海海洋工程勘察设计院有限公司,上海 200137
- 折叠
摘要
基于合成孔径雷达(SAR)图像的海冰分类已经成为海冰监测的重要基础,但现有方法往往利用图像空间特征,很少考虑时间特征.提出了一种融合时空特征的SAR图像海冰分类网络SE-ConvLSTM.首先使用ConvLSTM对HH和HV极化图像分别提取时空特征,然后将提取的不同层次和通道的时空特征进行拼接,并利用SE通道注意力进行通道特征响应的自适应重新校准,最后利用SoftMax函数进行图像分类.将SI-STSAR-7数据集6个时间步长的图像块作为输入对所提方法与其他分类方法进行了对比实验.结果显示:SE-ConvLSTM在总体情况和分类困难的厚一年冰上分别达到了 97.06%和 90.01%的精度,表明加入时间信息有助于提高分类准确率.同时,所提网络在生成海冰分布图时对主要冰类型密集度较低的区域和SAR影像的边界位置都具有更好的识别能力.
Abstract
Sea ice classification using Synthetic Aperture Radar(SAR)images is a crucial aspect of sea ice mon-itoring.Existing methods have mainly relied on spatial features of SAR images,but rarely consider temporal fea-tures,which can potentially provide additional information.A novel approach called SE-ConvLSTM has been developed to combine both spatial and temporal features for sea ice SAR image classification.Firstly,ConvL-STM is used to extract the spatial-temporal features of HH and HV polarization SAR images respectively.Then,the spatial-temporal features of different layers and channels are concatenated,and the channel feature response is adaptive recalibrated by using SE channel attention.Finally,SoftMax function is used for image classification.To evaluate the effectiveness of the SE-ConvLSTM method,six time-step image blocks of SI-STSAR-7 dataset were used for comparison with other classification methods.The results indicate that SE-ConvLSTM achieved an overall accuracy of 97.06%and 90.01%for the thick one-year ice which is difficult to classify.This suggests that adding temporal information can significantly improve classification accuracy.Addi-tionally,the proposed network has better recognition ability for regions with low density of main ice types and for boundary positions of SAR images,making it an effective tool for generating sea ice distribution maps.
关键词
SAR/海冰分类/时空特征/ConvLSTM/SE通道注意力Key words
SAR/Sea ice classification/Spatial-temporal features/ConvLSTM/SE channel attention引用本文复制引用
基金项目
国家重点研发计划项目(2021YFC3101602)
上海市科委部分地方高校能力建设项目(20050501900)
出版年
2023
NETL
NSTL
维普
万方数据