.png){kind=icon}
.png){kind=icon}
Return
RDANet: ResNeXt-based dual attention network for accurate sea and land segmentation in remote sensing image
.png){kind=icon}
PRE
AI
Y
J
X
X
D
P
S
U
DOI：10.1016/j.rsma.2026.104914.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Sea and land segmentation is critical for autonomous coastline monitoring and management. The complexity and diversity of coastline types present challenges in accurate extraction of sea and land areas. In addition, the weak sea-land features of images such as beach wetlands, narrow channels and artificial structures exacerbate the difficulty of accurate sea-land segmentation. To address these issues, we proposed a ResNeXt-Based Dual Attention Network (RDANet) for accurate sea and land segmentation in this study. The RDANet adopts ResNeXt50 to extract more local features of images, we designed a novel Multi-scale Segmentation Attention (MSA) module of CNNs and combined MSA with Hierarchical Cascade Multi-Head Self-Attention (H-MHSA) of Transfromer. The MSA module can capture multi-scale spatial information more efficiently, establish dependencies of long-distance channels, and enhance features of weak sea-land boundaries. The H-MHSA module can realize global information interaction. We conducted experiments by using the self-create SDD dataset and the publicly available BSD dataset to evaluate the sea-land segmentation performance of the RDANet. The results show that the RDANet model achieves the most accurate sea-land segmentation with 98.52% and 98.41% Mean Intersection over Union (MIoU) on the SDD and BSD, respectively. This study has great significance for the research and application with different types of coastlines in complex marine environments.
Keywords:
Remote sensing
Sea-land segmentation
Dual attention
Transformer
CNNs
GF-6 satellite data
Journal
IF:
2.4
Papers:
486
Citations:
6.1K