.png){kind=icon}
.png){kind=icon}
Return
Cross-Domain SAR Floating Raft Aquaculture Extraction via Intradomain Contrastive Adaptation and Style-Guided Pseudo-Labeling
.png){kind=icon}
PRE
AI
Y
C
D
X
DOI：10.1109/tgrs.2026.3716917.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Unsupervised domain adaptation (UDA) is widely used to address domain shifts in synthetic aperture radar (SAR) imagery. In mariculture monitoring, floating rafts differ markedly in shape, spatial distribution, and density across different sea regions, leading to physical-level domain shifts that existing methods struggle to address. In addition, the quality of pseudo-labels is highly dependent on interdomain alignment, and accumulated errors hinder the effectiveness of self-training and contrastive learning. To overcome these issues, this article proposes a UDA framework based on intradomain contrastive adaptation and style-guided pseudo-labeling (CASP-DA) for cross-domain mariculture area extraction. CASP-DA improves segmentation performance by generating reliable pseudo-labels to support contrastive adaptation in the target domain. First, a three-stage pretraining strategy is designed to progressively enhance feature representation and cross-domain generalization of the residual feature alignment network (RFA-Net) under varying domain discrepancies. Second, a style-guided pseudo-label (SGPL) generation module reduces error propagation by decreasing the reliance on interdomain alignment in pseudo-label generation, thereby generating high-quality pseudo-labels for the following module. In addition, an intradomain contrastive adaptation (IDCA) module is designed to mitigate the negative transfer from variations in raft morphology and spatial distribution. This module enables self-correction within the target domain via pixel-level contrastive learning and further boosts performance through iterative self-training. Extensive experiments on GF-3 SAR images from four coastal regions in China demonstrate the effectiveness of the proposed method.
Keywords:
Contrastive learning
floating raft aquaculture
semantic segmentation
synthetic aperture radar (SAR)
unsupervised domain adaptation (UDA)
Journal
IF:
8.6
Papers:
2.1W
Citations:
10.7W
Organization
Cited Papers