Task-Guided CycleGAN for Domain-Adapted Coastal Boundary Segmentation in Remote Sensing Images

Deep-learning-based segmentation techniques for coastal erosion monitoring have emerged as a promising solution to this challenge. However, existing approaches remain limited by insufficient data diversity and poor cross-dataset generalization. Although fine-tuning segmentation models can improve generalization performance, it inherently modifies the learned coastal boundary representation. This work investigates style transfer as a strategy to improve cross-domain coastal boundary segmentation while maintaining a fixed segmentation model. CycleGAN and two novel task-guided CycleGAN variants were employed to transfer styles from six publicly available coastal datasets (CoastTrain, SANet, SLS, SWED, YTU and HRNB) to the SNOWED target domain. The translated images were subsequently segmented using a SegFormer model trained exclusively on SNOWED to evaluate the impact of style transfer on downstream segmentation performance. The results demonstrate that the effectiveness of style transfer depends strongly on source-domain characteristics, including spectral composition, spatial resolution and visual similarity to the target domain. Datasets containing near-infrared (NIR) information and exhibiting clear spectral differences from SNOWED, particularly YTU and SANet, benefited the most from style transfer. The proposed task-guided CycleGAN variants improved the F1-score from 75.20% to 98.90% on YTU and from 35.07% to 53.39% on SANet, while consistently outperforming conventional normalization methods across the majority of the evaluated datasets. In contrast, high-resolution RGB source imagery that already achieved strong raw segmentation performance exhibited limited or no benefit from style transfer. HRNB and SLS showed modest performance improvements, with the F1-score increasing from 96.09% to 97.52% and from 59.20% to 61.41%, respectively. Overall, the results demonstrate that task-guided style transfer can improve cross-domain coastal boundary segmentation while maintaining a static segmentation model, making it a promising approach for long-term coastal monitoring across heterogeneous remote sensing datasets.