Recurrent blooms of Ulva prolifera (U. prolifera) in the South Yellow Sea (SYS) have become a significant ecological and socio-economic challenge, disrupting marine ecosystems, aquaculture, and coastal tourism. Traditional methods for detecting and managing these blooms face notable limitations, especially in complex marine environments and under adverse observation conditions. To address these issues, this study employs Sentinel-1 synthetic aperture radar (SAR) imagery and deep learning (DL) techniques. A comprehensive dataset, SYSUPD-SAR, was constructed, containing over 440,000 annotated U. prolifera patches alongside lookalike samples. Pre-training was conducted using the Contrastive Mask Image Distillation (CMID) framework, while the Swin Transformer model was enhanced with multi-head self-attention mechanisms and deep supervision strategies to improve segmentation accuracy and robustness. Key results indicate that the refined model achieved an Intersection over Union (IoU) of 93.24% and a Dice loss of 18.13%, demonstrating its effectiveness in reducing false positives and enhancing detection precision. Additionally, the integration of texture features and consideration of incidence angle variations further strengthened the model’s performance. This study provides a robust framework for U. prolifera detection, offering valuable insights and tools for mitigating the environmental and economic impacts of green tides.