Multi-scale drift characteristics of Ulva prolifera in the Yellow Sea derived from deep learning-based MODIS and Sentinel-1 observations.

Accurately quantifying the spatial coverage of Ulva prolifera (U. prolifera) and its drift characteristics across different temporal scales is crucial for understanding its spatiotemporal dynamics and improving prediction capability in the Yellow Sea. However, existing remote sensing methods have difficulty in reliably capturing the short-term emergence, expansion, and drift processes of U. prolifera due to missing observations, cloud cover, and limited time coverage. Moreover, previous studies have primarily focused on annual or monthly dynamics, while short-term variability, particularly short-term drift, which is critical for accurate prediction of U. prolifera, has received comparatively limited attention. In this study, we propose a novel AttFusionViT-UNet integrating attention mechanisms and a Vision Transformer for detecting U. prolifera using MODIS data from 2008 to 2024 and Sentinel-1 data from 2015 to 2024. The detection results were further fused to generate weekly, monthly, and annual spatiotemporal distributions of U. prolifera, enabling a multi-scale analysis of drift characteristics. The results indicate that the proposed model achieves mean IoU values of 81.81% and 84.81% on MODIS and Sentinel-1 data, respectively. The fusion of MODIS and Sentinel-1 data increases the effective weekly observation coverage to 71%, significantly enhancing the availability of effective observations and the reliability of detection results compared with single-sensor observations. Drift patterns of U. prolifera exhibit relative stability at the interannual scale and increased variability at the monthly scale, whereas weekly observations effectively capture short-term variations. These results provide high-frequency and high-reliability observations, supporting future prediction of U. prolifera drift in the Yellow Sea.