AI-revealed long-term spatio-temporal characteristics of yellow sea green tides

Amid accelerating global environmental change over recent decades, floating macroalgae blooms have expanded worldwide. Notably, the annual Ulva prolifera bloom in the Yellow Sea has emerged as the world's largest green tide disaster, significantly impacting the marine environment and economy. However, the complex ocean–atmosphere environment presents challenges for high-accuracy automatic remote sensing detection. Consequently, a long-term, high-resolution daily dataset remains absent, which limits understanding of spatiotemporal characteristics and hinders disaster prevention. In this study, we designed a high-accuracy automated model, Comprehensive Algae Net (CANet), for green tide detection in satellite images using deep learning. The model enables high-accuracy green tide detection in cloud edges and turbid waters, with a Kappa coefficient of 0.875, and outperforms traditional models such as U-Net and FCN. Based on CANet, we constructed a daily 250-m-resolution multivariate dataset (2007–2024) capturing unmixed area and biomass dynamics. Key findings indicate that over the past 18 years, green tides have cumulatively affected 23% (70,914 km²) of the South Yellow Sea, spanning 526 km east–west and 682 km north–south, with an average annual duration of 82 ± 14 days. Despite interannual variations, the disaster scale can be predicted from the initial area (R² = 0.66, p < 0.05). The model and dataset provide a valuable basis for mechanism analysis and improved operational monitoring.