Remote sensing-driven machine learning models for spatiotemporal analysis of coastal phytoplankton blooms under climate change scenarios

Abstract

Coastal phytoplankton blooms pose significant environmental challenges, yet spatiotemporal analyses of bloom dynamics under ocean warming and eutrophication remain limited. To address this, we developed machine learning-based regression and classification models for predicting bloom areas and warning levels. These models incorporate remote sensing data and key environmental variables from Coupled Model Intercomparison Project Phase 6 (CMIP6) outputs under different climate change scenarios. We evaluated multiple machine learning approaches including Logistic Regression (LR), Linear Discriminant Analysis (LDA), Classification and Regression Tree (CART), Extreme Gradient Boosting (XGboost), and Light Gradient Boosting Machine (LightGBM) for their predictive capabilities. The LightGBM model, incorporating multi-season remote sensing data and key variables, achieved the highest accuracy, with R-values of 0.95 for warning level classification and 0.6 for bloom area regression. The spatial autocorrelation analysis validated the robustness of our models, demonstrating minimal cross-correlation between training and testing datasets. Furthermore, pixel-level analysis identified the East China Sea as the most bloom-prone region, with consistently higher bloom frequency and magnitude, particularly during summer. Under the historical scenario (incorporating both anthropogenic and natural forcings), we observed higher bloom frequencies and broader area variations compared to scenarios with isolated forcings. Notably, there was a trend toward more frequent yet smaller-scale blooms, with an increase in minor bloom occurrences despite a decrease in extreme events. Critical factors influencing bloom dynamics included sea surface temperature, air temperature, wind speed, sea level pressure, salinity, and nutrient concentrations. Our findings highlight satellite data's importance in understanding anthropogenic-natural factor interactions on coastal blooms, emphasizing the need for targeted nutrient management in vulnerable areas.