Abstract
This research investigates the spatio-temporal dynamics of chlorophyll-a (Chl-a) in the Northern Bay of Bengal, merging satellite-derived data with in-situ measurements from 1990 to 2024. We employed advanced deep learning and machine learning frameworks—LSTM, CNN, ANN, SVM, RF, and ARIMA—to analyze and predict seasonal fluctuations throughout pre-monsoon, monsoon, and post-monsoon periods. During pre-monsoon, Chl-a concentrations were low and erratic (mean ≈ 2.1 mg m−3), primarily due to thermal stratification and limited nutrients. In contrast, the monsoon season witnessed significant increases (mean ≈ 13.6 mg m−3), driven by nutrient influxes and enhanced vertical mixing. Post-monsoon conditions exhibited moderate productivity (mean ≈ 6.2 mg m−3), which was supported by residual nutrients and enhanced light availability. Model validation revealed that LSTM and CNN provided exceptional predictive accuracy (R2 = 0.91–0.94; RMSE <0.04 mg m−3), effectively capturing nonlinear temporal and spatial trends. ANN and RF showed solid regression performance (R2 = 0.88–0.90), while ARIMA accurately captured seasonal periodicity (R2 = 0.87). Although SVM offered moderate accuracy (R2 = 0.82), it was sensitive to variations in data and kernel selection. Overall, the AI-driven framework proved robust and transferable for forecasting estuarine primary productivity and evaluating eutrophication risks in monsoon-affected coastal ecosystems. The findings underscore the pivotal role of monsoons in supporting phytoplankton productivity and highlight the vulnerability of pre-monsoon conditions under warming-induced stratification. This study aligns with UN SDGs 6 (Clean Water and Sanitation), 13 (Climate Action), 14 (Life Below Water), and 17 (Partnerships for the Goals) by promoting open-access data integration, collaborative AI methodologies, and climate-resilient monitoring initiatives for sustainable coastal management.
