Deep Learning Models to Identify Seasonal Drivers of Chlorophyll Changes in the Atlantic Ocean

Understanding the seasonal dynamics of plankton in the Atlantic Ocean is the first step towards the proper assessment of marine ecosystem health and productivity. Ocean colour and surface chlorophyll (chl-a) distribution serve as proxies for phytoplankton biomass, providing insights into marine food web dynamics and biogeochemical cycles. This study examines the response of the total chlorophyll concentration to physical drivers observable by remote sensing in the Atlantic Ocean using a combination of multivariate Principal Component Analysis (PCA) and deep learning models. The results show that the Sea Surface Salinity (SSS), Absolute Dynamic Topography (ADT), and Sea Surface Temperature (SST) are found to be the predominant drivers of physical variability across the ocean, with distinct spatial patterns. The clustering of the principal components identifies regions characterised by distinct physical processes. Based on these clusters, we devised a Transformer Encoder model to predict chl-a concentrations in three distinct regions. The model outperformed climatological baselines, especially in the temperate and tropical regions, though accuracy varied seasonally, with higher accuracy in winter months and increased complexity in summer due to more dynamic oceanographic conditions. A SHAP-based sensitivity analysis showed that ADT and SSS dominate chl-a variability, particularly during summer months, while SST and wind stress also contribute significantly during transitional periods. The study highlights the necessity to account for both seasonal and regional differences in predictive modelling, and it underscores the importance of continuing to develop spatio-temporal models to improve forecasting accuracy for marine ecosystem management and conservation.