Deep learning prediction of subsurface chlorophyll-a and analysis of basin-specific surface drivers in the Northern Indian Ocean using multi-source satellite and reanalysis data

The Northern Indian Ocean (NIO), comprising Arabian Sea (AS) and Bay of Bengal (BoB), is ecologically important, with primary productivity assessments crucial for fisheries and coastal livelihoods. Persistent gaps in subsurface observations limit productivity assessments in the NIO. Inferring subsurface chlorophyll-a (chl-a), a key proxy for primary productivity from surface observations offers a practical pathway to overcome this. This study reconstructs subsurface chl-a profiles using exclusively multi-source surface satellite observations combined with ocean–atmosphere reanalysis data. Existing regional studies in NIO rely mainly on subsurface information alongside surface predictors and therefore depend heavily on in-situ subsurface measurements. We employ a Bi-GRU framework to predict subsurface chl-a, linking surface and subsurface, and enabling estimation where in-situ observations are sparse but surface information is available. Models were trained using chl-a measurements within the upper 200 m obtained from BGC-Argo profiles. We incorporate a comprehensive set of surface predictors relevant to subsurface chl-a variability. The models demonstrated strong predictive skill, achieving RMSEs of 0.1067, 0.1097 and 0.0928 mg m−3 for the NIO, AS and BoB, respectively, with corresponding correlations of 0.8469, 0.859 and 0.7579. Seasonal analysis showed higher skill during stable, stratified periods, with slightly reduced performance during the AS summer monsoon and BoB fall inter-monsoon, when enhanced mixing weakens surface–subsurface coupling. Feature-importance analysis provided a quantitative assessment of the relative influence of surface drivers, highlighting biological–thermal dominance in the AS (surface chl-a and SST) and stronger SLA-driven physical control in the BoB with biological–thermal drivers playing a comparatively smaller role.