Evaluating Machine Learning Model Performance in Predicting Seasonal Sea Surface Salinity Across Regions of Differing Variability

This study develops a machine learning framework with an objective to predict monthly Sea Surface Salinity (SSS) between 2011-2024 in ocean regions with distinctly different variability characteristics. SSS affects ocean circulation, heat storage, global hydrologic cycle, marine ecosystems and coastal hazards. Thus, accurate prediction of SSS can be important for both scientific and operational decision-making. The study defined three standardized 3°×3° variability regimes – low, medium, and high, to explain regional differences in the behavior of SSS. Four models are evaluated (ElasticNet, Random Forest, SVR, and XGBoost) using satellitederived atmospheric, hydrological, and oceanographic predictors, including lagged terms that capture ocean memory. It was found that the predictability of SSS is very regime-specific. Random Forest performed best in the low-variability subtropical regions, whereas SVR performed best in freshwater-influenced regions such as the Amazon plume. Medium variability regions were a challenge to model effectively due to weak intrinsic signals, causing climatology to outperform ML models. The results provide actionable guidance for aquaculture and climate monitoring. The region-tailored SSS forecasts can support aquaculture site selection, stocking, and risk management. It can also assist agencies in climate monitoring and coastal hazard preparedness.