Advanced deep learning technique for estimating global surface ocean calcium carbonate saturation (Ωcal)

Abstract

The accurate estimation of surface ocean calcium carbonate saturation (Ω_cal) is crucial for understanding the impacts of ocean acidification (OA) on marine ecosystems, particularly for calcifying organisms. This study investigates the estimation of global surface ocean Ω_cal using machine learning (ML) models and satellite-derived data. Three ML models such as feed-forward neural networks (FFNN), random forests (RF), and Tabularnet (TabNet) were employed to estimate Ω_cal, utilizing in-situ and satellite measurements of sea surface temperature (SST), sea surface salinity (SSS), and Chlorophyll-a concentration (Chla). Among these, the TabNet model exhibited superior performance, with a root-mean-square error (RMSE) of 0.39, mean relative error (MRE) of 0.019, mean normalized bias (MNB) of 0.0058 and coefficient of determination (R²) of 0.96. SST showed a strong positive correlation with Ω_cal (r = 0.95), while SSS and Chla exhibited moderate positive (r = 0.49) and weak negative (r = −0.27) correlations, respectively. The study revealed significant spatiotemporal variability in Ω_cal, driven by seasonal changes and ocean circulation patterns. Sensitivity analysis highlighted the robustness of the TabNet model, maintaining high predictive capability despite variations in SST, SSS, and Chla. The TabNet model high accuracy provides a valuable tool for monitoring and forecasting changes in ocean chemistry, informing conservation efforts and policy-making. This study emphasizes the importance of advanced ML models in marine science and their potential for enhancing our understanding of global oceanic processes.