Estimating surface pCO2 and its long-term variability in the East China Sea using machine learning with satellite data
Abstract
Accurate estimation of sea surface partial pressure of CO 2 ( p CO 2 ) in optically complex coastal waters remains challenging due to significant biogeochemical heterogeneity. This study develops a machine learning approach to reconstruct p CO 2 in the East China Sea (ECS) from 2003 to 2023, utilizing MODIS-derived optical classification. Using normalized water-leaving radiance at 555 nm (nL w 555) as a threshold (1.5 mW cm -2 µ m -1 sr -1 ), the ECS was classified into two distinct regimes: Clear Water (CW) and Turbid Water (TW). Then, the p CO 2 was reconstructed using a Categorical Boosting (CatBoost) model that integrates multi-spectral satellite variables as numerical predictors with an optically-derived Water Case as a categorical input feature to account for heterogeneity. Evaluated on the test dataset, this approach achieved high overall accuracy ( R 2 = 0.86, RMSE = 16.55 µ atm), notably reducing prediction errors (RMSE) in TW by up to ~81% compared to conventional models. The model maintained robust regime-level generalization (CW: R 2 = 0.82, RMSE = 15.64 µ atm; TW: R 2 = 0.82, RMSE = 22.01 µ atm). The 21-year climatology reveals distinct mean p CO 2 levels for CW (361.2 µ atm) and TW (415.1 µ atm). While CW exhibits strong summer biological drawdown, TW functions as a highly variable net CO 2 source driven by positive non-thermal anomalies (+28.09 µ atm). Over the 21-year period, the atmospheric CO 2 growth rate (+2.41 µ atm yr -1 ) significantly outpaced the surface p CO 2 increases in both regimes (CW: +0.66 µ atm yr -1 ; TW: +1.11 µ atm yr -1 ). This growing thermodynamic gradient overwhelmingly drives an increasingly negative flux anomaly (−0.182 mmol m -2 d -1 yr -1 ), statistically intensifying the CW carbon sink. Crucially, recognizing the TW as a weak net source (+0.50 Tg C yr -1 ) rather than a sink reveals that conventional models overestimate ECS carbon uptake by ~46% (~4.04 Tg C yr -1 ). Ultimately, this study provides a highly refined long-term carbon budget for the ECS, demonstrating that accounting for spatial heterogeneity is key to accurately assessing coastal carbon dynamics.