Splitting the East China Sea into clear and turbid regimes by nLw555, a CatBoost reconstruction finds the turbid zone is a weak CO2 source and that conventional models overestimate the ECS sink by about 46%

Converting nLw555 from a continuous predictor into a categorical water-type input lets a single model learn regime-specific relationships, sharply reducing prediction error in turbid water. Earlier satellite pCO2 approaches (empirical regressions, MeSAA, and machine learning) mostly treated optical proxies as continuous numerical inputs without explicitly partitioning the domain; this study used the bimodal nLw555 distribution to set a 1.5 mW cm-2 µm-1 sr-1 threshold and fed water type as a categorical feature into CatBoost. Independent test set: overall R2 = 0.86, RMSE = 16.55 µatm; regime-level CW R2 = 0.82, RMSE = 15.64 µatm and TW R2 = 0.82, RMSE = 22.01 µatm; relative TW RMSE reductions of about 81%, 65%, and 52% versus Liu et al. (2023b), Yu et al. (2023), and Chen et al. (2016); removing the water-type variable raised TW test RMSE from 22.01 to 26.62 µatm.

The 21-year reconstruction shows two carbon systems with opposing controls: CW is dominated by summer biological drawdown, while TW is non-thermally driven and acts as a weak net source. Prior understanding of the ECS rested largely on short observational windows or unpartitioned unified models that could not separate drivers of opposite sign; this study provides a partitioned 2003–2023 climatology with thermal/non-thermal decomposition. Annual mean non-thermal anomalies of -23.15 µatm in CW and +28.09 µatm in TW (Cohen's d = -0.75, p < 0.001); TW flux means of 8.55 ± 6.43 mmol m-2 d-1 in summer and 7.40 ± 4.07 mmol m-2 d-1 in autumn, switching to a sink in winter (-3.96 ± 3.22 mmol m-2 d-1); annual TW net source of +0.40 ± 6.66 mmol m-2 d-1.

Atmospheric CO2 grew faster than surface pCO2 in both regimes, widening the thermodynamic gradient and significantly strengthening the CW sink while steadily weakening the TW source. The study attributes long-term flux trends to atmospheric forcing rather than wind change and reports regime-specific trend slopes with significance tests. Atmospheric CO2 rose +2.41 µatm yr-1 versus +0.66 µatm yr-1 in CW and +1.11 µatm yr-1 in TW (TW trend p = 0.06); CW flux anomaly trend -0.182 mmol m-2 d-1 yr-1 (Mann-Kendall Z = -5.405, p < 0.001) and TW -0.166 mmol m-2 d-1 yr-1 (Z = -4.460, p < 0.001); CW wind speed anomaly trend -0.0004 m s-1 yr-1 (p = 0.976).

Identifying the turbid zone as a weak source rather than a sink substantially changes the regional carbon budget of the ECS. Compared with the framework of Yu et al. (2023), conventional approaches misclassify the turbid nearshore source as a strong sink, inflating shelf-scale uptake. Yu et al. (2023) estimated a strong TW sink of -5.50 mmol m-2 d-1, whereas this study obtained a weak source of +0.78 mmol m-2 d-1, a difference of about 4.04 Tg C yr-1; total ECS uptake here is 8.05 Tg C yr-1 versus 14.80 Tg C yr-1 in Yu et al. (2023), about 46% lower; a 10,000-iteration Monte Carlo gave 8.64 ± 0.92 Tg C yr-1 and a TW budget of +0.63 ± 0.06 Tg C yr-1, with the 95% upper bound of 10.44 Tg C yr-1 still below 14.80 Tg C yr-1.

The framework targets river-influenced nearshore shelves with strong optical contrasts, and applies to reconstructing monthly surface pCO2 and air-sea CO2 fluxes from ocean-colour satellite data such as MODIS, serving regional carbon accounting, shelf carbon-cycle mechanism studies, and nearshore monitoring design. The authors note that nLw555 has already proven useful as an optical proxy for turbid river plumes beyond the ECS, such as off the Oregon and Washington coasts, and that observations in the Mississippi River and Amazon mouth indicate high turbidity and terrestrial organic input create biogeochemical regimes distinct from adjacent clear-water shelves, so the regime-aware approach could extend to other river-dominated margins; they also state that the framework was developed and validated solely within the ECS, so global transferability remains a hypothesis requiring in-situ data from diverse marine systems.

The surface pCO2 increase in the turbid regime did not reach statistical significance over 2003–2023 (p = 0.06), which the authors attribute to the large spatiotemporal variability of that zone, so the oceanic trend alone warrants caution while the long-term flux trend is argued to be robust because it is driven by atmospheric CO2 forcing. The model underestimates extreme pCO2 values above 500 µatm in the turbid regime, and this bias appears consistently across training, validation, and test sets, which the authors read as a generalized response to sparse extreme in-situ observations rather than overfitting, implying the turbid zone's future role as a source may be larger than currently estimated. Reconstructing the 21-year climatology partly relies on reanalysis salinity products, and the 0.1° (about 11.1 km) spatial block is slightly smaller than the 14.3 km decorrelation length, though the authors report sensitivity analyses showing minimal practical impact. In addition, the text available here is an incomplete version that does not include the figures or supplementary materials (such as Supplementary Table S1, S2 and Supplementary Figure S1–S3), so specific values for threshold sensitivity, hyperparameter settings, and ablation details rest on the main-text narrative and could not be checked item by item.