The sea surface total alkalinity (AT) and total dissolved inorganic carbon (CT) are two essential carbonate variables to understand the marine carbon cycle, yet it has been a big challenge to retrieve AT and CT from space for the Yellow and East China Seas (YECS) owing to they are affected coupling by physical and biogeochemical processes and the heterogeneous coastal environments. To address these challenges, we developed multilayer perceptron neural network (MPNN)‐based AT and CT models with the field measured environmental variables as the model predictors, and obtained a root mean square difference (RMSD) of 27.05 μmol/kg and coefficient of determination (R²) of 0.91 for AT (N = 1,520) and a RMSD was 28.31 μmol/kg and R² was 0.88 for CT (N = 513). Further, the MPNN‐based model showed much promise in remotely retrieving surface AT and CT with the spatial resolution of ∼1 km in the YECS with a RMSD of 26.59 μmol/kg, R² of 0.76 for AT and a RMSD of 37.14 μmol/kg, R² of 0.79 for CT. Applying the MPNN‐based model to the Moderate Resolution Imaging Spectroradiometer (MODIS) products, retrieved the monthly distributions of AT and CT over the past 20 years for the first time, demonstrated strong linkages to water masses circulations, upwelling and biological processes with seasonal cycles. Also, the interannual variations of AT and CT had significant relationships with the environmental proxies, as well as climate indices (North Pacific Gyre Oscillation). This work advances understanding of coastal carbon cycling and offers a valuable tool for large‐scale, high spatial‐temporal resolution monitoring of carbon dynamics.
Plain Language Summary
Understanding the coastal carbon cycle is critical for addressing climate change, but continuously measuring key carbonate variables total alkalinity (AT) and total dissolved inorganic carbon (CT) in dynamic coastal regions like the Yellow and East China Seas (YECS) has been challenging. To overcome this challenge, we developed a machine learning based AT and CT models based on extensive field data set with uncertainties of 27.05 μmol/kg for AT and 28.31 μmol/kg for CT, demonstrated that the environmental variables, such as sea surface temperature, salinity and chlorophyll a allow for retrieving AT and CT. Further validation showed that the model has much promise in retrieving surface AT and CT from space. Applying the model to satellite products, this work provides the monthly distributions and interannual dynamics of AT and CT in the YECS over the past 20 years for the first time. We found the physical and biogeochemical processes shape the monthly distribution of AT and CT in the YECS, and its interannual variations are strongly correlated with environmental proxies, as well as climate forcings. The integration of satellite-derived products and machine learning based approach enables cost-effective, high-resolution monitoring of coastal carbonate system dynamics.
Key Points
- A machine learning-based model was developed to retrieve the surface carbonate variables from space for the Yellow and East China Seas
- The multilayer perceptron neural network-based model has high tolerance to these uncertainties in satellite-derived environmental proxies
- The interannual and seasonal variations of carbonate variables are affected by coupling physical and biogeochemical processes
Liu J., Zhu Q., Bellerby R. G. & Liu J., 2025. Remote estimations of total alkalinity and total dissolved inorganic carbon in the Yellow and East China Seas using machine learning approach. Journal of Geophysical Research: Oceans 130(8): e2025JC022708. doi: 10.1029/2025JC022708. Article (subscription required).
