High-Resolution Coastal Carbon Dynamics in the Belgian Part of the North Sea Using Machine Learning

Coastal seas are vital players in the global carbon cycle, acting as both sinks and sources of atmospheric carbon dioxide (CO₂). However, their carbon dynamics remain poorly quantified at the spatial and temporal resolutions necessary for regional carbon budget assessments. Enhanced insights are critical for detecting anthropogenic impacts on the carbon cycle and for monitoring the effectiveness of CO₂ removal strategies. The Belgian Part of the North Sea (BPNS), equipped with advanced monitoring infrastructure, offers a unique platform to address these gaps through in-situ CO₂ measurements collected via buoys, research vessels, and other sources. This data collection provides high spatial and temporal coverage, enabling near-real-time estimation of the exchange of CO₂ with the atmosphere in the region.

To estimate the baseline carbon budget for the BPNS at an unprecedented local scale, we applied a feedforward neural network approach capable of achieving a spatial resolution of 1 km and a temporal resolution of 1 day. This analysis spans the period from 2014 to 2024 and incorporates an extensive dataset of sea surface partial pressure of CO₂ (pCO₂) measurements. These in-situ observations were sourced from the Surface Ocean CO₂ Atlas (SOCAT) and the Integrated Carbon Observation System (ICOS) databases. Additionally, we integrated a suite of predictor variables derived from satellite data and oceanographic reanalysis products, including sea surface temperature, salinity, chlorophyll-a concentrations, and suspended particulate matter, all of which are recognized as key factors influencing pCO₂ variability in the BPNS. By combining this calculated sea surface pCO₂ with atmospheric CO₂, we also estimated the air-sea CO₂ flux.

Initial findings reveal that the sea surface pCO₂ reconstruction achieves strong predictive ability for coastal zones, with an R² exceeding 0.80, successfully capturing both local spatial heterogeneity and seasonal variations. Sensitivity analyses highlight sea surface temperature as the dominant predictor, followed by chlorophyll-a and suspended particulate matter, emphasizing the interplay of thermal and non-thermal processes in shaping pCO₂ variability across the BPNS. The seasonal cycle of pCO₂ decreases after winter primarily due to increased CO₂ solubility in cold water and biological spring uptake, and peaks after summer, mainly driven by warming of the seawater and reduced biological activity, leading to an increased release of CO₂. While sea surface salinity exerts a relatively minor influence overall, its localized impact near the Scheldt estuary plume is significant, underscoring the critical role of riverine inputs in modulating regional carbon dynamics. Overall, our findings indicate that the region functions as a net CO₂ sink for the atmosphere.