Methane (CH₄) emissions from oceans represent a critical yet poorly quantified component of the global carbon cycle, largely due to the challenges in capturing their spatial and temporal variability. This study addresses these challenges by applying a time-series neural network approach to retrieve CH₄ concentrations at the ocean surface using multi-sensor satellite remote sensing data, including ocean color, sea surface temperature, and wind speed from OCCCI, OISST, and NBS datasets. The model integrates oceanic environmental parameters over daily scales, capturing the complex interactions between physical, chemical, and biological processes that influence CH₄ production and emissions. By incorporating these temporal dynamics, including the lag and accumulation effects of these interacting factors, the model aims to enhance the accuracy of CH₄ retrievals in the surface ocean.
An additional innovation in this work is including the quantified CH₄ photo-production as model inputs, which was simulated based on a series of laboratory photo-incubation experiments. Its variability in the global surface ocean was constrained by factors including water temperature, light history, and the characteristics of colored dissolved organic matter (CDOM). Unlike other parameters used by the neural network, this component is fully understood and provides a more transparent and physically grounded understanding of CH₄ dynamics.
Initial tests of the model have yielded a promising coefficient of determination (R²) of 0.93, indicating strong predictive capability for global CH₄ concentration estimates. However, further refinement is needed, including hyperparameter tuning, cross-validation, and in-situ data calibration. This research contributes to a deeper understanding of the spatial and temporal distribution of oceanic CH₄ emissions, offering significant insights into the marine contribution to global methane budgets and its response to climate change.
 

methane emissions, CH₄ photo-production, time-series neural network, CH4