Climate-driven controls on greenhouse gas emissions from global reservoirs inferred from satellite observations and machine learning

Reservoirs are increasingly recognised as dynamic components of the global carbon cycle. Yet, their greenhouse gas (GHG) emissions are still poorly understood due to strong spatiotemporal variations, seasonality and the scarcity of in-situ measurements. Climate-driven variability in thermal conditions, hydrodynamics and reservoir morphology is expected to control both the magnitude and temporal variability of carbon dioxide (CO₂) as well as methane (CH₄) emissions. However, these controls remain poorly understood at the global scale.

Here, we combine satellite observations and machine-learning models to examine climate-related patterns in reservoir GHG emissions across more than 21,000 reservoirs globally from 2020 to 2024. Average CO₂ and CH₄ emissions on a monthly scale are obtained by combining GHG concentration-based observation from the Greenhouse Gases Observing Satellite (GOSAT) with climate reanalysis data (ERA5) and relevant reservoir information such as surface area or catchment area. We employ tree-based ensembles of models to estimate monthly emissions and explore how emissions vary with season, location and reservoir characteristics among different hydroclimatic regions.

The resulting emission estimates exhibit clear global seasonal variations and show a strong seasonal phasing, with most emissions peaking during local seasonal extremes. Seasonal emissions show less variation in larger reservoirs while the smaller reservoirs show greater seasonal changes because they are strongly influenced by climate forcing and have less ability to moderate variability. Spatial aggregation reveals strong zonal differences and nonlinear relationships with thermal regimes, highlighting the complex interplay between climate variability and physical characteristics of the reservoir on GHG emissions regulation.

Together, these findings show that machine learning models using satellite-derived information can reveal physically consistent spatiotemporal patterns in reservoir GHG emissions at global scales. While comprehensive site-scale validation remains limited at the global scale, the observed consistency across temporal, spatial and physical characteristics reconfirms that satellite-enabled modelling could be useful to assess climate-driven variability in inland-water carbon emissions at larger scales and guide focused future observational efforts.