Machine Learning-Based Estimation of Clear-Sky Direct Aerosol Radiative Forcing Using Multisensor Satellite Observations

Aerosol forcing remains highly uncertain in climate change assessment. Observation-based forcing estimates have been typically given more weight as they do not rely on assumptions and parameterizations in models. Traditional estimation of direct aerosol radiative forcing (DARF) based on observation is realized by regressing the observed radiative fluxes against aerosol optical depth (AOD). Here we improve this procedure by considering more aerosol parameters such as single scattering albedo (SSA) and adopting a machine learning-based method (eXtreme Gradient Boosting (XGBoost)) to establish the relationship between top-of-atmosphere (TOA) radiative fluxes and aerosol properties. Our approach gives a global mean DARF of −0.80 ± 0.73 W/m² under clear sky, which largely agrees with previously reported results. The DARF uncertainty includes the impact of measurement errors and the XGBoost model, in which AOD and SSA uncertainties cause ∼0.66 and ∼0.14 W/m² changes in the DARF respectively.