Towards Goes Retrievals of Near-Cloud Aerosols: Machine Learning Development and Assessment of Radiative Effects

June 29, 2026

Olivia Pierpaoli

Committee: Christine Chiu (Advisor); Steven Miller; Haonan Chen (Electrical and Computer Engineering)

Abstract

Atmospheric aerosols play a central role in Earth’s radiation budget and hydrological cycle, yet they remain a major source of uncertainty in estimates of anthropogenic climate forcing. Passive satellite observations have become a primary means of monitoring and characterizing aerosols. However, their applicability is limited in near-cloud regions. Current operational retrieval methods exclude these pixels because reflectance at the top of the atmosphere is strongly influenced by three-dimensional (3D) radiative effects from adjacent clouds, with enhancements often exceeding 50% within several cloud-edge diameters.

This limitation would be less consequential for radiative forcing estimates if aerosol properties near clouds were similar to those in far-from-cloud regions, where retrievals are generally more reliable. However, extensive observational evidence indicates that near-cloud aerosol populations are physically and optically distinct. Strong humidity gradients near cloud edges promote hygroscopic growth, increasing particle size and scattering efficiency over kilometer scales. In addition, interactions with clouds alters aerosol size distributions through activation, collection, and evaporation, producing larger and more internally mixed particles. As a result, excluding near-cloud pixels not only biases radiative forcing estimates toward clear-sky conditions, but also represents a missed opportunity to characterize aerosol properties that encode detailed aerosol processes and aerosol–cloud interactions.

To address this limitation, we develop a machine learning method that jointly retrieves aerosol optical depth (AOD) and particle size from multi-channel reflectance observations in near-cloud environments, enabling us to disentangle aerosol number concentration and size and investigate their underlying processes. The retrieval model is trained using diverse cloud fields from large-eddy-simulation and 3D radiative transfer calculations, explicitly accounting for cloud-induced radiative effects and hygroscopic growth. Based on test data, the retrieval uncertainties are approximately 0.002 plus 3.9% of AOD for aerosol optical depth and 0.021 µm plus 1.1% of effective radius for aerosol size. These uncertainties meet the accuracy targets defined for NASA missions. In particular, the requirement for aerosol effective radius is demanding, and the method developed here achieves performance comparable to polarization-based retrievals.

Consistent with the previous studies, near-cloud aerosols substantially modify estimates of the aerosol direct radiative effects, enhancing both shortwave cooling and longwave warming. With decreasing distance to clouds, the shortwave radiative effect increases from approximately −5 W m–2 in far-from-cloud regions to about −17 W m–2 near clouds, while the longwave effect increases from around 0.2 to 0.7 W m–2. This 0.5 W m–2 increase in longwave effect is comparable to that produced by 40 ppm of additional atmospheric CO2, assuming a baseline concentration of 425 ppm. Overall, both shortwave and longwave effects are amplified by roughly a factor of 2.5, underscoring the importance of including near-cloud aerosols in radiative forcing assessments to improve the accuracy of current estimates.

The newly developed retrieval model will be applied to GOES-16 observations for a case study during the Elucidating the Role of Clouds-Circulation Coupling in Climate (EUREC4A) campaign, leveraging aircraft measurements for independent evaluation of the retrieval. Although the impact of parallax shifts associated with off-nadir viewing angles has been minimized, preliminary results indicate that the primary limitation for improving GOES aerosol retrievals is the relatively coarse spatial resolution (~1 km). At this scale, sub-grid clouds are likely present within individual pixels, making their effects difficult to account for. Future work will focus on alleviating this limitation by incorporating high-resolution simulations and additional collocated observations.