Near-Cloud Aerosol Retrieval Using Machine Learning Techniques, and Implied Direct Radiative Effects.

Abstract

There is a lack of satellite-based aerosol retrievals in the vicinity of low-topped clouds, mainly because reflectance from aerosols is overwhelmed by three-dimensional cloud radiative effects. To account for cloud radiative effects on reflectance observations, we develop a Convolutional Neural Network and retrieve aerosol optical depth (AOD) with 100-500m horizontal resolution for all cloud-free regions regardless of their distances to clouds. The retrieval uncertainty is 0.01+5%AOD, and the mean bias is approximately -2%. In an application to satellite observations, aerosol hygroscopic growth due to humidification near clouds enhances AOD by 100% in regions within 1km of cloud edges. The humidification effect leads to an overall 55% increase in the clear-sky aerosol direct radiative effect. Although this increase is based on a case study, it highlights the importance of aerosol retrievals in near-cloud regions, and the need to incorporate the humidification effect in radiative forcing estimates.