Validating Machine Learning Retrievals of the Cloud‐Top Droplet Effective Radius Over Oceans That Account for 3D Radiative Transfer Effects

Abstract Bispectral retrievals of the droplet effective radius ( r e ) from instruments such as MODIS are widely utilized to study cloud microphysics in marine boundary layer clouds. These retrievals are known to have systematic errors due to cloud heterogeneity. Here, we develop a neural network regression to retrieve cloud‐top r e at a solar zenith angle of 30° and nadir viewing using MODIS. The neural network regression is trained on 3D radiative transfer simulations of quasi‐adiabatic stochastically generated clouds and corrects relative errors in r e with respect to cloud‐top with an r 2 of 0.88. The neural network regression produces unbiased retrievals of cloud‐top r e against large eddy simulation cloud fields where the bispectral retrieval has biases reaching +100%. The neural network regression reduces retrieval biases against airborne observations of cumulus from CAMP 2 Ex from +100% to +40%, and marginally improves already good consistency against stratocumulus sampled during VOCALS. A cross‐comparison technique for assessing statistical remote sensing retrievals is introduced. The neural network regression explains 63% and 91% of the variance in the differences between MODIS 1.6 and 2.1 μm retrievals for Overcast and partially cloudy pixels (PCL) and 42% and 76% for the 2.1 and 3.7 μm differences, respectively. Residual spectral inconsistency is partially attributed to precipitation‐sized particles using radar observations. Regional averages of the operational MODIS r e exceed the cloud‐top r e predicted by the neural network by +50% for Overcast pixels in the tropics and a consistent +70% for PCL pixels. Errors in bispectral retrievals due to heterogeneity are nonrandom at both the cloud and climate scale.