Inferring Vertical Cloud Structure From Passive Top‐of‐Atmosphere Observations and Environmental Context With a Mixture Density Network

Abstract

Passive satellite observations provide global coverage of cloud‐top and column‐integrated properties but offer limited direct information about vertical cloud structure. Here, we evaluate the extent to which passive cloud‐top, optical, and microphysical retrievals, together with large‐scale environmental variables, constrain CloudSat radar reflectivity profiles. We first establish a cloud‐type‐based baseline using cloud‐top pressure and cloud optical depth, demonstrating that substantial vertical variability remains within fixed cloud types. We then introduce a probabilistic Mixture Density Network that predicts the full conditional distribution of reflectivity profiles from Moderate Resolution Imaging Spectroradiometer cloud‐top, optical, and microphysical properties and European Center for Medium‐Range Weather Forecasts environmental fields. The model reduces both level‐by‐level and joint structural uncertainty relative to cloud‐type compositing, with test correlations between observed and predicted mean profiles ranging from 0.79 to 0.94 across cloud types and mean continuous ranked probability score reductions ranging from 38.17% in shallow cumulus to 67.07% in deep convective clouds. However, non‐negligible uncertainty persists, reflecting residual ambiguity in mapping top‐of‐atmosphere observations and environmental context to full vertical radar‐observed structure. These results quantify the predictive constraints provided by the selected passive cloud properties and environmental variables, while showing that substantial ambiguity remains when inferring vertical cloud structure from top‐of‐atmosphere observations.