ABSTRACT Cloud horizontal heterogeneity within instrument footprints constitutes a major error source in ice cloud retrievals. We present a novel algorithm considering cloud horizontal heterogeneity based on lightweight convolutional neural network (LCNN) designed for Ice Cloud Imagers (ICI), which retrieves high‐resolution ice water path (IWP) by leveraging spatial correlations in low‐resolution satellite observations. Our approach employs a prior database generated through three‐dimensional radiative transfer simulations, containing collocated high‐resolution atmospheric profiles and cloud parameters, as well as the corresponding brightness temperature differences (BTDs). Then, the LCNN is trained and tested by the simulated prior database. Compared to Bayesian Monte Carlo integration (BMCI), the LCNN reduces mean absolute error (MAE) by 12.05 g/m 2 while tripling spatial resolution (from 16 to 5.3 km). This resolution enhancement mitigates errors induced by ice cloud horizontal heterogeneity, achieving a 23.6% reduction in MAE attributing to plane‐parallel cloud assumptions within footprints. The framework demonstrates the viability of deep learning for submillimeter‐wave cloud sensing, providing a pathway to resolve sub‐footprint cloud variability in future satellite missions.
Zhuo et al. (Sun,) studied this question.