CNN-Based Retrieval of 3D Cloud Structures Solely From Geostationary Satellite Imagery

The Cloud vertical structure (CVS) is important, yet operational CVS products depend on active observation or reanalysis fields, limiting high-frequency monitoring. In this study, we propose a lightweight and satellite-only model that reconstructs volumetric cloud masks from geostationary multispectral imagery. This method employs a compact one-dimensional convolutional neural network that combines three convolutional layers, channel attention and L1 regularization, which is trained on CALIPSO/CloudSat joint profiles and Himawari-8 multispectral observations. The network produces per-pixel 38-layer cloud masks at 500 m vertical resolution and attains strong performance (Intersection over Union = 0.8609; mean absolute error = 0.4851; cloud top height bias ≈ 470 m). Ablation experiments demonstrate that the chosen architecture and regularization considerably improve layer discrimination. A case study of Typhoon Yutu shows that the reconstructed three-dimensional structure is consistent with active-sensor profiles. Because this approach only requires geostationary imagery (without reanalysis input), it is inherently well-suited for near-real-time monitoring.