Multi-scale U-Net with Kolmogorov-Arnold network for joint retrieval of cloud properties

Clouds play a pivotal role in atmospheric radiation balance and climate change. However, the accurate joint retrieval of cloud optical and microphysical properties from geostationary satellite thermal infrared observations remains challenging due to their complex multi-scale structures and nonlinear radiative interactions. We propose a multi-scale U-Net with Kolmogorov-Arnold network (MSKAN-UNet) for joint retrieval of cloud mask, cloud optical thickness (COT), cloud effective radius (CER), and cloud top height (CTH) from Himawari-8/Advanced Himawari Imager (AHI) infrared channel observations, with MODIS Level-2 products serving as training labels. The proposed framework integrates an Adaptive Spatial Multi-scale Enhancement (ASME) module in the shallow layers to strengthen multi-scale feature representation and a KAN Adaptive Residual (KAR) module in the deep layers to model high-order nonlinear interactions among cloud properties. Extensive experiments demonstrate that MSKAN-UNet outperforms comparative methods across all retrieval tasks. Compared to AHI operational products, MSKAN-UNet reduces RMSE from 19.35 to 14.28 for COT, from 10.84 to 8.66 µm for CER, and from 2.80 to 1.93 km for CTH, while cloud mask OA improves substantially from 0.693 to 0.888. These findings demonstrate the efficacy of the proposed method for accurate and continuous cloud property retrieval from geostationary satellite data.