Improving radar echo reconstruction via deep learning of Himawari-8 and Goes-R

While weather radar systems remain indispensable for meteorological forecasting and atmospheric studies, their operational efficacy is constrained by geographical obstructions and restricted spatial coverage, leading to suboptimal characterization of localized climatic dynamics. To address these limitations, this study introduces a novel multispectral data integration framework leveraging the Swin-MoE-GAN-UNet (SMG-UNet) architecture for radar data reconstruction using synergistic observations from Himawari-8 and GOES-R geostationary satellites. The proposed methodology specifically resolves spectral discrepancies arising from heterogeneous echo wavelengths and multivariate data complexity through a dual-path feature alignment mechanism embedded within the hybrid transformer-CNN network. Experimental validation across diverse meteorological regimes demonstrates that the multimodal fusion approach enhances model generalizability for underrepresented climate patterns. Quantitative evaluations reveal statistically significant improvements, with a 14.7% reduction in root mean square error (RMSE) and 19.3% lower false alarm rate (FAR) relative to single-source baselines, accompanied by 22.1% and 17.9% enhancements in critical success index (CSI) and probability of detection (POD), respectively. This advancement establishes a robust paradigm for satellite-radar synergy in high-resolution atmospheric modeling.