Reconstructing radar composite reflectivity from geostationary satellite observations using multiscale deep learning network

Radar reflectivity is vital for monitoring severe convective weather. The limited coverage of ground-based weather radars often leaves mountainous and oceanic regions without effective observations. Geostationary satellites (GEO) provide near-global cloud-top observations, enabling the retrieval of radar reflectivity from satellite data as a practical approach. This study introduces a multi-scale deep learning network (MS-Net) to reconstruct radar reflectivity from GEO satellite observations, delivering reliable reflectivity estimates for areas beyond radar coverage. By extracting fine-grained multi-scale features from satellite cloud imagery, MS-Net enhances nonlinear fitting and retrieval accuracy. Experimental results demonstrate robust performance across key metrics, including Probability of Detection (POD), False Alarm Ratio (FAR), Critical Success Index (CSI) and Mean Absolute Error (MAE). Case studies further validate its efficacy in reconstructing radar reflectivity, offering strong support for severe convective weather monitoring.