A feature-enhanced UNet for radar composite reflectivity reconstruction from satellite observations
Abstract
Ground-based radars suffer from limited detection ranges and beam blockages, leaving substantial observation blind spots in mountains and oceans. To address this, this study proposes a Feature-Enhanced UNet (FEU-Net) to reconstruct radar composite reflectivity from wide-coverage Geostationary Operational Environmental Satellite-R (GOES-R) observations. The proposed Feature Enhancement Module (FEM) of FEU-Net integrates dense connections and channel attention mechanisms to improve multi-scale features extraction and fusion for strong radar echoes. Experimental results demonstrate that FEU-Net outperforms the baseline U-Net across various categorical skill scores, especially for strong echoes. Ablation experiments confirm the effectiveness of the FEM design. Additional experiments on FEU-Net further show that adding lightning to infrared inputs reduces false alarms of strong echoes. This method provides a reliable deep learning solution for real-time severe weather monitoring in radar blind spots.