Beyond Blurriness and Artifacts: A Synergistic Deterministic-Probabilistic Approach for Radar Reconstruction

Jun CHEN, Lin CHEN, Yan Fang, Minghui Qiu, SHUXIN ZHONG, Yu ZHANG, Binghong CHEN, Kaishun Wu

Abstract: Synthesizing high-fidelity radar reflectivity from satellite observations is crucial for weather forecasting and climate hazard tracking, particularly in regions with sparse radar coverage such as mountainous and oceanic areas. Existing methods either ensure global consistency but over-smooth details, or capture details while deviating from the true distribution. To address this challenge, we propose \N\, a new cascaded framework consisting of deterministic and probabilistic components. The framework operates through the synergy of two meticulously designed modules: a \textit{Physics-Constrained Structural Extractor}, which captures the macro-scale structure of the radar field that conforms to physical statistical distributions; and a \textit{Flow-based Distribution Adaptor}, precisely maps the initial macroscopic result to the fine-grained distribution of radar observations. Comprehensive experiments on two large-scale, real-world datasets demonstrate that \N\ significantly outperforms existing baseline models, not only improve the CSI$_{50}$ by over 16\% but also achieving a breakthrough in key structural fidelity metrics.