PhyGroup-UNet: Dynamic Physically Correlated Channel Grouping in a Lightweight UNet for Efficient Precipitation Nowcasting

Short-term precipitation nowcasting is critical for urban disaster prevention and emergency response, as it directly impacts flood mitigation and public safety. Existing deep learning approaches encounter three core limitations: high model redundancy, underutilization of multimodal physical correlations, and insufficient robustness in complex meteorological conditions. To address these challenges, we propose an enhanced ECA-LC-UNet framework integrating radar and wind speed data. The architecture innovatively embeds physically-correlated channel grouping mechanism in skip connections and develops a wind speed-reflectivity correlation-based dynamic channel grouping strategy, significantly improving the modeling of precipitation cloud boundaries and convective features. A progressive width attenuation dilated convolution scheme is designed to exploit precipitation field spatiotemporal sparsity, achieving 68.6% parameter reduction while maintaining equivalent receptive fields. Experiments on the MeteoNet dataset demonstrate that the proposed model achieves 8% and 10% F1-score improvements over radar-only UNet and optical flow methods, respectively, with a 12.7% enhancement in critical success index for heavy precipitation. Ablation studies reveal that dynamic channel grouping increases cloud-edge feature response by a factor of 1.8, while the lightweight design ensures robust performance under complex weather conditions. This work establishes a high-precision, high-efficiency solution for short-term precipitation prediction and pioneers a new paradigm for multimodal meteorological data-driven modeling.