A multiscale dynamic network-based method for radar precipitation nowcasting: A case study in the Yiluo River Basin

Study region Yiluo River Basin, China Study focus Accurate and timely precipitation nowcasting is essential for mitigating rainstorms and floods. This study proposes a Multiscale Dynamic Nowcasting Network (MDN-Net) that integrates temporal and motion features for radar echo extrapolation and quantitative precipitation estimation. A four-dimensional evaluation framework is also developed to comprehensively assess nowcasting performance from multiple perspectives. New hydrological insights for the region (1) MDN-Net significantly improves strong echo extrapolation and outperforms the other models at lead times of 1.5–2 h. (2) For quantitative precipitation estimation, MDN-Net accurately captures the observed temporal evolution of rainfall, although cumulative precipitation is slightly overestimated. It achieves the smallest RBIAS magnitude (-2.05% to −3.51%), indicating superior robustness and generalization capability. (3) Module analysis demonstrates that MDN-Net with the feature fusion module maintains stable performance across thresholds and lead times. Specifically, F1-score increases by 44.3% and 35.1% at 1.5 h and 2 h lead times compared with U-Net and Flow U-Net, respectively. Overall, by integrating temporal sequence and motion features, MDN-Net provides a more comprehensive and accurate representation of precipitation systems, significantly enhances spatiotemporal perception, and improves nowcasting performance, offering a promising framework for enhancing the accuracy and reliability of precipitation nowcasting and supporting disaster prevention and water resource management.