PEDNet: A Predictive Encoder–Decoder Network with Multi-Scale Global–Local Modeling for Radar Precipitation Nowcasting

Radar precipitation nowcasting remains challenging because a model must not only represent the overall motion trends of large-scale precipitation systems, but also capture the fine-grained structural variations of localized strong echo regions while maintaining stable temporal evolution in multi-step forecasting. To address this issue, this paper proposes PEDNet, a predictive encoder–decoder network for radar precipitation nowcasting, and evaluates its performance under a unified 6-to-12 frame forecasting setting. The proposed framework jointly models global contextual perception, local structural refinement, and temporal dependencies within a unified architecture. Specifically, the designed multi-scale global–local spatial modeling strategy is used to simultaneously capture large-scale precipitation organization patterns and local echo details, while the temporal modeling module introduced at the bottleneck stage enhances sequence representation across multiple future lead times. Experimental results on the KNMI radar dataset and the SEVIR VIL benchmark dataset show that PEDNet achieves competitive overall performance across multiple categorical and continuous metrics under the adopted evaluation protocol. Meanwhile, the model maintains a practical computational cost, with 14.84 M parameters, 15.63 G FLOPs, and an inference throughput of 2.12 complete forecast samples per second. These results indicate that PEDNet provides a competitive balance between predictive accuracy and computational efficiency for short-term radar precipitation nowcasting.