MultiPredNet: Multi-Source Fusion and Multi-Objective Learning for Radar Echo Extrapolation

Abstract

Radar echo extrapolation is essential for weather nowcasting and early warning of severe convective events. However, existing deep learning approaches mainly rely on radar observations and insufficiently exploit environmental information, limiting their capability to capture convective initiation and evolution under complex meteorological conditions. This paper proposes MultiPredNet, a multi-source environmental fusion network for radar echo extrapolation. The proposed framework adopts a multi-branch encoder to extract and integrate radar temporal features, meteorological variables, and static topographic information. A residual ConvLSTM decoder is further developed to progressively predict future radar echoes, facilitating feature propagation and maintaining spatiotemporal continuity. In addition, a composite loss function combining intensity reconstruction, structural consistency, and precipitation discrimination objectives is introduced to improve prediction performance across different echo intensities. Experiments conducted on a radar dataset from Shanxi Province demonstrate that MultiPredNet achieves CSI values of 0.702, 0.621, and 0.522 at thresholds of 10, 20, and 30 dBZ, respectively, outperforming representative baseline models. Visualization results and ablation studies further verify the effectiveness of multi-source fusion, residual decoding, and composite optimization. The proposed framework provides an effective approach for improving radar echo extrapolation and supporting severe weather nowcasting.