SA-DGNet: Precipitation Nowcasting with Attention Mechanism and dual-path residual

Precipitation nowcasting is crucial for disaster prevention and mitigation. This paper proposes a novel encoder-decoder model—Spatial Attention and Dual-Path Gated Network (SA-DGNet) , which better identifies precipitation regions and estimates intensity. First, we introduce the Bidirectional Path Residual Block (BPRB), which integrates standard and dilated convolutions to capture multi-scale structural evolution features of convective precipitation systems, addressing a key limitation of existing models in representing their formation, dissipation dynamics, and shape deformation. Furthermore, the Local Window Self-Attention Block (LWSA) uses window-based self-attention with relative position bias to preserve local precipitation details while modeling cross-region spatial dependencies. It mitigates the high computational cost of global self-attention on high-resolution radar data and balances local detail with global context. In addition, we address the issues of low-level spatial detail loss, blurred precipitation prediction boundaries, and morphological distortion caused by traditional skip connections in UNet architectures through a Gated Skip Attention Block (GSA). We evaluated SA-DGNet against baselines on radar precipitation maps from the Netherlands, it achieved a 6.6\% improvement in critical success index (CSI) over UNet on the NL-50 dataset, reaching 0.7427.