Scholarlink

<p>Short-term rainfall nowcasting using weather radar plays a critical role in early warning systems for natural disasters, particularly in regions with complex terrain such as Vietnam. However, the existing models still face limitations in preserving fine-scale structures, modeling complex motion patterns, and maintaining temporal consistency. In this study, we propose S2R-GAN++, a spatiotemporal GAN based on a sequence-to-sequence framework for radar nowcasting. Compared to its predecessor, the proposed model incorporates three key components: (1) a Multi-Branch Intensity Separation architecture combined with a 3D multi-scale feature extractor (MFF3D) to enhance intensity-aware representation, (2) an Adaptive Gating Fusion mechanism for effective feature selection and integration, and (3) a 3D Temporal Attention module to improve temporal dependencies across frames. Various experiments conducted on two radar datasets, PhaDin (Vietnam) and DWD (Germany), demonstrate that S2R-GAN++ consistently outperforms baseline methods. Specifically, the model reduces average MSE by approximately 20.3% on PhaDin and 43.6% on DWD compared to S2R-GAN, while also achieving notable improvements in SSIM, indicating superior preservation of spatial structures. These results demonstrate that S2R-GAN++ not only improves prediction accuracy but also enhances temporal stability and generalization capability across different meteorological scenarios.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A novel spatiotemporal GAN for radar nowcasting with multi-branch intensity separation and 3D temporal attention

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月29日

查看全文
  • Ha Gia Son,
  • Hoang Duc Trung,
  • Tran Manh Tuan,
  • Le Hoang Son

摘要

Short-term rainfall nowcasting using weather radar plays a critical role in early warning systems for natural disasters, particularly in regions with complex terrain such as Vietnam. However, the existing models still face limitations in preserving fine-scale structures, modeling complex motion patterns, and maintaining temporal consistency. In this study, we propose S2R-GAN++, a spatiotemporal GAN based on a sequence-to-sequence framework for radar nowcasting. Compared to its predecessor, the proposed model incorporates three key components: (1) a Multi-Branch Intensity Separation architecture combined with a 3D multi-scale feature extractor (MFF3D) to enhance intensity-aware representation, (2) an Adaptive Gating Fusion mechanism for effective feature selection and integration, and (3) a 3D Temporal Attention module to improve temporal dependencies across frames. Various experiments conducted on two radar datasets, PhaDin (Vietnam) and DWD (Germany), demonstrate that S2R-GAN++ consistently outperforms baseline methods. Specifically, the model reduces average MSE by approximately 20.3% on PhaDin and 43.6% on DWD compared to S2R-GAN, while also achieving notable improvements in SSIM, indicating superior preservation of spatial structures. These results demonstrate that S2R-GAN++ not only improves prediction accuracy but also enhances temporal stability and generalization capability across different meteorological scenarios.