Scholarlink

<p>Global Navigation Satellite System (GNSS) provides valuable atmospheric information for short-term rainfall forecasting, such as zenith tropospheric delay, precipitable water vapor, and other tropospheric products. Existing single-station GNSS rainfall prediction approaches ignore the inter-station spatiotemporal dependencies and fail to forecast the dynamic evolution of rainfall. This paper proposes a spatiotemporal Flux Transformer (ST-FluxFormer) that achieves accurate rainfall nowcasting with lead time of up to 3 h based on GNSS tropospheric products. The FluxFormer is designed to explicitly associate historical atmospheric fluctuations with their future evolution through the flux awareness attention mechanism, providing stronger temporal dependency modeling than state-of-the-art time-series approaches. The ST-FluxFormer introduces the attention-based graph neural network (GNN) to characterize the spatial interactions and propagation of atmospheric moisture and rainfall. Evaluations based on measurements from the Hong Kong satellite positioning reference station network demonstrate that the integration of GNSS-derived tropospheric products and their temporal difference significantly enhance the forecasting performance of ST-FluxFormer. Compared with ST-iTransformer, a spatiotemporal baseline that replaces FluxFormer with iTransformer while retaining the same GNN for spatial dependency modeling, the ST-FluxFormer improves the Critical Success Index (CSI) for heavy rainfall by 36.6%. Case studies of two extreme rainfall events in 2024 further confirm the ability of ST-FluxFormer to capture the spatiotemporal evolution of advective storms. Overall, the ST-FluxFormer provides an effective solution for leveraging GNSS reference networks to support short-term rainfall nowcasting, with the potential to be flexibly deployed wherever GNSS infrastructure is available.</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 spatiotemporal framework for GNSS-based rainfall nowcasting with FluxFormer and attention-based graph neural networks

  • 期刊论文

  • 发表日期：

    2026年7月8日

查看全文
  • Xiangyang Han,
  • Xianwei Wang,
  • Jinhua Wu,
  • Yuli Wang,
  • Joseph Awange,
  • Bo Zhang,
  • Ting On Chan

摘要

Global Navigation Satellite System (GNSS) provides valuable atmospheric information for short-term rainfall forecasting, such as zenith tropospheric delay, precipitable water vapor, and other tropospheric products. Existing single-station GNSS rainfall prediction approaches ignore the inter-station spatiotemporal dependencies and fail to forecast the dynamic evolution of rainfall. This paper proposes a spatiotemporal Flux Transformer (ST-FluxFormer) that achieves accurate rainfall nowcasting with lead time of up to 3 h based on GNSS tropospheric products. The FluxFormer is designed to explicitly associate historical atmospheric fluctuations with their future evolution through the flux awareness attention mechanism, providing stronger temporal dependency modeling than state-of-the-art time-series approaches. The ST-FluxFormer introduces the attention-based graph neural network (GNN) to characterize the spatial interactions and propagation of atmospheric moisture and rainfall. Evaluations based on measurements from the Hong Kong satellite positioning reference station network demonstrate that the integration of GNSS-derived tropospheric products and their temporal difference significantly enhance the forecasting performance of ST-FluxFormer. Compared with ST-iTransformer, a spatiotemporal baseline that replaces FluxFormer with iTransformer while retaining the same GNN for spatial dependency modeling, the ST-FluxFormer improves the Critical Success Index (CSI) for heavy rainfall by 36.6%. Case studies of two extreme rainfall events in 2024 further confirm the ability of ST-FluxFormer to capture the spatiotemporal evolution of advective storms. Overall, the ST-FluxFormer provides an effective solution for leveraging GNSS reference networks to support short-term rainfall nowcasting, with the potential to be flexibly deployed wherever GNSS infrastructure is available.