Scholarlink

<p>Accurate radar echo extrapolation is crucial for severe weather prevention. However, the nonlinear evolution and complex spatiotemporal dynamics of convective systems pose significant challenges to existing methods. Traditional methods rely on linearized approximations or convolution and attention mechanisms, which often fail to capture the inherent chaotic interactions and long-range spatiotemporal dependencies of these systems, resulting in ambiguous forecasts and underestimation of extreme precipitation areas. To overcome these limitations, we propose a self-supervised learning framework, ST-MIM. By combining mask image reconstruction with frequency loss, we integrate nonlinear spatiotemporal modeling to better reveal the underlying dynamic processes that control echo evolution. By reconstructing echo cubes of random masks at spatiotemporal scales, our method captures atmospheric evolution patterns more effectively than traditional intra-frame techniques. In addition, the Fourier enhancement loss function preserves high-frequency features and enhances the model’s representation of nonlinear convective structures and chaotic transition zones. To support spatiotemporal modeling, we introduce AH-Rain, the largest open-source radar dataset available. It covers the range of 27-36 degrees north latitude and 113-121 degrees east longitude, containing a variety of synoptic and mesoscale patterns. Extensive evaluations show that ST-MIM outperforms state-of-the-art methods in terms of meteorological metrics, especially with improvements in the localization of extreme rainfall.</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

Spatiotemporal dynamic modeling for nowcasting rainfall prediction using a dual-dimensional masked autoencoder

  • 期刊论文

  • 发表日期：

    2026年1月20日

查看全文
  • Jie Dong,
  • Kaichao Miao,
  • Zhize Wu,
  • Teng Li

摘要

Accurate radar echo extrapolation is crucial for severe weather prevention. However, the nonlinear evolution and complex spatiotemporal dynamics of convective systems pose significant challenges to existing methods. Traditional methods rely on linearized approximations or convolution and attention mechanisms, which often fail to capture the inherent chaotic interactions and long-range spatiotemporal dependencies of these systems, resulting in ambiguous forecasts and underestimation of extreme precipitation areas. To overcome these limitations, we propose a self-supervised learning framework, ST-MIM. By combining mask image reconstruction with frequency loss, we integrate nonlinear spatiotemporal modeling to better reveal the underlying dynamic processes that control echo evolution. By reconstructing echo cubes of random masks at spatiotemporal scales, our method captures atmospheric evolution patterns more effectively than traditional intra-frame techniques. In addition, the Fourier enhancement loss function preserves high-frequency features and enhances the model’s representation of nonlinear convective structures and chaotic transition zones. To support spatiotemporal modeling, we introduce AH-Rain, the largest open-source radar dataset available. It covers the range of 27-36 degrees north latitude and 113-121 degrees east longitude, containing a variety of synoptic and mesoscale patterns. Extensive evaluations show that ST-MIM outperforms state-of-the-art methods in terms of meteorological metrics, especially with improvements in the localization of extreme rainfall.