Scholarlink

<p>Spatiotemporal prediction has become a widely studied topic in recent years. It is crucial for many applications, particularly in meteorology. Atmospheric fields have diverse geometric structures and complex local variations that fixed receptive field architectures cannot capture well. Radar echo sequences make modeling more complicated due to low temporal resolution (usually 1 frame per 6&#xa0;min) and sudden physical processes such as initiation, intensification, and dissipation. Temporal dependencies therefore include long-term smooth trends and highly transient variations. Recently, Transformers have shown strong performance in spatiotemporal prediction, especially in modeling inter-frame correlations. However, excessive patch sequences in high-resolution scenarios make practical deployment challenging. By contrast, Mamba shows higher efficiency and accuracy in handling long-sequence data. Our model combines Mamba and Transformer to propose GSMFormer, which is designed to better capture complex geometric variations in weather forecasting. The Geometric Structure Mamba Block (GSMBlock) is introduced in the spatial domain to enhance the modeling of complex non-rigid local structures. In the temporal domain, considering the short input sequences and large inter-frame intervals, we employ multi-head self-attention (MHSA) and pair it with causal convolution (CausalConv) to capture long-term trends and short-term transient variations without recursive reasoning. Multiple experiments on four public video prediction benchmarks and two radar echo datasets under different scenarios show that GSMFormer is always more advantageous than existing methods and demonstrates strong robustness and generalization ability in challenging spatio-temporal conditions.</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

GSMFormer: a hybrid model based on geometric structure mamba and transformer for complex spatiotemporal prediction

  • 期刊论文

  • 发表日期：

    2026年7月11日

查看全文
  • Wei Tian,
  • Rongjia Guo,
  • Chenxu Zhou,
  • Jing Wang

摘要

Spatiotemporal prediction has become a widely studied topic in recent years. It is crucial for many applications, particularly in meteorology. Atmospheric fields have diverse geometric structures and complex local variations that fixed receptive field architectures cannot capture well. Radar echo sequences make modeling more complicated due to low temporal resolution (usually 1 frame per 6 min) and sudden physical processes such as initiation, intensification, and dissipation. Temporal dependencies therefore include long-term smooth trends and highly transient variations. Recently, Transformers have shown strong performance in spatiotemporal prediction, especially in modeling inter-frame correlations. However, excessive patch sequences in high-resolution scenarios make practical deployment challenging. By contrast, Mamba shows higher efficiency and accuracy in handling long-sequence data. Our model combines Mamba and Transformer to propose GSMFormer, which is designed to better capture complex geometric variations in weather forecasting. The Geometric Structure Mamba Block (GSMBlock) is introduced in the spatial domain to enhance the modeling of complex non-rigid local structures. In the temporal domain, considering the short input sequences and large inter-frame intervals, we employ multi-head self-attention (MHSA) and pair it with causal convolution (CausalConv) to capture long-term trends and short-term transient variations without recursive reasoning. Multiple experiments on four public video prediction benchmarks and two radar echo datasets under different scenarios show that GSMFormer is always more advantageous than existing methods and demonstrates strong robustness and generalization ability in challenging spatio-temporal conditions.