ST-TriMambaUNet: A Weather Radar Echo Extrapolation-Based Spatiotemporal Sequence Prediction Network for...

Area of Science:

• Meteorology
• Artificial Intelligence
• Computer Vision

Background:

• Precipitation nowcasting is crucial for mitigating extreme weather impacts.
• Existing methods struggle with temporal dependencies and global spatiotemporal modeling.
• Radar echo sequence analysis often overlooks multi-scale directional features of rainbands.

Purpose of the Study:

• To develop an advanced model for accurate precipitation nowcasting.
• To address limitations in temporal dependency modeling and global feature extraction.
• To improve the prediction of heavy rainfall events using radar data.

Main Methods:

• Proposed ST-TriMambaUNet with encoder-decoder architecture and feature enhancement module.
• Introduced Spatiotemporal Fusion Attention (STFA) for parallel spatial and temporal dependency learning.
• Developed Multi-Scale Interaction Mamba (MSIM) integrating Mamba and Multi-Scale Directional Convolution (MSDC).

Main Results:

• ST-TriMambaUNet demonstrated superior performance in overall accuracy and heavy rainfall prediction.
• Significant improvements observed in Critical Success Index (CSI) on SEVIR dataset (up to 10.19%).
• Enhanced CSI, Probability of Detection (POD), and Heidke Skill Score (HSS) on CIKM dataset for heavy rainfall scenarios.

Conclusions:

• ST-TriMambaUNet effectively models long-range spatiotemporal correlations and radar echo dynamics.
• The model enhances representation of strong-echo regions and precipitation band structures.
• Achieved state-of-the-art results in precipitation nowcasting, particularly for extreme events.