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

Area of Science:

• Meteorology and Atmospheric Science
• Artificial Intelligence and Machine Learning
• Computer Vision

Background:

• Precipitation nowcasting is crucial for mitigating extreme weather impacts.
• Existing methods struggle with temporal dependencies and global spatiotemporal modeling.
• Limitations include information loss and difficulty in capturing multi-scale rainband features.

Purpose of the Study:

• To develop an advanced precipitation nowcasting model addressing limitations of current methods.
• To improve the accuracy of radar echo sequence prediction, especially for heavy rainfall events.
• To enhance the integration of spatiotemporal information and multi-scale directional features.

Main Methods:

• Proposed ST-TriMambaUNet with an encoder, decoder, 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 on SEVIR and CIKM datasets.
• Achieved up to 10.19% improvement in Critical Success Index (CSI) for high-threshold precipitation scenarios.
• Showcased significant enhancements in CSI, Probability of Detection (POD), and Heidke Skill Score (HSS) for heavy rainfall events.

Conclusions:

• ST-TriMambaUNet effectively models long-range spatial correlations and temporal dependencies in radar echo sequences.
• The model's architecture enhances representation for strong-echo regions and multi-scale precipitation structures.
• Results confirm the model's advantage in accurate and reliable precipitation nowcasting, particularly for extreme events.