Multi-Scale Fourier Temporal Network for Multi-Source Precipitation Nowcasting

Area of Science:

• Meteorology
• Hydrology
• Computer Science

Background:

• Accurate precipitation nowcasting is crucial for disaster prevention and hydrometeorological applications.
• Current deep learning models struggle with complex precipitation dynamics and multi-source data integration.

Purpose of the Study:

• To develop an advanced deep learning framework for improved precipitation nowcasting.
• To address limitations in existing models regarding multi-source observations and physical representations.

Main Methods:

• Proposed a Multi-Scale Frequency-Temporal Network (MS-FTNet) utilizing Fourier transform for frequency-domain modeling.
• Decomposed precipitation dynamics into low-frequency (stratiform) and high-frequency (convective) components.
• Introduced Global Feature Collaboration (GFC) and Adaptive Temporal Fusion (ATF) modules for enhanced feature integration and temporal modeling.

Main Results:

• MS-FTNet demonstrated superior performance over baseline models on the SEVIR dataset.
• Significant improvements were observed in Mean Squared Error (MSE), Critical Success Index (CSI), and Learned Perceptual Image Patch Similarity (LPIPS).
• The model excelled particularly in nowcasting heavy precipitation events and for longer forecast lead times.

Conclusions:

• The MS-FTNet framework offers a novel and effective approach to precipitation nowcasting.
• Frequency-domain analysis combined with advanced deep learning modules enhances the prediction of complex precipitation patterns.
• The proposed method shows promise for operational hydrometeorological forecasting and disaster management.