MF-UFNO: A Fourier Neural Operators-Based Model for Radar Echo Extrapolation

Severe convective weather events, such as heavy rainfall and flooding, are serious threats to human society. However, accurate nowcasting of these events using radar remains challenging due to the complex dynamics and nonlinear physical processes involved. Moreover, existing radar echo extrapolation methods primarily rely on time-domain analysis, leaving a considerable gap in capturing frequency-domain features. Therefore, this paper proposes a Multi-variable Fusion UNet-Fourier Neural Operator (MF-UFNO) model, which combines multiple radar variables through a late-fusion strategy for extrapolation tasks. The model integrates Fast Fourier Transform to extract frequency-domain features, enhancing the representation of the spatiotemporal evolution of radar echoes. The MF-UFNO model is trained and validated on polarimetric radar variables collected from an S-band dual-polarization radar between April 2020 and June 2021. Experimental results indicate that the MF-UFNO model achieves high forecasting accuracy, with statistical Threat Scores exceeding 0.5 and 0.4 for 15 and 25 dBZ thresholds, respectively, over the forecast period. Compared to existing models such as the SmaAt-UNet model and the Rainymotion model, the proposed model demonstrates superior performance in radar echo extrapolation, particularly in accurately predicting fine-scale structures within a 60-min forecast window. With the precise nowcasting capabilities, the MF-UFNO model can improve the generation of severe convective weather warnings and enhance short-term weather guidance.