MFC-RFNet: A Multi-scale Guided Rectified Flow Network for Radar Sequence Prediction
In the realm of weather forecasting, accurate and high-resolution precipitation nowcasting from radar echo sequences is of paramount importance for disaster mitigation and economic planning. However, this task poses several significant challenges, including the modeling of complex multi-scale evolution, correcting inter-frame feature misalignment due to displacement, and effectively capturing long-range spatiotemporal context without sacrificing spatial fidelity.
To tackle these pressing issues, researchers have introduced the Multi-scale Feature Communication Rectified Flow (RF) Network, or MFC-RFNet. This innovative generative framework synergizes multi-scale communication with guided feature fusion, presenting a robust solution for radar-based nowcasting.
Key Innovations of MFC-RFNet
- Multi-scale Fusion with Wavelet-Guided Skip Connection (WGSC): To enhance multi-scale fusion while retaining fine detail, the MFC-RFNet incorporates a WGSC that preserves high-frequency components. This allows the model to maintain crucial details that are often lost in standard fusion processes.
- Feature Communication Module (FCM): The FCM promotes bidirectional cross-scale interaction, enabling the model to share information effectively across different scales. This interaction is essential for understanding the complexities involved in precipitation patterns.
- Condition-Guided Spatial Transform Fusion (CGSTF): To address the challenges of inter-frame displacement, the CGSTF module learns spatial transforms from conditioning echoes to align shallow features. This ensures that the model accurately represents the spatial relationships within the radar data.
- Rectified Flow Training: The backbone of MFC-RFNet employs rectified flow training to learn near-linear probability-flow trajectories. This approach allows for few-step sampling while maintaining stable fidelity, which is crucial for generating reliable forecasts.
- Lightweight Vision-RWKV Blocks: Positioned strategically at the encoder tail, the bottleneck, and the first decoder layer, lightweight Vision-RWKV blocks capture long-range spatiotemporal dependencies at low spatial resolutions. This design offers a balance between computational efficiency and performance.
Evaluation and Results
The MFC-RFNet was rigorously evaluated on four public datasets: SEVIR, MeteoNet, Shanghai, and CIKM. The results consistently indicated significant improvements over strong baseline models. Key findings include:
- Clearer echo morphology at higher rain-rate thresholds.
- Sustained skill and accuracy at longer lead times, crucial for timely decision-making in weather-related scenarios.
These promising results suggest that the integration of RF training with scale-aware communication, spatial alignment, and frequency-aware fusion presents an effective and robust approach for radar-based precipitation nowcasting.
The MFC-RFNet not only addresses existing limitations in radar echo sequence prediction but also paves the way for future advancements in meteorological forecasting technologies. By leveraging the strengths of multi-scale communication and guided feature fusion, this framework represents a significant step forward in the quest for accurate and timely weather predictions.