SRViT: Vision Transformers for Estimating Radar Reflectivity from Satellite Observations at Scale (2406.16955v2)

Abstract: We introduce a transformer-based neural network to generate high-resolution (3km) synthetic radar reflectivity fields at scale from geostationary satellite imagery. This work aims to enhance short-term convective-scale forecasts of high-impact weather events and aid in data assimilation for numerical weather prediction over the United States. Compared to convolutional approaches, which have limited receptive fields, our results show improved sharpness and higher accuracy across various composite reflectivity thresholds. Additional case studies over specific atmospheric phenomena support our quantitative findings, while a novel attribution method is introduced to guide domain experts in understanding model outputs.

• The paper introduces a transformer-based approach that generates high-resolution (3 km) synthetic radar reflectivity fields from satellite data, enhancing convective-scale forecasts.
• SRViT employs vision transformers with convolutional refinement to overcome CNN limitations, achieving an RMSE of 3.09 dBZ and an R² of 0.572 compared to baseline models.
• Case studies on severe weather events demonstrate SRViT’s ability to provide sharper, more interpretable forecasts, aiding effective data assimilation for numerical weather prediction.

SRViT: Vision Transformers for Estimating Radar Reflectivity from Satellite Observations at Scale

The paper "SRViT: Vision Transformers for Estimating Radar Reflectivity from Satellite Observations at Scale" proposes a transformer-based neural network designed to generate high-resolution (3 km) synthetic radar reflectivity fields from geostationary satellite imagery. This research aims to enhance short-term convective-scale forecasts of high-impact weather events and facilitate data assimilation for numerical weather prediction (NWP) across the United States.

Key Contributions

1. Transformer Architecture: Unlike conventional convolutional neural networks (CNNs) typically employed for radar reflectivity estimation, the authors leverage vision transformers (ViTs) to address the limitations of constrained receptive fields in CNNs. This novel application of transformers exploits their ability to capture long-range dependencies and synoptic-scale atmospheric phenomena.
2. Dataset and Preprocessing: The paper utilizes observational data from the Geostationary Operational Environmental Satellite (GOES) series, incorporating infrared bands from the Advanced Baseline Imager (ABI) and real-time lightning data from the Geostationary Lightning Mapper (GLM). The target dataset comprises composite radar reflectivity from the Multi-Radar Multi-Sensor (MRMS) product. The data spans the contiguous United States (CONUS) over several years and undergoes rigorous quality control.
3. SRViT Model Architecture: The proposed Satellite to Radar Vision Transformer (SRViT) extends the standard ViT architecture for image-to-image translation tasks. The model's design includes multiple transformer blocks followed by convolutional layers to refine the output, effectively smoothing boundary artifacts resulting from token-based reconstruction.
4. Quantitative and Qualitative Evaluation: SRViT outperforms a baseline UNet model—a fully-convolutional encoder/decoder architecture—especially at low- to mid-range reflectivity thresholds. Quantitative metrics include the Root Mean Squared Error (RMSE) and coefficient of determination (R²). SRViT achieves an RMSE of 3.09 dBZ and an R² of 0.572, compared to the UNet's RMSE of 3.21 dBZ and R² of 0.488. The model also demonstrates superior sharpness in predicted radar reflectivity fields, verified through gradient magnitude analysis.
5. Case Studies and Model Attribution: The paper includes case studies of severe weather events such as a Northern Plains Derecho and Midwest Squall Lines, showing that SRViT produces more accurate and sharper reflectivity fields. Additionally, the authors introduce a gradient-based attribution method called Token (Re)Distribution to help domain experts interpret the model's outputs, enhancing the transparency and usability of the predictions.

Implications and Future Directions

Practical Implications

The enhanced accuracy and sharpness of SRViT's predictions offer significant practical benefits for weather forecasting and NWP. The high-resolution synthetic radar reflectivity generated by SRViT can support real-time decision-making for severe weather warnings and improve the initialization of NWP models, leading to better forecasts of convective weather events.

Theoretical Implications

From a theoretical perspective, this work exemplifies the potential of transformer architectures in addressing complex geospatial prediction tasks traditionally dominated by convolutional approaches. The successful application of transformers in this domain underscores their versatility and capability to model long-range dependencies effectively.

Future Developments

Future research directions include fine-tuning SRViT’s hyperparameters to enhance performance further and exploring probabilistic diffusion models to improve uncertainty quantification and sharpness. Additionally, expanding the model to estimate 3D radar reflectivity fields could provide more comprehensive data for NWP models, potentially transforming operational weather forecasting. Investigating the integration of SRViT with other NWP data assimilation techniques could also yield substantial advancements in predictive accuracy and efficiency.

Conclusion

The research delineates a transformative approach to generating synthetic radar reflectivity fields using vision transformers. By addressing the limitations of traditional convolutional methods, SRViT presents a promising advancement in the field of meteorological forecasting, paving the way for more accurate and timely weather predictions. The incorporation of novel attribution methods further enhances the model’s applicability and interpretability, making it a valuable tool for operational meteorology.