Forecasting large-scale circulation regimes using deformable convolutional neural networks and global spatiotemporal...

Area of Science:

• Atmospheric Science
• Machine Learning
• Climate Science

Background:

• Classifying atmospheric states into circulation regimes aids in understanding weather patterns and climate change.
• Predicting severe weather events and teleconnections is crucial for climate research.

Purpose of the Study:

• To forecast North Atlantic-European weather regimes using a supervised machine learning approach.
• To evaluate the performance of deformable convolutional neural networks (deCNNs) with transfer learning for medium-range weather prediction.

Main Methods:

• Utilized deformable convolutional neural networks (deCNNs) for weather regime forecasting.
• Applied transfer learning and state-of-the-art interpretation techniques.
• Compared deCNN performance against meteorological benchmarks, logistic regression, and random forests.

Main Results:

• Demonstrated superior forecasting performance of deCNNs over classical benchmarks and other machine learning models.
• Observed deCNNs outperforming regular CNNs at lead times beyond 5-6 days due to a wider field of view.
• Confirmed the critical role of transfer learning in achieving high forecasting accuracy.

Conclusions:

• Deformable convolutional neural networks with transfer learning offer a powerful tool for medium-range weather regime forecasting.
• The approach provides insights into teleconnections and regions relevant to weather predictions.
• This method advances the predictability of atmospheric circulation patterns.