Enhancing Tropical Cyclone Formation Prediction Using Graph Neural Networks

Sabbir Rahman, Dr Nusrat Sharmin, Md. Mahbubur Rahman, Md Mokhlesur Rahman

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Graph Neural Network (GNN), Temporal Evolution, Node classification, Cyclone Prediction

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract:

Tropical cyclones are among the most powerful and destructive weather events on Earth, and the formation and evolution of these systems is crucial to the resilience and safety of coastal populations. Although physical models have historically been used to research tropical cyclones, these models frequently fail to capture
the complex interactions between many atmospheric and oceanic factors that influence cyclonic systems’ behavior. In this research, we suggest a unique method of employing graph neural networks (GNNs) to analyze the development and evolution of tropical cyclones. GNNs are an effective machine learning technique
that can learn from huge and complex datasets, which makes them well-suited to capture the underlying patterns in the behavior of tropical cyclones. In our method, a GNN is used to estimate cyclone formation, forecast whether it will become stronger or weaker in the following time step, and match the evolution pattern of cyclones in the training set. We tested our method on a substantial
dataset of tropical cyclones and showed that it outperformed conventional physical models in predicting the genesis of tropical cyclones. Our research also shown that the intricate connections between atmospheric and oceanic factors that affect
tropical cyclones are better captured by the GNN-based method, leading to a better understanding of their behavior. As a result of our research, better early warning systems and disaster response planning will be possible, allowing for more
precise forecasts of tropical cyclone development and behavior. Our work also shows how machine learning methods may improve our comprehension of intricate meteorological processes, presenting new avenues for research in atmospheric science.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 7475