Predicting 60-Minute Hail Swaths Using Advanced Machine Learning Models on AWSSageMaker

Our project focuses on developing and testing advanced machine learning models to predict
60-minute Maximum Estimated Size of Hail (MESH) swaths, which serve as a proxy for actual
hail occurrence. We utilize a variety of U-Net architectures, including attention-based and
transformer-based variants, as well as Recurrent U-Nets, to model this complex meteorological
phenomenon. The data collection process is extensive, involving the retrieval of meteorological
data from AWS, including radar, satellite, and environmental fields. This data is preprocessed
and normalized before being used to train our models on AWS SageMaker, leveraging the
platform's scalability and computational power to handle large datasets and complex models.
To ensure the effectiveness of our models, we employ a rigorous evaluation framework using
case studies. Each model is tested against real-world hail events, with a focus on how well the
predicted 60-minute MESH swaths align with observed hail occurrences. The case studies are
carefully selected to cover a range of storm types and intensities, providing a comprehensive
assessment of each model's performance. By utilizing AWS's infrastructure, we efficiently train,
test, and refine our models, ultimately aiming to improve hail prediction accuracy and advance
the state of severe weather forecasting.