Overview
Research investigated the influence of input data characteristics on the skill of a deep-learning hail nowcasting model employing spatiotemporal neural networks. The study aimed to identify methods for improving model performance without modifying the underlying architecture. Deep-learning approaches have emerged as an alternative to conventional nowcasting methods, yet the impact of input data selection on their performance has not been extensively explored.
Research Context
Hail events pose substantial financial risks and public safety concerns. Consequently, reliable nowcasts are critical for providing timely warnings. This research addresses the need for enhanced nowcasting capabilities, specifically focusing on optimizing deep-learning models for hail prediction.
Approach
The investigation utilized sensitivity experiments to evaluate the impact of various input data choices on the nowcasting model's skill. The experiments focused on three key areas:
- Training-Data Volume: The quantity of data used for model training.
- Random Data Augmentation: The application of mirroring and rotation techniques to the training data.
- Number of Input Timesteps: The quantity of temporal data points provided as input to the model.
These experiments were conducted while maintaining a fixed model architecture to isolate the effects of data-related interventions.
Findings
- Training Data Volume: Increasing the number of years represented in the training data led to a substantial improvement in forecast skill. This enhancement was observed to be up to 25 minutes at later lead times.
- Data Augmentation: The application of random data augmentation (mirroring and rotation) demonstrated a beneficial effect on performance when applied to larger datasets. Conversely, for smaller datasets, augmentation surprisingly resulted in a degradation of performance.
- Input Timesteps: The model's sensitivity to the number of input timesteps was found to be weaker compared to its sensitivity to the volume of training years.
These findings collectively indicate that significant gains in deep-learning nowcasting skill can be achieved through judicious data selection and preprocessing strategies, even when the model architecture remains unchanged.
Why This Matters
The findings provide guidance for future development of deep-learning nowcasting systems. Understanding the influence of input data characteristics, such as volume and augmentation strategies, can inform optimization efforts, potentially leading to more accurate and timely warnings for hail events.