The Implementation of 3D Radar Data in a Deep Learning Precipitation Nowcasting Model

By Niek Bongers

Abstract
The observed increase in more frequent and intense precipitation events demands accurate short-term weather forecasts. The primary method for forecasting in use today, numerical weather prediction (NWP) models, is not suitable for nowcasting as their inference time is often longer than the lead time. Machine learning-based weather prediction (MLWP) models have emerged as a powerful alternative to NWPs, offering faster and potentially more accurate forecasts. Despite the availability of detailed 3D data, MLWPs have yet to fully leverage this rich information. Recent developments in deep learning enable the integration of altitude as a third dimension in extreme precipitation nowcasting. Since large portions of the radar-observed atmosphere result in clear-sky returns, which do not contribute significantly to prediction, sparse methods offer a more efficient approach by considering only the informative features.

In this study, the Minkowski Engine was adapted to predict dynamic spatiotemporal processes by applying dilation to the input data. Data from two radar stations were combined to create a detailed 3D dataset. The model was tested with both 2D and 3D representations to assess the impact of the third dimension on prediction accuracy. The results show that while the inclusion of 3D data offers a more detailed representation, the performance of the 3D model did not surpass that of the 2D model, especially for multi-timestamp predictions. For single-timestamp predictions, the models exhibited similar performance, whereas the 12-timestamp predictions of the 3D model were significantly worse.

The findings suggest that the integration of 3D as opposed to 2D data offers potential, further improvements are needed. Further exploration should focus on different CNN architectures, and consider implementation of temporal dynamics through RNNs. Additionally, incorporating functionality such as wind velocity and direction could enhance the dilation process, improving computational efficiency and prediction accuracy. Despite these challenges, the research highlights the potential of sparse data-driven models for precipitation nowcasting and offers insights for further advancements in dynamic spatiotemporal weather prediction.