Developing Deep Learning Models for Storm Nowcasting

Storm nowcasting relies on reasonably fast sampled radar data, and deep learning (DL) can be used to harness this vast amount of data. Despite all the publications on this topic over the past five years, there are still ad hoc assumptions and a lack of standardization. This work addresses aspects that have not yet been analyzed on the development of DL models for nowcasting systems, such as the effects of different history lengths or using non-convex metrics during the training phase. For example, we show that even if the loss function is varied, it does not significantly influence the predictions, and that the number of predicted frames has a significant impact. We used the experiments' results to propose different models and compare their performance against other DL models. The results show that the proposed models outperform, in many aspects, the existing implementations.