Preprints
https://doi.org/10.5194/egusphere-2024-1945
https://doi.org/10.5194/egusphere-2024-1945
29 Jul 2024
 | 29 Jul 2024

Brief Communication: Training of AI-based nowcasting models for rainfall early warning should take into account user requirements

Georgy Ayzel and Maik Heistermann

Abstract. In the field of precipitation nowcasting, deep learning (DL) has emerged as an alternative to conventional tracking and extrapolation techniques. However, DL struggles to adequately predict heavy precipitation, which is essential in early warning. By taking into account specific user requirements, though, we can simplify the training task and boost predictive skill. As an example, we predict the cumulative precipitation of the next hour (instead of five minute increments), and the exceedance of thresholds (instead of numerical values). A dialogue between developers and users should identify the requirements to a nowcast, and how to consider these in model training.

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this preprint. The responsibility to include appropriate place names lies with the authors.
Georgy Ayzel and Maik Heistermann

Status: final response (author comments only)

Comment types: AC – author | RC – referee | CC – community | EC – editor | CEC – chief editor | : Report abuse
  • RC1: 'Comment on egusphere-2024-1945', Anonymous Referee #1, 20 Aug 2024
    • AC1: 'Reply on RC1', Maik Heistermann, 06 Sep 2024
  • RC2: 'Comment on egusphere-2024-1945', Remko Uijlenhoet, 27 Aug 2024
    • AC2: 'Reply on RC2', Maik Heistermann, 06 Sep 2024
Georgy Ayzel and Maik Heistermann

Model code and software

The RainNet2024 family of deep neural networks for precipitation nowcasting Georgy Ayzel and Maik Heistermann https://doi.org/10.5281/zenodo.12547127

Georgy Ayzel and Maik Heistermann

Viewed

Total article views: 361 (including HTML, PDF, and XML)
HTML PDF XML Total BibTeX EndNote
177 81 103 361 7 7
  • HTML: 177
  • PDF: 81
  • XML: 103
  • Total: 361
  • BibTeX: 7
  • EndNote: 7
Views and downloads (calculated since 29 Jul 2024)
Cumulative views and downloads (calculated since 29 Jul 2024)

Viewed (geographical distribution)

Total article views: 377 (including HTML, PDF, and XML) Thereof 377 with geography defined and 0 with unknown origin.
Country # Views %
  • 1
1
 
 
 
 
Latest update: 05 Nov 2024
Download
Short summary
Forecasting rainfall over the next hour is an essential feature of early warning systems. Deep learning has emerged as a powerful alternative to conventional nowcasting technologies, but it still struggles to adequately predict impact-relevant heavy rainfall. We think that DL could do much better if the training tasks were defined more specifically, and that such a specification presents an opportunity to better align the output of nowcasting models with actual user requirements.