Thursday, 31 August 2023: 11:15 AM
Great Lakes A (Hyatt Regency Minneapolis)
Nowcasting (0-1 hour) initiation and rapid development of convective storms is of great importance. Current nowcasting methods are mostly developed based on single-polarized radar echo extrapolation, which suffers from both ineffectiveness of nowcasting models and insufficiency of input information. Recently, deep-learning (DL) based methods have been vastly used in convective nowcasting task. Despite massive proposed work, the models are still limited in nowcasting initiation and rapid development of convective storms, mostly owing to the lack of input information.
As an advanced observing tool, polarimetric weather radars provide much more microphysics information for convective systems, and the spatial patterns of the polarimetric variables can potentially disclose dynamical structure of convective storms. Therefore, a deep learning architecture, Fusion and Reassignment Networks (FURENet), specially designed for fusing and extracting information from multiple polarimetric input variables is developed. The model uses U-Net backbone, and is strengthened by newly-designed late-fusion branches and squeeze-and-excitation attention blocks. The model is thus capable of capture higher-level dependencies of multiple input variables as well as focusing on the crucial and informative data features by weight-redistribution mechanisms.
Nowcasts of two representative cases indicate that KDP and ZDR can help the DL model better forecast convective organization and initiation. Quantitative nowcasting skills (CSI score) on ~1000 samples are improved by 13.2% (17.4%) for 30- and 60-minute with FURENet inputting KDP and ZDR. The results indicate that microphysical information provided by the polarimetric variables can enhance the DL forecasting model to make more trustable nowcasts of convective evolution.
Furthermore, using explainable AI (XAI) methods, our recent results show that compared to only inputting 2D variables, incorporating 3D polarimetric variables to an advanced DL models (FURENetV2) leads to more accurate and physically-consistent forecasts.
Fig1 https://drive.google.com/file/d/18Wxa__AZdmIWhPS_Nwy5yd5eUuhhbSWE/view?usp=sharing
As an advanced observing tool, polarimetric weather radars provide much more microphysics information for convective systems, and the spatial patterns of the polarimetric variables can potentially disclose dynamical structure of convective storms. Therefore, a deep learning architecture, Fusion and Reassignment Networks (FURENet), specially designed for fusing and extracting information from multiple polarimetric input variables is developed. The model uses U-Net backbone, and is strengthened by newly-designed late-fusion branches and squeeze-and-excitation attention blocks. The model is thus capable of capture higher-level dependencies of multiple input variables as well as focusing on the crucial and informative data features by weight-redistribution mechanisms.
Nowcasts of two representative cases indicate that KDP and ZDR can help the DL model better forecast convective organization and initiation. Quantitative nowcasting skills (CSI score) on ~1000 samples are improved by 13.2% (17.4%) for 30- and 60-minute with FURENet inputting KDP and ZDR. The results indicate that microphysical information provided by the polarimetric variables can enhance the DL forecasting model to make more trustable nowcasts of convective evolution.
Furthermore, using explainable AI (XAI) methods, our recent results show that compared to only inputting 2D variables, incorporating 3D polarimetric variables to an advanced DL models (FURENetV2) leads to more accurate and physically-consistent forecasts.
Fig1 https://drive.google.com/file/d/18Wxa__AZdmIWhPS_Nwy5yd5eUuhhbSWE/view?usp=sharing
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner