guest ::
| Home > Publications database > GAN-based video prediction model for precipitation nowcasting |
| Home > Publications database > GAN-based video prediction model for precipitation nowcasting |
| Conference Presentation (Other) | FZJ-2022-02385 |
; ; ; ; ; ;
2022
This record in other databases:
Please use a persistent id in citations: doi:10.5194/egusphere-egu22-12086
Abstract: Detecting and predicting heavy precipitation for the next few hours is of great importance inweather related decision-making and early warning systems. Although great progress has beenachieved in convective-permitting numerical weather prediction (NWP) over the past decades,video prediction models based on deep neural networks have become increasingly popular overthe last years for precipitation nowcasting where NWP models fail to capture the quickly varyingprecipitation patterns. However, previous video prediction studies for precipitation nowcastingshowed that heavy precipitation events are barely captured. This has been attributed to theoptimization on pixel-wise losses which fail to properly handle the inherent uncertainty. Hence,we present a novel video prediction model, named CLGAN, embedding the adversarial loss isproposed in this study which aims to generate improved heavy precipitation nowcasting. Themodel applies a Generative Adversarial Network (GAN) as the backbone. Its generator is a u-shaped encoder decoder network (U-Net) equipped with recurrent LSTM cells and its discriminatorconstitutes a fully connected network with 3-D convolutional layers. The Eulerian persistence, anoptical flow model DenseRotation and an advanced video prediction model PredRNN-v2 serve asbaseline methods for comparison. The models performance are evaluated in terms of application-specific scores including root mean square error (RMSE), critical success index (CSI), fractions skillscore (FSS) and the method of object-based diagnostic evaluation (MODE). Our model CLGAN issuperior to the baseline models for dichotomous events, i.e. the CSI, with a threshold of heavyprecipitation (8mm/h), is significantly higher, thus revealing improvements in accurately capturingheavy precipitation events. Besides, CLGAN outperforms in terms of spatial scores such as FSS andMODE. We conclude that the predictions of our CLGAN architecture match the stochasticproperties of ground truth precipitation events better than those of previous video predictionmethods. The results encourage the applications of GAN-based video prediction architectures forextreme precipitation forecasting.
|
The record appears in these collections: |
