Deep Learning Improves GFS Wintertime Precipitation Forecast Over Southeastern China

Wintertime precipitation, especially snowstorms, significantly impacts people's lives. However, the current forecast skill of wintertime precipitation is still low. Based on data augmentation (DA) and deep learning, we propose a DABU‐Net which improves the Global Forecast System wintertime prec...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Geophysical research letters 2023-07, Vol.50 (14), p.n/a
Hauptverfasser: Sun, Danyi, Huang, Wenyu, Yang, Zifan, Luo, Yong, Luo, Jingjia, Wright, Jonathon S., Fu, Haohuan, Wang, Bin
Format: Artikel
Sprache:eng
Schlagworte:
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:Wintertime precipitation, especially snowstorms, significantly impacts people's lives. However, the current forecast skill of wintertime precipitation is still low. Based on data augmentation (DA) and deep learning, we propose a DABU‐Net which improves the Global Forecast System wintertime precipitation forecast over southeastern China. We build three independent models for the forecast lead times of 24, 48, and 72 hr, respectively. After using DABU‐Net, the mean Root Mean Squared Errors (RMSEs) of the wintertime precipitation at the three lead times are reduced by 19.08%, 25.00%, and 22.37%, respectively. The threat scores (TS) are all significantly increased at the thresholds of 1, 5, 10, 15, and 20 mm day−1 for the three lead times. During heavy precipitation days, the RMSEs are decreased by 14% and TS are increased by 7% at the lead times within 48 hr. Therefore, combining DA and deep learning has great prospects in precipitation forecasting. Plain Language Summary In this paper, we propose a deep learning‐based method to improve the forecast performance of Global Forecast System wintertime precipitation over southeastern China. Due to the imbalanced distribution of precipitation data, we use data from the three other seasons as an augmented data set for wintertime precipitation to train the deep neural network. The results show that the method can reduce the Root Mean Squared Error and improve the TS, a metric of precipitation forecast performance, of the precipitation. In particular, TS at the threshold of 20 mm day−1 are increased by 69.23%, 90.00%, and 100.00% at lead times of 24, 48, and 72 hr. The proposed method performs well during heavy precipitation days at lead times within 48 hr. Combining data augmentation with deep learning provides a successful approach to predicting precipitation. Key Points A deep learning model based on data augmentation (DA) is proposed to improve the Global Forecast System wintertime precipitation forecast The deep learning model improves the heavy precipitation forecast at lead times within 48 hr DA plays a critical role in the heavy precipitation forecast
ISSN:0094-8276
1944-8007
DOI:10.1029/2023GL104406