Enhancing Extreme Rainfall Nowcasting with Weighted Loss Functions in Deep Learning Models

With the increasing frequency and intensity of extreme rainfall events, the importance of nowcasting to minimize damage from disasters such as flash floods is becoming ever more prominent. However, most nowcasting models use loss functions aimed at minimizing the average prediction error. As a result, they tend to underestimate extreme rainfall—which has relatively low occurrence frequency but significant impact. In this study, we applied various types of weighted loss functions to a ConvLSTM-based nowcasting model to more accurately predict extreme rainfall. In particular, we varied parameters within these weighted loss functions and conducted sensitivity analyses to identify the optimal weighting strategies. We also categorized extreme rainfall types and evaluated the models’ predictive performance with weighted loss functions, thereby examining both the accuracy and stability of the model’s forecasts under extreme conditions from multiple perspectives. The results showed that the model employing a weighted loss function achieved significantly improved accuracy in predicting extreme rainfall, compared to an unweighted model. Furthermore, depending on the type of weighted loss function and parameter settings, the model demonstrated notably strong performance for specific types of extreme rainfall. This finding suggests that, in a rainfall environment characterized by high variability, dynamically selecting weighted loss functions according to forecasting objectives and conditions can enhance both the efficiency and reliability of extreme rainfall prediction. The approach presented in this study can be applied to flood forecasting and is anticipated to contribute to the advancement of deep learning–based disaster response systems, reducing the potential damage caused by natural disasters.

 

Acknowledgements

This work was supported by Korea Environment Industry & Technology Institute(KEITI) through R&D Program for Innovative Flood Protection Technologies against Climate Crisis Program(or Project), funded by Korea Ministry of Environment(MOE)(RS-2023-00218873).