Unveiling the Role of Weighted Loss Functions in Deep Learning-Based Nowcasting of Extreme Rainfall Events

Abstract

Nowcasting plays a crucial role in responding to disasters such as flash floods by predicting rainfall in real time. However, existing nowcasting models struggle to accurately predict extreme rainfall events, which, although rare, can have devastating impacts. This challenge primarily arises because typical loss functions focus on minimizing average prediction errors rather than emphasizing the importance of extreme events, leading to their underestimation. To address this issue, this study introduces various weighted loss functions that impose greater penalties on prediction errors as rainfall intensity increases. These weighted loss functions were applied to a convolutional long short-term memory (ConvLSTM)-based nowcasting model to assess their impact on model performance. Recognizing that weighted loss functions may influence the learning of spatial patterns, and this study categorized extreme rainfall events based on their spatial characteristics and conducted a detailed performance evaluation for each type. The results demonstrated that models using weighted loss functions significantly improved the accuracy of extreme rainfall predictions compared to unweighted (UW) models. Notably, depending on the applied weighted loss functions, each model clearly exhibited its strengths and weaknesses across various extreme rainfall types. This finding suggests that selecting the best-performing weighted model based on prediction goals can lead to optimal results. Furthermore, this study revealed that the effectiveness of prediction methods varies significantly depending on the type of extreme rainfall event, indicating the need for dynamic selection of prediction methods tailored to specific condition. This article provides valuable insights into improving extreme rainfall nowcasting and is expected to contribute to enhancing disaster response systems in the future.