Predicting Rapid Intensification of Tropical Cyclones in the Western North Pacific Using TabNet Classifier: Performance Improvement through Data Augmentation and Loss Function Optimization

Abstract

In this study, the TabNet model was employed to effectively predict Rapid Intensification (RI) cases of Tropical Cyclones (TCs) in the Northwestern Pacific region. By leveraging the model’s interpretability, key features influencing predictions were analyzed, revealing Sea Surface Temperature (SST) as the most significant factor, followed by Temperature and Specific Humidity at 850 hPa. These findings align with existing research, underscoring TabNet’s ability to accurately identify variables critical to ocean-atmosphere interactions. To address the severe class imbalance between RI and non-RI cases, the Synthetic Minority Oversampling Technique (SMOTE) was utilized with ratios of 4:1, 3:1, 2:1, and 1:1, and three loss functions—Cross Entropy, Balanced Cross Entropy, and Focal Loss—were combined to design twelve experimental setups. The combination of Balanced Cross Entropy and 2:1 SMOTE achieved the highest Critical Success Index (CSI) of 0.780, effectively balancing class imbalance and maximizing prediction performance. The optimized TabNet model demonstrated approximately 46.15% higher Probability of Detection (POD) and 42.31% higher CSI compared to the Hurricane Weather Research and Forecasting (HWRF) model, highlighting TabNet’s superior performance in predicting minority classes. Yearly performance evaluation revealed significant performance degradation in 2020, attributed to the scarcity of RI cases rather than atypical meteorological conditions. This emphasizes the critical role of data quantity and diversity in enhancing RI prediction accuracy. Additionally, the high interpretability of the TabNet model facilitates quantitative evaluation of influential variables, enhancing the understanding and reliability of prediction results. In conclusion, this research demonstrates that TabNet, combined with data augmentation and optimized loss functions, significantly improves RI prediction performance, surpassing traditional Numerical Weather Prediction (NWP) models. Future work should focus on further addressing data imbalance and enhancing model robustness under diverse environmental conditions to develop more reliable weather prediction systems. MS/EN 20231076