Machine Learning Based Prediction of Tropical Cyclone Rapid Intensification in the Western North Pacific: Importance of Data Augmentation and Loss Function

Abstract In this study, we developed a TabNet‐based machine learning model to predict tropical cyclone (TC) rapid intensification (RI) in the Western North Pacific. The most significant challenge in predicting RI is the severe class imbalance between rapid and non‐rapid intensification cases, typically 4.2:1 ratio based on 1977–2021 records. To overcome this, the synthetic minority oversampling technique (SMOTE) 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 examined. The combination of balanced cross entropy and 2:1 SMOTE achieved the highest performance. Leveraging the TabNet's interpretability, sea surface temperature was identified as the most important factor, followed by temperature and specific humidity at 850 hPa. However, performance showed clear year‐to‐year variation. The year 2020 exhibited relatively poor performance likely due to the sparsity of RI cases rather than abnormal weather conditions, which highlights the important role of sample size in our model. In conclusion, TabNet combined with data augmentation and optimized loss functions significantly improves forecast per