When Does a Typhoon Form? A Deep Learning Approach using Synthetic Images Constructed from Climate Reanalysis Data

Reliable prediction of typhoon (TY) formation is crucial for disaster preparedness and mitigation. This study presents a deep learning approach leveraging ResNet-18 to predict future TY formation using synthetic satellite images derived from the National Aeronautics and Space Administration (NASA) reanalysis data. We preprocess this data by applying a feature enrichment method and normalizing features to create a comprehensive multi-channel dataset for training. Our model is evaluated using precision, recall, F1 score, and AUROC metrics, which demonstrate promising results in predicting TY formation in the western North Pacific basin. This study presents a new methodology for TY prediction and highlights the effectiveness of deep learning in meteorological applications.