Item

Abstract

Effective seasonal prediction of tropical cyclones (TCs) is essential for disaster prevention and mitigation in the Western North Pacific (WNP), one of the most active basins globally. While dynamical and statistical approaches have been widely used, seasonal forecast that provide spatially explicit formation regions remain limited. Here, we develop a Tropical Cyclone Formation Region Prediction Model (TCFRPM) based on Convolutional Neural Network (CNN) to predict TC formation region from 1 to 9-month leads during 1979–2021. The model incorporates 10 ocean-atmospheric variables and the Oceanic Niño Index (ONI) to capture both local environmental variables and large-scale climate signals. Results show that TCFRPM successfully reproduces most observed formation locations, with capture rates exceeding 0.5 at 1-month lead under stringent possibility threshold. Interpretability analysis using Shapley Additive Explanations (SHAP) reveals that sea surface temperature, mid-level humidity, and vertical wind shear consistently dominate the model’s prediction, with important contributions across lead times. Contributions from regions near formation centers are particularly pronounced. These findings highlight both the predictive potential and physical consistency of TCFRPM. Our work demonstrates the feasibility of applying deep learning for seasonal-scale, spatial TCs formation prediction, offering a promising tool for risk assessment and early preparedness in the WNP.