Strong dependence of tropical cyclone initial size on its genesis position

Tropical cyclone (TC) size variation involves complex physical processes that remain only partially understood despite considerable research efforts. Previous studies have explored TC size prediction using satellite infrared imagery, environmental factors, and intrinsic TC characteristics. However, key forecasting mechanisms continue to elude researchers. This study employs a Convolutional Neural Network (CNN) to investigate the relationship between TC initial size (upon first reaching tropical storm intensity) and genesis position (where tropical storm intensity is first attained) in the western North Pacific. Results show a strong correlation, enabling effective prediction of initial size from genesis position alone. Despite having a straightforward structure, the model achieves a correlation coefficient ( r ) of 0.41 and a Mean Absolute Error (MAE) of 66.19 km, comparable to models with many more predictors. Notably, under El Niño conditions, performance improves markedly ( r = 0.82, MAE = 47.42 km) when using only genesis position information. Further analysis indicates that key environmental fields—including long-term mean sea surface temperature, latent heat flux, and low-level vorticity—are strongly embedded in the spatial information of genesis position, offering insights into the physical mechanisms governing TC size variation.