Principal Research Scientist, Ocean Space Development & Energy Research Department, Korea Institute of Ocean Science & Technology
Corresponding author:
Young Hyun Park ,Tel: +82-51-664-3521, Email: yhpark@kiost.ac.kr
Received: November 11, 2025; Revised: December 3, 2025. Accepted: December 3, 2025.
Share :
ABSTRACT
Global climate change is altering the characteristics of tropical cyclones, and numerous studies aim to characterize these changes and apply the findings to coastal hazard mitigation. Identifying statistical trend shifts in tropical cyclone characteristics is essential for return-period analyses used in disaster prevention and for understanding the long-term variability of the climate system. Focusing on tropical cyclones in the northwestern Pacific adjacent to the Korean Peninsula over the 80 years from 1945 to 2024, this study detects long-term regime shifts in tropical cyclone behavior. The trend shift detection framework combines two deep-learning models—Long Short-Term Memory (LSTM) and Transformer-based autoencoders—with statistical techniques designed to capture nonlinear interactions among tropical cyclone attributes. We then apply three trend shift detection algorithms, suited to both short- and long-term variability, and define long-term regime shifts as periods consistently identified across methods, thereby enhancing reliability through a multi-architecture ensemble. To address limitations of conventional AI approaches, we employ SHapley Additive exPlanations (SHAP) to interpret model decisions and diagnose the dominant factors driving long-term shifts. The ensemble and voting procedure identify 1997 as a significant long-term trend shift. SHAP analysis reveals that the longitudinal shift of maximum intensity is the primary contributor to explaining the observed trend shift. This AI-based framework provides a quantitative methodology for identifying long-term changes in tropical cyclone characteristics, supporting coastal engineering applications.
PDF Links
Full text via DOI
Download Citation 
