Artificial Intelligence-Based Prediction Model for Tropical Cyclone Intensity with Inner-Core and Environmental Features

Predicting the intensity of Tropical Cyclones(TC) is has become a major concern for both research and operational centers over the decades. The Korean Peninsula has repeated threats to life and property due to the adverse effects of TCs, making accurate prediction of TC intensity changes crucial. we present a novel TC intensity prediction model that utilizes various machine learning techniques, including decision tree, random forest, and support vector machine (SVM) and deep-learning algorithm such as Multi-Layer Perceptron (MLP) and integrates geostationary satellite and numerical reanalysis data.

Our dataset comprises Best-Track data from the Regional Specialized Meteorological Center (RSMC) Tokyo, focusing on TC cases affecting the Korean Peninsula as identified by the Korea Meteorological Agency's criteria for strong wind areas (17 m/s) over the four-year period from 2016 to 2019. Cloud features representing changes in convective intensity, area, and organization within the inner core were extracted using brightness temperature data from the Himawari-8 Advanced Himawari Imager(AHI) geostationary satellite. Environmental parameters at the synoptic scale, including vorticity, relative humidity, vertical wind shear and sea surface temperature obtained from the Global Forecast System(GFS), were considered essential input parameters. Through a time series analysis, we examined the relationship between TC intensity changes and the input parameters across a variety of TC intensity.

A prediction model was constructed to forecast TC intensity at 6, 12, and 24-hour lead-time based on MLP, decision tree, random forest, and SVM, utilizing prediction parameters derived from both satellite and reanalysis data. Model performance was evaluated through training and verification, employing mean square root error calculations for estimated TC intensity. Through evaluation between developed models, performance evaluation between machine learning models and comparison with deep learning models were performed. The integration of diverse scale information in our TC intensity prediction model is anticipated to contribute significantly to reducing forecast errors in TC intensity.