AN ADVANCED Artificial Intelligence System for Investigating the Tropical Cyclone Rapid Intensification



Journal Title

Journal ISSN

Volume Title



Tropical cyclones (TCs) can cause heavy casualties due to storm surge, high wind gusts, heavy rainfall and flooding, and landslides, so predicting TC is important. There are mainly two elements of TC forecasting: tracking prediction and intensity prediction. So far, it is found that tracking prediction is more mature than the intensity prediction. Various models are developed for TC intensity prediction and can be simple enough to run for a few seconds or complex enough to run for a couple of hours on a supercomputer. Although with so many models are developed, the intensity prediction accuracy is still very low, and one primary reason is the existence of Rapid Intensification (RI). Currently, most RI prediction studies are conducted based on a subset of the SHIPS database using a relatively simple model structure. However, variables (features) in the SHIPS database are built upon expert knowledge in TC intensity studies, and the variable values are derived from gridded model outputs or satellite observations. Are there any more important variables in TC intensity predictions but not identified in the SHIPS dataset? In this study, two AI-based techniques are used to extract new features from a widely used comprehensive gridded reanalysis data set. The original SHIPS data, and the newly derived features are used as inputs to an artificial intelligence (AI) for the RI prediction. This study first constructs a complicated artificial intelligence (AI) system, the COR-SHIPS model, based on the complete SHIPS dataset that handles feature engineering and selection, imbalance, prediction, and hyper parameter-tuning simultaneously. The COR-SHIPS model is derived to improve the performance of the current researches in RI prediction and to identify other essential SHIPS variables that are ignored by previous studies with variable importance scores. COR-SHIPS is also used as the baseline model in the dissertation. To distill new variables from vast amounts of gridded data, two models, with a similar structure to the COR-SHIPS model but with an additional data filters, are designed in the dissertation to identify new features related to TC intensity changes in general and RI in particular. Here, we adopt the Local linear embedding (LLE) and deep learning (DL) techniques respectively to filter the near center and large-scale spatial data of the European Centre for Medium-Range Weather Forecasts (ECMWF) ERA-Interim reanalysis data, one of the best reanalysis products at the moment, for identifying new variables related to RI, and term the corresponding LLE-SHIPS model and DL-SHIPS model, respectively. The result of the three models outperforms most of the earlier studies by at least approximately 30%, 60%, and 75%, respectively. In addition to the well-known SHIPS database, we specify the 400 and 450 hPa wind speeds, identify 1000 hPa potential vorticity and vertical pressure speed, and differentiate humidity southeast, vorticity north, and eastward wind north to the TC centers that could help the prediction and understanding the occurrence of RI.