Interpretable tropical cyclone intensity estimation using Dvorak-inspired machine learning techniques. (May 2021)

• Record Type: Journal Article
• Authors: Lee, Yu-Ju; Hall, David; Liu, Quan; Liao, Wen-Wei; Huang, Ming-Chun
• Is Part Of: Engineering applications of artificial intelligence. Volume 101(2021)
• Publication Date: 2021-05
• Subjects: Tropical cyclone -- Random walk with restart -- Machine learning -- Meteorological data -- Satellite images; Engineering -- Data processing -- Periodicals; Artificial intelligence -- Periodicals; Expert systems (Computer science) -- Periodicals; Ingénierie -- Informatique -- Périodiques; Intelligence artificielle -- Périodiques; Systèmes experts (Informatique) -- Périodiques; Artificial intelligence; Engineering -- Data processing; Expert systems (Computer science); Periodicals; 620.00285
• DOI: 10.1016/j.engappai.2021.104233
• Languages: English
• ISSNs: 0952-1976

Abstract

Abstract: The intensity of a tropical cyclone is correlated strongly to the damage it causes when it makes landfall. Most of the time, tropical cyclones are located over the open ocean, where direct intensity measurements are difficult to obtain. An alternative approach is to estimate the tropical cyclone intensity indirectly from satellite images. In this case, there are two key points to consider: spatial and temporal relationships. For spatial relationships, the basic assumption is that cyclones with similar intensities have similar patterns. Thus, researchers can estimate intensity using pattern extraction and investigating similarities. For temporal relationships, the intensity of the cyclone is assumed to change smoothly, as a tropical cyclone is a continuous weather phenomenon. Thus, satellite images belonging to the same tropical cyclone should have a temporal (chronological) relationship with one another, meaning that the estimated intensity value of subsequent images should not change too drastically. In this research, we take advantage of these two key points and use random walk with a restart model to discover hidden correlations between target and historical cyclone images to estimate their intensity. We then use machine learning models to determine the temporal relationships among cyclone images, smoothing the prediction of the tropical cyclone event as a whole. Finally, our results show 15.77-knot root-mean-square error (RMSE) for the intensity estimation of tropical cyclones in the West Pacific Basin area.