Hurricanes, which are growing in complexity and strength in an increasingly warming world, are one of the worst natural disasters of the twenty-first century. Therefore, it is crucial to conduct research that integrates the changing characteristics of hurricanes, as well as methods that allow learning complex patterns to predict future events.

The AI4OCEANS group of the Image Processing Laboratory (IPL) at the University of Valencia has developed a study that provides new data related to the anatomy and temperature of the cloud system, and analyses them using machine learning algorithms –automatic learning– for diagnosing the possible transition to a major hurricane status. The final accuracy of the prediction is 79% for up to 54 hours in advance for severe hurricane events over the Atlantic and Pacific oceans.

“Machine learning allows us to analyse the characteristics of the prominent clouds of a tropical cyclone at the early stage and establish non-linear relationships between the process variables to diagnose its possible progression to a severe hurricane”, explains climatologist Verónica Nieves, CIDEGENT researcher, responsible for AI4OCEANS and project leader. For this study, the team used data from the GOES spacecraft, a geostationary satellite system developed by NASA for climate and weather prediction in the Western Hemisphere.

Results have been obtained using a spatial pattern extraction algorithm –k-means– applied to the GOES images. In addition, the parameters have been optimally combined using an algorithm called ‘random forest’, a machine learning technique widely applied to diverse problems to classify events. “This is the first time that this unique set of variables has been combined using a hybrid AI-based approach,” says PhD student Javier Martínez-Amaya, a member of the team.

The evaluation and diagnosis framework proposed in this study published in Remote Sensing allows the integration of other characteristics or factors of the effects of environmental variations, such as seasonality, for future analysis. “We have created an adaptable system capable of adjusting parameters as new processes are incorporated”, adds Cristina Radin, also a member of the AI4OCEANS group and co-author of the article.

This study was carried out within the framework of project MALOPH: A novel Machine Learning based perspective to identify and model Ocean Precursors to extreme Hurricane development, funded by the European Space Agency (ESA).

Verónica Nieves’ team focuses its work on the next generation of advanced algorithms for analysing Earth observation data. The researcher has recently participated in the preparation of an international report by the Group of Experts on Artificial Intelligence for the Prediction of the Earth System (Chapter: AI4ESP Coastal, ocean and ice dynamics). Prepared by more than 740 participants from 178 institutions around the world, the text refers to the important role that AI and machine learning play in improving integrated models that incorporate complex natural processes to aid decision-making, one of the emerging scientific challenges in this field.

Reference:

Advanced Machine Learning Methods for Major Hurricane Forecasting. Javier Martinez-Amaya, Cristina Radin, Veronica Nieves. Remote Sens. 2022, 15(1), 119; https://doi.org/10.3390/rs15010119