Evrim Yazgin
Cosmos science journalist
A new artificial intelligence (AI) model based on “contrastive learning” has shown promise as a tool that can forecast rapid intensification of tropical cyclones.
Rapid intensification is when cyclone’s maximum sustained wind speed increases by 13m/s within 24 hours. While rapid intensification periods only account for about 5% of the total life of the tropical cyclone, these bursts of cyclone energy can pose a serious threat because they are difficult to forecast.
Tropical cyclones in the southern hemisphere, called hurricanes in the north, can cause devastation to lives, homes and businesses.
Cyclone season in Australia is November to April. On January 19, a tropical low off the coast of Western Australia had developed into Tropical Cyclone Sean – the second tropical cyclone of the 2024–25 season in Australia according to NASA.
Traditional forecasting methods often fail to predict rapid intensification.
AI has been considered before to sharpen predictions but has previously struggled with false alarms and limited reliability.
The new tool, developed at the Institute of Oceanology of the Chinese Academy of Sciences (IOCAS) might change that. The results are published in the Proceedings of the National Academy of Sciences journal.
“Forecasting RI TC [rapid intensification tropical cyclone] periods requires the analysis of atmospheric and oceanic environmental factors, such as wind, relative humidity, air temperature, and sea surface temperature (SST), as well as the development processes of TCs, such as eyewall replacement, and RI of deep convection,” the authors write.
Their model has 2 inputs: a known sample of tropical cyclone rapid intensification periods and an unknown sample to be forecast. Each unknown sample is compared with 10 known samples.
The model also learns to differentiate periods of rapid intensification using satellite imagery and atmospheric and oceanic data.
It was tested on data from the Northwest Pacific between 2020 and 2021.
The AI predicted rapid intensifications with an accuracy of 92.3% – an 11.7% improvement on existing techniques.
False alarms were reduced to just 8.9%. Current methods have a false alarm rate of 27.2% – more than 3 times higher than the new tool.
“This study addresses the challenges of low accuracy and high false alarm rates in RI TC forecasting,” says corresponding author LI Xiaofeng. “Our method enhances understanding of these extreme events and supports better defences against their devastating impacts.”
