New machine learning method accurately estimates extreme tropical cyclone rainfall with lower computational cost

01 Oct 2026

Tropical cyclones are the most damaging meteorological hazard that we have in the Philippines and most damage is attributable to the extreme rainfall they bring. The more intense a tropical cyclone is, by definition, the stronger the winds. Also, it will bring more extreme rain in general.

One important way to mitigate the impending impacts of tropical cyclones is through forecasting their rainfall. One way to do this is with numerical weather models. Weather models make use of observations and calculate complex mathematical equations to simulate future weather. However, the complexity of the equations requires these models to run on high-performance computers, which we lack in the country.

Another method to predict rainfall induced by tropical cyclones is the analog method. It is based on the idea that if there is an impending typhoon, we look for analogous historical typhoons and assume their rainfall will be similar to that of the impending typhoon.

In this study, we developed a new method using machine learning to estimate rainfall generated by a tropical cyclone by training it to look at historical typhoon events and the corresponding rainfall recorded. We developed a rainfall forecasting model based on the analog method using two machine learning techniques, namely the self-organizing maps to group tropical cyclones with similar tracks and the random forest to calculate rainfall projections. We trained our machine learning model with tropical cyclones from 1951 to 2015, together with the associated rainfall for each cyclone. We then tested and evaluated the model for typhoons from 2016 to 2020. So the next time there is a typhoon, the model looks at similar events in the past and calculates what the likely rainfall will be from that typhoon. Moreover, the more cyclones we put in its database, the more it learns and changes the way it computes and forecasts rainfall.

The model assessment shows that the rainfall prediction skills of the model are comparable to a dynamical model while requiring significantly lower computational requirements (dynamical models require hours on advanced computers, while this model runs on a laptop within a few minutes or less). In addition, the skill of our model is better for the most extreme rainfall, which can potentially make it useful in determining extreme events.

Authors: Cris Gino Mesias and Gerry Bagtasa (Institute of Environmental Science & Meteorology, University of the Philippines Diliman)

Read the full paper: https://doi.org/10.1002/met.70083

Image by Jobelle Meana from Pexels

New machine learning method accurately estimates extreme tropical cyclone rainfall with lower computational cost

Tropical cyclones are the most damaging meteorological hazard that we have in the Philippines and most damage is attributable to the extreme rainfall they bring. The more intense a tropical cyclone is, by definition, the stronger the winds. Also, it will bring more extreme rain in general.

One important way to mitigate the impending impacts of tropical cyclones is through forecasting their rainfall. One way to do this is with numerical weather models. Weather models make use of observations and calculate complex mathematical equations to simulate future weather. However, the complexity of the equations requires these models to run on high-performance computers, which we lack in the country.

Another method to predict rainfall induced by tropical cyclones is the analog method. It is based on the idea that if there is an impending typhoon, we look for analogous historical typhoons and assume their rainfall will be similar to that of the impending typhoon.

In this study, we developed a new method using machine learning to estimate rainfall generated by a tropical cyclone by training it to look at historical typhoon events and the corresponding rainfall recorded. We developed a rainfall forecasting model based on the analog method using two machine learning techniques, namely the self-organizing maps to group tropical cyclones with similar tracks and the random forest to calculate rainfall projections. We trained our machine learning model with tropical cyclones from 1951 to 2015, together with the associated rainfall for each cyclone. We then tested and evaluated the model for typhoons from 2016 to 2020. So the next time there is a typhoon, the model looks at similar events in the past and calculates what the likely rainfall will be from that typhoon. Moreover, the more cyclones we put in its database, the more it learns and changes the way it computes and forecasts rainfall.

The model assessment shows that the rainfall prediction skills of the model are comparable to a dynamical model while requiring significantly lower computational requirements (dynamical models require hours on advanced computers, while this model runs on a laptop within a few minutes or less). In addition, the skill of our model is better for the most extreme rainfall, which can potentially make it useful in determining extreme events.

Authors: Cris Gino Mesias and Gerry Bagtasa (Institute of Environmental Science & Meteorology, University of the Philippines Diliman)

Read the full paper: https://doi.org/10.1002/met.70083

Image by Jobelle Meana from Pexels