EGU25-13891, updated on 15 Mar 2025
https://doi.org/10.5194/egusphere-egu25-13891
EGU General Assembly 2025
© Author(s) 2025. This work is distributed under
the Creative Commons Attribution 4.0 License.
Ensemble Random Forest for Tropical Cyclone Tracking
Pradeebane Vaittinada Ayar1, Stella Bourdin2, Davide Faranda1, and Mathieu Vrac1
Pradeebane Vaittinada Ayar et al.
  • 1CNRS, LSCE-IPSL, Gif-sur-Yvette, France (pradeebane@laposte.net)
  • 2Atmospheric, Oceanic and Planetary Physics, Department of Physics, University of Oxford, Oxford, United Kingdom


Even though tropical cyclones (TCs) are well documented from the moment they reach a certain intensity to the moment they start to evanesce, many physical and statistical properties governing them are not well captured by gridded reanalysis or simulated by earth system models. Thus, the tracking of TCs remain a matter of interest for the investigation of observed and simulated tropical. Many cyclone tracking schemes are available. On the one hand, there are trackers that rely on physical and dynamical properties of the TCs and users prescribed thresholds, which make them rigid, and need numerous variables that are not always available in the models. On the other hand, there are trackers leaning on deep learning which, by nature, need large amounts of data and computing power. Besides, given the number of physical variables needed for the tracking, they can be prone to overfitting, which hinders their transferability to climate models. In this study, the ability of a Random Forest (RF) approach to track TCs with a limited number of aggregated variables is explored. Hence, it becomes a binary supervised classification problem of TC-free (zero) and TC (one) situations. Our analysis focuses on the Eastern North Pacific and North Atlantic basins, for which respectively 514 and 431 observed tropical cyclones tracks record are available from the IBTrACS database over the 1980-2021 period. For each 6-hourly time step, RF associates TC occurrence or absence (1 or 0) to atmospheric situations described by predictors extracted from the ERA5 reanalysis. Then situations with TC occurrences are joined for reconstructing TC trajectories. Results show good ability of the method for tracking of tropical cyclones over both basins and good ability for spatial and temporal generalization as well. It also shows similar TC detection rate as trackers based on TCs' properties and significantly lower false alarm rate. RF allows us to detect TC situations for a range of predictor combinations, which brings more flexibility than threshold based trackers. Last but not least, this study shed light on the most relevant variables allowing to detect tropical cyclone.

How to cite: Vaittinada Ayar, P., Bourdin, S., Faranda, D., and Vrac, M.: Ensemble Random Forest for Tropical Cyclone Tracking, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-13891, https://doi.org/10.5194/egusphere-egu25-13891, 2025.