Ayar et al. (2025) Ensemble random forest for tropical cyclone tracking

Identification

• Journal: Natural hazards and earth system sciences
• Year: 2025
• Date: 2025-11-24
• Authors: Pradeebane Vaittinada Ayar, Stella Bourdin, Davide Faranda, Mathieu Vrac
• DOI: 10.5194/nhess-25-4655-2025

Research Groups

• Laboratoire des Sciences du Climat et de l’Environnement, UMR 8212 CEA-CNRS-UVSQ, Université Paris-Saclay & IPSL, CEA Saclay, Gif-sur-Yvette, France
• Atmospheric, Oceanic and Planetary Physics, Department of Physics, University of Oxford, Oxford, United Kingdom
• London Mathematical Laboratory, London, United Kingdom
• Laboratoire de Météorologie Dynamique/IPSL, École Normale Supérieure, PSL Research University, Sorbonne Université, École Polytechnique, IP Paris, CNRS, Paris, France

Short Summary

This study develops and evaluates an Ensemble Random Forest (ERF) approach for tracking tropical cyclones (TCs) using a limited set of aggregated atmospheric variables. The ERF method demonstrates good performance in detecting TCs over the Eastern North Pacific and North Atlantic basins, achieving similar detection rates but significantly lower false alarm rates compared to a physics-based tracker.

Objective

• To explore the ability of a Random Forest (RF) approach to track tropical cyclones (TCs) by associating atmospheric situations, described by a limited set of aggregated predictors, with the presence or absence of TCs.
• To provide physical interpretation of the most relevant variables for TC detection using feature importance and SHAP values.

Study Configuration

• Spatial Scale: Eastern North Pacific (ENP) and North Atlantic (NATL) basins. Atmospheric variables extracted at 0.25° x 0.25° resolution and aggregated within 20° x 10° overlapping boxes.
• Temporal Scale: 1980–2021 period, with data processed at 6-hourly time steps.

Methodology and Data

• Models used:
  • Ensemble Random Forest (ERF) for binary classification of TC occurrence.
  • UZ algorithm (physics-based detection scheme) for comparison.
• Data sources:
  • Observed TC tracks: International Best Track Archive for Climate Stewardship (IBTrACS) "since 1980" set.
  • Atmospheric variables: ERA5 reanalysis (0.25° x 0.25° resolution, 6-hourly data). Five variables used:
    • Mean sea level pressure (MSLP, in Pa)
    • 10 m wind intensity (UV10, in m s⁻¹)
    • Total column water vapour (TCWV, in kg m⁻²)
    • Relative vorticity at 850 hPa pressure level (RV850, in s⁻¹)
    • Geopotential thickness between 300 and 500 hPa pressure level (THZ300_Z500, in m)
  • For each variable within a box, four statistics were computed: minimum, mean, maximum, and standard deviation, resulting in 20 predictors.

Main Results

• The ERF tracker achieved a Probability of Detection (POD) of 77.5 % and a False Alarm Rate (FAR) of 8.7 % for the ENP basin, and a POD of 77.8 % and FAR of 7.9 % for the NATL basin in validation experiments.
• Compared to the UZ physics-based tracker, ERF showed similar PODs but significantly lower FARs (UZ: 24.1 % for ENP, 15 % for NATL).
• ERF-detected tracks exhibited similar duration frequencies to IBTrACS tracks, with discrepancies primarily for short-duration (1 to 3 days) and lower-intensity TCs, which were often missed or constituted false alarms.
• ERA5 reanalysis systematically underestimated TC intensity, leading to detected TCs being weaker than observed ones. Missed tracks were often associated with minimum pressures above 1.005 x 10⁵ Pa and maximum winds below 16 m s⁻¹ in ERA5.
• The six most important predictors for TC detection were RV850 standard deviation, MSLP minimum, UV10 maximum, RV850 maximum, THz300_z500 maximum, and TCWV maximum, all physically relevant for TCs.
• SHAP-based partial dependency plots revealed that TCs can be detected through diverse combinations of predictor values, offering greater flexibility than rigid threshold-based approaches.
• Regional ERF trackers showed specificities between basins, suggesting potential benefits of basin-specific models. Ablation experiments indicated that while removing less important predictors had minimal impact on POD, it significantly degraded FAR, highlighting their role in controlling false alarms.

Contributions

• Introduction of an Ensemble Random Forest (ERF) approach for tropical cyclone tracking that effectively handles imbalanced data through subsampling.
• Demonstration of ERF's superior performance in terms of False Alarm Rate compared to established physics-based trackers (UZ algorithm), while maintaining comparable detection rates.
• Development of a computationally efficient tracking method that uses a limited number of aggregated atmospheric variables, enhancing transferability to climate models and reducing computational demands compared to deep learning approaches.
• Provision of physical interpretability of TC detection through feature importance and SHAP values, identifying key atmospheric drivers and their non-linear contributions to TC presence.
• Validation of the method's good temporal and spatial generalization capabilities across different basins.

Funding

• European Union’s Horizon 2020 research and innovation program, project XAIDA: Extreme Events – Artificial Intelligence for Detection and Attribution (grant agreement no. 01003469).
• National Research Agency under France 2030, reference ANR-22-EXTR-0005 (TRACCS-PC4-EXTENDING project).
• INSU-CNRS-LEFE-MANU grants (projects CROIRE and COESION).
• Institut Pascal at Université Paris-Saclay, program TROPICANA, reference ANR-11-IDEX-0003-01.
• NERC-NSF research grant no. NE/W009587/1 (NERC) & AGS-2244917 (NSF) HUrricane Risk Amplification and Changing North Atlantic Natural disasters (Huracan).
• EUR IPSL-Climate Graduate School, ICOCYCLONES2 project, managed by the ANR under the “Investissements d’avenir” programme, reference ANR-11-IDEX-0004-17-EURE-0006.

Citation

@article{Ayar2025Ensemble,
  author = {Ayar, Pradeebane Vaittinada and Bourdin, Stella and Faranda, Davide and Vrac, Mathieu},
  title = {Ensemble random forest for tropical cyclone tracking},
  journal = {Natural hazards and earth system sciences},
  year = {2025},
  doi = {10.5194/nhess-25-4655-2025},
  url = {https://doi.org/10.5194/nhess-25-4655-2025}
}