Hang et al. (2026) CycloneMAE: A Scalable Multi-Task Learning Model for Global Tropical Cyclone Probabilistic Forecasting

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: arXiv (Cornell University)
• Year: 2026
• Date: 2026-04-14
• Authors: Renlong Hang, Zihao Xu, Jiuwei Zhao, Runling Yu, Leye Cheng, Qingshan Liu
• DOI: None

Research Groups

Not specified in the provided text.

Short Summary

CycloneMAE is a multi-task masked autoencoder model that leverages multi-modal data to provide scalable, probabilistic, and interpretable tropical cyclone forecasts. It outperforms leading numerical weather prediction (NWP) systems in predicting pressure, wind, and track across five global ocean basins.

Objective

• To develop a scalable, multi-task forecasting framework that overcomes the computational intensity of NWP models and the lack of generalization/probabilistic output in existing deep learning models for tropical cyclones.

Study Configuration

• Spatial Scale: Global (evaluated across five global ocean basins).
• Temporal Scale: Forecast horizons up to 120 hours.

Methodology and Data

• Models used: CycloneMAE (a TC structure-aware masked autoencoder), discrete probabilistic gridding mechanism, and a pre-train/fine-tune paradigm.
• Data sources: Multi-modal data, specifically including satellite imagery.

Main Results

• Performance: Outperforms leading NWP systems in pressure and wind forecasting for horizons up to 120 hours, and in track forecasting for horizons up to 24 hours.
• Interpretability: Attribution analysis using integrated gradients demonstrates that short-term forecasts rely on internal core convective structures, while long-term forecasts shift focus toward external environmental factors.
• Capability: Simultaneously delivers both deterministic forecasts and probability distributions.

Contributions

• Establishes a scalable and interpretable pathway for operational TC forecasting.
• Introduces a multi-task approach that generalizes across different forecasting variables, moving beyond variable-specific deterministic models.
• Integrates a structure-aware masked autoencoder with probabilistic gridding to bridge the gap between deep learning and traditional NWP.

Funding

Not specified in the provided text.

Citation

@article{Hang2026CycloneMAE,
  author = {Hang, Renlong and Xu, Zihao and Zhao, Jiuwei and Yu, Runling and Cheng, Leye and Liu, Qingshan},
  title = {CycloneMAE: A Scalable Multi-Task Learning Model for Global Tropical Cyclone Probabilistic Forecasting},
  journal = {arXiv (Cornell University)},
  year = {2026},
  url = {https://openalex.org/W7154655542}
}