Guo et al. (2026) Interval-Based Tropical Cyclone Intensity Forecasting with Spatiotemporal Transformers

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

Identification

• Journal: Remote Sensing
• Year: 2026
• Date: 2026-04-02
• Authors: Tao Guo, Hua Zhang, Tao Song, Shiqiu Peng
• DOI: 10.3390/rs18071069

Research Groups

Not explicitly stated in the paper.

Short Summary

This paper proposes TC-QFormer, an interval-based probabilistic framework for 24 h tropical cyclone intensity forecasting, which combines transformer-based spatiotemporal modeling with scalar conditioning to achieve improved deterministic accuracy and well-calibrated prediction intervals.

Objective

• To develop an interval-based probabilistic framework for 24 h tropical cyclone intensity forecasting that improves accuracy and provides well-calibrated prediction intervals, particularly for longer forecast lead times and rapidly evolving intensity regimes.

Study Configuration

• Spatial Scale: Tropical cyclone regions, inferred from geostationary satellite imagery.
• Temporal Scale: 24 h forecast lead time for intensity prediction.

Methodology and Data

• Models used: TC-QFormer (adapted PredFormer with Scalar–Image Fusion Block), ConvLSTM (baseline), PredRNN (baseline), SimVP (baseline).
• Data sources: TCIR dataset, geostationary infrared and water vapor satellite imagery, aligned historical intensity records.

Main Results

• The proposed TC-QFormer framework achieves improved deterministic accuracy in tropical cyclone intensity forecasting.
• It produces well-calibrated 80% prediction intervals.
• The method demonstrates particular effectiveness at longer forecast lead times.
• It performs strongly during rapidly evolving intensity regimes.

Contributions

• Introduction of TC-QFormer, a novel interval-based probabilistic framework for 24 h tropical cyclone intensity forecasting.
• Integration of transformer-based spatiotemporal modeling with scalar–image conditioning for enhanced forecasting.
• Adaptation of the PredFormer video prediction model for multi-horizon scalar regression and the development of a lightweight Scalar–Image Fusion Block.
• Demonstration of improved deterministic accuracy and well-calibrated probabilistic forecasts compared to representative recurrent and non-recurrent baselines.
• Provision of an effective and interpretable approach for probabilistic tropical cyclone intensity forecasting.

Funding

Not explicitly stated in the paper.

Citation

@article{Guo2026IntervalBased,
  author = {Guo, Tao and Zhang, Hua and Song, Tao and Peng, Shiqiu},
  title = {Interval-Based Tropical Cyclone Intensity Forecasting with Spatiotemporal Transformers},
  journal = {Remote Sensing},
  year = {2026},
  doi = {10.3390/rs18071069},
  url = {https://doi.org/10.3390/rs18071069}
}