Articles In Press
Back to the articles list |
Back to browse issues page
1- RTU MIREA (Russian Technological University), Moscow, Russia
Abstract: (14 Views)
Rapid intensification (RI) of tropical cyclones remains a key source of forecast uncertainty and a major driver of rapid hazard escalation in early warning systems. We propose a lightweight temporal mixing neural model for probabilistic RI prediction from a compact set of physically interpretable, storm center predictors: Sea Surface Temperature (SST), Mean Sea Level Pressure (MSLP) and vertical wind shear, augmented with 6-24 h lagged values. We also consider storm-wise splits by cyclone identifier and report discrimination with ROC-AUC and average precision, probabilistic accuracy via Brier score and decision-oriented verification with POD and FAR metrics. We apply post-hoc Platt scaling on the validation split and assess calibration with reliability diagrams. On the held-out test split the model achieves strong discrimination with ROC-AUC approximately 0.79 and AP metric approximately 0.20, Brier score approximately 0.054. Based on SHAP analysis, MSLP- and SST- related features identified as most the impactful features for the forecast. All experiments were carried out on a single graphics processing unit (GPU) NVIDIA RTX 5080 Ti. Therefore, the proposed approach is suitable for environments with limited computational resources.
Keywords: tropical cyclone, rapid intensification, early warning, disaster risk reduction, intensity forecast, TimeMixerDrop, ERA5, IBTrACS.
| Rights and permissions | |
 |
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License. |