01-05 December 2025
INCOIS, Hyderabad, India.
| Abstract Submission No. | ABS-05-0316 |
| Title of Abstract | Interpretable Machine Learning for Tropical Cyclone Genesis and Track Prediction over the Bay of Bengal |
| Authors | Anuj Gupta*, Debasis Kundu, Sreenivas Pentakota, Pushpalatha T., P. Gopi Krishna |
| Organisation | University of Hyderabad |
| Address | 1400 Master Colony, O/S Bada Goan Gate Jhansi, Uttar Pradesh, India Pincode: 284002 E-mail: thisisanujgupta@gmail.com |
| Country | India |
| Presentation | Oral |
| Abstract | Forecasting the onset and path of tropical cyclones remains a substantial hurdle in operational meteorology, particularly over the Bay of Bengal (BoB), where storms often develop quickly and impact densely populated coastlines. The present research investigates the use of three distinct machine learning (ML) models, viz. Random Forest, Support Vector Machine, and Neural Network-to anticipate both the formation and movement of tropical cyclones at lead times of 24, 36, 60, and 84 hours. Initial detection of tropical disturbances is achieved using a refined Kalman Filter approach, with relevant atmospheric and oceanic parameters extracted from reanalysis data. The ML models exhibit strong classification performance in identifying developing systems. At 24- and 36-hour lead times, the F1-score approaches 0.90, and remains robust (~0.86) at longer forecasts of 60 and 84 hours. In addition to genesis prediction, storm trajectories are estimated by predicting the future positions of disturbances, yielding accurate displacement forecasts over all tested lead times. To improve interpretability, SHapley Additive exPlanations (SHAP) are applied to evaluate the role of individual predictors. It is well established that, CAPE emerges as the most influential variable, with wind shear and mid-level vorticity also playing important roles in both genesis and track prediction. To illustrate practical utility, a case study of Cyclone Titli is presented, where we have shown role of distinct cyclogenesis parameters in affecting both the probability of genesis and projected track. The present results enhance transparency, making ML outputs more accessible for forecasters. The study also discusses key challenges such as data imbalance and generalization across seasons, and recommends combining ML with physical modeling to further improve reliability. Overall, the results support the integration of interpretable ML frameworks into tropical cyclone forecasting over the BoB. |
| Are you part of IIOE-2 endorsed project | no |
| Keywords | ML, tropical cyclones, Bay of Bengal, ocean and atmosphere data |
| For Awards | no |