Cyclone RI Tracking And Predication

Engineered an automated end-to-end Machine Learning pipeline to predict Tropical Cyclone Rapid Intensification (RI) — defined as a ≥30kt wind increase in 24 hours.

Ingested and processed decades of real-world historical data including NOAA NCEI IBTrACS global best-track records, SHIPS environmental predictors (SST, wind shear, OHC), and GridSat-B1 satellite IR brightness metrics.

Trained ensemble models (Random Forest and XGBoost) using a leakage-proof GroupShuffleSplit validation architecture, preventing chronological contamination of storm events across datasets.

Evaluated model integrity against the most extreme RI event on record (Hurricane Patricia), achieving an excellent AUC-ROC score of 0.916 and a Probability of Detection (POD/Recall) of 75.0%.

• Built resilient predictive models using XGBoost and Random Forest, mitigating class imbalance via scale_pos_weight.
• Integrated geospatial (Haversine) and temporal calculus to track intensity and pressure velocity changes over 6h/12h windows.
• Designed a live inference tier capable of reconstructing feature matrices from live coordinates for real-time risk assessment.
• Achieved 0.916 AUC-ROC and 75% Recall on blind test datasets.

Prediction Dashboard View

Dashboard showcasing multi-source data predictions and risk probabilities.

Data Visualization

Detailed historic geospatial tracking analysis.