Prediction of tropical cyclone in Bangladesh using ENSO index through ensemble learning technique

Tropical cyclones have catastrophic impacts on life and infrastructure in many places on the earth. Accurate prediction is therefore crucial for cyclone-prone regions like Bangladesh. This study employs an ensemble learning approach to predict tropical cyclones using relationships between El Niño Southern Oscillation (ENSO) indices and cyclone occurrences. Cyclone data from 1977 to 2022 reveal severe class imbalance, with only 26 cyclone events in 540 months. A voting ensemble model with 17 Random Forest classifiers was developed through random under-sampling, each trained with a Random Forest classifier. Predictions were aggregated using majority voting to improve robustness. Model performance was evaluated with accuracy, precision, recall, and negative predictive value (NPV), supplemented with Wilson Score Intervals and a sensitivity analysis across probability thresholds from 10–90%. Results show that the ensemble model achieved 75% overall accuracy, with high recall (80%) for cyclone months but relatively low precision (22%), favoring sensitivity to minimize missed events. Correlation analysis reveals strong seasonal associations between cyclone activity and ENSO indices, particularly ONI and Niño 3.4 SST during monsoon and post-monsoon periods. These findings demonstrate that ENSO-based ensemble learning can capture non-linear climate–cyclone relationships and enhance early warning capabilities. The proposed framework provides a data-driven tool that can support meteorological agencies and disaster managers in reducing cyclone-related impacts on vulnerable coastal communities.