Ulfa Qurrotul Aini, Halmar Halide
Tornadoes, locally known as puting beliung, are small-scale extreme weather phenomena that frequently occur in Indonesia and rapidly develop from strong convective systems. Their short duration and limited spatial scale make them difficult to predict using conventional deterministic approaches. This study aims to develop a probabilistic classification model based on machine learning using surface atmospheric observational variables and to evaluate its capability in distinguishing between confirmed and probable/possible event categories.
Hourly observational data for the period 2015–2024 were obtained from SYNOP reports at the BANDUNG/HUSEIN station (WMO 96781), including air temperature, dew point temperature, relative humidity, pressure tendency, wind speed, total cloud cover, and mean sea level pressure. Event data were referenced from the Indonesia Tornado Database (2025) and used as the basis for labeling. The models were developed using logistic regression and ensemble decision tree approaches with several forecast lead-time scenarios. Model performance was evaluated using ROC curves, AUC values, and skill scores suitable for rare event prediction.
The findings of this study are expected to contribute to strengthening observation-based weather warning systems in the study area.
Keywords: strong convection, probabilistic classification, logistic regression, ensemble methods, forecast verification, rare events, short-term forecasting.
Â
