Challenges in radar-based non-supercell tornado detection using machine learning approaches

Abstract

Tornado detection in Indonesia remains challenging as most areas are monitored by single-polarization weather radar, while dual-polarization systems offer superior detection capabilities. This study presents a novel approach by applying random forest (RF) and XGBoost machine learning algorithms to detect tornadoes using single-polarization radar data, addressing a critical gap in tropical tornado monitoring where dual-pol infrastructure is limited. Four tornado cases in Surabaya during 2024 were analyzed. Radar features including reflectivity, radial velocity, vorticity, and angular momentum were extracted through a multi-elevation sliding window technique. Spatial labels were assigned based on reports from the Indonesian National Meteorological Services (BMKG) with a 7.5 km radius from the event center. The dataset was balanced using synthetic minority over sampling technique (SMOTE). Evaluation was performed using the leave one-case-out (LOCO) scheme. Within-case evaluation showed strong performance with area under the curve (AUC) >0.94 for both models. XGBoost achieved higher probability of detection (POD 0.67-0.72) but with elevated false alarm rates (FAR up to 70%). RF demonstrated more balanced performance (POD 0.61-0.65, FAR 0.34-0.35). LOCO evaluation revealed significant POD reduction and FAR increase when tested on new cases. This indicates generalization challenges due to variability in tornado characteristics. This study demonstrates the potential of machine learning for tropical tornado early detection using readily available single-polarization radar.