Application of Random Forest Algorithm on Tornado Detection

Tornadoes are highly destructive small-scale extreme weather processes in the troposphere. The radar is one of most effective remote sensing devices for monitoring and early warning tornadoes. existing tornado detection algorithms based on data unsupervised have strict multi-altitude constraints, such as algorithm vortex signatures (TDA-TVS), which may lead to high false alarm rates, performance greatly affected by quality control algorithm. A novel TDA-RF random forest (RF) classification proposed real-time identification S-band China new generation Doppler (CINRAD-SA). uses velocity features identify tornadoes adds related reflectivity spectrum width level-II data. Historical CINRAD-SA from 2006–2015 used construct dataset train model. evaluated using five 2016–2020 with enhanced Fujita(EF) scale ratings ranging EF0 EF4 distances 10 130 km radar. performs well overall probability (POD), ratio (FAR), critical success index (CSI) 71%, 29%, 55%, respectively. Moreover, improves POD CSI, reduces FAR compared TDA-TVS. maximum early-warning time 17 min, average 6 min; can provide according development process facilitate tracking ability.