Zonda wind classification using machine learning algorithms

Zonda wind is a typical downslope windstorm over the eastern slopes of Central Andes, in Argentina, which produces extremely warm and dry conditions creating substantial socioeconomic impacts. To achieve the Zonda wind classification, objective methods based on supervised machine learning (ML) algorithms are used. ML training and supervision is based on the subjective Zonda wind classification assessing the total hourly data that correspond to Zonda wind observations for three surface stations longtime series. ML algorithms includes; the linear discriminant analysis (LD), linear support vector machine (SVM), k nearest neighbours (k-NN), logistic regression (LR) and classification trees. Metrics obtained from the confusion matrix are used to compare the models' skills in class separation. Considering event-based statistics, the obtained probability of detection values locate all models above 85% with a probability of false detection lower than 0.523% and a missing ratio below 15%. From an alarm-based perspective, algorithms show values below 11.42% in false alarm rate, lower than 0.7% in missing alarm ratio and higher than 88.85% in correct alarm ratio. The false negative rate occurs mostly from August to December, where the onset time of the events presents greater difficulty in the classification than the offset, while the false alarm increases in June and October months. Models skills reveal that k-NN, SVM and LR are better discriminators than LD and classification tree. The high efficiency of these models indicates that ML classification models could be used for the phenomenon diagnosis.