Convective weather characterization and prediction using Machine Learning algorithms: analysis for Amazon Region

This study presents an objective tool for predicting convective storms (CS) in the Terminal Maneuverer Area of Manaus (TMA-Manaus), Amazon, Brazil, using Machine Learning (ML) algorithms. The occurrence and severity of CS events were characterized by atmospheric discharges (AD) using thresholds. The prediction models leverage AD thresholds as target and 12Z radiosonde data as predictors. The AD climatology between 2012 and 2017 for TMA-Manaus revealed that AD occurs in every month of the year, in this period there were only in 21 days without lightning. The analysis of feature importance for classifying CS stages revealed that the Showalter index, Bulk Richardson Number, Convective Available Potential Energy, Lifted index, and Equilibrium Level are the most relevant thermodynamic indices for classifying the convective state in the region. Results indicate that, for a small amount of AD (69/day), the mean POD and FAR for the ten selected models were 0.92±0.06 and 0.19±0.01, respectively. The QDA algorithm showed the best performance with a POD of 0.99 and a FAR of 0.19. However, as the AD threshold increased up to 5,000 AD per day, a decrease in model performance was observed, because Severe CS is rare compared to CS. The findings suggest that ML models using radiosonde data as predictors are only capable of predicting the occurrence or not of CS with relative accuracy, but the models are not capable of classifying whether it will be severe. The development of this tool marks a significant step towards improving the accuracy and timeliness of CS forecasts, thereby enhancing safety and efficiency of flights in the region.