Convective Weather Characterization and Prediction Using Machine Learning Algorithms: An Analysis for the Amazon Region

Abstract This study develops an objective tool for predicting convective storms (CS) in the Terminal Maneuvering Area of Manaus (TMA-Manaus), Brazil, using machine-learning (ML) algorithms applied to thermodynamic indices derived from the 12 UTC Manaus radiosonde. Atmospheric discharges (AD) are used as both a proxy and a severity metric for CS activity, and five AD-based thresholds are defined to characterize increasing levels of convective intensity. A 4-h forecast horizon was adopted, consistent with the interval between the pre-convective sounding and the afternoon lightning peak over the region. Feature-selection analysis identified a physically meaningful subset of 16 sounding-derived predictors, with Precipitable Water, Lifted Index, CAPE, and the Bulk Richardson Number among the most relevant variables for CS prediction. At the lowest threshold (DR-AD > 75.5 day -1 ), the Quadratic Discriminant Analysis (QDA) model achieved the best overall result with the 24-feature configuration, reaching a Probability of Detection (POD) of 0.99 and a False Alarm Ratio (FAR) of 0.20. Models based on reduced predictor sets also retained competitive skill, confirming that dimensionality reduction can preserve the essential physical information required for operational forecasting. A permutation-importance analysis further supported the physical consistency of the selected predictors and provided model-agnostic interpretability. Although predictive performance gradually decreased as the AD threshold increased, useful skill was retained even under more imbalanced and severe-event conditions. This represents the first radar-free, AD-based convective forecast framework for the Amazon region that satisfies Brazil’s aeronautical nowcasting lead-time standard (ICA 100–12). The proposed method is suitable for real-time implementation wherever a 12 UTC sounding and automated feature computation are available.