Forecasting Intense Radar Reflectivity Using Machine Learning and Deep Learning Algorithms

Abstract

Over the past 50 years, numerous studies have been conducted in the Cuyo region of Argentina, South America, investigating the relationship between meteorological variables and hail precipitation. These studies have led to the development of various models aimed at classifying hydrometeors, determining their precipitation, size, and the resulting surface damage. Based on 16 years of observations using a three-radar network in the Cuyo region, this paper presents preliminary results from a hail prediction study employing machine learning and deep learning techniques applied to radar data. Algorithms random forest (RF), gradient boosting (GB) and logistic regression (LR) in addition to a recurrent neural network, were used to predict hail occurrence based on radar data. Storm cells were classified as hail or no-hail when their reflectivity reached or exceeded 55 dBZ during their evolution. Reflectivity was found to be the most suitable variable among over 50 radar variables for studying hail occurrence. Results showed that considering the temporal evolution of radar observations, by including data at different time steps (from Results showed that considering the temporal evolution of radar observations, by including data at different time steps (from t = 1 to t = 5), significantly improved the algorithms ability to predict hail occurrence). This can be attributed to both a reduction in forecast lead time and the relevance of the temporal evolution of the variables. The inclusion of global model data, such as reanalysis from ECMWF (ERA5) did not demonstrate any significant improvement in our predictions. Models such as recurrent neural networks (RNN) have the potential to deliver enhanced performance since they explicitly account for temporal dynamics.