A Comparison of Machine Learning Approaches for Classification of Radar-Derived Convective Clusters

Parker and Johnson (2000) define a mesoscale convective system (MCS) as an assemblage of thunderstorms that persists for at least 3 hours and contains a contiguous or semi-contiguous convective (i.e., ≥40 dBZ) area of at least 100 km along the system’s major axis.  Although this definition is widely used, it includes a large subset of precipitation clusters that would not typically be considered MCSs, such as tropical storms and synoptic systems with areas of embedded convection.  Differentiation between MCSs and non-MCSs that fulfill the Parker and Johnson (2000) criteria could depend on several visual features and combinations of visual features (i.e., intensity gradient, size, eccentricity, etc.)  This work presents an exploration of this classification problem.  First, feature selection and summary statistics of human-labeled MCS and non-MCS clusters is discussed.  Next, results from the training and testing of several of machine learning algorithms are presented.  Finally, a visual and objective exploration of the best performing approaches are provided.