This paper presents a scheme that classifies reflectivity measurements from ground-based volume scanning radar into convective and stratiform precipitation types. The algorithm developed in this paper uses the neural network approach as a tool for classification. The algorithm is trained and validated on radar data obtained from 18 Weather Surveillance (WSR-88D) Doppler Radar sites in the southern United States, sampled over the period December 1997 to October 1999. For the training and validation of the neural network, classification information from a hybrid approach based on Tropical Rainfall Measurement Mission (TRMM) Precipitation Radar (PR) classification products and low-level ground radar reflectivity fields is taken as reference. The approach utilises PR's ability to observe the vertical structure of storms with high resolution (250 metres) and WSR-88D's ability to observe the low-level horizontal reflectivity magnitude and spatial variability. The method was applied to an extended dataset consisting of storm cases associated with squall line convection, mesoscale convective system, and widespread stratiform precipitation with embedded convection. The scheme is shown to have an overall 75% probability of detection, 18% false alarm rate, 67% critical success ratio, and 0.554 in prediction correlation (Cramer's V coefficient). It has an overall 4% overestimation and 27% underestimation of stratiform and convective rain areas, respectively. The classification results are compared to two current methods and improvements are shown in terms of the above statistical measures and through visual comparison with PR classification fields. The improved classification scheme derived herein will help establish more accurate radar reflectivity-rainfall relationships and improve the retrieval of diabatic beating in different cloud systems.
