An automated procedure for classifying rainfall systems (meso-a scale and larger) was developed using an operational analysis of hourly precipitation estimates from radar and rain gauge data. The development process followed two main phases: a training phase and a testing phase. First, 48 hand-selected cases were used to create a training dataset, from which a set of attributes related to morphological aspects of rainfall systems were extracted. A hierarchy of classes for rainfall systems, in which the systems are separated into general convective (heavy rain) and nonconvective (light rain) classes, was envisioned. At the next level of classification hierarchy, convective events are divided into linear and cellular subclasses, and nonconvective events belong to the stratiform subclass. Essential attributes of precipitating systems, related to the rainfall intensity and degree of linear organization, were determined during the training phase. The attributes related to the rainfall intensity were chosen to be the parameters of the gamma probability distribution fit to observed rainfall amount frequency distributions using the generalized method of moments. Attributes related to the degree of spatial continuity of each rainfall system were acquired from correlogram analysis. Rainfall systems were categorized using hierarchical cluster analysis experiments with various combinations of these attributes. The combination of attributes that resulted in the best match between cluster analysis results and an expert classification were used as the basis for an automated classification procedure. The development process shifted into the testing phase, where automated procedures for identifying and classifying rainfall systems were used to analyze every rainfall system in the contiguous 48 states during 2002. To allow for a feasible validation, a testing dataset was extracted from the 2002 data. The testing dataset consisted of 100 randomly selected rainfall systems larger than 40 000 km(2) as identified by an automated identification system. This subset was shown to be representative of the full 2002 dataset. Finally, the automated classification procedure classified the testing dataset into stratiform, linear, and cellular classes with 85% accuracy, as compared to an expert classification.
