ABSTRACT

A K-means clustering algorithm was used to classify Tropical Rainfall Measuring Mission (TRMM) Precipitation Radar (PR) scenes within 1° square patches over the tropical (15°S-15°N) oceans. Three cluster centroids or "regimes" that minimize the Euclidean distance metric in a five-dimensional space of standardized variables were sought [convective surface rainfall rate; ratio of convective rain to total rain; and fractions of convective echo profiles with tops in three fixed height ranges (<5, 5-9, and >9 km)]. Independent cluster computations in adjacent ocean basins return very similar clusters in terms of PR echo-top distributions, rainfall, and diabatic heating profiles. The clusters consist of shallow convection (SHAL cluster), with a unimodal distribution of PR echo tops and composite diabatic heating rates of 2 K day-1 below 3 km; midlevel convection (MID-LEV cluster), with a bimodal distribution of PR echo tops and 5 K day-1 heating up to about 7 km; and deeper convection (DEEP cluster), with a multimodal distribution of PR echo tops and >20 K day-1 heating from 5 to 10 km. Each contributes roughly 20%-40% in terms of total tropical rainfall, but with MID-LEV clusters especially enhanced in the Indian and Atlantic sectors, SHAL relatively enhanced in the central and east Pacific, and DEEP most prominent in the western Pacific. While the clusters themselves are quite similar in rainfall and heating, specific cloud types defined according to the PR echo top and surface rainfall rate are less similar and exhibit systematic differences from one cluster to another, implying that the degree to which precipitation structures are similar decreases when one considers individual precipitating clouds as repeating tropical structures instead of larger-scale cluster ensembles themselves.