HIERARCHICAL-CLUSTERING OF ATMOSPHERIC SOUNDINGS

This is a procedural paper that compares six commonly used hierarchica l clustering techniques applied to atmospheric soundings. The followin g techniques are compared: single-, complete-, and two average-linkage techniques, the Ward's technique, and one centroid type technique. Ve rtical profiling of the atmosphere is based on common soundings sample d on synoptic scales over the Northern Hemisphere. At each analysis po int, the atmospheric profile is represented by an 'atmospheric state v ector' consisting of surface pressure, total column (precipitable) wat er, and temperature, wind and relative humidity values at 12 discrete vertical levels. An evaluation of the techniques is performed by compa ring the mean state vectors of final clusters. Four techniques produce comparable results, with the largest differences between the techniqu es appearing for the surface pressure component of the cluster state v ector. Also, large differences are found in the temperature and wind v ector components at the 850 hPa pressure level, and the smallest diffe rences are found for the relative humidity components. The results sup port findings of studies on air-mass typing on local or regional scale s: the average-linkage method based on group average provides the most distinct and homogeneous clusters.