Pattern classification of typhoon tracks and related large-scale circulations

Tropical cyclone (TC) tracks in the western North Pacific (WNP) are classified using the fuzzy c-means clustering method (FCM). The FCM is distinctive in that it does not do hard partitioning directly, but yields membership coefficients to whole clusters. The membership coefficient is a distance measure how close a data object is located from a cluster center. The FCM is suitable especially for the dataset of which cluster boundaries are so vague that conventional clustering methods (doing hard partitioning) may not be appropriate, e.g., the set of TC tracks.

A total of 855 tracks during the TC season (June through October) for the period from 1965 through 2006 reveals optimum partition when the cluster number is set to seven. This decision is made objectively by cross-checking five widely used validity measures. If each TC is put into the cluster where the membership coefficient is the largest, the seven hard clusters (C1–C7) show distinctive features in their tracks; TCs striking Korea and Japan (C1), those affecting Japan with long lifetime (C2), those towards Taiwan and East China (C3), those recurving east of Japan (C4) and farther east thereof (C5), those over the South China Sea (C6) and those moving straight across the Philippines (C7).

As for each cluster, the anomalous large-scale environments, such as sea surface temperatures, 850-hPa flows and relative vorticity, 500-hPa geopotential height and tropospheric layer-mean flows are examined. We also investigate the roles of the natural oscillatory phenomena, such as El Niño/Southern Oscillation (ENSO), central Pacific El Niño, the stratospheric quasi-biennial oscillation (QBO) and the Madden–Julian oscillation (MJO). While C1, C2 and C3 are apparently modulated by the ENSO-type tropical SST patterns, C4 is strongly modulated by the QBO. The MJO also modulates the TCs in each cluster. The impact of each cluster on the coastal areas is shown through TC-induced rainfall as well as TC landfalls. Also, the seasonal forecasting of the TC tracks patterns is attempted using the statistical-dynamical method based on the statistical relationships between large-scale circulations and track patterns and the NCEP Climate Forecast System results.