Statistical-dynamical typhoon intensity predictions in the western North Pacific based on track pattern clustering, decision tree and ocean coupling predictors

Abstract

A statistical-dynamical model for predicting tropical cyclone (TC) intensity has been developed using a track-pattern clustering (TPC) method and ocean-coupled potential predictors. Based on the fuzzy c-means clustering method, TC tracks during 2004-2012 in the western North Pacific (WNP) were categorized into five clusters and their unique characteristics were investigated. The predictive model uses multiple linear regressions, where the predictand or the dependent variable isthe change in maximum wind speed relative to initial time. To consider TC-ocean coupling effect due to TC induced vertical mixing and resultant surface cooling, we also developed new potential predictors for maximum potential intensity (MPI) and intensification potential (POT) using depth-averaged temperature (DAT)instead of sea surface temperature (SST). All together, we used six static, 11 synoptic, and three DAT-based potential predictors. Results from a series of experiments for the training period of 2004 -2012 showed that the use of TPC and the DAT-based predictors improved TC intensity predictions remarkably. The model was tested on predictions of TC intensity for 2013 and 2014 which are not used in the training samples. Relative to the non-clustering approach, the TPC and DAT-based predictors reduced the prediction errors up to 10~28% at most lead time. The present model is also comparedwithfouroperationaldynamical forecast models. At short leads (up to 24 hours)the present model has the smallest mean absolute errors. After 24-hour lead times, the present model still shows a good skill comparable to the best operational model. The developed model which uses TPC and DAT-based predictors, CSTIPS-DAT, led to a significant improvement in intensity prediction of TC, but it still showed relatively large errors in a specific cluster, in which strongly-developing TCs and non-developing TCs coexist, particularly in the central WNP where environmental conditions are the most favorable for TC intensification. In order to improve the prediction skill of CSTIPS-DAT, here we employed a decision tree algorithm which is most popular classification method, to classify the type of TCs in early stage, based on whether they will develop to a strong intensity (maximum wind speed of 70 kt) during their lifetime or not. For the binary classification, a decision tree with four leaf nodes was built using the Classification And Regression Tree (CART) algorithm. According to the discovered rules from the trained tree, DAT, latitude, and DAT-based MPI were major factors in determining the two types. The decision tree has classification accuracy of 92.5% for training period, and of 80.5% for test period. The fact that DAT and DAT-based MPI are selected in the decision tree implies that TC-induced vertical mixing process and pre-existing ocean thermal structures along the track play a major role in determining type of TCs. The present results finally suggest that the TC intensity prediction skill of CSTIPS-DAT can be further improved by establishing independent statistical models for the classified groups.