Analysis of the Interdecadal and Interannual Variability of Autumn Extreme Rainfall in Taiwan Using a Deep-Learning-Based Weather Typing Approach

Abstract

This study sought to assess the interdecadal and interannual variability of autumn extreme rainfall (ER) in Taiwan from 1979 to 2019. Three types of ER events were identified based on a clustering analysis augmented by a deep autoencoder-based neural network model. This method outperforms other methods in obtaining the optimal number of clusters by extracting the synoptic features in advance. The patterns associated with these three types include a tropical cyclone covering Taiwan (TC), a TC-like circulation in the South China Sea (SCS) accompanied by northeasterly near northern Taiwan (TC-NE), and northeasterly near northern Taiwan (NE). The differences in the rainfall pattern caused by the three types were discernable over Taiwan. How the PDO or ENSO modulates the regional large-scale environment to favor the occurrence of these ER events was investigated. The occurrence of TC-NE events was simultaneously correlated with the negative phases of PDO/ENSO in the interdecadal/interannual scale. In the negative phases of PDO/ENSO, a low-level anomalous cyclone over SCS accompanied by background northeasterly favored the regional TC activities and may cause more TC-NE events. The occurrence of NE events is simultaneously correlated with the cold phase of ENSO. An anomalous low-level anticyclone in Northeast Asia strengthened the northeasterly toward northern Taiwan, and with the seasonal background moisture, provided favorable conditions for the occurrence of the NE events. Overall, the occurrence of the TC events did not correlate with the PDO or ENSO signals; the reasons for the lack of correlation were discussed herein.