Annual variation in heavy rainfall frequency in Kyushu, Japan, linking to a synoptic field pattern classified by Self-Organizing Map

The aim of this study is to investigate annual variation in heavy rainfall frequency in Kyushu area (target area) located in the west of Japan and, subsequently, to demonstrate annual variation in heavy rainfall frequency relating to each synoptic field pattern recognized by using the Self-Organizing Map (SOM) developed by Kohonen (1995). The pattern recognition technique is a kind of unsupervised artificial neural networks (ANNs) in the field of information science. The SOM provides useful information for helping the interpretation of non-linear complicated features by classifying a set of high-dimensional data into the units (patterns) arranged regularly on a two-dimensional space that can be easily and visually recognized by ‘human eye'. Synoptic fields treated in this study are represented by the spatial distribution of wind components at the 850 hPa level and Precipitable Water (PW) using NCAR-NCEP reanalysis data (4 times per a day) for detecting a Low-Level Jet (LLJ) and associated convective activity. A data sample consists of 48 dimensions (16 grid points, 3 variables). On the other hand, for detecting heavy rainfall features, this study uses the rainfall observation system called as the Automated Meteorological Data Acquisition System (AMeDAS), which has been maintained by the Japan Meteorological Agency (JMA). In this study, to combine synoptic field and heavy rainfall, assuming that a synoptic field at T (=0, 6, 12, 18UTC) is related to heavy rainfalls observed during 6 hours between T-3 and T+3, heavy rainfall frequency is calculated by summing up AMeDAS hourly rainfalls (>=30mm/h) recorded during 6 hours in the target area. The SOM training in this study uses input samples (total_num=14648) obtained from the outputs (NCAR-NCEP reanalysis data) of 4 times (T=0, 6, 12, 18UT) per a day in the warm season (June-September) for 30 years. As a result of SOM training, the input samples are classified into the 30 (x-axis) ×30 (y-axis) hexagonal units, in other words, 900 synoptic field patterns. Each unit includes a reference vector and the most similar input samples to it. The reference vector obtained by the SOM training shows a representative feature among the input samples classified in the unit. In addition, all the patterns formed by the SOM training are divided into 36 groups using the K-means clustering method for confirming an averaging feature of each group. From these results, it was found that dominant heavy rainfall activity is related to ten groups among 36 groups. Especially, the top 3 heavy rainfall groups (G31, 32 and 34) are characterized by high PW distribution and the Low Level Jet (LLJ). G31 shows typical synoptic field pattern in a rainy season in Japan. G32 is similar to G31 in that Kyushu is affected by the dominant area of high PW, accompanying strong LLJ. However, G32 is characterized by eastward LLJ and a steep gradient of PW in the northern area of Kyushu. The steep PW gradient is related to frontal activity. G31 and 34 have common feature of high PW area in Kyushu with dominant northeastward LLJ. However, wind patterns of synoptic field in G34 are characterized by cyclonic motion in the northwest edge of the target area. Annual variation in heavy rainfall frequency is characterized by high frequency (especially, 1999, 2006, and 2007) in the recent period after 1999. The annual variation was divided into all the 36 groups obtained according to the SOM and the K-means method. The result shows that G32 contributes to high frequency of heavy rainfall in 2006. G32 is characterized by northern dry region and southern wet region, which means the existence of a front with a steep gradient of water vapor amount. Actually, dominant synoptic patterns causing heavy rainfall in Kyushu in 2006 were related to frontal activities. From these results, it could be clearly understood that what kinds of synoptic field groups contributed to decadal trend of heavy rainfall frequency by using the SOM classification technique.