JournalApril 2019 of the Meteorological Society of Japan, 97(2),M. 387−401, FUJITA et2019. al. doi:10.2151/jmsj.2019-022 387 Special Edition on Tropical Cyclones in 2015 − 2016

Analyses of Extreme Precipitation Associated with the Kinugawa River Flood in September 2015 Using a Large Ensemble Downscaling Experiment

Mikiko FUJITA

Japan Agency for Marine-Earth Science and Technology, Kanagawa, Japan

Tomonori SATO

Faculty of Environmental Earth Science, Hokkaido University, Hokkaido, Japan

Tomohito J. YAMADA

Faculty of Engineering, Hokkaido University, Hokkaido, Japan

Sho KAWAZOE, Masuo NAKANO

Japan Agency for Marine-Earth Science and Technology, Kanagawa, Japan

and

Kosuke ITO

University of the Ryukyus, Okinawa, Japan

(Manuscript received 1 April 2018, in final form 29 November 2018)

Abstract

We investigated extremely heavy precipitation that occurred around the Kinugawa River, Japan, in September 2015, and the probability of extreme precipitation occurrence, using data from a large ensemble forecast of more than 1,000 members that were dynamically downscaled to 1.6 km horizontal grid spacing. The observed event was statistically rare among simulated cases and the 3-day accumulated precipitation around the target area was equivalent to the 95th percentile among all simulated ensemble members. Our results show that this extreme precipitation event occurred under specific conditions: two coexisting typhoons at close proximity that produced a high atmospheric instability, and water vapor transported from the Pacific Ocean. We also assessed the- prob ability of extreme precipitation in mountainous areas other than the Kinugawa River case. Heavy precipitation also occurred southwest of the Kinugawa River region due to two typhoons, similar to the Kinugawa River case. The tracks of these typhoons shifted marginally; however, there was a difference in the water vapor supplied to the area, causing heavy precipitation. The large-ensemble downscaled data used in this study hence enabled us to evaluate the occurrence probability of a torrential rainfall event that was rarely observed, which may contribute to updating a disaster-mitigating plan for possible similar disasters in future.

Corresponding author: Mikiko Fujita, JAMSTEC, 3173-25 Showa-machi Kanazawaku, Yokohama, Kanagawa 236-0001, Japan E-mail: [email protected] J-stage Advance Published Date: 17 December 2018 ©The Author(s) 2019. This is an open access article published by the Meteorological Society of Japan under a Creative Commons Attribution 4.0 International (CC BY 4.0) license (https://creativecommons.org/licenses/by/4.0). 388 Journal of the Meteorological Society of Japan Vol. 97, No. 2

Keywords extreme precipitation; ensemble forecast; downscaling; disaster prevention

Citation Fujita, M., T. Sato, T. J. Yamada, S. Kawazoe, M. Nakano, and K. Ito, 2019: Analyses of extreme precipi- tation associated with the Kinugawa River flood in September 2015 using a large ensemble downscaling experiment. J. Meteor. Soc. Japan, 97, 387–401, doi:10.2151/jmsj.2019-022.

uncertainty in large ensemble datasets explains the 1. Introduction probability of even more extreme events and helps From September 7 to 11, 2015, very heavy continu- improve our understanding of the 2015 Kinugawa ous precipitation occurred in central Japan. The Kinu­ heavy rainfall event. gawa River, a tributary of the Tone River, which is a 2. Design of the large-ensemble numerical major river in Japan, experienced extensive flooding, experiment resulting in the worst disaster-related damage in this area over the past century (Japan Society of Civil The use of a large amount of data derived from Engineers Committee 2015; Kitabatake et al. 2017; operational ensemble forecast output, which subsumes Ushiyama et al. 2017). This extreme rainfall event uncertainty caused by chaotic atmospheric behavior, was indirectly affected by two typhoons. The remote can be helpful in the investigation of the mechanisms influence of tropical cyclones (TCs) on synoptic-scale causing record-breaking heavy precipitation and its phenomena is a key condition for producing an ex- probability of occurrence under similar conditions treme precipitation event (e.g., Galarneau et al. 2010; elsewhere. In this study, we used the output of a Tsuguti and Kato 2014). Nguyen-Le and Yamada global ensemble forecast to generate members of a (2017) investigated Typhoon Lionrock (TY1610) and large ensemble and performed dynamical downscaling its remote effect on heavy rainfall in Hokkaido, north- using the large ensemble forecast to investigate the ern Japan. Their numerical simulation, which removed 2015 flooding in the Kinugawa region. the vortex associated with the typhoon, showed that The Global Ensemble Forecast System of the remote rainfall in Hokkaido was largely suppressed, National Center for Environmental Prediction (GEFS– thereby highlighting the importance of moisture NCEP, Table 1) consists of 21 ensemble members at a transport toward Hokkaido by the synoptic-scale cir- horizontal resolution of 1° × 1° (Whitaker et al. 2008). culation related to the TC location. The forecast is performed with 21 different initial To better prepare for and mitigate similar hazards conditions at 6 h intervals, and 16 days of forecast of this type, the probability of extreme events and data are available. These data were adopted in view their contributing preconditions must be investigated. of their compatibility with the physics of the regional Although the flooding event in September 2015 im- model (see Section 3) and data accessibility. The target pacted only a limited region, similar events including period during which extreme precipitation occurred heavy precipitation may have occurred in adjacent in the real world (from 0000 UTC on September 8 to areas due to chaotic atmospheric behavior. The fre- 0000 UTC on September 11, 2015) was set for dy- quency of extreme precipitation events has increased namical downscaling by the regional model using the worldwide over the past century (e.g., Frich et al. GEFS–NCEP ensemble forecast as forcing data (see 2002), including in Japan (Fujibe et al. 2005); there- Section 3 for details of the regional model). Note that fore, the risk of disasters due to such events has also the dates and duration of the downscaling experiment increased. In addition, changes in the frequency of by the regional model are the same among all ensem- heavy precipitation events have been projected based ble members. However, the input data for the regional on future climate scenarios (e.g., Endo et al. 2017). model are different in each case due to variation in the In this pilot study, we applied statistical methods initial conditions of the GEFS–NCEP forecast (Fig. to investigate an extreme precipitation event and the 1a). We used lagged initial conditions for the forecast probabilities of such an event in the surrounding area to create the ensemble field, as performed in previous using ensemble forecast data with more than 1,000 studies (e.g., Hoffman and Kalnay 1983; Nakano et al. members. We used the extremely heavy rainfall that 2017). As shown in Figure 1a, 1,029 ensemble mem- occurred around the Kinugawa River as an example. bers were adopted from the GEFS–NCEP forecasts: Our results show that the consideration of chaotic 21 ensemble members × 49 initial conditions corre- April 2019 M. FUJITA et al. 389

Table 1. Data used for the dynamical downscaling experiment. Data name Horizontal resolution Type GEFS-NCEP 16-days forecast, four-times a day (Global Ensemble Forecast System of the National 1° × 1° with forecasts (00, 06, 12, 18 UTC), Center for Environmental Prediction) 21 ensemble members FNL 1° × 1° Operational Analysis (NCEP Final Operational Global Analysis data)

Fig. 1. (a) Schematic diagram of the ensemble forecast data for dynamical downscaling. (b) Regional model domains. The color shading in Domain 3 (enlargement) indicates the topography [m] and black lines show the rivers.

sponding to 6 h interval data (four times per day). event may depend on the model predictability. We The 49 sets of initial conditions were 12 days from assumed that both the ensemble spread (representing August 27 to September 7, plus 0000 UTC September uncertainty in the initial conditions) and lagged initial 8, resulting in 49 sets of initial conditions in total conditions (representing uncertainty due to prediction (i.e., 4 × 12 + 1). All the data of the 1,029 ensemble error) adequately represented uncertainty in the initial members from GEFS–NCEP were used as input data and boundary conditions for the dynamical downscal- for the regional model. ing experiment. It is likely that the calculated prob- Because the prediction ability of each forecast model ability of heavy rainfall would be different if other generally has limitations in terms of the synoptic fields parent forecast models or initial dates were selected. of the target event, these forecast data may contain However, because this was a pilot study, we used only model bias derived from errors in the particular model forecast data obtained from GEFS–NCEP to evaluate and its physical scheme. Therefore, the uncertainty (or the ability of large ensemble dynamical downscaling ensemble spread) of the synoptic fields at the target to assess a particular event. 390 Journal of the Meteorological Society of Japan Vol. 97, No. 2

Table 2. Names of the dynamical downscaling experiment. 3. Regional model settings and used data Input (initial condition and Experiment Number of To reproduce heavy rainfall events, we performed boundary data) for name experiments dynamical downscaling using the Weather Research the dynamical downscaling and Forecasting (WRF) model V3.5.1 (Skamarock DS-GEFS GEFS-NCEP 1029 et al. 2008). Each ensemble forecast output of GEFS– DS-FNL FNL 1 NCEP was used as initial conditions and boundary data for each of the 1,029 ensemble members during the target period, from 0000 UTC September 8 to of the typhoon center was determined as the grid 0000 UTC September 11, 2015 (hereafter, these cell with the lowest SLP below 1,000 hPa in the sub­ ensemble experiments are referred to as DS–GEFS; sequent area. This procedure was performed for two Table 2). The model domains are shown in Figure 1b; typhoons (KILO and ETAU), one in each of the two these were set up as triple-nested domains with hori- areas obtained by dividing the domain along 140°E. zontal resolutions of 24, 8, and 1.6 km, respectively, Domain 2 was used for analyses of regional atmo- in one-way nesting with 35 layers on the vertical spheric instability. The targeted heavy precipitation scale. The Kain–Fritsch cumulus parameterization (Ma event at Nikko occurred in a mountainous area and and Tan 2009) was used for domains 1 and 2. For all therefore is likely to have been affected by the terrain domains, the WRF double-moment six-class micro- (Rauscher et al. 2016). Taking into account the local physics scheme (Hong et al. 2010) was adopted, along orographic effects on precipitation, the simulated pre- with the Mellor–Yamada–Nakanishi–Niino level-2.5 cipitation and related parameters (equivalent potential (MYNN2) planetary boundary layer scheme (Naka­ temperature and precipitable water) were analyzed nishi and Niino 2006) and the Noah land surface model using the output from domain 3, which was expected scheme (Chen and Dudhia 2001). to realistically reproduce the circulation induced by The reproducibility of heavy precipitation in the the complex terrain at a fine horizontal resolution. regional model strongly depends on its horizontal res- 4. Results olution and cumulus parameterization (e.g., Kendon et al. 2012; Cruz et al. 2016). To quantify the repro- 4.1 WRF model performance in estimating ducibility of the simulated precipitation in the WRF precipitation around the Kinugawa River using the abovementioned settings, a hindcast exper- Figure 2 shows the target heavy precipitation event iment was performed using NCEP Final Operational from September 8 to 11, 2015. This event occurred Global Analysis data (FNL; Table 1; National Centers in the Nikko area (139.4 – 139.7°E, 36.6 – 37.0°N), for Environmental Prediction/National Weather around the upper reaches of Kinugawa River. Two Service/NOAA/US Department of Commerce 2000) typhoons were located near the region of interest with a horizontal resolution of 1° × 1°, because the during the heavy precipitation event in the Kinugawa forecast (GEFS–NCEP) data did not always corre- River catchment (KILO and ETAU; Fig. 2a). Figure spond to observations. The same initial settings and 2b shows the results of DS–FNL. The typhoon tracks time window were used (hereafter, this hindcast is and SLP at the centers of the typhoons were reason- referred to as DS–FNL; Table 2). We used these data ably close to the observations. to confirm the ability of the NCEP global model and Figure 2c shows the horizontal distributions of 72 h the WRF model to replicate the real event, because the (3-day) accumulated precipitation between 0000 UTC FNL data were generated from the same model as the September 8 and 0000 UTC September 11, 2015, NCEP Global Forecasting System, but also assimilat- as observed at AMeDAS stations. An areal mean of ed observational data. In addition, precipitation data > 300 mm precipitation was observed within 3 days from the Automated Meteorological Data Acquisition throughout the Nikko area. Figure 2d shows the down­ System (AMeDAS) weather stations and the weather scaled DS–FNL precipitation results. The horizontal map of the Japan Meteorological Agency were used to distribution of accumulated precipitation peaked validate the precipitation estimates. around the Nikko area in both the observations and In both experiments, the results for domain 1 downscaled data. The 3-day accumulated precipitation were used to analyze the synoptic-scale circulation. around Nikko during the target period estimated Typhoon tracks were determined from the sea level from the downscaled results was 390.3 mm, and the pressure (SLP) in domain 1 using methods described observed value was 319.4 mm. Another precipitation by Sakai and Yamaguchi (2005). The first position peak was found in the simulated results along 139°E; April 2019 M. FUJITA et al. 391

Fig. 2. (a) Observed typhoon tracks and sea level pressure (SLP) [hPa] map at 0000 UTC on September 9, 2015. (b) Simulated typhoon tracks and SLP [hPa] of DS-FNL. (c) Accumulated precipitation observed at by AMeDAS stations [mm] between 0000 UTC on September 8 and 0000 UTC on September 11, 2015, and (d) simulated ac- cumulated precipitation in domain 3 [mm (3 days)−1] of DS-FNL. The gray dashed lines in (c) and (d) show the elevation over 500 m in 100 m intervals. The dashed white boxes indicate the Nikko area (139.4 – 139.7°E, 36.6 – 37.0°N) and Okutama area (139.0 – 139.3°E, 35.9 – 36.2°N), these were used in the analysis.

however, it was difficult to evaluate this peak because values for comparison with other ensembles. the area is mountainous and fewer observations were available (AMeDAS station locations are shown in 4.2 Precipitation intensity of the downscaled Fig. 2c). The downscaled precipitation of the south- ensemble members western Nikko area showed a positive bias, while that As described in Section 2 and Fig. 1, 1,029 sets of the southern coast had a negative bias. Based on of GEFS–NCEP data were dynamically downscaled this comparison, we considered that the WRF setting using the WRF model. Figure 3 shows the percentiles was appropriate to simulate the extreme precipitation of the 3-day precipitation totals in the Nikko area, around Nikko. The downscaled data used reference simulated in domain 3 of the 1,029 DS–GEFS ensem- 392 Journal of the Meteorological Society of Japan Vol. 97, No. 2 ble members. The observed precipitation and DS–FNL simulation results were most closely approximated by the 95th percentile of all DS–GEFS ensemble mem- bers (1,029 samples). Figure 4 shows the movement of the typhoon cen- ters between 0000 UTC September 8 and 0000 UTC September 11, as well as the 3-day precipitation totals in Nikko. The locations of the typhoons were detected using the 6-hourly SLP in domain 1 throughout each experiment (see also Section 3). Figure 4a shows all 6-hourly typhoon locations among the 1,029 DS– GEFS ensemble members. Figure 4b reveals that extreme precipitation (> 300 mm per 3-day at Nikko, corresponding to the 90th percentile among all en- Fig. 3. Percentiles of the 3-day precipitation [mm] semble members; Fig. 3) was reproduced only if the estimated from all downscaled ensemble mem- simulated typhoon tracks were close to the real tracks bers of DS-GEFS. Red value indicates the 3-day (Fig. 2a). These results show that such extreme precip- precipitation simulated in DS-FNL. itation requires a highly specific set of circumstances, i.e., typhoons passing through a limited area over the Pacific Ocean and Sea of Japan. We also analyzed variation in the DS–GEFS water these distance criteria, 75 cases were selected (Fig. 6a, vapor flux around the target area (Figs. 5a – c). For red circles) and the associated typhoon tracks (Fig. 4c) forecasts in which the 3-day accumulated rainfall and moisture transport (Figs. 5d – f) were plotted. The estimates exceeded 300 mm in Nikko (Fig. 5a), a selected ensemble members were not perfectly coin- large amount of water vapor was transported from cident with extreme precipitation cases (> 300 mm the Pacific Ocean, driven by the airflow of the two per 3-day; Fig. 4b); however, > 70 % of members that typhoons. The horizontal pattern in water vapor flux produced such heavy precipitation at Nikko were in- differences (Fig. 5c) between the heavy precipitation cluded. Water vapor transport from the Pacific Ocean case (Fig. 5a) and others (Fig. 5b) indicates that a large was also important for heavy precipitation and the water vapor supply from the Pacific Ocean, forming synoptic-scale circulation formed by the two typhoons a narrow zone of highly precipitable water vapor produced moist airflow toward the Nikko area. toward the Nikko area, was a key factor contributing Next, we investigated the heavy precipitation to the unprecedented precipitation levels during this around Nikko from the perspective of the convective event. instability criterion computed from the difference in equivalent potential temperature between 500 and

4.3 Factors controlling heavy precipitation at Nikko: 850 hPa (i.e., θ e, 500hPa – θ e, 850hPa , where negative values inter-typhoon distance and atmospheric instability indicate convective instability). According to previous The results reported above indicate that the distance reports (e.g., Kitabatake et al. 2017), such atmospheric between the two typhoons was a key factor contribut- instability develops in response to moist airflow in ing to heavy precipitation at Nikko. The 3-day mean the lower atmosphere from the Pacific Ocean and distances between the typhoons and between Nikko dry air supplied to the upper atmospheric layer from (139.5°N, 36.7°E) and each typhoon were calculated. regions to the west and north of the target area. This Figure 6 shows the relationship between these dis- circulation pattern is strongly linked to the presence of tances and the 3-day precipitation at Nikko calculated the two typhoons and therefore is likely to have been from the DS–GEFS results. Figure 6a shows that a key contributor to the precipitation event. Figure heavy precipitation (> 300 mm) occurred when dis- 7a shows the 3-day mean of differences in equivalent tances of 1120 – 1440 km separated the two typhoons. potential temperature derived from DS–FNL (domain A further condition for the generation of heavy precip- 2). These mean estimates show high instability above itation was that the distance between ETAU (the west- central Japan. ern typhoon) and Nikko was in the range 320 – 710 To determine the horizontal pattern of high atmo- km, and that between KILO (the eastern typhoon) and spheric instability, we attempted to deconvolve the Nikko was in the range 830 – 1090 km (Fig. 6b). Using effects of convective instability in driving the heavy April 2019 M. FUJITA et al. 393

Fig. 4. Six-hourly plots of the typhoon center locations during the target period of DS-GEFS, employing (a) all ensemble members, (b) ensemble members that simulated 3-day accumulated precipitation over 300 mm at Nikko, (c) ensemble members that reproduced two typhoons satisfying the distance criteria (see Section 4.3), and (d) en- semble members categorized by SOM (see Section 4.3). The dots indicate the 6-hourly typhoon center locations and these are colored based on the 3-day accumulated precipitation at Nikko for each member (each 6-hourly track is dotted with the color of the 3-day precipitation at Nikko). Track lines are shown where the simulated precipita- tion exceeds 300 mm within the 3-day window. precipitation event, using a self-organizing map (SOM; the 3-day mean differences in equivalent potential Kohonen 1982). SOMs utilize an unsupervised learn- temperature between 500 and 850 hPa in domain 2 ing process to create a user-determined, two-dimen- as input. After testing several array sizes and SOM sional array of nodes to segregate the major pattern learning parameters, we opted for a 7 × 6 array, which characteristics of the input data (e.g., Nishiyama et al. yielded 42 different spatial patterns. On average, each 2007; Nguyen-Le et al. 2017). In this study, we used node contained about 25 (1,029 members/42 nodes) 394 Journal of the Meteorological Society of Japan Vol. 97, No. 2

Fig. 5. Composite maps of the 3-day mean precipitable water vapor (shading), equivalent potential temperature at 850 hPa (contours), and mean wind at 850 hPa (vectors) in domain 1 of DS-GEFS. (a) – (c) Means of the ensemble members simulated the 3-day precipitation over 300 mm at Nikko in domain 3. (d) – (f) Means of the ensemble members satisfied the criteria of distance. (g) – (i) Means of the ensemble members categorized by SOM(see Section 4.3). Center panels show the mean value of the outside of criteria, and the right panels show the differenc- es (left minus center).

members and appeared to discriminate between two-dimensional representation of the relationship prominent spatial patterns without overgeneralizing or between each node and its closest neighbor. Such flat over-segregating the input data. Although the SOMs Sammon maps are often preferred, to avoid convolut- may give the impression that each node is regularly ed relationships between nodes (Cassano et al. 2015), spaced, this is usually not the case. Sammon maps as we confirmed in the present study (Fig. S1). The (Sammon 1969) are commonly used to provide a SOM nodes calculated by the instability are shown April 2019 M. FUJITA et al. 395

Fig. 6. The relationship between the 3-day mean distance of two typhoons and the precipitation at Nikko simulated in domain 3 of DS-GEFS. (a) Accumulated precipitation versus distance between the two typhoons. Red circles indicate the cases satisfying the distance criteria (see Section 4.3). (b) Scatter plot of the distance between each typhoon and Nikko. Color indicates the precipitation amount in the Nikko area.

Fig. 7. Horizontal distributions of mean atmospheric instability (color) derived from the difference of the equivalent potential temperatures; 500 hPa (green contour) minus 850 hPa (pink contour) in domain 2. (a) The results of DS- FNL, (b) DS-GEFS ensemble mean that was selected by the distance criteria, and (c) DS-GEFS ensemble mean that was categorized by SOM (see Section 4.3). The bold black boxes in (c) indicate the Nikko area and southern area of Nikko used in the analysis. in Fig. S2 and the corresponding typhoon tracks are FNL (Fig. S2, top left panel). The spatial correlation shown in Fig. S3. coefficient between DS–FNL and the category 1 mean The SOM node matrix was clustered again using the was 0.36, and those of other categories were < 0.10. unified distance matrix (U-matrix) method (Ultsch and We used the five nodes in category 1 (Fig. S4, hatched Siemon 1990), which evaluates the distances between area) for subsequent analyses. different parts of neighboring nodes. Figure S4 shows The composited map of typhoon tracks categorized the calculated U-matrix, revealing that, where the by these five nodes is shown in Fig. 4d. This wasa differences between neighboring values were large, looser criterion than the distance criterion; however, the spatial pattern differed between the nodes. The more than 50 % of tracks that produced heavy precip- matrix was separated into three categories. Category 1 itation (> 300 mm per 3-day) at Nikko were included shows similar instability distributions to those of DS– in this selection. Moreover, a similar pattern of the 396 Journal of the Meteorological Society of Japan Vol. 97, No. 2 water vapor supply from the Pacific Ocean was de- found in mountainous areas as well as the Kinugawa tected using SOM (Fig. 5g). The horizontal patterns of River region (Figs. 9a – c). In the composite map of atmospheric instability composited using the selected all ensemble members, higher precipitation was sim- distances and SOM category 1 are shown in Fig. 7. ulated around Nikko and an area southwest of Nikko Compared to that of DS–FNL (Fig. 7a), the distance (hereafter, Okutama; 139.0 – 139.3°E, 35.9 – 36.2°N; selection pattern was quite similar (Fig. 7b). The area Fig. 9a). The most extreme precipitation (99th percen- with high atmospheric instability in SOM category 1 tile) in each grid cell is shown in Figs. 9d – f. Heavy tended to shift to the west compared to that of DS– rain (> 500 mm per 3-day) was found around both FNL (Fig. 7c); moreover, atmospheric instability was Nikko and Okutama (the mountainous area and an ad- weaker in the eastern region. This difference is likely jacent area). The probability of extreme precipitation due to the selected typhoon tracks (Fig. 4d); some of around Nikko and Okutama exceeded 50 % in both the western TC (ETAU) tracks shifted to the south- the distance and SOM cases. The probability of heavy west, which should weaken the circulation from the precipitation calculated from all ensemble members Pacific Ocean to the Nikko area. The SOM category 1 also exceeded 15 % in both Nikko and Okutama. node results demonstrate the difficulty of categorizing This numerical experiment was performed for only heavy precipitation at Nikko using only the horizontal one particularly extreme event; however, our results pattern of mean atmospheric instability, but they can demonstrate the specific nature of the conditions re- be used to coarsely sort cases with the potential for quired for extreme precipitation at Nikko. unstable phenomena. To investigate the differences in heavy precipitation The relationship between areal mean atmospheric between Nikko and Okutama, composite analyses instability and precipitation at Nikko and the wind- were performed using the five most extreme precipita- ward area (139.0 – 139.7°E, 35.1 – 35.5°N) are shown tion cases in the two regions, as determined from the in Fig. 8. Atmospheric instability was highly variable distance selection. The ensemble members in the dis- in both cases, with heavy precipitation (> 300 mm tance selection were required to possess the character- per 3-day) found at an instability range of −8.0 to 1.5 istic conditions for producing extreme precipitation at K. Instability tended to be higher in windward areas Nikko. Figure 10a shows typhoon tracks for extreme (Fig. 8c). Heavy precipitation occurred at Nikko only precipitation cases at Nikko and Okutama. The tracks when the amount of precipitable water was relatively of the western typhoon (ETAU) were similar at both large, and all cases defined by the distance criteria Nikko and Okutama. However, the eastern typhoon (see Section 4.3) contained a large amount of water (KILO) in the Okutama case tended to be located vapor (Figs. 8b, d). Atmospheric instability generally farther southwest than in the Nikko case. In addition, predicted heavy precipitation; however, water vapor there was a westerly shift in the high moisture zone supply was also essential for 3-day accumulated heavy in the Okutama case (Figs. 10b – d). Precipitation precipitation, especially during this event. composites showing heavy rain events are presented in Figs. 10e and 10f (> 500 mm per 3-day). The max- 5. Probability of heavy rain in nearby regions imum possible precipitation can occur when abundant The criteria that allowed us to simulate extreme water vapor is supplied at the foot of a mountainous precipitation in terms of the locations of typhoons area and dynamically lifted to the upper layer by the supplying large amounts of water vapor under strong terrain and strong air flow (Smith et al. 2010). atmospheric instability applied only to precipitation A typhoon track shift of even < 100 km can change events over the Kinugawa region. Extreme precipita- the location of a typhoon-related disaster. In this study, tion events can also occur over different topography factors controlling extreme precipitation at Nikko in- under similar atmospheric conditions. Therefore, we cluded the locations of the typhoons and the distance tried to estimate the heavy precipitation caused by a between them, as these factors control the synoptic two-typhoon system in nearby regions. Our findings circulation supplying water vapor from the Pacific could be applied to improve mitigation measures for Ocean during periods of high atmospheric instability extreme weather events over a wide area. in the Kinugawa basin region. This type of study can Figure 9 shows maps of precipitation statistics in be applied to other low-frequency events and is useful domain 3 of DS–GEFS calculated from all ensemble for understanding the high likelihood of heavy rainfall members, ensemble members selected using the in a particular area. distance criteria, and ensemble members selected using SOM categories. Mean precipitation peaks were April 2019 M. FUJITA et al. 397

Fig. 8. The relationship on mean precipitation at Nikko area. (a), (c) show the scatter diagram of atmospheric insta- bility, and (b), (d) show the scatter diagram of precipitable water and atmospheric instability (color). Upper panels are estimated from Nikko area mean, lower panels are estimated from the southern windward area (139.0 – 139.7°E, 35.1 – 35.5°N, southern black box in Fig. 7c). Red and black circles indicate the cases satisfying the distance crite- ria (see Section 4.3).

quate separation that led to the development of strong 6. Conclusion atmospheric instability and water vapor transport from We investigated the extreme precipitation event that the Pacific Ocean, ultimately causing extreme precip- occurred in the Kinugawa River area in September itation. The observed heavy precipitation was equiva- 2015 using ensemble forecast data (> 1,000 members). lent to the 95th percentile of the distribution among all Our results showed that this event occurred as a result ensemble members. The precipitation distribution was of highly unstable atmospheric conditions caused by nonlinearly affected by local circulation induced by the close proximity of two typhoons. These typhoons the topography. passed through a narrow area while maintaining ade- We also assessed the probability of extreme precip- 398 Journal of the Meteorological Society of Japan Vol. 97, No. 2

Fig. 9. Horizontal distributions simulated in DS-GEFS, (a) – (c) Mean precipitation, (d) – (f) max (99th percentile) precipitation, and (g) – (i) probability of heavy precipitation (percentage of number of members exceeding 100 mm 3 day−1). Left panels were calculated from all ensembles, center panels were calculated from the ensemble members satisfying the distance criteria (see Section 4.3), and right panels were calculated from ensemble mem- bers that categorized by SOM (see Section 4.3). The white contours (at 100 m intervals) and numbers in (a) show the altitude over 500 m. The red boxes in (a) indicate the Nikko and Okutama area used in the analysis. April 2019 M. FUJITA et al. 399

Fig. 10. Horizontal distributions of the mean of the top 5 of 3-day precipitation for Nikko and Okutama selected from the distance selection. (a) Typhoon tracks for the Nikko (black line) and Okutama (red line) cases. (b) – (d) Distributions of precipitable water (color), equivalent potential temperature at 850 hPa (contours), and horizontal wind at 850 hPa (vectors) in domain 1; (b) and (c) shows the composite in each case, (d) shows the difference (c minus b). (e) – (f) Distributions of composited 3-day accumulated precipitation (shading), precipitable water (white contours), and equivalent potential temperature at 850 hPa (pink contours) in domain 3. Light gray contours show altitude, as in Fig. 9. 400 Journal of the Meteorological Society of Japan Vol. 97, No. 2 itation in mountainous areas other than the Kinugawa References River region. Around Okutama, southwest of the Kinu­ gawa River region, heavy precipitation also occurred Cassano, E. N., J. M. Glisan, J. J. Cassano, W. J. Gutowski, due to the two typhoons, similar to the Kinugawa Jr., and M. W. Seefeldt, 2015: Self-organizing map River case. However, the typhoon tracks shifted mar- analysis of widespread temperature extremes in ginally (< 100 km); this shift supplied water vapor to Alaska and Canada. Climate Res., 62, 199–218. the area, causing heavy precipitation. Chen, F., and J. Dudhia, 2001: Coupling an advanced land surface-hydrology model with the Penn State–NCAR Our results demonstrate that the proposed method MM5 modeling system. Part I: Model implementation is useful for investigating the mechanisms of low- and sensitivity. Mon. Wea. Rev., 129, 569–585. frequency events. Ensemble numerical weather Cruz, F. T., H. Sasaki, and G. T. Naria, 2016: Assessing forecast estimates simulate perturbed synoptic fields the sensitivity of the non-hydrostatic regional cli- that can occur under similar large-scale forcing. Our mate model to boundary conditions and convective method allowed us to assess the Kinugawa disaster schemes over the Philippines. J. Meteor. Soc. Japan, by quantifying the probability of its occurrence and 94A, 165–179. intensity under specific atmospheric conditions. Thus, Endo, H., A. Kitoh, R. Mizuta, and M. Ishii, 2017: Future the present study provides an important approach changes in precipitation extremes in East Asia and for modeling hypothetical disaster scenarios based their uncertainty based on large ensemble simulations on actual disasters and it will be helpful in designing with a high-resolution AGCM. SOLA, 13, 7–12. measures against future similar disasters without rely- Frich, P., L. V. Alexander, P. Della-Marta, B. Gleason, M. ing on limited observational data. Haylock, A. M. G. Klein Tank, and T. Peterson, 2002: Observed coherent changes in climatic extremes dur­ Supplements ing the second half of the twentieth century. Climate Res., 19, 193–212. The file Supplement includes four figures used for Fujibe, F., N. Yamazaki, M. Katsuyama, and K. Kobayashi, SOM analysis. Figure S1: Sammon map for the SOM. 2005: The increasing trend of intense precipitation in The four corner nodes were labelled as shown in Figs. Japan based on four-hourly data for a hundred years. S2 and S3. Figure S2: Sorted maps of the mean atmo- SOLA, 1, 41–44. spheric instability (difference of the equivalent poten- Galarneau, J. T., Jr., L. F. Bosart, and R. S. Schumacher, tial temperatures, 500 hPa minus 800 hPa) in domain 2010: Predecessor rain events ahead of tropical cy- 2 of DS-GEFS with the number of samples in each clones. Mon. Wea. Rev., 138, 3272–3297. SOM nodes. Top-left panel shows the DS-FNL result. Hoffman, R. N., and E. Kalnay, 1983: Lagged average fore­ Figure S3: Composite maps of the typhoon track in casting, an alternative to Monte Carlo forecasting. domain 1 corresponding to the 7 × 6 nodes in Fig. S2. Tellus A, 35, 100–118. Hong, S.-Y., K.-S. S. Lim, Y.-H. Lee, J.-C. Ha, H.-W. Kim, Colors show the precipitation in Nikko (same as Fig. S.-J. Ham, and J. Dudhia, 2010: Evaluation of the 4). The track lines indicate that the simulated precipi­ −1 WRF double-moment 6-class microphysics scheme tation exceeds 300 mm 3 day . Figure S4: U-matrix for precipitating convection. Adv. Meteor., 2010, of the SOM corresponding to the 7 × 6 nodes in Fig. 707253, doi:10.1155/2010/707253. S2. Color shows the Euclidean distances between the Japan Society of Civil Engineers (JSCE) committee, 2015: neighboring nodes. The hatched nodes were used for Reports of the Kanto and Tohoku heavy rain disaster the analysis (see Section 4.3). in September 2015 (in Japanese). [Available at http:// committees.jsce.or.jp/report/taxonomy/term/48.] Acknowledgments Kendon, E. J., N. M. Roberts, C. A. Senior, and M. J. Roberts, This study was supported by a Grant-in-Aid for 2012: Realism of rainfall in a very high-resolution Young Scientists B (Grant Nos 15K16316 and 17K regional climate model. J. Climate, 25, 5791–5806. 13010) and Challenging Research Exploratory Kitabatake, N., H. Tsuguti, and T. Kato, 2017: Effects of (Grant No. 17K18963), promoted by the Ministry of synoptic-scale environmental flows on heavy rainfall event in the Kanto and Tohoku district in September Education, Culture, Sports, Science and Technology 2015. Tenki, 64, 887–890 (in Japanese with English (MEXT), Japan. This work was also supported in abstract). part by the Program for Risk Information on Climate Kohonen, T., 1982: Self-organized formation of topological- Change, and the Social Implementation Program on ly correct feature maps. Biol. Cybern., 43, 59–69. Climate Change Adaptation Technology (SI-CAT), Ma, L.-M., and Z.-M. Tan, 2009: Improving the behavior of MEXT, Japan. the cumulus parameterization for tropical cyclone pre- diction: Convection trigger. Atmos. Res., 92, 190–211. April 2019 M. FUJITA et al. 401

Nakanishi, M., and H. Niino, 2006: An improved Mellor– comparison of tropical cyclone track forecasts by Yamada level-3 model: Its numerical stability and operational global models. Research Activities in application to a regional prediction of advection fog. Atmospheric and Oceanic Models, CAS/JSC WGNE Bound.-Layer Meteor., 119, 397–407. Res. Activ. Atmos. Oceanic Modell., 35, 2.7–2.8. Nakano, M., H. Kubota, T. Miyakawa, T. Nasuno, and M. [Available at http://bluebook.meteoinfo.ru/uploads/ Satoh, 2017: Genesis of Super Cyclone Pam (2015): 2009/individual-articles/02_Onoda_Hirokatsu_02_ Modulation of low-frequency large-scale circulations onoda_hirokatsu_TCverification.pdf.] and the Madden–Julian oscillation by sea surface tem- Sammon, J. W., Jr., 1969: A nonlinear mapping for data perature anomalies. Mon. Wea. Rev., 145, 3143–3159. structure analysis. IEEE Trans. Comput., C-18, 401– National Centers for Environmental Prediction/National 409. Weather Service/NOAA/U.S. Department of Com- Skamarock, W. C., J. B. Klemp, J. Dudhia, D. O. Gill, D. M. merce, 2000: NCEP FNL Operational Model Global Barker, M. G. Duda, X.-Y. Huang, W. Wang, and J. G. Tropospheric Analyses, continuing from July 1999. Powers, 2008: A description of the Advanced Research Research Data Archive at the National Center for WRF version 3. NCAR Technical Note, NCAR/TN475 Atmospheric Research. Computational and Informa- + STR, 113 pp. tion Systems Laboratory, Boulder, Colorado, USA. Smith, B. L., S. E. Yuter, P. J. Neiman, and D. E. Kingsmill, [Available at https://doi.org/10.5065/D6M043C6.] 2010: Water vapor fluxes and orographic precipitation Nishiyama, K., S. Endo, K. Jinno, C. B. Uvo, J. Olsson, and over northern California associated with a landfalling R. Berndtsson, 2007: Identification of typical synoptic atmospheric river. Mon. Wea. Rev., 138, 74–100. patterns causing heavy rainfall in the rainy season Tsuguti, H., and T. Kato, 2014: Objective extraction of heavy in Japan by a Self-Organizing Map. Atmos. Res., 83, rainfall events and statistical analysis on their charac- 185–200. teristic features. Tenki, 61, 455–469 (in Japanese with Nguyen-Le, D., and T. J. Yamada, 2017: Simulation of trop- English abstract). ical cyclone 201610 (Lionrock) and its remote effect Ultsch, A., and H. P. Siemon, 1990: Kohonen’s self organiz- on heavy rainfall in Hokkaido. J. Japan Soc. Civ. Eng., ing feature maps for exploratory data analysis. Pro- 73, I_199–I_204. ceedings of International Neural Network Conference Nguyen-Le, D., T. J. Yamada, and D. Tran-Anh, 2017: Clas- (INNC), 305–308. sification and forecast of heavy rainfall in northern Ushiyama, T., T. Sayamam, and Y. Iwami, 2017: Devel- Kyushu during Baiu season using weather pattern opment of a flood forecasting system using regional recognition. Atmos. Sci. Lett., 18, 324–329. ensemble prediction – Application to Kinugawa flood Rauscher, S. A., T. A. O’Brien, C. Piani, E. Coppola, F. in 2015. J. Japan Soc. Civ. Eng., 73, I_193–I_198 (in Giorgi, W. D. Collins, and P. M. Lawston, 2016: A Japanese with English abstract). multimodel intercomparison of resolution effects on Whitaker, J. S., T. M. Hamill, X. Wei, Y. Song, and Z. Toth, precipitation: Simulations and theory. Climate Dyn., 2008: Ensemble data assimilation with the NCEP 47, 2205–2218. global forecast system. Mon. Wea. Rev., 136, 463–482. Sakai, R., and M. Yamaguchi, 2005: The WGNE inter-