The present study utilized several datasets: (1) weather station data, (2) TC track, intensity, and size data, (3) national survey data of TC damage, and (4) national survey data of regional wealth. This section will describe how different datasets were obtained and processed before the statistical risk analysis.

First, from 60 weather stations throughout South Korea (see Fig. 4 for station locations), daily maximum near-surface wind speed and daily accumulated precipitation values were gathered (Korea Meteorological Administration, 2015).

Second, TC information including track, intensity, and size was obtained from the Regional Specialized Meteorological Center (RSMC) best-track data (RSMC Tokyo, 2015). For intensity, we used the maximum wind speed and central pressure data. For TC size, we used the largest radius of 30 knot winds, which is specifically provided by RSMC. RSMC best-track data in the 6 h interval were interpolated to a 1 h interval to obtain precise hazard values at landfall (Park et al., 2011). The interpolated RSMC best-track data were used to select the TCs that made landfall on South Korea (TCs entering the area within 3∘ of the coastline of Korea). Then, we verified these TCs with the official influential TC record in the Typhoon White Book issued by the National Typhoon Center (NTC, 2011), as in our previous study (refer to Park et al., 2016, for more details on NTC Typhoon White Book). We chose only the TCs that had maximum wind speeds greater than or equal to 17 m s-1 (above or equal to tropical storm (TS) class) at the time of entering the 3∘ line. As a result, 85 TCs were determined to be influential in South Korea from 1979 to 2010 (see Supplement Table S3 for the list of the names and years of all the target TCs).

Third, we used damage data from the National Disaster Information Center (NDIC) of the Korean government (NDIC, 2015) after following procedures. NDIC property loss data consist of data on monetary damages to industrial, public, and private facilities, standardized to the value of money in 2005 by accounting for inflation. The loss data were collected by local governmental offices, and therefore most losses could be reported regardless of whether the victims were insured or uninsured. There may, however, be some cases of minor losses that were not reported to the local offices by the victims. The raw dataset included damage data caused by all types of extreme weather such as TCs, heavy monsoon rainfall, heavy snowfall, or high waves. Some cases were not classified by specific damage sources. Some cases were categorized under high wave damage, although they were also caused by TCs as the high waves were induced by TCs. Therefore, we matched all the raw loss data to the TC data using the NTC White Book (NTC 2011) and RSMC best-track data (RSMC Tokyo, 2015). To be more specific, we compared the three relevant periods: (1) the period of warning issued recorded in the NTC White Book (NTC, 2011), (2) the number of days that the TC stayed within 3∘ of the Korean coastline based on the RSMC dataset (RSMC Tokyo, 2015), and (3) the period of damage occurrence recorded in the NDIC dataset (NDIC, 2015).

The NDIC cannot differentiate the damage from multiple hazards when there are multiple successive extreme phenomena. For example, if a heavy rainfall watch started on 15 July, and then a TC came to South Korea on 20 July and decayed on 22 July, and there was no gap between the rainfall and TC advisories, NDIC aggregates the damage amounts and records the damage period as 15 to 22 July. Therefore, to confine the origin of the loss data to one TC, we excluded cases whose damage period exceeded 5 days from landfall.

Fourth, province-level aggregated wealth data were obtained from government statistical surveys (Korean Statistical Information Service, 2015). We aggregated the 17 districts of South Korea into 5 provinces because the administrative division had been changed between 1979 and 2010, and the size of the 17 districts varies from city size to province size. The names of the provinces are Gyeong-gi (GG), Chung-cheong (CC), Jolla (JL), Gang-won (GW), and Gyeong-sang (GS) (See Fig. 2 for the distribution of the provinces). These five provinces have independent records of damage for every influential TC case and annual regional wealth. The temporal variation of wealth was considered through the normalization of damage data to the reference year, 2005, with wealth per capita. In general, the wealth of South Korea has consistently increased. However, there are significant differences in the growth rates among provinces, which affect the TC damage records. Through normalization, the potential impact of regional differences in wealth trends was eliminated. The spatial disparity of wealth at a certain time (i.e. 2005) should be addressed when mapping the damage distribution.

In order to objectively evaluate the effect of each TC track on damage, a total of 85 TCs which influenced South Korea during 1979–2010 were grouped into four track patterns. The four TC track patterns can be characterized as (1) east-short, (2) east-long, (3) west-long, and (4) west-short types based on the position and length of the TC tracks around the Korean Peninsula (Fig. 2). Although the average zonal distance between the mean tracks of east-types (i.e. east-short and east-long) and west-types (i.e. west-short and west-long) was only about 250 km, hazards (both potential and active) and damages caused by the TCs are significantly different depending on the four TC track patterns at the 99 % confidence level based on the Kruskal–Wallis test (Fig. 3). This highlights the importance of the track in TC risk assessment because the 250 km distance is not long considering that the average errors of track forecasting in the western North Pacific are about 200 and 400 km for 24 and 48 h, respectively (Roy and Kovordanyi, 2012). Meanwhile, the high sensitivity of damage on the track shown in Fig. 3 suggests that the current skill of TC track forecasting may not be enough to exactly estimate TC risk distribution over South Korea in advance of 1 day and over.

As shown in Fig. 3, potential hazards display different results from active ones, although both originate from the same TC. Potential hazards are stronger in longer tracks, i.e. east-long and west-long, while active ones are stronger in west-types than east-types (compare Fig. 3a–c to d–f). In addition, even if potential hazard parameters have been widely used in TC risk analysis (Nordhaus, 2010; Hsiang and Narita, 2012; Czajkowski and Done, 2014; Zhai and Jiang, 2014), they show worse accordance with damage than active ones. For all potential hazards, the ranking is in order of east-long, west-long, west-short, and east-short. It is natural for a TC with a longer track to have higher wind speed, deeper central pressure, and larger size, since a TC with a stronger intensity should be more durable compared to a weaker TC under the same environmental conditions, such as friction, vertical wind shear, and sea surface temperature (Kim et al., 2011). In contrast, for active hazards, the ranking is in order of west-types, east-long, and east-short track patterns. This relationship between track and active hazard parameters (i.e. near-surface wind, rainfall, and influence duration) cannot be simply explained unlike potential hazard parameters.

Figure 4 shows the spatial distribution of active hazard parameters with the topography for each track pattern. Focusing on the near-surface winds (Fig. 4a–d), the near-surface winds of west-type tracks were comparable to near-surface winds of east-long track, particularly along the coast, even though potential hazards of west-type tracks are significantly weaker than those of east-long tracks (Fig. 3a–f). This can be attributed to the concept of dangerous and navigable semicircles. In the case of west-type tracks, South Korea falls within a dangerous semicircle (right-hand side of the direction of TC movement) in which the TC translation speed and rotational wind field are additive, and hence, strong wind speed is observed therein. In contrast, in the case of east-type tracks, the country is located under a navigable semicircle (left-hand side of the direction of TC movement), in which the TC translation is counter-directional to the rotational wind. Therefore, weaker wind speeds are found there than that in the dangerous semicircle.

In terms of rainfall, much heavier rainfall was found in the west-type tracks than that in east-type tracks along mountainous areas, particularly the Sobaek mountains (Figs. 2c and 4e–h). Heavy rainfall along the mountains can be explained by the orographic lifting effect. When a TC is located in the southwest of South Korea, the eastern sides of the Sobaek mountains become the upstream slope of the tangential wind of the TC, causing more torrential rainfall than the inherent rain band of the TC (Park and Lee, 2007; Lin et al., 2002). Thus, the orographic lifting effect can be maximized by west-type tracks but not east-type ones.

Finally, the influence duration was distinctly longer for west-type tracks compared to east-type tracks. Only a TC with a west-long track penetrates the country, and hence, the west-approaching TCs could affect a more extensive area for a longer time. The long influence duration of the west-short tracks was possibly because a TC with a west-short track moves the slowest, and it could stay in the same area for the longest time (significant at the 95 % confidence level, Kruskal–Wallis test). Note that we calculated influence duration for each station by applying the same criteria for wind and rainfall. A station was marked as “influenced” if either the daily accumulated precipitation or daily maximum sustained wind speed recorded at that station on the specific day exceeded the station's critical thresholds, which we set as the 90th percentile of each station's precipitation and wind speed records over the analysis period (1979–2010).

Looking at Fig. 3, the ranking of active hazards was exactly same as that of damage. In addition, the spatial distribution of damage also matched well with those of active hazards (compare Figs. 4 and 5). The area where active hazards are high, exhibited high risks. There was only one exception, for west-types, the southwestern province (Jolla, JL) recorded less damage than the southeastern province (Gyeong-sang, GS), although stronger active hazards are found in JL than GS. This discordance was partly explained by exposure disparity. GS possesses higher wealth compared to JL. After the damage was divided by regional wealth (parentheses of Fig. 5), the spatial distribution of damage became more analogous to that of active hazards, even if the damage in GS is still slightly higher than in JL for the west-short tracks (Fig. 5d). This may be related to different vulnerability to TCs between the two provinces. Since GS is more mountainous than JL, the vulnerability of GS to TC rainfall can be higher than that of JL. Nevertheless, all the active hazard parameters (r=0.62) showed much higher correlations with damages than potential hazard parameters (r=0.29), although most of the potential hazards displayed statistically significant correlations at the 95 % or 99 % confidence levels (Table 1).

All the results in this section suggest that active hazards are better indicators of TC risk than potential ones. In other words, if we want to a predict whether there would be damage to a city or not and if the active hazard information is available, we may not need to gain any additional information of potential hazards. Nevertheless, this does not mean that potential hazards are not important. We have to utilize a climate model with a fine resolution of at least less than 10 km for realistic simulation of active hazards, i.e. wind and rainfall (Park and Lee, 2007; Lee and Choi, 2010), which is a difficult task and requires high computing power. Hence, active hazards seem not to be optimal for risk forecasting to help emergent decision making, as well as climate change research with large spatiotemporal scale. In this respect, it can be more valuable to use potential hazards for risk assessment if we identify an additional factor that can fill the gap between active and potential hazards.

Here, through the decision tree analysis, the importance of track in TC risk was investigated. Through the decision tree analysis, the following three questions could be answered: (1) what is the most effective factor for classifying “damage” and “no damage” cases for TCs making landfall on South Korea, (2) how do different factors in combination determine damage occurrence, and (3) what critical values of the factors can be used as quantitative guidelines related to TC damage occurrence? Here, the decision tree model was designed to objectively classify whether a TC will bring damage to a province or not; the decision tree used potential hazards and track as input variables (see Table S1 for more information about the input variables). Overall, we had 355 effective cases, comprising of 160 “damage” cases and 195 “no damage” cases; we only considered damage occurrence in each province by a TC (see Table S2 for detailed information of damage cases).

According to the decision tree, track information acts as the primary determinant of TC risk. Information about the track pattern was nominated as the first splitting attribute (Fig. 6). This means all 355 cases should be classified by the track group prior to all other decision nodes in order to reach the end nodes. In other words, the most important factors of TC risk may be neglected when performing a risk analysis without track information. The detailed process is as follows. First, the model simply sent all west-type TCs to the end damage node. Next, the east-type TC cases were assessed according to province and TC intensity (maximum wind speed). For a TC in the east-long group, damage could occur in JL, GS, and GW provinces. Particularly, for the GW province, damage will be inflicted only if the maximum wind speed is greater than 41.1 m s-1. East-short cases, unlike east-long cases, were sent to the intensity criteria before the province criteria. East-short TCs with weak intensity (maximum wind speed less than 25.7 m s-1) were directly linked to the no damage node. The east-short TCs with satisfactory intensity (maximum wind speed greater than 25.7 m s-1) were sent to the province criterion; even though it is a small portion, the critically strong east-short TCs can cause damage in the southern provinces.

The relative importance of the variables in each decision tree was offered quantitatively in terms of the usage rate by the See5/C5.0 algorithm. When an attribute is the most-related variable to the target variable, the attribute should be used most frequently for classification by a decision tree model. In our decision tree (Fig. 6), the track group variable was used in 100 % of the cases; province and potential hazards then follow with usage rates of 48 % and 37 %, respectively. Therefore, we can say that for risk determination, TC track was the most important attribute, which gives essential information on TC risk analysis. The use of the province variable as the second most important variable was mainly related to the relative location of the province with respect to the TC centre along the track. Southern provinces are generally closer to the TC centre regardless of the four track types because TCs move from the south (low latitude) to the north (high latitude). Potential hazards were the third most important attribute. Maximum wind speed was utilized as an effective classifier, but TC size was not used.