1. Introduction
Landslide disasters occur very frequently worldwide, causing substantial human loss and property damage [
1,
2,
3,
4]. Such events are predisposed by various physical factors inherent to the slope in question, such as geology, geomorphology, steepness, drainage system, and others [
5,
6], but the major triggering factor in most of the cases is rainwater infiltration [
7]. Most of the landslide disasters around the world happen during the rainy seasons of the respective region, such as around January in South America [
8] or July in Japan [
9]. The problem of rainfall-induced landslides in Japan was exemplified during the July 2018 heavy rain-induced disasters in west Japan. One of the most heavily affected areas was the city of Kure, in Hiroshima Prefecture (
Figure 1).
During the disasters, landslides and massive floods were caused by heavy rains in an event officially referred to as “Heavy Rains of July 2018”. In the course of about 10 days, from 28 June to 8 July, rainfall records reached as high as 1800 mm on the island of Shikoku and 1200 mm in the Tokai region. Many cities recorded more than 400 mm of rainfall over the course of 72 h [
12].
In Hiroshima Prefecture, one of the most affected areas was Kure City, with 10 people became deceased due to landslides during the disasters. Additionally, most transportation lines into the city (except maritime ways) were cut off, and 760 houses were damaged (e.g.,
Figure 2).
The southern part of Hiroshima Prefecture is occasionally affected by heavy rain-induced landslides and flood disasters. Other than the July 2018 event analyzed in this study, other heavy rain-induced landslide disasters were recorded in July 1967 [
14], June 1999 [
9,
15,
16,
17], and August 2014 [
18]. A common factor for these landslide disasters is that they occurred during periods of continuous heavy rain between June and August. Although some predisposed factors (geology, soil condition, slope geometry, etc.) present significant relevance for the occurrence of a slope failure, a triggering factor is always necessary to spark the final break of stability and, consequently, mass movement [
19]. The most common trigger in most landslide-affected areas around the world is an increase in soil water saturation, led by unusually heavy rainfall [
15,
19,
20,
21,
22,
23,
24,
25].
It is also suspected that long-term antecedent rainfall may be related to landslide activation [
26], and that rocks and soil that have been weathered due to high precipitation rates might be more susceptible to landslide activation. Moreover, the recurrence of certain rainfall localization patterns has been noted in various areas of the world, a phenomenon that is usually attributed to the regional topography and subsequent circulation of air around the mountains [
27,
28,
29,
30,
31,
32]. This aspect is expected to be especially remarkable in a study area such as Kure City (and the majority of Japan’s coastline), with its rugged terrain and abrupt altitude variability. Therefore, it is expected that event precipitation localization patterns might be similar to long-term precipitation patterns, making it so that knowing the mean annual precipitation patterns for long-term data may allow for knowledge of future landslide-activation-specific rainfall events.
In view of the need for thorough analysis studies of rainfall patterns and their relationship and relevance with landslide disasters to produce more efficient hazard and risk mapping methods, this analytical work aims to investigate rainfall data in the study area of Kure City (Southern Hiroshima Prefecture, Southwestern Japan) in the context of the July 2018 landslide disasters, using innovative XRAIN (eXtended RAdar Information Network) radar-acquired rainfall data. In Hiroshima, XRAIN started its activities in 2016, roughly two years before the landslide disasters of July 2018. Yokoe et al. [
33], when using XRAIN data to investigate the same July 2018 heavy rain disasters in Southern Hiroshima, evidenced the quantitative accuracy of XRAIN radar data to capture and measure precipitation values in point data when compared to conventional rain gauge methods, with the exception of rainfall intensity exceeding 80 mm/h. Marc et al. [
34] correlated radar-acquired rainfall data in Japan with a specific landslide event and discussed how long-term rainfall and event rainfall localization patterns correlate to slope failure occurrence. They show that rainfall anomalies in a specific rainfall event are closely relatable to landslide activation, thus evidencing the importance of long-term rainfall analysis. Likewise, Moriyama and Hirano [
35] use XRAIN data to investigate the relationship between maximum three-hour cumulative rainfall and landslide occurrence, evidencing that slope failure activation occurs in peak rainfall timing. Cremonini and Tiranti [
36] exemplified a case study allying radar and gauge rainfall measurement methods for forecasting and early warning systems for landslide disasters, showing that radar data, though susceptible to quantitative error due to miscalibration and other technical difficulties, are viable for the detection of landslide hazards. Other uses of XRAIN data include investigation of the development of small-scale guerilla rain clouds [
37], detailed rainfall measurement methods for forecasting [
38], and snowfall precipitation measurement methods [
39].
This study aims to correlate landslide occurrence (during the July 2018 disasters) with rainfall volume distribution in search of localization patterns, as well as investigate whether long-term mean annual precipitation localization patterns (in this case, from 2016 to 2021) both before and after the analyzed landslide event may be co-relatable with precipitation localization patterns of specific rainfall-induced landslide events (in this case, July 2018). The identification of such localization patterns in different time windows may evidence the effectiveness of XRAIN radar-acquired mean annual precipitation data as a landslide conditioning factor in landslide hazard and risk mapping, as well as lead to a better understanding of the effects of rainfall in landslide activation and probability, which may contribute to better strategies in landslide disaster prevention methods.
2. Materials and Methods
The rainfall-induced landslide GIS inventory referent to the July 2018 disasters was provided by the Geospatial Information Authority of Japan [
10]. The landslides were mapped from aerial photographs of the analyzed areas taken by the Geospatial Information Authority of Japan directly after the July 2018 disasters, from 9 to 16 July 2018. Since the whole of the study area was entirely encompassed by the surveyed frame, the landslide distribution of the event is complete and unbiased.
The utilized DEM is provided by the Geospatial Information Authority of Japan [
11] and was acquired with airborne laser survey with 0.2″ interval (5 m resolution) and has 0.3 m vertical accuracy. Other miscellaneous GIS data were also provided by the Geospatial Information Authority of Japan [
11].
Other details regarding the study area and the utilized database and analysis methods are presented in the following sections.
2.1. Study Area
The area analyzed in this research comprises a 390.5 km
2 rectangle around the municipality of Kure in Southern Hiroshima (
Figure 1). Stranded between the Hiroshima Mountains in the north and the Seto Inland Sea to the south, the city was a small shipbuilding town that experienced rapid growth in the first half of the 20th century, which forced urbanization in areas at or adjacent to the mountainous terrain.
The mountains are mostly composed of volcanic rocks, namely rhyolites and granites/granodiorites from the Hiroshima Group [
40]. When weathered, these rocks erode into a soil commonly referred to as Masado, which is known to be highly permeable and brittle when wet, configuring a material prone to slope failure during rainfall events [
25]. This setting, common around the country’s coastline, makes the city a potential high-risk area for landslide disasters.
The Seto Inland Sea (to which Kure is adjacent) has little rainfall compared to the surrounding oceanic coastal areas in Japan, like the Sea of Japan and the Pacific Ocean. Although a substantial part of the oceanic precipitation clouds is blocked either by the Chugoku Mountains northward or the Shikoku Mountains southward, and the region is relatively dry [
41], heavy rainfall is particularly concentrated in mountainous areas. In Kure, the average annual precipitation ranges from 1000 to 1600 mm, characterizing a relatively mild rainy zone. Mountain areas around the Seto Inland Sea, however, reach annual average precipitation of 2000 mm to 3000 mm. The period of the year with the heaviest rainfall occurs between June and July every year when the average precipitation reaches 227 mm in a month [
42].
2.2. XRAIN Radar-Acquired Rainfall Data
XRAIN radar technology started to be utilized in Japan in the year of 2014, operated by the country’s Ministry of Land, Infrastructure, Transport and Tourism. The technology used in the measurements differs from common radar rainfall data since it uses Multi-parameter (MP) radars, which allows for more accurate measurements of rainfall volume.
Although not as quantitatively accurate as regular rain gauge measurements, radar-based rainfall measurement methods have the advantage of being performed over a bi-dimensional “planar” area, where each pixel in the area’s grid represents a specific value, whereas rain gauge methods extrapolate the value of a single measurement station over extensive regions. This means that radar-acquired data allow for rainfall distribution analysis on a larger scale.
In landslide hazard assessment studies, rainfall data usually comprise mean annual precipitation data collected via rain gauge stations. Typically, each station is referent to a whole municipality and is located near city centers. In the case of the study area, Japan’s Meteorological Agency (JMA) has a measurement station in downtown Kure, with the other nearest stations being at Kurahashi, 16 km southwards, and at Hiroshima, about 18 km northwest.
Slope failure assessment and analysis using rainfall data have been widely performed in the literature, including for the study area of this work [
9,
16]. However, rain gauge stations gather information on the scale of whole municipalities, which may not be representative of the actual spatial distribution of precipitation in a degree of detail considered ideal for various methods of slope failure assessment. In the case of Kure City, for example, since the nearest rain gauge measurement station is located about 16 km south of the Kure station, mean rainfall data would have, at best, 16 km of accuracy. In reality, rainfall intensity values may vary in the order of less than hundreds of meters, especially in areas with rugged mountainous terrain or coastal regions. However, recent advancements in radar technology, such as XRAIN (eXtended RAdar Information Network), provided by the Data Integration and Analysis System [
43], have allowed for instant measurement of rain intensity in much more detailed scales of spatial distribution.
XRAIN data are represented in 287 × 230 m pixel grids, where each pixel value represents the rainfall intensity in mm/h for the referred location at the time of measurement, which occurs every 1 min. The data are obtainable in the Data Integration and Analysis System (DIAS) platform, which is operated by the University of Tokyo and sponsored by the Ministry of Education, Culture, Sports, Science and Technology (MEXT). The measurement spacing can be set as 1, 5, 10, 15, 30, and 60 min.
Although XRAIN raw data express rainfall intensity, precipitation volume can be estimated by the calculation of intensity during a specified period either by averaging the measurements over that period and multiplying it by the number of times that the measurement interval is repeated in it, or simply by summing the intensity values in the case of a 1 h interval analysis with no missing measurements.
In this work, the XRAIN data were downloaded as .zip packed .csv files spaced in 5-min, 30 min, or 1 h intervals. The .csv files comprise tables with cells spatially organized so that each cell represents a 287 × 230 m pixel in a north-oriented grid representing the designated area, and each cell value expresses the rainfall intensity in mm/h at the time of measurement. For the study area, each of the files comprised of a 97 × 67 grid with 4999 pixels. XRAIN data collection for the area of the Chugoku region (where Kure is situated) started in 2016, so the range of collected data spans from 2016 to 2021, summing up to more than 60,000 .csv files. These files were combined in single worksheets in Microsoft Excel for calculations, analysis, and then conversion into ArcGIS shapefiles.
2.3. Data Analysis
After proper conversions, the data were jointly analyzed in ArcGIS Pro for localization and spatial distribution inspections and in Microsoft Excel for other statistical analysis and data visualization.
The investigation methods include determining average precipitation throughout the analyzed time windows (from the year 2016 to 2021) for the whole study area, as well as its spatial distribution. With the use of XRAIN data, precipitation is calculated by acquiring hourly rainfall intensity for each cell of the study area throughout the analyzed period, with a one-hour interval between each intensity measurement in the case of annual precipitation. The hourly rainfall was then converted to precipitation data by multiplying the intensity values by the number of hours in the desired period (e.g., 8760 for a common year). The general flow and steps of the research are illustrated in
Figure 3.
2.3.1. Localization Patterns
The reason why accumulation data were analyzed in the format of one year is that meteorological patterns tend to repeat themselves in the cycle of a year. It is expected that rainfall localization and quantification will generally recur after the interval of a year, while the same is not expected in monthly or semiannual intervals. The repetition of rainfall localization patterns over the years evidences the tendency of certain areas to receive more precipitation and, thus, be more susceptible to landslides, both in the sense of soil and rock weathering by rainfall and also in the sense of slope failure triggered by soil saturation during high precipitation events. Thus, it is considered that annual rainfall analysis is appreciated in looking for localization patterns, even in years previous and subsequent to when the case study landslide disaster event occurred. This hypothesis is judged by visualizing these localization patterns in map view along with the actual landslide occurrence points of the July 2018 disasters.
The relationship between localized precipitation and landslide occurrence was further explored by calculating landslide density based on precipitation class zones. The precipitation class zones vary for each analysis period, depending on the localization and quantity of precipitation for the analyzed period. For each analysis (total rainfall from 2016 to 2021 and event rainfall between 5 and 7 July 2018), five precipitation zones (very low, low, medium, high, and very high) and their boundaries were determined by calculating five equal intervals based on the total precipitation value of the analyzed period. The landslide density for each zone is determined with the use of the “Spatial Join” analysis tool in ArcGIS.
In order to investigate the differences and similarities between the annual rainfall patterns and to understand the triggering of slope failures in the July 2018 disasters, along with the map visualization of localized precipitation in the same fashion as the annual rainfall analysis, a timeline of rainfall intensity in the study area was constructed during the 2-day period of rainfall pertaining to the July 2018 disasters. This was carried out by arranging 30 min interval XRAIN rainfall intensity measurement data from the whole study area in a 47 h (period of continuous rainfall during the disasters) span with 30 min accuracy. The timing of rainfall intensity was then compared to disaster documentation, as well as localization patterns along the area.
2.3.2. Intensity–Duration Rainfall Threshold
Empirical slope failure thresholds are based on the statistical analysis of previous rainfall history that has not resulted in landslides. These thresholds are usually expressed by plotting lines in Cartesian, semi-logarithmic, or logarithmic coordinates, where values above the line are considered prone to slope failure occurrence. In empirical slope failure thresholds, the use of data reflecting rainfall conditions that did not result in slope failure triggering is also necessary for efficient sampling [
22]. Extensive research and attempts of empirically based threshold models evidence that the threshold values cannot globally represent the situation for slopes and climate settings in any part of the world, which leads to the necessity of developing different models depending on the geographical situation or at least co-relatable settings [
44].
Intensity–duration (ID) thresholds are the most widely adopted type of threshold analysis and are considered efficient for making a relationship between the intensity of event rainfall and the accumulation of precipitation from the start of rainfall to the failure event. ID thresholds are usually expressed as a line plotted over Cartesian coordinates with logarithmic values, where rainfall duration until the event is represented by the horizontal (X) axis and event rainfall intensity is represented on the vertical (Y) axis. The threshold lines are usually expressed in the form of the power law equation:
where
I is rainfall intensity,
D is rainfall duration, and
c,
α and
β are parameters based on the study area. Since ID thresholds always exhibit situations where higher intensity requires less duration for landslide activation,
β is always a negative value, characterizing a negative power law. According to the collection of threshold values summarized by Guzzetti et al. [
22], the majority of the ID threshold equations across the bibliography exhibit
c = 0,
α between 4 and 176.4, and
β between −2 and −0.19.
The intensity–duration threshold analysis always makes use of a defined duration range, given by the duration between the first analyzed landslide event and the last (first event duration < D < last event duration). Local threshold values usually have smaller ranges due to the narrower sampling data available for analysis, which is also reflected in higher threshold values.
2.3.3. Event Precipitation and Mean Annual Precipitation Comparison
In order to investigate whether there is a correlation between spatial and localization patterns of rainfall when comparing event precipitation (in this case, along the 48 h between 5 and 7 July of 2018, when the landslide disasters of that event occurred) and long-term precipitation (for the years between and including 2016 and 2021, when XRAIN data were available for the study area), relative precipitation values for event precipitation and mean annual precipitation and their differences were calculated, and Pearson’s product-moment correlation coefficient (PPMCC) was employed to compare these two datasets.
What is here defined as relative precipitation is the percentage relative to the maximum precipitation which the sample cell represents for a given period. This results in a normalized value representing the relative precipitation of XRAIN precipitation cells in the study area, allowing for a quantitative comparison of rainfall localization patterns between different time periods (i.e., event precipitation over 47 h and mean annual precipitation between 2016 and 2021). Thus, the relative precipitation difference is obtained simply by subtracting the mean annual precipitation from the relative event precipitation, both in percentage values. A negative relative precipitation difference value indicates that event precipitation was less intense than the usual mean annual precipitation in that cell, while a positive value indicates that it was more intense, and a value near 0% suggests that there was no big deviation of localized precipitation in that area and that there is a good fit between the two datasets.
The PPMCC value, represented by
r, is a measure of linear correlation between two sets of data, given by the ratio between the covariance of two variables and the product of their standard differences [
45]. A value closer to 1 represents a good fit between compared datasets. It is given by the following formula:
where
x and
y are the sample means for each dataset, which in the case of this study are the collection of XRAIN-measured precipitation pixels for the event precipitation and the long-term mean annual precipitation, both normalized to percentage values. Only pixels above land were considered in the calculation.
4. Conclusions
In order to better comprehend the relationships between precipitation and landslide occurrence and investigate the recurrency of rainfall localization patterns throughout the years, rainfall data were analyzed along different ranges of intervals in terms of intensity, volume, and localization using the landslide events around Kure City (Hiroshima Prefecture) during the July 2018 heavy rain disasters as a study case.
Considering the rainfall events of the July 2018 disasters, it was observed that Kure City experienced heavy continuous rainfall starting at 8:30 AM on 5 July, a condition which continued for about 47 h until the cessation of rainfall at 7:30 AM on 7 July. XRAIN data show that the precipitation accumulated up to 368 mm and that the mean rainfall intensity was 7.8 mm/h. There were two peaks of rainfall intensity, one 35 h into the event at 7:30 PM of 6 July, when rainfall intensity reached 47 mm/h, and another 44 h into the event, at 4:30 AM of 7 July, when rainfall intensity reached 40 mm/h. These peaks are associated with high landslide activity according to records of the disasters. Considering the rainfall localization during the July 2018 disasters in Kure City, it was observed that landslide density peaks in high precipitation class zones, with 16 landslides per km2 in the 403.10 mm to 434.45 mm precipitation zone. There is, however, a decrease in landslide activity in maximum precipitation zones, which is attributed to the flat mountain peak topography or hard rock material associated with elevated areas.
XRAIN data were also utilized together with reports of landslide occurrences in eight locations of Southern Hiroshima in order to calculate an intensity–duration threshold for the area, resulting in the threshold of I = 133.44 × D
−0.841, where I is the average rainfall intensity until failure, and D is rainfall duration until failure. When compared with Guzzetti et al.’s [
22] collections of rainfall thresholds for landslide activation throughout the world, it is noted that the calculated threshold for Southern Hiroshima is found slightly above the world average.
Analysis of rainfall data from 2016 to 2021 demonstrated that the mean annual precipitation amounts to about 2300 mm in the study area. Considering the spatial distribution of rainfall volumes around the study area, the XRAIN data show that precipitation volumes are highly localized, with intense rainfall values being concentrated in locations of elevated topography. However, peak landslide density is found in areas associated with intermediate precipitation volumes, peaking on 9.2 landslides per km2 in the 2427.43 mm to 2628.42 mm precipitation class zone. Maximum precipitation class zones, however, show decreased landslide activity when compared to the intermediate zones.
The decrease in landslide activity in maximum precipitation areas, observed in the mean annual precipitation localization maps and also, to some extent, in the event precipitation localization maps, may be explained by the fact that the maximum precipitation zones are associated with high topographical elevation areas, usually referent to bedrock-weathered mountain peaks and extremely steep slopes, which are not prone to landslide activity.
Comparing the rainfall localization patterns in the study area of the event precipitation during the 47 h of the July 2018 disasters and the mean annual precipitation from 2016 to 2021 by checking the relative precipitation differences in the two datasets indicates that although not many XRAIN cells show differences higher than 30%, particularly high landslide activity is closely related to areas of peak positive relative precipitation difference, that is, where event precipitation was more intense than mean annual precipitation. This conclusion of strong correlation between landslide occurrence and extreme rainfall anomalies is also pointed out by Marc et al. [
34].
Using Pearson’s correlation coefficient, an r value of 0.55 was found, which is considered a moderate correlation. Although the correlation is not perfect or even very high, a positive relationship is found, which points out that mean annual precipitation localization patterns may indeed be used to forecast what the localized rainfall may be in a specific future event. It is judged that these patterns are controlled by the topographical features of the area (which is shown in this study by how maximum rainfall volumes are usually concentrated in peak topographical areas) or by meteorological dynamics of rainfall movement.
In this research, it was found that landslide activity is more co-relatable with high precipitation volumes in event rainfall and not so much with mean annual precipitation volumes, where peak landslide density is found in intermediate precipitation zones, though maximum rainfall volumes show a decrease in landslide activity in both observations. Relative precipitation difference, however, was found to be closely related to landslide activity in a directly proportional behavior. An ID rainfall threshold for landslide activation was calculated for the Southern Hiroshima area using XRAIN radar-acquired rainfall data. Finally, it was evidenced that long-range localized precipitation patterns are moderately co-relatable with event precipitation localization patterns. Recommendations for future studies on the subject include using mean annual precipitation data in landslide susceptibility mapping approaches, as well as further investigation into what factors influence rainfall localization.