Spatial analysis of historical extreme rainfall characteristics using regionalization in the Lake Biwa and Yodo River Basin, Japan

The analysis of historical extreme rainfall characteristics is important for flood management planning. Flood management planning is usually conducted on a regional scale, e.g., basin-scale, taking into consideration any historical extreme rainfall characteristics and their spatiotemporal distributions in the target basin. This is because rainfall is the main driving force of the water cycle as described by general hydrology. Many studies have focused on the spatial rainfall characteristics on the scale of a country or even a still broader scale (e.g., in South America (Haylock et al. 2006), Malaysia (Zin et al. 2010), Taiwan (Su et al. 2012), Montenegro (Burić et al. 2015), and over the Indochina Peninsula (Muhammad & Usa 2015)). Basin-scale rainfall characteristics have also been evaluated in recent studies (for example, in Peru (Casimiro et al. 2012), China (Song et al. 2015), and Malaysia (Yazawa et al. 2019)).

Most of these studies used rainfall data obtained from satellites or simulated through climate models to support the missing information caused by the sparse distribution of rain gauge stations or the lack of observation data by incorporating the modeled rainfall data with the observed ones in the target area. The underlying problem of this is the limitation of access to said simulated weather data, particularly for local flood management planners, because the data are mainly available for and limited to research purposes. If local planners do not have enough knowledge or the skills to gather data from the satellite and/or climate models, it becomes difficult to apply recent research methods to the actual flood management planning. In Japan, the weather data obtained from the rain gauge stations are managed by the Japan Meteorological Agency (JMA) and are available to the public via its website. Flood management planning in Japan is performed by estimating the probable maximum rainfall/precipitation and simulating peak flows through hydrological frequency analysis (HFA) as directed by the Ministry of Land, Infrastructure, Transport and Tourism in Japan (Ministry of Land, Infrastructure, Transport & Tourism 2005). In the Lake Biwa and Yodo River Basin, which is the target area of this research, extreme rainfall characteristics, such as probable maximum precipitation (PMP), have been analyzed using the rainfall data obtained from the JMA. The spatial distribution of PMP in the Lake Biwa and Yodo River Basin was analyzed using the previous studies (Alias et al. 2013a, 2013b) and a spatial interpolation method.

This research focuses on the spatial extreme rainfall characteristics of the Lake Biwa and Yodo River Basin. While the aforementioned studies (Alias et al. 2013a, 2013b) have only estimated PMP, this research analyzes the other rainfall indices and relevant spatial distributions using a spatial interpolation technique to provide detailed information for the purpose of flood management planning in the basin. The investigation of the spatial distribution of hydrological information using a spatial interpolation technique is usually facilitated today by a Geographic Information System (GIS), known as regionalization (Takara & Oka 1992; Mano & Nakayama 2008; Rau et al. 2017). Regionalization can generate hydrological information in regions where there are no/fewer rain gauge stations according to the data obtained from the existing rain gauge stations surrounding the target regions. Thus, the regionalization of the hydrological information enables us to make regional flood management plans in areas where the rain gauge stations are sparsely distributed or absent.

Several data-driven methodologies that could also be used for regionalization have been applied to the areas where the rainfall data are missing. For example, Sattari et al. (2020) used the support vector machine (SVM) regression method to predict the missing monthly precipitation in Turkey. The methodology was applied to the meteorological stations under similar climatic conditions. Thus, the applicability of the methodology to areas with different climatic conditions is not yet clear. For the areas with different climatic conditions, Chen et al. (2022) applied a deep-learning-based long short-term memory model to the monthly rainfall forecast. Since the model requires long-term observation records as the training data, it would be difficult to make an accurate prediction for the areas with only a short-term observation record. Wang et al. (2021) compared the performance of three data-based models, the back-propagation neural network (BPNN), the group method of data handling (GMDH), and the autoregressive integrated moving average (ARIMA), to forecast the monthly rainfall forecast in China. They concluded that the BPNN has good performance when the Wavelet Packet Decomposition was combined. Overall, the data-based methodologies used in these studies need long-term records for the training and testing of the models. In addition, the purpose of developing these new methodologies was mainly for rainfall prediction, not for flood risk assessment or decision-making.

This research examines the historical changes in both the general and extreme rainfall characteristics in the basin. More specifically, the return period of an extreme rainfall event, like a planning scale, is used as one indicator since it is one of the most important factors in flood management schemes (Ministry of Land, Infrastructure, Transport & Tourism 2005). For example, changes in the return periods of extreme rainfall events have been analyzed in South Korea (Seo et al. 2015) and Malaysia (Yazawa et al. 2019), respectively. Understanding the time series changes in the return periods of extreme rainfall events is also important for flood control. The return period of an extreme rainfall event decreases as the frequency of extreme rainfall events in the basin increases. For example, even if the basin sets the planning scale as a 100-year rainfall event based on the historical rainfall data, the rainfall event might drop to 80 years based on the updated rainfall data because of additional rainfall events. The decrease in return period simply means that the area experiences rainfall events with a higher intensity as time passes because the estimation method of probable rainfall commonly uses the annual maxima and calculates the rainfall amount corresponding to a specific return period, such as 100 years, by extrapolation. Understanding the changes in the return periods is necessary to update the regional flood management plan using the data collected annually (Yazawa et al. 2019).

The application of GIS is still a popular approach for disaster risk assessment because of its efficiencies in terms of time, cost, and resources (Montoya 2003) and its applicability for decision-making (Tomar et al. 2021). GIS has been applied to flood risk assessments through the geoprocessing and/or modeling approaches (Chakraborty & Mukhopadhyay 2019; Mahmood et al. 2019; Waghwala & Agnihotri 2019; Cabrera & Lee 2020; Tomar et al. 2021). However, this GIS-based flood risk assessment research required a variety of spatial information data such as elevation, land cover/use, soil distribution, etc. This research, on the other hand, technically used only rainfall data and utilized the spatial interpolation technique through GIS. The methodology employed in this research, which is the combination of utilizing rainfall indices and spatial interpolation, has the advantage of enabling the spatial analysis of extreme rainfall in the area using both short-term and long-term rainfall observations. Thus, the methodology and findings obtained from this research could be beneficial for both regional flood management planning and decision-making that was not considered in the previous studies of extreme rainfall analysis. Providing insights into the extreme rainfall characteristics as part of the flood management planning for the Lake Biwa and Yodo River Basin has been done in two steps in this research. First, there were the historical changes in general and the extreme rainfall characteristics checked using the rainfall indices to understand the rainfall trends in the basin. Then the spatial distributions of the rainfall indices were investigated using the spatial interpolation method of GIS to regionalize each rainfall index.

Table 1 shows the catchment areas of the main lake and tributaries in the Lake Biwa and Yodo River Basin (Lake Biwa-Yodo River Water Quality Preservation Organization 2021). The catchment area of the Lake Biwa and Yodo River Basin is 8,240 km². The population in the basin was approximately 12.1 million as of 2015. The basin is the main water resource for about 14.5 million people in the Japanese western region (Lake Biwa-Yodo River Water Quality Preservation Organization 2021). The rivers and their tributaries in the basin supply water for the main western cities such as Osaka and Kyoto.

The Lake Biwa and Yodo River Basin have a decades-long history of integrated management. The basin is referred to not only as Japan's largest lake–river system but also historically as the most successful basin from the perspective of water resource management involving local governments, private sectors, NGOs, and citizens of the six prefectures from upstream to downstream (Bamba 2011; Sharip et al. 2021). Thus, the basin is one of the largest study areas regarding integrated water resource management in Japan (Nakamura & Rast 2014; Nakatsuka et al. 2020).

In the Lake Biwa and Yodo River Basin, the planning scales (i.e., return periods) for flood management differ depending on the importance of the rivers as classified by the local governments. This classification is based on the area, population, assets, etc. in the basin. The goals of the planning scales in basin flood management are currently set at 200 years for the Yodo River and 150 years for the Uji, Katsura, and Kizu Rivers. However, the actual planned scales for most of these rivers are relatively lower than the goals in place as extreme rain and flood events have been occurring. Thus, achieving sustainable and integrated water resource management poses a significant challenge in the Lake Biwa and Yodo River Basin (Union of Kansai Governments 2015; Nakatsuka et al. 2020).

For the purpose of this research, daily rainfall data were collected from 61 rain gauge stations (24 rain gauge stations inside and 37 stations surrounding the Lake Biwa and Yodo River Basin) as shown in Figure 1 using the Automated Meteorological Data Acquisition System (AMeDAS) of JMA until the year 2019. The details, such as longitude, latitude, elevation, and the year the observation started, of the rain gauge stations are shown in Table 2. The year when the rainfall observation started varied depending on the rain gauge station (with the oldest at the Kyoto observatory dating from 1881 and the most recent at the Shuuchi observatory dating to 1982). There are four more rain gauge stations (two in Shiga prefecture and two in Osaka prefecture) inside the basin according to JMA. However, these stations were excluded from the data collection and analysis in this research since they are relatively new as the observations started in 1991.

The collected rainfall data were then divided into the determined Historical 1 (hereafter, His 1) and Historical 2 (hereafter, His 2) scenarios for the rainfall analysis conducted in this research. The His 1 scenario used data from the first half of the observation period at each rain gauge station. The His 2 scenario used the data from the second half of the observation period at each station. Changes in the general and extreme rainfall characteristics of the basin were investigated by comparing the changes in the rainfall indices between the His 1 and His 2 scenarios.

To grasp the comprehensive rainfall characteristics of the basin, nine rainfall indices were analyzed including two general rainfall indices and seven rainfall indices representing extreme rainfall events (Yazawa et al. 2019). This was based on the daily rainfall data from the His 1 and His 2 scenarios at each rain gauge station. Table 3 shows the definitions of the nine rainfall indices investigated in this research. The first two indices, the average annual rainfall (AAR) and the average number of rainy days (ANORD), describe the general rainfall characteristics. The other seven are used to represent the extreme rainfall characteristics. These were part of the simple precipitation intensity index (SDII) (Kitoh et al. 2013): very and extreme wet day intensities exceeding 95th and 99th percentiles (I95 and I99); very and extreme wet day proportions (R95 and R99); and the number of very and extreme wet days exceeding the 95th and 99th percentiles (N95 and N99). In the previous studies, there were other indices used to explain heavy/extreme rainfall such as days of heavy (the number of days that the rain rate is more than 10 mm) and very heavy (the number of days that the rain rate is more than 20 mm) precipitation (Haylock et al. 2006; Sheikh et al. 2015; Kruger & Nxumalo 2017). This research alternatively used the 95th and 99th percentiles to represent these heavy and extreme rainfall conditions and to determine the thresholds of extreme events that represent very wet days and extremely wet days, respectively (Zin et al. 2010). Percentile-based thresholds could make the relative comparison by considering the different lengths of the records at each rain gauge station. The analysis in this research used the 1.0 mm/day threshold to extract rainy (wet) days since this is the threshold defined by JMA.

There are two main types of method, physical (hydrometeorological) and statistical, that are able to estimate probable rainfall (Singh et al. 2018; Sarkar & Maity 2020). The physical methods, such as the moisture maximization and the storm transposition technique, require meteorological data other than rainfall (Sarkar & Maity 2020) and are constrained by the record length. Thus, it could be useful and accurate when enough meteorological data are available. The Hershfield method is one of the more popular statistical methods and it is similar to the HFA used to estimate PMP when long precipitation records are available (Singh et al. 2018; Sarkar & Maity 2020). The enveloping technique of the parameters used in the Hershfield method varies depending on the region of interest (Singh et al. 2018). With the consideration of data availability and the different length of the data records at each rain gauge station, this research employed the HFA fitting the existing probability density function (PDF).

The HFA was conducted using the historical annual maximum daily rainfall data at each rain gauge station in the basin to estimate the probable rainfall amounts corresponding to the representative return periods. For the traditional parametric method applying an existing PDF to HFA, the uncertainty of the results would become more extensive for the extended return period when a smaller sample size was used (Mishra et al. 2007; Hu et al. 2020). In addition, if the target return period exceeds 100 years, using a sample size of less than 30 is not recommended by Takara & Kobayashi (2009). In most of the rain gauge stations handled in this research, the sample size of the His 1 and His 2 scenarios was less than 30. Therefore, 50- and 100-year return periods were selected as the representative return periods of this research considering the average number of samples.

The utilization of probability plotting paper with a representative plotting formula, such as the Weibull (Makkonen & Pajari 2014) formula, is the traditional method used to estimate the non-exceedance probability. It has been reported that selecting the formula of the probability plotting paper might cause underestimation (Makkonen 2006) or a biased estimation of the risks (Makkonen & Pajari 2014) when linear interpolation and extrapolation are attempted. On the other hand, CDF mapping is a flexible method which can directly plot the rainfall values based on the corresponding non-exceedance probability (Pierce et al. 2015; Maraun 2016). The CDF mapping method has been thus widely used for assessing hydrological data such as the central atmospheric pressure of typhoons (Ide et al. 2017) and extreme rainfall (Shibuo & Kanae 2010; Yazawa 2017) when there are multiple scenarios. In this research, CDFs were made for both the His 1 and His 2 scenarios after the parameters were estimated at each rain gauge station. The probable rainfall value in the His 1 scenario (⁠⁠) with the specific non-exceedance probability (⁠⁠) was then assigned to the CDF of the His 2 scenario. After that, the non-exceedance probability of in the CDF of the His 2 scenario (⁠⁠) was estimated. Finally, the corresponding return period with the non-exceedance probability (⁠⁠) was calculated using Equation (9). The advantage of the CDF mapping method is its flexibility (Yazawa 2017). Although only the GEV distribution was used in this research, the method can be applied even when the optimum PDFs for two scenarios are different.

In Figure 2, for example, a return period of a 100-year probable rainfall event (⁠ = 0.99) in one scenario becomes 80 years (⁠ = 0.9875) in the other scenario. This means that the possibility of the occurrence of a rainfall event becomes higher, implying that the risk of extreme rainfall relatively increases. If the return period of a probable rainfall value in one scenario becomes longer in the other scenario, it implies that the risk of extreme rainfall relatively decreases. In this research, the rainfall events with 50- and 100-year return periods at all rain gauge stations were selected in the His 1 scenario. The changes in the return periods of those rainfall events were then investigated in the His 2 scenario using the CDF mapping method.

The spatial distributions of the nine rainfall indices, the estimated probable rainfall, and the return period changes were analyzed using the Kriging method and the ArcGIS 10.8 software since the rain gauge stations in the Lake Biwa and Yodo River Basin were unevenly and sparsely distributed. Regionalization via spatial interpolation has been conducted to support the basin flood risk assessment, as in some of the previous studies (Takara & Oka 1992; Mano & Nakayama 2008). The Kriging method is one of the geostatistical methods used for spatial interpolation. Its performance and applicability to the regional analysis of extreme rainfall have been already confirmed (Hashimoto et al. 1994; Das 2019; Catalini et al. 2021). In comparison with other deterministic spatial interpolation methods, such as the inverse distance weighing (IDW), natural neighbor, and spline, the Kriging method is more applicable to unevenly and sparsely distributed sampling points (Mantzafleri et al. 2009).

The ordinary Kriging method with a spherical semivariogram was used for the spatial interpolation of the rainfall characteristics in this research. The relative change (%) of all rainfall characteristics between the His 1 and His 2 scenarios was then analyzed using the raster calculator tool in ArcGIS 10.8. As for the historical changes in the return periods from the His 1 to the His 2 scenarios, there were some results (outliers) that did not fit the GEV distribution obtained in the His 2 scenario applied with the CDF mapping process. Since too high/low results affect the spatial interpolation process, the Grubbs test (Grubbs 1969) with a confidence level of 95% was applied to the log-transformed return period values to remove any outliers before conducting the spatial interpolation. Other methods have also been suggested for outlier detection. For example, the Grubbs–Beck and multiple Grubbs–Beck tests could be practical to support low outlier detection when applied to the Weibull distribution (Cohn et al. 2013; Tiwari et al. 2019; Tiwari & Tripathi 2022). The Dixon test was compared with the Grubb's test by Mohammed et al. (2021) and showed a similar performance in terms of detection results. This research employed the Grubbs test because of its broad applicability for outlier detection in the hydrological analysis (e.g., Vivekanandan 2015; Batalini de Macedo et al. 2019). The Grubbs test was iterated until no outliers were detected.

In Figure 3, the higher AAR in the northern area and lower AAR in the southwestern area (i.e., downstream of the Yodo River) of the basin have been confirmed in both the His 1 and His 2 scenarios. The relative change in AAR shows that the annual rainfall amount has clearly increased from the His 1 to His 2 scenarios in the basin. Specifically, 98.2% of the cells show a positive relative change (Table 4). In particular, the southern part of the basin, i.e., near the Kizu River Basin, showed a higher increment than the other areas. AAR in this area during the His 1 scenario was relatively low; however, it was in the His 2 scenario that it became high. Since the Kizu River Basin is one of the historical typhoon-prone and flood-affected areas in the Lake Biwa and Yodo River Basin (Chakraborty 2013), this large increment even in AAR could not be disregarded as a regional rainfall characteristic for use in flood management.

The spatial distribution of ANORD (Figure 3(d) and 3(e)) showed a similar distribution to the AAR results; that is, ANORD is higher in the northern area and lower in the southwestern area of the basin in both the His 1 and His 2 scenarios. As for the relative change of ANORD, 53.4% of the cells show an increment from the His 1 to His 2 scenario (Table 4). In Figure 3(f), the area with the increment in ANORD goes from the southern area of Lake Biwa to the eastern area of the Kizu River Basin. In these areas, there have been more rainy days recently compared to the period of the His 1 scenario. On the other hand, the northern and southern areas of the basin have had fewer rainy days than in the His 2 scenario.

The relative change in SDII, which is the total rainfall divided by the number of rainy days, was derived from the results of Figure 3(g) and 3(h) and showed an increment of 96.8% in the cells between the His 1 and His 2 scenarios as shown in Figure 3(i) and Table 4. A high increment of SDII is observed particularly in the southern area of the Kizu River Basin. This means that the area demonstrates a relatively high increase in rainfall amount per rainy day between the His 1 and His 2 scenarios. This result was expected because a relatively large increment of AAR (Figure 3(c)) compared to a small change in ANORD (Figure 3(f)) was observed. Since most of the basin areas (cells) showed a positive change, it was confirmed that the rainfall intensity per rainy day in the Lake Biwa and Yodo River Basin tends to increase from the His 1 to His 2 scenario.

In Figure 4, majority increases in I95 (84.2% of the cell ratio as shown in Table 4) and I99 (73.8% of the cell ratio) are observed. In Figures 5 and 6, on the other hand, R95, R99, and N95 show a majority decrease (i.e., 99.9% of the cell ratio for R95, 86.9% for R99, and 99.3% for N95 as shown in Table 4). The spatial distribution of N99 in Figure 6(d) and 6(e) shows higher values for the northern part of the basin, i.e., in Shiga prefecture. N99 exhibits a negative change in the northeastern part. However, the southern part of the Lake Biwa and Yodo River Basin, i.e., the Katsura, Kizu, and Yodo River Basins, shows there to be a slight increase in N99 from the His 1 to His 2 scenario.

The areas with an increase in SDII, wet day intensities (I95 and/or I99), and the number of wet days (N95 and/or N99) have higher combined risks due to heavy rainfall events and pluvial floods. The spatial analysis of the nice rainfall indices revealed potentially high-risk areas like the northern, western, and southern areas of the Lake Biwa and Yodo River Basin. This is because some of the rainfall indices showed relatively large changes between the His 1 and His 2 scenarios.

The spatial distribution of the 50-year probable rainfall changes from the His 1 to the His 2 scenario in Figure 7. The 50-year probable rainfall increases are particularly found in the north of Lake Biwa and around the midstream of the Yodo River Basin. The changes in the spatial distribution of the 100-year probable rainfall display a similar tendency. The relative changes to the probable rainfall amount within both the 50- and 100-year return periods [Figure 7(c) and 7(f), respectively] describe both the increment and decrement of the recorded annual maximum rainfall values from the period of the His 1 scenario through to the His 2 scenario. The increment of the probable rainfall values means that the recorded annual maximum rainfall values are increasing, i.e., the pluvial flood risk has relatively increased as time has passed. Similarly, the decrement of the probable rainfall values means that the recorded annual maximum rainfall values have not changed so much and even fallen slightly, i.e., the pluvial flood risk has not changed/relatively decreased as time has passed.

Table 5 shows the ratio of the number of cells for the relative change in probable rainfall for the 50- and 100-year return periods based on the results from Figure 7(c) and 7(f). In Table 5, 15.8% of the cells show the decrement of the 50-year probable rainfall. This area is mainly located in the south of Lake Biwa based on Figure 7(c). The relative change to the 50-year probable rainfall is high in the north, towards the western area of the basin. Although the probable rainfall values are not high in this area, as shown in Figure 7(a) and 7(b), the pluvial flood risk in the His 2 scenario has increased compared to the His 1 scenario which can be observed by the relative change in probable rainfall. As for the relative change in the 100-year probable rainfall in Figure 7(f), the spatial distribution shows a similar tendency to the results of the relative change of the 50-year probable rainfall; that is, a large increment was observed in the north-western area while the south of Lake Biwa exhibited a decrement.

The results of the spatial distribution of the 50- and 100-year probable rainfall indicate the relatively high pluvial flood risk in the southeast area of the basin. The areas that have historically shown there to be a high flood risk, such as the Kizu River Basin, already have flood control measures in place. However, the potentially high flood risk in the northern areas of the basin was also confirmed based on the large increment of probable rainfall from the His 1 to His 2 scenarios. It has been also reported that, in particular, the northern area of Lake Biwa has a high possibility of frequent flooding (Taki 2022). The results of this subsection suggest the need for further updates to the regional flood management plans for the areas with a potential high flood risk.

The areas with less than a 50-year return period in Figure 8(a) and 100-year return period in Figure 8(b) mean that the occurrence of rainfall events in the His 2 scenario becomes more frequent with the incremental increase of probable rainfall amounts compared to the His 1 scenario. As demonstrated in Figure 8 and Table 6, 56.1 and 40.4% of the cells show less than a 50- and 100-return period. In these areas, rainfall events have become more frequent; that is, these areas have a relatively high risk of pluvial flooding in the His 2 scenario compared to the His 1 scenario. There is a difference in the spatial distribution of the high-risk areas that can be observed between the rainfall events of the two representative return periods. In the His 2 scenario, the return period of the 50-year rainfall events for the His 1 scenario dropped to less than 10 years in the northern area of Lake Biwa. A decremental drop is also observed in the western area of the basin as shown in Figure 8(a). This tendency is similar to the distribution of changes in the return period for the 100-year rainfall events (Figure 8(b)). However, the decrement is large, particularly in the western area of the basin. The return period of the 100-year rainfall event for the His 1 scenario becomes less than 70 years in the His 2 scenario. This means that the safety level of the flood control in this area has decreased compared to the period in the His 1 scenario. In some areas of Lake Biwa and Yodo River Basin, the current flood planning scale is still based on a 100-year return period using historical rainfall data and has not yet been updated. There are still concerns and discussions about the flood control such as an increase in flood damage due to urbanization and the decrease in safety levels because of changing climate for Lake Biwa and the Yodo River Basin in the context of integrated basin management (Bamba 2011; Nakatsuka et al. 2020). More frequent updates of the regional flood management plan are needed, particularly for the areas where the probable rainfall amounts increase and/or the return periods decrease as part of the integrated management of the Lake Biwa and Yodo River Basin.

This research conducted a spatial analysis of the basin-scale extreme rainfall characteristics and investigated the historical changes by setting up two historical scenarios, His 1 (the first half of the observation period at each rain gauge station) and His 2 (the second half of the observation period at each rain gauge station) for the Lake Biwa and Yodo River Basin. The results of the changes in the rainfall indices show that there is a potentially high risk of extreme rainfall events and pluvial floods, particularly in the northern, western, and southern areas of the Lake Biwa and Yodo River Basin. It was also confirmed that the 50- and 100-year probable rainfall has shown an increase in the north through to the western area of the basin according to the His 1 and His 2 scenarios. In the His 2 scenario, the return periods of the 50- and 100-year rainfall events in the His 1 scenario decreased in the northern area of Lake Biwa and the western area of the basin.

There are at least three technical limitations that should be handled in future work. First, rainfall data from different sources, such as satellite observed and/or modeled data, could be compared with the regionalization results of this research, which used data from the rain gauge stations. The verification of the results by other rainfall data is also a necessary step in providing accurate hydrological information for use in flood management. In addition, this research applied only one PDF, the GEV distribution, for HFA to simplify the comparison of results between two scenarios. However, other PDFs also need to be applied as part of the optimum PDF selection so then the methodology follows the flood design guidelines in Japan. In this research, the Kriging method was the only method used for regionalization because of its ease of application. However, there are more variables (such as elevation) that should be considered when spatial interpolation is conducted for rainfall. There are already some studies that have used a multivariate technique. Therefore, the application of multivariate techniques, such as co-Kriging (Rau et al. 2017) and artificial neural networks (El Alaoui El Fels et al. 2021), has to be considered in future work.

The regionalization of the basin-scale extreme rainfall characteristics can provide useful information for use in flood management planning in the target basin, particularly in the areas with sparsely distributed rain gauge stations. Through regionalization, it was revealed that there are regional differences in the historical changes in rainfall characteristics, even within the same basin. Although it is important to manage the whole basin in the context of integrated basin management, the local/regional flood management planning needs to be updated more frequently with the latest understanding of the regional weather characteristics. Since a river flow simulation using a hydrological model was also outside the scope of this research, a hydrological simulation needs to be conducted to further update the flood management planning in the Lake Biwa and Yodo River Basin.

All authors contributed to the conception and design of this research. Taishi Yazawa was in charge of conceptualization, methodology, formal analysis, writing – review & editing, project administration, and funding acquisition. Ayane Shoji contributed to conceptualization, data collection, methodology, investigation, visualization, and writing the original draft.

This research was supported by the AY2020 research promotion program of Ritsumeikan University.

All relevant data are available from https://www.data.jma.go.jp/obd/stats/etrn/index.php

The authors declare there is no conflict.