At-Site Versus Regional Frequency Analysis of Sub-Hourly Rainfall for Urban Hydrology Applications During Recent Extreme Events

Abstract

Accurate rainfall quantile estimation is critical for urban flood management, particularly given the escalating climate change impacts. This study comprehensively compared at-site frequency analysis and regional frequency analysis for sub-hourly rainfall quantile estimation, using data from 27 sites across Seoul. The analysis focused on Seoul’s disaster prevention framework (30-year and 100-year return periods). Employing L-moment statistics and Monte Carlo simulations, the rainfall quantiles were estimated, the methodological performance was evaluated, and Seoul’s current disaster prevention standards were assessed. The analysis revealed significant spatio-temporal variability in Seoul’s precipitation, causing considerable uncertainty in individual site estimates. A performance evaluation, including the relative root mean square error and confidence interval, consistently showed regional frequency analysis superiority over at-site frequency analysis. While at-site frequency analysis demonstrated better performance only for short return periods (e.g., 2 years), regional frequency analysis exhibited a substantially lower relative root mean square error and significantly narrower confidence intervals for larger return periods (e.g., 10, 30, 100 years). This methodology reduced the average 95% confidence interval width by a factor of approximately 2.7 (26.98 mm versus 73.99 mm). This enhanced reliability stems from the information-pooling capabilities of regional frequency analysis, mitigating uncertainties due to limited record lengths and localized variabilities. Critically, regionally derived 100-year rainfall estimates consistently exceeded Seoul’s 100 mm disaster prevention threshold across most areas, suggesting that the current infrastructure may be substantially under-designed. The use of minute-scale data underscored its necessity for urban hydrological modeling, highlighting the inadequacy of conventional daily rainfall analyses.

1. Introduction

Urban areas worldwide are facing increased flood risks due to rapid urbanization and changing patterns of extreme precipitation, due to climate change [1]. The expansion of impervious surfaces leads to an increase in the runoff volume and reduces infiltration, making cities vulnerable to flash floods, even from short durations of intense rainfall [2]. This issue of urban imperviousness is especially pronounced in densely populated metropolitan areas, where concrete and asphalt may cover 70–90% of the land surface [3]. Such extensive impermeable coverage significantly alters the rainfall–runoff dynamics by decreasing lag times and increasing peak flows, often by a factor of 2–5 compared to conditions prior to development. As a result, even moderate rainfall events of 15–30 mm/h can induce localized flash flooding in these urban environments, as the natural capacity of soil to absorb water is effectively circumvented [4,5,6].

Moreover, traditional urban drainage systems frequently become overwhelmed when rainfall intensities surpass their design capacities, further intensifying flood risks in particularly susceptible low-lying or topographically constrained areas [7]. These hydrological changes exemplify the direct structural impacts of urbanization, leading to a fundamental disparity between the water conveyance capabilities of the built environment and the rapid accumulation of surface water during precipitation events [8,9].

To effectively address the challenges associated with urban flooding, the precise estimation of rainfall quantiles is crucial for hydraulic engineering and water resource management [10,11]. These quantiles are typically represented through the use of intensity–duration–frequency curves, which form the basis for the design of hydraulic structures and the planning of flood defense systems [12,13]. The effective design of stormwater infrastructure, encompassing elements such as pipes, retention basins, culverts, and flood control systems, is significantly dependent on reliable rainfall quantile data [10,11]. Furthermore, these estimations are critical for flood hazard mapping, which serves as a vital resource for urban planning, emergency response strategies, and flood insurance evaluations [14]. Consequently, the efficacy of urban infrastructure and planning tools is directly linked to the precision of rainfall frequency analysis [13,15].

Two principal methodologies are employed for the estimation of rainfall quantiles: at-site frequency analysis (AFA) and regional frequency analysis (RFA). AFA utilizes probability distributions applied to long-term rainfall records from individual monitoring stations, typically focusing on annual maxima or peaks-over-threshold data [15,16,17]. While this approach yields site-specific estimates, its reliability is frequently constrained by the availability and granularity of historical data, particularly in urban areas where continuous sub-hourly rainfall records are often limited. This constraint introduces significant uncertainty into the estimation of short-duration extremes, which are crucial for effective urban flood management [18,19].

In contrast, RFA seeks to mitigate some of these limitations by aggregating data from multiple stations located in hydrologically similar regions, thereby enhancing the sample size and potentially increasing the robustness of quantile estimates. However, the efficacy of RFA is contingent upon the accurate delineation of homogeneous regions, a particularly challenging endeavor for sub-hourly rainfall, due to its pronounced spatial and temporal variability [20,21]. This study aims to evaluate the performance of AFA and RFA in the context of sub-hourly rainfall within urban environments, highlighting the necessity for improved methodologies in light of evolving climate conditions. RFA is posited to provide more accurate and stable rainfall quantile estimates in urban settings characterized by data scarcity, especially when considering recent extreme events and the nonstationarity of precipitation patterns.

This study investigates three key questions: (1) Which method provides more accurate rainfall quantile estimates for sub-hourly urban precipitation? (2) Which approach yields more reliable and trustworthy results, particularly when analyzing the characteristics of recent extreme rainfall events? (3) How do the differences between these methods impact urban hydrological infrastructure design? The research evaluates both methodologies using 28 years of sub-hourly rainfall data from 27 automatic weather stations within a metropolitan area, employing the index-flood method procedure for RFA and implementing bootstrap resampling techniques for uncertainty quantification. The performance assessment is based on multiple metrics, including the root mean square error, bias, confidence interval width, and cross-validation, supplemented by case studies of recent extreme weather events and trend analyses to address nonstationarity concerns. Additionally, the spatial distribution of rainfall is analyzed using the Thiessen polygon method to calculate areal rainfall estimates, providing insights into how methodological differences translate into practical design considerations for urban drainage systems, detention basins, and other hydrological infrastructure. The findings of this research are anticipated to yield significant insights for urban planners, engineers, and policymakers by providing an evidence-based comparison of rainfall frequency analysis methodologies. Ultimately, the results aim to enhance the reliability of rainfall quantile estimates, thereby fortifying the design of urban drainage systems and facilitating more informed and cost-effective decision making in light of increasing climate variability.

2. Study Area and Data

2.1. Study Area

This research examines Seoul in South Korea, recognized as one of the most densely populated and urbanized areas in East Asia. Seoul has an exceptionally high population density, estimated at approximately 16,000 individuals per square kilometer, alongside significant impermeability, with an average exceeding 70%. The city has been the focus of numerous studies aimed at assessing urban vulnerability [22,23,24,25,26]. The topography of Seoul is notably intricate, ranging from elevations of 5 m along the Han River to approximately 300 m in the adjacent mountainous regions, which creates varied drainage patterns that influence the region’s rainfall–runoff dynamics considerably. The climate of Seoul is characterized by a distinct East Asian monsoon system, with around 60–70% of the annual precipitation occurring during the summer monsoon months (June to September). The area is particularly susceptible to intense, short-duration convective storms during the rainy season, which can lead to extreme rainfall rates exceeding 100 mm per hour. The interplay of high levels of urbanization, complex topography, and a monsoonal climate renders Seoul particularly vulnerable to urban flooding events. This vulnerability has been highlighted by several significant flooding incidents in recent decades, notably in 2010, 2011, and 2022 [27].

2.2. Rainfall Monitoring Network and Datasets

The rainfall observation network employed in this research consists of 28 Automated Weather Stations (AWSs) operated by the Korea Meteorological Administration, which are strategically located across the 25 administrative districts of Seoul. Figure 1 depicts the spatial distribution of these AWSs, underscoring the extensive coverage afforded by the observation network throughout the study area. This network provides significant spatial representation, with an average density of approximately one observation station per 23 km², thereby ensuring adequate representation of rainfall variability across various urban settings.

The AWSs have been in operation from 1997 to 2024. This provides a continuous dataset that is suitable for rigorous frequency analysis, with individual station records ranging from 12 to 28 years and an average data length of 25.7 years. Rainfall data are recorded at a sub-hourly resolution, specifically as 15 min accumulated rainfall depths, which effectively capture the rapid variability associated with convective storms, while providing sufficient data density for statistical evaluation. For a comprehensive analysis of rainfall quantiles, the 15 min base data were systematically aggregated into multiple durations relevant for urban drainage design: 15 min, 30 min, 60 min, and 120 min accumulations. The selection of these specific durations is crucial because Seoul’s disaster prevention performance target is defined by probable rainfall for return periods of 30 to 100 years at hourly durations. Furthermore, urban watersheds are characterized by a short time of concentration, making them highly susceptible to flash floods triggered by short-duration, high-intensity rainfall events. Therefore, analyzing data for these precise, short durations is paramount for accurate hydrological modeling and effective urban flood risk management.

The aggregation methodology employed a moving-window technique to determine the maximum rainfall depths for various durations on a 15 min basis, thereby ensuring the capture of peak intensities regardless of their occurrence throughout the day. The annual maximum series were derived for each duration and station by identifying the highest aggregated value for each calendar year. Years exhibiting more than 10% missing data during the monsoon season (June–September) were excluded from the analysis to ensure data quality, while maintaining adequate sample sizes. Consequently, complete annual maximum datasets were obtained for all the stations, yielding an average of 24.2 years of usable data per station.

Comprehensive information regarding the stations utilized in this study is presented in

Table 1. Detailed information on the rainfall observation stations used in this study.

Code	Name	Latitude	Longitude	Altitude	Type
108	Seoul	37.57	126.97	85.67	ASOS/AWS
400	Gangnam	37.50	127.08	12.66	AWS
401	Seocho	37.48	127.03	33.05	AWS
402	Gangdong	37.56	127.15	55.29	AWS
403	Songpa (excluded)	37.47	127.13	54.05	AWS
404	Gangseo	37.57	126.83	9.30	AWS
405	Yangcheon	37.53	126.88	22.75	AWS
406	Dobong	37.67	127.03	56.65	AWS
407	Nowon	37.62	127.09	25.30	AWS
408	Dongdaemun	37.58	127.06	53.96	AWS
409	Jungnang	37.59	127.09	39.09	AWS
410	KMA (Korea Meteorological Administration)	37.49	126.92	41.76	AWS
411	Mapo	37.55	126.93	100.67	AWS
412	Seodaemun	37.57	126.94	103.08	AWS
413	Gwangjin	37.53	127.09	29.91	AWS
414	Seongbuk	37.61	127.00	128.62	AWS
415	Yongsan	37.52	126.98	31.73	AWS
416	Eunpyeong	37.65	126.94	55.00	AWS
417	Geumcheon	37.47	126.90	45.00	AWS
418	Hangang	37.52	126.94	10.66	AWS
419	Jung	37.55	126.99	267.05	AWS
421	Seongdong	37.55	127.04	34.73	AWS
423	Guro	37.49	126.83	56.08	AWS
424	Gangbuk	37.64	127.01	69.80	AWS
425	Namhyeon	37.46	126.98	113.00	AWS
509	Gwanak	37.45	126.95	141.64	AWS
510	Yeongdeungpo	37.53	126.91	25.38	AWS
889	Seoul National Cemetery	37.50	126.98	16.23	AWS

, while Figure 2 depicts the spatial distribution of the 28 stations in conjunction with the Thiessen polygons. In the context of South Korea, the relationship between rainfall and runoff is employed to estimate the design flood discharge. The areal rainfall derived from the Thiessen polygon method is a critical component in the estimation of design flood discharge, as it utilizes rainfall quantiles obtained from observation stations as input data for hydrological models. As illustrated in Figure 2, this methodology is frequently applied to estimate design flood discharge within the study area.

2.3. Comparison of Extreme Events Using Daily Precipitation Data

Figure 3 presents comparative analysis of the rankings of the daily maximum values for site 108, as officially documented by the Seoul Metropolitan Government, alongside the rankings of the daily maximum values from an additional 27 AWSs. This comparison is grounded in daily time-series data spanning from 1997 to 2024.

Figure 3 illustrates distinct variations between the daily maximum precipitation recorded at site 108, which represents the official Seoul record (indicated by blue stars), and the maximum values documented at the AWSs, denoted by red circles. The precipitation amounts, ranked from first to tenth, along with their respective dates, are presented in the figure and are further detailed in

Table 2. Comparison of daily maximum precipitation records between the official Seoul observation station (site 108) and AWSs from 1997 to 2024.

	Seoul (108)		AWS
Rank	Prcp. (mm)	Date	Prcp. (mm)	Date	Station
1	332.8	08/08/98	403.5	08/08/98	402
2	301.5	07/27/11	381.5	08/08/22	410
3	273.4	07/15/01	375.5	08/03/97	420
4	261.6	08/02/99	359.0	07/27/11	425
5	259.5	09/21/10	310.0	07/15/01	409
6	241.0	07/16/06	302.0	07/12/06	406
7	190.0	07/09/09	293.0	09/21/10	400
8	178.0	08/07/02	286.5	08/02/99	406
9	177.0	08/24/03	249.5	08/29/18	406
10	176.2	06/30/22	226.0	08/07/02	423

, which summarizes the discrepancies observed.

The comparative examination of the daily maximum precipitation data from the official Seoul observation station (site 108) and the 27 AWSs spanning the years 1997 to 2024 reveals significant discrepancies between these two observational frameworks. A significant event transpired on 8 August 1998, when the official station recorded a precipitation level of 332.8 mm, in contrast to the AWS site 402, which reported a measurement of 403.5 mm. This discrepancy of approximately 70.7 mm is noteworthy, particularly given that both measurements were obtained on the same day.

Throughout the ten significant precipitation events detailed in Table 2, the maximum precipitation values recorded at the AWSs consistently surpassed those documented at the official station. For example, during the second-ranked event, the official station reported 301.5 mm of precipitation on 27 July 2011, whereas AWS site 410 recorded a markedly higher value of 381.5 mm on 8 August 2022, leading to a difference of nearly 80 mm. These findings underscore the pronounced spatial variability of extreme rainfall events within the region.

It is noteworthy that, although the dates of the daily precipitation maxima frequently align between the official and AWSs, certain rankings, specifically the 2nd, 3rd, and 9th events, indicate that the peak precipitation occurred on different days across the network. This observation suggests that the locus of extreme precipitation events can vary locally, resulting in a highly heterogeneous distribution of rainfall, even within a single urban area.

Additionally, AWS site 406 appears multiple times in the rankings (5th, 8th, and 9th place), indicating that specific locations may experience recurrent episodes of particularly intense rainfall, thereby rendering them more susceptible to hydrometeorological hazards. In the third-ranked event, the discrepancy is particularly pronounced: the official station recorded 273.4 mm of precipitation on 15 July 2001, while AWS site 420 documented 375.5 mm on 3 August 1997, reflecting a substantial difference of 102 mm.

In summary, these findings illustrate that reliance solely on the official observational record is inadequate for capturing the full spatial extent and intensity of extreme rainfall events in Seoul. The integration of data from spatially distributed AWS networks is, therefore, crucial for more accurate estimations of areal rainfall, hydrological modeling, and disaster risk management. This study emphasizes the critical importance of a dense and distributed observational network in regard to urban hydrometeorology, both for advancing scientific understanding and for the effective formulation of mitigation strategies against extreme weather events.

2.4. Disaster Prevention Performance Targets in Seoul

The Ministry of the Interior and Safety of the Republic of Korea has established disaster prevention performance targets every five years since 2012 under the relevant national disaster management legislation, with local governments implementing region-specific objectives. Following the August 2022 extreme precipitation event (141.5 mm/h) that caused significant flooding and casualties in southern Seoul districts, particularly affecting underground spaces, the Seoul Metropolitan Government implemented its first target revision in a decade. This policy revision elevated the design rainfall intensity from a 30-year return period (95 mm/h capacity) to a 50–100-year return period (100 mm/h capacity), with enhanced targets of 110 mm/h for highly vulnerable areas, such as Gangnam Station (site 400). Consequently, this study prioritized the analysis of probable rainfall intensities associated with 30-year and 100-year return periods for a duration of 60 min, corresponding to the critical design parameters established within Seoul Metropolitan Government’s disaster prevention framework.

3. Methodology

Spatial and temporal inconsistencies in Seoul’s rainfall data significantly compromise the accuracy of rainfall quantile estimations. Discrepancies between single-point measurements and distributed AWS network data create divergent interpretations of rainfall events, reducing the reliability of design flood estimations in hydrological analyses. These inconsistencies, most pronounced during extreme rainfall events, increase uncertainties in rainfall quantile estimations, critical inputs for hydraulic structure design and urban disaster prevention planning, thereby affecting water resource management and disaster response effectiveness.

This study addresses these challenges by investigating frequency analysis methodologies to enhance rainfall quantile estimation reliability, using data from Seoul’s comprehensive observation network. To overcome spatial heterogeneity and limited observation periods, this study conducted an L-moment-based RFA approach and quantitatively compares it with traditional AFA.

3.1. Analytical Framework for Frequency Analysis

Frequency analysis is a fundamental statistical method employed to estimate the probability of the occurrence of extreme hydrological events and to determine their corresponding magnitudes for specific return periods [10,28,29]. In this study, it specifically pertains to the analysis of sub-hourly rainfall data. The core principle of this methodology involves fitting a theoretical probability distribution to observed extreme event data, thereby enabling extrapolation beyond historical records. This process can be undertaken through two primary approaches: AFA, which focuses on individual gauging stations with sufficient historical records, and RFA, which pools data from hydrologically homogeneous regions to enhance the reliability of estimates, particularly in data-sparse areas or for ungauged sites. This comprehensive approach is paramount in urban hydrology applications, serving as a foundational framework for the design and evaluation of hydraulic structures and flood protection measures in numerous countries worldwide.

While traditional frequency analysis typically assumes stationarity, meaning the statistical properties of the data remain constant over time, recent extreme weather events and the discernible impacts of climate change have increasingly challenged this assumption. Numerous studies are actively exploring nonstationary frequency analysis methods; however, integrating these approaches into established design criteria still presents considerable practical difficulties. Consequently, this study deliberately focuses on stationarity-based at-site and regional frequency analysis, consistent with the prevailing methodologies utilized for establishing design criteria in South Korea.

The methodological procedure for frequency analysis involves extracting annual maximum rainfall data from the observed records and subsequently fitting this data to a selected theoretical probability distribution. Rainfall quantiles for various return periods are then estimated by utilizing the inverse cumulative distribution function of the fitted distribution. A multitude of probability distributions and parameter estimation methods exist for this process, and research continues to identify the most appropriate and robust techniques. As highlighted in the introduction, this study specifically adopts the Generalized Extreme Value (GEV) distribution, recognized for its effectiveness in characterizing South Korean rainfall patterns and explicitly referenced in the national flood discharge estimation guidelines [30,31]. For the crucial step of parameter estimation, the L-moment method was applied, primarily due to its demonstrated robustness and efficiency in regard to extreme value analysis.

3.2. At-Site and Regional Frequency Analysis Based on the L-Moment Approach

3.2.1. L-Moments: Mathematical Framework and Principles

The L-moments methodology, originally developed by Hosking [32], represents a sophisticated statistical framework that enhances probability distribution estimation and frequency analysis procedures. Contemporary research on extreme rainfall statistics increasingly employs the L-moments estimator approach for analyzing annual maximum series (at hourly, daily, or monthly intervals), particularly in regional analytical contexts. This methodology delivers statistically efficient parameter estimation for hydrological datasets and distributional characterization.

The L-moments approach offers several substantive advantages in regard to statistical hydrology: enhanced performance with limited sample sizes, superior dispersion representation, and reduced bias in higher-order statistics (specifically skewness and kurtosis) compared to conventional moment-based techniques. L-moments are mathematically derived from probability weighted moments (PWMs), which can be formally expressed as [32]:

$${\lambda }_r={\displaystyle \frac{1}{r}}{\sum }_{k=0}^{r-1}{\left(-1\right)}^k\left(\begin{array}{c}r-1\\ k\end{array}\right){\rm E}\left\{X_{r-k:r}\right\}$$

(1)

where ${\lambda }_r$ represents a linear function of the $r$-th L-moment for the distribution of random variable $X$, where $r$ = 1, 2, 3, … denotes a non-negative integer. Derived from Equation (1), the first four L-moments can be explicitly formulated as:

$${\lambda }_1=EX$$

(2)

$${\lambda }_2=\left({\displaystyle \frac{1}{2}}\right)E\left(X_{2:2}-X_{1:2}\right)$$

(3)

$${\lambda }_r=\left({\displaystyle \frac{1}{3}}\right)E\left(X_{3:3}-{2X}_{2:3}+X_{1:3}\right)$$

(4)

$${\lambda }_r=\left({\displaystyle \frac{1}{4}}\right)E\left(X_{4:4}-{3X}_{3:4}+3X_{2:4}-X_{1:4}\right)$$

(5)

Hosking [32] demonstrates the analytical effectiveness of employing L-moment ratio estimators for extreme hydrological event characterization, which can be expressed as:

$${\tau }_2={\displaystyle \raisebox{1ex}{${\lambda }_2$}\!\left/ \!\raisebox{-1ex}{${\lambda }_1$}\right.} \left(L\text{-}{C}_v\right)$$

(6)

$${\tau }_3={\displaystyle \raisebox{1ex}{${\lambda }_3$}\!\left/ \!\raisebox{-1ex}{${\lambda }_2$}\right.} \left(L\text{-}skewness\right)$$

(7)

$${\tau }_4={\displaystyle \raisebox{1ex}{${\lambda }_4$}\!\left/ \!\raisebox{-1ex}{${\lambda }_2$}\right.} \left(L\text{-}kurtosis\right)$$

(8)

where ${\tau }_2$ quantifies the relative dispersion (scale parameter), ${\tau }_3$ represents the distribution’s asymmetry (shape parameter) with values constrained between 0 and 1, and ${\tau }_4$ characterizes the probability concentration in the tails (peakedness). Significantly, these ratio estimator formulations and their corresponding graphical representations prove to be exceptionally valuable for characterizing distributional properties of asymmetrically distributed datasets. Accordingly, by implementing these equations, this research presents the L-moment ratio analysis for 15 min, 30 min, 45 min, 60 min, and 120 min rainfall patterns across the studied area.

A discordancy measure ($D_i$) is employed to identify and filter out anomalous site data, ensuring appropriate dataset selection for regional analysis. When vector $u_i={\left[t^{\left(i\right)},t_3^{\left(i\right)},t_4^{\left(i\right)}\right]}^T$ contains the L-moment ratios for a specific site i (as established by Hosking and Wallis [33]), the discordancy measure can be mathematically expressed as:

$$D_i={\displaystyle \frac{1}{3}}{\left(u_i-\overline{u}\right)}^T{S}^{-1}\left(u_i-\overline{u}\right)$$

(9)

where $u_i$ represents a vector comprising $L\text{-}{C}_v$, $L\text{-}skewness$, and $L\text{-}kurtosis$ values; $S$ denotes the covariance matrix of $u_i$; and $\overline{u}$ signifies the mean vector of ui across all the sites under consideration.

The identification of homogeneous regions constitutes a critical phase of regional frequency analysis. Statistical methods comparing L-moment sample distributions across multiple sites enable the delineation of homogeneous regions. Hosking and Wallis [33] formulated a statistical evaluation for regional homogeneity, defined as the heterogeneity measure ($H$). To establish expected heterogeneity parameters, Monte Carlo simulations of rainfall data with equivalent record lengths to observed datasets were executed, an approach widely recognized in hydrological assessments. $H$ is calculated as follows:

$$H={\displaystyle \frac{V_{obs}-{\mu }_v}{{\sigma }_v}}$$

(10)

where ${\mu }_v$ and ${\sigma }_v$ represent the mean and standard deviation of the simulated dataset, respectively. $V_{obs}$ is derived from the regional data and can be calculated using three distinct $V$-statistics ($V_1$, $V_2$, and $V_3$), which are formulated as follows:

$$V_1={\left[\sum _{i=1}^N{N}_i{\left\{t^{\left(i\right)}-t^R\right\}}^2/\sum _{i=1}^N{N}_i\right]}^{1/2}$$

(11)

$$V_2=\sum _{i=1}^N{N}_i{\left\{{\left(t^{\left(i\right)}-t^R\right)}^2+{\left(t_3^{\left(i\right)}-t_3^R\right)}^2\right\}}^{1/2}/\sum _{i=1}^N{N}_i$$

(12)

$$V_3=\sum _{i=1}^N{N}_i{\left\{{\left(t_3^{\left(i\right)}-t_3^R\right)}^2+{\left(t_4^{\left(i\right)}-t_4^R\right)}^2\right\}}^{1/2}/\sum _{i=1}^N{N}_i$$

(13)

According to the $H$-statistic criteria, regional classification follows these thresholds: a region is considered sufficiently homogeneous when $H<1$, potentially homogeneous when $1\le H<2$, and definitively heterogeneous when $H>2$ [20].

The discordancy measure ($D$) was computed for each of the initial 28 sites, following the methodology outlined previously. Hosking and Wallis [20] establish a critical $D$ threshold of 3 for identifying discordant sites within a homogeneous region comprising 15 or more stations. This analysis revealed that site 403 exhibited a $D$ value of 5.152 for the 15 min rainfall duration, significantly exceeding this threshold. As depicted in Figure 2, site 403’s location on the outskirts of Seoul, combined with its statistically inconsistent rainfall characteristics, led to its exclusion from the study. This decision aimed to preserve the homogeneity of the study area. After removing site 403, a re-evaluation was conducted on the remaining 27 sites. All $D$ values for these sites fell below the critical values across all the examined durations.

Figure 4 presents a comparative analysis of heterogeneity measures ($H$ values) derived from two distinct datasets, namely the complete set comprising all 28 observation stations and a reduced set of 27 stations with the exclusion of station 403, which exhibited an anomalously high discordancy value ($D$ value). The elimination of this statistical outlier resulted in only a marginal reduction in the $H$ value from 2.441 to 2.411 for the 15 min rainfall duration series. Notably, the $H$ statistic for the 15 min duration exceeds the threshold of 2, placing this duration interval in the “definitely heterogeneous” category, according to the classification criteria [20]. Conversely, the $H$ values calculated for all of the other duration intervals remained below 1, indicating that these temporal scales fall within the “acceptably homogeneous” classification. This pattern suggests that regional homogeneity is duration dependent, with short-duration extreme rainfall exhibiting greater spatial heterogeneity across the study region compared to longer duration events.

3.2.2. Probability Distribution

This research evaluates six probability distributions commonly employed in precipitation frequency analysis: Generalized Logistic (GLO), Generalized Normal (GNO), Pearson Type III (PE3), Generalized Pareto (GPA), Gumbel (GUM), and GEV distributions. These selections are based on their statistical characteristics, which are conducive to modeling extreme precipitation events, and their ability to accommodate a diverse range of skewness and kurtosis patterns observed in hydrological data.

Figure 5 illustrates the L-moment ratio diagrams ($L\text{-}skewness$ versus $L\text{-}kurtosis$), a widely accepted graphical tool for visually assessing distributional goodness of fit. These diagrams facilitate the comparison of sample L-moments with the theoretical lines of the aforementioned distributions for five rainfall durations: 15, 30, 45, 60, and 120 min. The adequacy of a particular distribution model is indicated by the proximity of the sample L-moment coordinates to its corresponding theoretical curve. Within these diagrams, the ${\mathrm{R}}_{\mathrm{A}\mathrm{v}\mathrm{g}.}$ value represents the average L-moment ratio calculated across all the study sites for each specified duration.

For the 15 min duration (Figure 5a), the ${\mathrm{R}}_{\mathrm{A}\mathrm{v}\mathrm{g}.}$ value closely aligns with both the GEV and PE3 distribution curves. As the duration increases (from Figure 5b–e), the ${\mathrm{R}}_{\mathrm{A}\mathrm{v}\mathrm{g}.}$ values progressively converge towards and follow the curve of the GEV distribution more prominently. This trend indicates a clear shift towards the improved suitability of the GEV distribution for characterizing longer-duration rainfall events. This pattern underscores the flexibility and statistical robustness of the three-parameter GEV distribution in effectively capturing a wide range of rainfall characteristics across varying durations.

These results, derived from the objective framework of L-moment-based analysis, strongly support the application of the GEV distribution in estimating rainfall quantile from a statistical goodness-of-fit perspective. The selection of an appropriate distribution is particularly critical in the context of RFA, wherein preserving spatial consistency while accurately reflecting the complexity of rainfall patterns is essential. The demonstrated utility of L-moment diagrams in guiding these distributional assumptions validates its application for both at-site and regional hydrological analyses.

The selection of regional frequency distributions employs L-moment ratio diagrams and goodness-of-fit measures, which are determined by comparing the sample regional average $L\text{-}kurtosis$ with theoretical $L\text{-}kurtosis$ values. For each candidate distribution, the goodness-of-fit measure is quantified using the following formula:

$$Z^{Dist}={\displaystyle \frac{t_4^R-{\tau }_4^{Dist}}{{\sigma }_4}}$$

(14)

where $t_4^R$ represents the regional mean $L\text{-}kurtosis$ value derived from datasets within a specific region, ${\tau }_4^{Dist}$ denotes the theoretical $L\text{-}kurtosis$ value corresponding to a fitted probability distribution, and ${\sigma }_4$ indicates the standard deviation calculated from simulated datasets. At approximately a 90% confidence level, the distribution demonstrates acceptable goodness of fit when the absolute value of $Z^{Dist}$ does not exceed 1.64 ($\left|Z^{Dist}\right|\le 1.64$).

Figure 6 presents the Z values for various probability distributions, derived from the RFA results for the 15 min duration. The Z-test is employed to assess the statistical significance of the model fit, where distributions with |Z| ≤ 1.64 are generally regarded as appropriate at the 10% significance level. According to the analysis, the GLO and GPA distributions exhibited Z values exceeding the threshold of 1.64, indicating statistical inadequacy at the specified significance level. In contrast, the GEV distribution yielded a Z-value very close to zero, suggesting the most stable and superior fit among the distributions considered. Additionally, most distributions, including GNO, PE3, and LN3, were found to have Z values within the acceptable range, demonstrating statistically meaningful goodness of fit to the data.

These results support the reliability of L-moment-based distribution selection and reaffirm that the GEV distribution is theoretically the most suitable for modeling extreme events, such as short-duration extreme rainfall. Therefore, the GEV distribution should be given priority in the frequency analysis of short-term extreme rainfall events in urban areas, including Seoul.

3.3. Performance Evaluation

As one of the main objectives of this study is to assess and compare the ability of the AFA and RFA methods to provide more reasonable rainfall quantile estimates using sub-hourly rainfall data in urban areas, the use of a robust performance evaluation framework is essential. This section details the methodologies employed to rigorously assess the reliability and accuracy of the rainfall quantiles derived from both AFA and RFA. Given the inherent uncertainties in extreme value estimations from limited data, this framework is crucial for carrying out a comprehensive comparison and for quantifying the confidence in the estimated rainfall quantiles.

3.3.1. Monte Carlo Simulation

Monte Carlo simulation is a widely used computational technique for evaluating the statistical properties of L-moment estimators, probability distribution parameters, and the bias and variability associated with rainfall quantiles. This method offers the advantage of assessing estimator performance in a controlled setting, where the underlying probability distribution is known.

In this study, the simulation procedure is structured as follows. First, multiple independent synthetic rainfall datasets are generated by sampling from a predefined theoretical distribution, specifically, the kappa distribution, which serves as a versatile generative tool capable of approximating a wide range of distributional forms across the L-moment ratio space. While the kappa distribution theoretically merits consideration as a candidate fitting distribution, due to its demonstrated ability to produce quantiles with reduced bias and enhanced representation of hydrological reality, practical constraints, including numerical complexity in solving four coupled nonlinear equations and potential parameter instability with our limited sample sizes (12–28 years), led us to employ it primarily as a generative tool within this Monte Carlo framework. Following the methodology established by Hosking and Wallis [20], the kappa distribution thus serves as a flexible four-parameter family that can represent diverse underlying conditions, including those resembling GLO, GEV, PE3, GPA, and GUM distributions. The parameters of the kappa distribution are selected to reflect the rainfall characteristics specific to each observation site. Using these synthetic datasets, both the AFA and RFA methods are employed to estimate rainfall quantiles. The relative root mean square error (RRMSE) is then computed by treating the AFA estimates for each site as the ‘true’ reference values. This approach enables us to evaluate the robustness and performance of our selected distributions under diverse underlying conditions without being constrained to a single assumed population distribution. By employing this comparative framework, it becomes possible to quantitatively evaluate the strengths and limitations of the AFA and RFA methodologies under controlled, idealized conditions.

3.3.2. Uncertainty Analysis

Uncertainty analysis is crucial for quantifying the reliability of estimated rainfall quantiles in regard to hydrological frequency analysis, particularly in light of the significant variability associated with extreme hydrological events and the challenges of extrapolating data from limited observational data. In the present study, uncertainty is evaluated by comparing the 95% confidence intervals (CIs) for rainfall quantile estimates produced by each method, thereby enabling a direct assessment of their respective accuracies. These confidence intervals are constructed using the distribution of quantile estimates generated from the Monte Carlo simulations detailed in Section 3.3.1. Specifically, for each return period, the lower and upper bounds of the interval correspond to the 2.5th and 97.5th percentiles of the simulated quantile distribution, respectively. This interval represents the range within which the true rainfall quantile is expected to lie with 95% confidence. Narrower confidence intervals indicate estimates with greater precision and reliability, facilitating a robust comparison between quantile estimates derived from AFA and RFA. This analysis is essential for understanding the practical implications of employing each method in urban hydrological design and for identifying which approach offers more reasonable and robust estimates.

Statistical computations and visual representations in this research were executed using R statistical software version 4.4.3 and Python 3.13.3. Furthermore, the analysis incorporated the L-moment methodology through the use of the lmomRFA package (version 3.8) for R, developed by Hosking [34].

4. Results

4.1. Comparison of Rainfall Quantiles for AFA and RFA

To assess the suitability of site 108 as a representative observation point for Seoul, a comparative analysis of the rainfall quantiles derived from both the AFA and RFA for 26 other sites within Seoul was conducted. This enabled the visualization of the distributional characteristics of probable rainfall across various durations and return periods, thereby enabling the quantitative evaluation of the relative position of site 108 within these distributions.

Figure 7 illustrates a quantitative comparison of the results obtained from the AFA and RFA at site 108. Specifically, Figure 7a presents the AFA results for site 108, with lines representing the 2-year (green), 10-year (yellow), 30-year (red), and 100-year (purple) return periods. The shaded range in Figure 7a delineates the spread of the results for the remaining 26 sites.

The analysis revealed that the distribution range (shaded area) generated from the rainfall quantiles of all the sites exhibited considerable deviation under most conditions. This finding strongly indicates that even under identical return period conditions, rainfall characteristics within Seoul possess significant spatial variability. Notably, rainfall quantiles under high-frequency conditions (e.g., 30 year, 100 year) showed differences of up to 90% or more between the minimum and maximum values, suggesting substantial spatial uncertainty in rainfall quantile estimation. Within this broad distribution, site 108 did not consistently align with the mean or median values; instead, its position shifted depending on the return period. This variability serves as a compelling argument against considering site 108 as a stable representative of Seoul’s spatially averaged rainfall characteristics.

Figure 7b displays the results from the RFA. Although the distributional range in this case tended to be somewhat reduced compared to the AFA, residual errors and spatial dispersion attributable to the statistical estimation method were still evident. Even when subject to regional analysis, site 108 was observed to be situated within the upper or middle quantiles of the overall values, failing to secure consistency as a truly representative site for the entire region.

These findings collectively underscore the inherent limitations of generalizing rainfall characteristics across the entire Seoul metropolitan area based solely on the frequency analysis results of a single, specific site, namely 108. Consequently, future approaches for estimating representative rainfall values should incorporate methodologies capable of reflecting the spatial dispersion among sites, such as regional frequency analysis or spatial statistics-based techniques. Particularly for design-oriented rainfall estimation, a single-site approach risks underestimating the spatial heterogeneity of rainfall distribution, which could lead to either under- or over-design of urban infrastructure under extreme rainfall conditions.

4.2. Comparative Performance of AFA and RFA Approaches

Figure 8 presents a comprehensive comparative evaluation of the RRMSE between the AFA and RFA across multiple return periods (2, 10, 30, and 100 years). The empirical results demonstrate distinct performance differentials between these methodologies across the temporal spectrum of the return periods and durations. Overall, the RFA consistently exhibited lower RRMSE values than the AFA, indicating its superior estimation accuracy. While the RRMSE values for both methods generally increased with longer return periods, suggesting higher uncertainty for rarer events, the performance gap between the AFA and RFA became more pronounced with increasing return periods. For instance, the AFA showed an average RRMSE of 0.214 for the 100-year return period, which was approximately 42% higher than the RFA’s average of 0.151, suggesting a rapid increase in the AFA’s estimation error under high-return-period conditions.

As demonstrated by the 100-year return period results shown in Figure 8d, the difference between the AFA and RFA becomes even more pronounced compared to the RRMSE results for the 2-year return period shown in Figure 8a. In particular, for short durations such as 15 min, the RRMSE values for the AFA and RFA were 0.160 and 0.126, respectively, representing a difference of approximately 27.0%. This discrepancy tended to further increase with longer durations, reaching 44.4% for the 60 min duration (AFA: 0.231, RFA: 0.160). For the longest duration considered, 120 min, the difference increased again to 41.3%. In contrast, under the 2-year return period (Figure 8a), the differences between the AFA and RFA were generally small. For example, for the 15 min duration, the RRMSE values for the AFA and RFA were 0.060 and 0.062, respectively, corresponding to only a marginal difference of 3.3%. These results suggest that while the performance gap between the two methods is negligible for short return periods, the prediction error associated with the AFA increases sharply with longer return periods, whereas the RFA maintains a relatively stable performance.

Examining the performance across different durations, the AFA’s RRMSE generally increased across all the durations, peaking at 0.291 for the 120 min duration combined with a 100-year return period. In contrast, the RFA maintained a relatively stable estimation performance, showing an RRMSE of 0.206 under the same conditions. While the RRMSE values were either negligible or the AFA was consistently slightly superior for shorter recurrence intervals (specifically the 2-year return period), the RFA consistently demonstrated a superior performance with lower RRMSE values for return periods of 10 years and longer. This systematic improvement in estimation accuracy with increasing return period provides compelling evidence that single-site approaches introduce substantial uncertainty into the estimation of low-frequency, extreme precipitation events. Consequently, the regional methodology, which leverages the collective statistical power of multiple sites within a homogeneous precipitation region, demonstrates a superior capability in regard to characterizing the probabilistic behavior of extreme rainfall events.

Therefore, this study evaluates the RFA technique as a more reliable and accurate method overall, with its superiority becoming more evident under long-duration and long-return-period conditions. These results underscore the importance of incorporating regional rainfall characteristics into frequency analysis and suggest the greater validity of using the RFA for extreme rainfall analysis and design rainfall estimation.

4.3. Uncertainty Assessment with Confidence Intervals

Figure 9a illustrates the rainfall quantiles and their corresponding CIs, derived from both the AFA and RFA, for a representative site in Seoul (site 108), for a 60 min duration. The analysis revealed that for both methods, the rainfall quantiles proportionally increased with longer return periods, and the width of the CIs also progressively widened. Notably, the AFA consistently exhibited wider CIs across all the return periods compared to the RFA. This suggests that the inherent uncertainties stemming from limitations in site-specific data are more profoundly reflected in the AFA’s distribution estimates. Conversely, the RFA, by integrating information from multiple sites, demonstrates higher statistical stability, resulting in generally narrower CIs and a more gradual pattern of increase in quantile estimates.

Figure 9b further supports these findings by presenting the CIs of rainfall quantiles for all the 27 sites under the same conditions: a 60 min duration and a 100-year return period. This visualization clearly shows that the AFA method leads to considerable inter-site variability, with some sites exhibiting excessively high rainfall quantiles. In stark contrast, the RFA results display relatively smaller CI widths. This reaffirms that the RFA effectively ensures greater consistency in extreme value estimation.

Beyond individual site comparisons, the RFA demonstrably improves the overall statistical reliability by integrating information from multiple points. This results in consistently narrower CIs and reduced variability across locations. Figure 9 (referring to the overall trend across the sites, as previously discussed or as would be visually represented if combined with an average CI plot) further illustrates this trend by presenting the results at each point in ascending order of the CI width. While the AFA shows marked variability between points, the RFA produces a more consistent pattern. These combined results strongly support the use of the RFA as a reliable estimation method for establishing robust design criteria in urban areas characterized by complex spatial variability.

Furthermore, both methodologies consistently show a tendency for CI widths to increase with longer return periods. This reflects the inherent statistical property of increased uncertainty for extreme event periods, particularly for high return periods, such as 100 years. This result emphasizes the need for conservative considerations when establishing design criteria based on high-frequency quantiles.

Collectively, these findings indicate that while the AFA offers the advantage of reflecting localized characteristics, its susceptibility to observational data length and the influence of outliers introduces significant uncertainty into rainfall quantile estimates. Conversely, the RFA, building upon its statistical robustness, provides narrower CIs and enhanced spatial consistency. Therefore, the RFA is evaluated as a more suitable and robust methodology for establishing design criteria for urban infrastructure and developing climate adaptation plans, particularly in complex urban environments where the reliable estimation of extreme rainfall is paramount.

4.4. Evaluation of Seoul’s Disaster Prevention Target

Figure 10 visualizes the spatial distribution of rainfall quantiles, incorporating spatial representativeness through the use of Thiessen polygons, generated based on the 27 sites within Seoul. For each station, the rainfall quantiles derived from both the AFA and RFA are presented, and compared with the urban disaster prevention target of 100 mm and the intensive management zone criterion of 110 mm.

Figure 10a illustrates the AFA results for a 30-year return period, showing that five of the 27 stations exhibit rainfall quantiles exceeding the 100 mm threshold. This highlights significant localized variations in the estimated rainfall quantiles when using the AFA, with values ranging from a minimum of 71.3 mm to a maximum of 108.2 mm (a difference of 36.9 mm). Conversely, Figure 10b, which presents the RFA results, shows that all the stations indicate values below 100 mm (ranging from a minimum of 75.6 mm to a maximum of 94.5 mm, with a difference of 18.9 mm). This suggests that the RFA significantly enhances the spatial consistency of rainfall quantile estimates compared to the AFA. This notable divergence underscores the inherent sensitivity of AFA to the specific characteristics of individual site data. As previously discussed, AFA is particularly susceptible to data scarcity or the presence of outliers, which can lead to either overestimation or underestimation of probable rainfall quantiles. In contrast, RFA integrates data from multiple stations, thereby achieving improved spatial consistency and statistical stability. This characteristic positions RFA as a potentially more reliable foundational dataset for establishing urban-scale disaster prevention plans and design criteria.

Figure 10c,d illustrate the spatial distribution of the 100-year rainfall quantile derived from the AFA and RFA, respectively. In the figure, pronounced spatial variability is evident, with nine stations predominantly in southern Seoul exceeding the 110 mm threshold. The maximum value of 148.1 mm was observed at site 410, which exceeded the minimum value (79.3 mm at site 424) by 86.8%. This substantial range indicates significant spatial variability in rainfall quantile estimation when using site-specific approaches. Figure 10d presents the RFA results, which exhibit a notably more moderate spatial distribution compared to the AFA outcomes. The maximum rainfall quantile value of 114.2 mm at site 425 and the minimum of 91.4 mm at site 419 yield a range of 22.8 mm, representing 24.9% of the minimum value. This reduced spatial variability aligns with theoretical expectations, as regional approaches inherently moderate extreme local variations through information sharing across homogeneous regions.

Critically, the RFA results, which demonstrated superior accuracy and reduced uncertainty in this comparative analysis, indicate that 22 of the 27 stations exceed the current disaster prevention performance standard of 100 mm, with five stations surpassing 110 mm. This finding has profound implications for Seoul’s urban flood management framework, as it strongly suggests the current disaster prevention performance threshold may substantially underestimate actual disaster risk. Given that the RFA provides more reliable estimates for extreme rainfall events, these results underscore the urgent need for a comprehensive reassessment of Seoul’s existing disaster prevention standards to ensure adequate protection against increasingly intense precipitation events.

In conclusion, while AFA offers advantages in reflecting localized rainfall characteristics, it exhibits limitations in establishing consistent design standards across an entire metropolitan area. The research findings consistently confirm that RFA provides a more practical and stable approach for designing urban infrastructure. This underscores the critical need for methodologies that can account for spatial variability in rainfall extremes when developing robust urban flood mitigation strategies.

5. Discussion

This study rigorously investigated the efficacy of AFA versus RFA for estimating sub-hourly rainfall quantiles in urban environments, specifically focusing on the Seoul metropolitan area in Korea. The findings reveal critical insights into the spatial variability of rainfall and the varying reliability of traditional and regional approaches. The analysis consistently demonstrated that RFA offers significantly more robust and reliable estimates for extreme rainfall events, particularly for longer return periods (e.g., 30 years and 100 years), which are directly relevant to urban disaster prevention targets. This enhanced reliability is evidenced by notably narrower confidence intervals and greater spatial consistency across the study area, especially when compared to the substantial uncertainties and localized biases observed with AFA.

The superior performance of RFA in this study aligns with and reinforces findings from numerous previous hydrological research endeavors that advocate for regionalization in extreme value analysis, particularly in regions with limited historical data or pronounced spatial variability [20,35,36]. The results obtained using high-resolution sub-hourly data in a complex urban setting corroborate the established statistical advantages of pooling information from homogeneous regions to overcome the inherent limitations of short individual site records and high data variability. This consistency with the existing literature supports the credibility and generalizability of these findings, confirming that L-moment-based RFA is a statistically sound and practical approach for urban rainfall frequency analysis.

The choice of distributions, such as the three-parameter GEV and two-parameter GUM, aligns with empirical evidence in Korea [31,37] and national guidelines [30]. This approach prioritizes a balance of flexibility, parameter stability, and interpretability for the available short-duration rainfall records [20], as incorporating more parameters risks overfitting and numerical instability in regard to limited sample sizes [38,39]. However, several critical limitations must be acknowledged that may affect the interpretation of these results. The relatively short record length at the monitoring stations (ranging from 12 to 28 years) introduces substantial uncertainty into the estimation of the distribution parameters and return level quantiles, especially for higher return periods. Recent studies [29,40,41] have emphasized that short rainfall records can significantly increase variability and bias in extreme value estimates, leading to wider confidence intervals and reduced reliability of quantile forecasts. Therefore, our quantile results should be interpreted with caution, particularly when applied to infrastructure design and urban flood risk management.

Additionally, this study has certain limitations that warrant clear delineation to prevent misinterpretation. First, the analysis primarily assumes stationarity in rainfall patterns, consistent with current design guidelines in South Korea. While recent extreme events indicate potential nonstationary behavior attributable to climate change, the comprehensive incorporation of nonstationary models into practical design criteria remains a significant challenge that is not fully addressed in this study. Consequently, the derived quantiles represent estimates based on historical stationary conditions. Second, the definition of homogeneous regions was based on the statistical properties of L-moments; meanwhile, robust, future studies could explore hydrological or geographical factors more deeply in regard to region delineation. Third, while sub-hourly data represents a significant advancement, the study’s scope was limited to Seoul’s specific site network. Extrapolation of these exact quantitative results to other urban areas should be conducted with caution, although the methodological insights are broadly applicable.

Despite these limitations, the expected contributions of this research are multi-faceted and significant. It provides empirical evidence for the necessity of shifting towards regional approaches to urban rainfall quantile estimation, particularly for critical high-return-period events. The generated rainfall quantile maps, derived from the more reliable RFA method, offer an evidence-based framework for re-evaluating Seoul’s existing disaster prevention targets and identifying vulnerable areas. This advancement can lead to more spatially differentiated and effective flood management strategies. The emphasis on minute-scale rainfall data also highlights its critical importance for precise urban hydrological modeling, which is essential for rapidly responding urban catchments.

The findings and methodology presented here have broad applicability across various fields. They can be directly applied to urban planning and infrastructure design (e.g., storm sewer systems, pumping stations, flood barriers), risk assessments for climate change adaptation strategies, and the development of national and local hydrological design guidelines. Furthermore, the framework for evaluating AFA vs. RFA performance can be adopted by other cities or regions facing similar challenges in regard to extreme rainfall and limited data.

Future research should prioritize several key areas to enhance the robustness and applicability of these findings. The integration of nonstationary approaches into regional frequency analysis frameworks, with particular emphasis on the direct incorporation of climate change projections, represents a critical advancement. Additionally, advancing dynamic regionalization methodologies that reflect evolving urban landscapes and climate impacts will be essential for long-term applicability. Other important areas include determining optimal rainfall monitoring network configurations for capturing extreme events and conducting comprehensive sensitivity analyses of higher-parameter distribution families, particularly the kappa and Wakeby distributions, which have demonstrated superior statistical properties in regard to generating quantiles with reduced bias and could enhance both theoretical rigor and practical accuracy. Collectively, these research directions will contribute to the development of more resilient and adaptive urban water management practices in the face of changing climatic and urbanization patterns.

6. Conclusions

This study comprehensively compares AFA and RFA methodologies for rainfall quantile estimation, utilizing data from 27 stations, strategically distributed across 25 administrative districts of Seoul in Korea. Employing L-moment statistics, the research estimated rainfall quantiles for various return periods and critically examined the adequacy of Seoul’s current disaster prevention standards.

The analysis revealed pronounced spatio-temporal variability in Seoul’s precipitation patterns. This finding suggests significant uncertainty when extrapolating rainfall estimates from individual stations representing the entire metropolitan area. When comparing methodological approaches, RFA consistently demonstrated markedly superior performance metrics, including enhanced estimation accuracy and reduced uncertainty, producing 26–35% narrower 95% confidence intervals and exhibiting RRMSE values for 100-year return periods that were approximately 42% lower than those from the AFA. The regional methodology’s information-pooling capabilities yielded more reliable estimates, even with limited data.

A notably important discovery was that the 100-year rainfall quantile estimates derived from regional data consistently surpassed Seoul’s existing disaster prevention threshold of 100 mm in the majority of the areas studied. This indicates that the current flood management infrastructure may be significantly inadequate in relation to the actual risks posed by such disasters. This vulnerability is expected to exacerbate as climate change continues to advance, and urbanization persists. The study’s utilization of minute-scale precipitation data represents a methodological advancement, particularly relevant for urban hydrology. Urban watersheds, characterized by high imperviousness and accelerated runoff response times, are especially sensitive to short-duration, high-intensity rainfall events. This highlights the inadequacy of conventional daily rainfall analyses for urban flood risk assessment and emphasizes the necessity of high-resolution temporal data for effective urban hydrological modeling.

The research results provide robust evidence in favor of adopting spatially differentiated flood management strategies across Seoul. The current standards for flood prevention, drainage system capacity, and emergency protocols warrant re-evaluation to accommodate potentially more extreme precipitation events. The rainfall quantile maps generated through regional frequency analysis provide an evidence-based framework for vulnerability assessments and infrastructure prioritization. Finally, this research demonstrates the critical importance of methodological rigor in rainfall quantile estimation and establishes the superiority of regional frequency analysis for urban flood management applications. The empirical evidence presented offers valuable guidance for policy development, addressing evolving precipitation patterns in the context of climate change and urban development.