Urban flood hazard analysis in present and future climate after statistical downscaling: a case study in Ha Tinh city, Vietnam

ABSTRACT Vietnamese cities are highly vulnerable to urban flooding as a consequence of climate change and rapid urbanisation. In this study, current and future pluvial urban flood hazard was assessed for Ha Tinh city. Climate scenarios were obtained after statistical downscaling by applying a quantile-perturbation approach on ensembles of 170 global and 20 regional climate models. Flood impact analysis was based on the 1D-2D dual drainage modelling approach. Extreme daily rainfall intensities are projected to increase by 5 to 20%, whereas wet day frequency will decrease with some uncertainty. Larger changes in rainfall intensities were obtained for the finer scale climate models. Under the 95% upper limit scenario for future rainfall intensities (2071–2100), a 20-year intensity in the current climate would become a 2-year storm in the future and the flood extent is projected to increase by 30–40%. This indicates a need for climate adaptation measures and sustainable future urban planning.


Introduction
Floods threaten human security and sustainable economic growth in many cities of the world. Climate change and urbanization trends will further increase urban flood hazards in the future. Urbanization leads to an increase in impervious surface areas, resulting in a significant decrease in infiltration of storm water, a faster concentration time and an increase in surface runoff, which puts a lot of strain on urban drainage systems and streams (Semadeni-Davies et al. 2008). Climate change has direct impacts on urban floods due to its potential to increase the frequencies and magnitudes of extreme rainstorms, surface runoff, and sewer overflow events (Arnbjerg-Nielsen et al. 2013;IPCC 2014). In deltas and coastal areas, these impacts are aggravated due to the influence of sea-level rise (IPCC 2014).
Vietnam is identified as one of the most vulnerable countries to climate change (IPCC 2014). In the last 50 years, the annual average surface temperature in Vietnam has increased by approximately 0.7°C. To date, global climate models (GCMs) and regional climate models (RCMs) are essential to assess the climate change impacts on local extreme rainfall intensity and urban flooding. These models do, however, have limitations in terms of spatial and temporal resolutions. Super-resolution climate models have been developed for some specific regions in the world, but they are not available for Vietnam. To overcome that problem, statistical downscaling methods have become extremely useful and valuable, if validated (Ahmadalipour, Moradkhani, and Rana 2018;Arnbjerg-Nielsen et al. 2013;Hundecha et al. 2016;Sunyer et al. 2015;Teutschbein and Seibert 2012;Willems 2013). However, these methods do require local observations on extreme rainfalls, which are often limited, especially in data-scarce regions such as the Ha Tinh province in Vietnam. This paper, therefore, focuses on that scarcity region. The objectives of this study are to analyze pluvial urban flood hazards in present and future climate conditions. The methodology includes, (i) the development of an 1D/2D urban flood model for Ha Tinh city, (ii) the analysis of current and future rainfall intensities for various return periods and of wet day frequencies, after applying statistical downscaling to the ensemble of available GCMs and RCMs, and (iii) the investigation of urban flood hazards in current and future climate conditions. Current and future pluvial flash flood hazard maps of Ha Tinh city were produced and compared by simulating rain storms for given return periods based on rainfall intensity-duration-frequency (IDF) curves. These maps are considered as valuable information for spatial planning, disaster preparedness, adaptation, and mitigation measures for the Ha Tinh city authorities. 56.55 km 2 with a population of approximately 104,000 inhabitants (HTSO 2019). The city is situated in a low-lying area of a coastal region, ranging from 1.5 to 2.5 m above sea level, and surrounded by three tidal rivers (VAWR 2017). Ha Tinh city center is only about 12 km from the coast, resulting in a strong impact of the tidal regime on the city's drainage capacity.
The combined effect of tropical depression, cold surges and topography induces extreme orographic rainfall in the central coast of Vietnam, including Ha Tinh province (Nguyen-Le and Matsumoto 2016;Tuan 2019). The province has a monsoon tropical climate with the annual mean temperature of 24.6°C. The amount of annual rainfall is relatively high ranging from 1142 mm to 4391 mm based on the observed rainfall from the period 1958-2018 at Ha Tinh station. The rainy season, from August to November, accounts for 65-80% of the total annual rainfall. The peak rainfall often occurs in September and October with a monthly average of 500 mm and 800 mm, respectively, which is considered very high in comparison with other provinces in Vietnam (Ngo-Duc 2014). During the rainy season, extreme rainfall events often cause severe flooding in the province.
Ha Tinh is the city in Vietnam with the fourth-highest vulnerability with regard to the consequences of climate change and of sea-level rise (Casse, Milhøj, and Nguyen 2015), which poses an increased burden to the urban drainage system. Ha Tinh's urban drainage system is insufficient in capacity and poor in design, urban planning, and management since the combined sewer system covers only 57% of the city area, mainly in the city center (VAWR 2017). As a result, this city has been faced with serious problems of pluvial flooding that occurred frequently almost every year.

Framework of climate change impact analysis on urban flooding
In order to assess the impact of climate change on urban flooding, the framework presented in Figure A1 (Appendix A) was applied for this study. The framework is divided into three parts as follows: (1) Model setup: Model setup: An urban flood model was set up based on the dual drainage concept (Leandro et al. 2009;Schmitt, Thomas, and Ettrich 2004), where full hydrodynamic models were implemented and linked for the 1D underground sewer network and the 2D surface system. Prior to model calibration and validation, sensitivity analysis of model parameters, such as infiltration rate (Ks), imperviousness, and Manning roughness coefficient (n), was performed to provide an understanding of the behaviour of key model parameters. This analysis does increase the confidence about the model performance and its predictions (Kleidorfer et al. 2009) and is presented in the supplementary material of this paper. The model was then built, calibrated, and validated against the observed inundation depths.
(2) Historical and current climate analysis: The historical hourly rainfall data at Ha Tinh station was used to develop IDF curves and design storms, which are storms for given return periods following the Chicago design storm concept (Keifer and Chu 1957). Hourly rainfall available for the study area is considered too coarse for urban flooding analysis; therefore, rainfall downscaling to 10 min was undertaken. This downscaling was based on the rainfall scaling properties, following the method developed by Willems (2000). The pluvial flood hazard under the current climate conditions was assessed by simulating the urban flooding model with design storms for different return periods. The flooded area and inundation depth maps were then produced. (3) Future climate analysis: Future rainfall intensity conditions were obtained after statistical downscaling by applying the quantile perturbation method (Hundecha et al. 2016;Ntegeka et al. 2014;Sunyer et al. 2015;Teutschbein and Seibert 2012;Willems 2013) on a large ensemble of Coupled Model Intercomparison Project Phase 5 Global Climate Models (CMIP5-GCMs) and South Asia Coordinated Regional Downscaling Experiment Reginal Climate Models (SA-CORDEX-RCMs) for different Representative Concentration Pathway (RCP) greenhouse gas concentration scenarios. Prior to statistical downscaling, bias analysis of all GCM and RCM runs was performed based on 15 rainfall stations. The sensitivity and uncertainty assessment of the RCP scenarios, ensemble sizes, and model resolutions on relative changes of daily rainfall intensity and wet day frequency were studied for different return periods in the range from 0.5 to 30 years. Changes in extreme rainfall intensities, IDF curves, and design storms, were then applied to obtain future pluvial flood hazard maps of Ha Tinh city.

MIKE URBAN and MIKE FLOOD
A coupled 1D/2D urban flood model was implemented in the MIKE FLOOD software of Danish Hydraulic Institute (DHI), widely used in numerous studies (Li, Xu, and Wen 2016;Tan et al. 2019;Xie et al. 2017). MIKE FLOOD integrates the hydrological rainfall-runoff model, the 1D hydrodynamic urban water modelling system MIKE URBAN, and the 2D hydrodynamic surface water model MIKE 21. The MOUSE Runoff model simulates the rainfall-runoff processes on the urban and peri-urban sub-catchments. It considers an initial loss, runoff routing based on the concentration time of the urban catchment, and time/area curves.
The MU MOUSE Pipe Flow model is a 1D hydrodynamic model of the Ha Tinh city's combined sewer system that simulates backwater effects and surcharge. The simulated drainage system comprises mainly of 873 manholes, 4 basins, 827 subcatchments, and 1,020 pipes and canals ( Figure A2, Appendix A). The MIKE 21 model is a spatially distributed 2D hydrodynamic model that simulates overland flow paths and velocities. The underlying high-resolution 2-m DEM data were created based on a topographic map of Ha Tinh city area at a scale of 1:2000 and on surveyed campaigns for main hydraulic structures. Based on the DEM, the rectangular mesh of 2 m resolution was created in the 2D model. Land use/land cover data as of 2011 were used from the geographical database ( Figure 1). Because of the rapid construction in the city center, the new buildings were also added using Google satellite images from 2017.

Model input data
The parameters of the urban water system were reviewed from previous studies (VAWR 2017). Time series of rainfall intensity with 10-min interval and daily meteorological data of Ha Tinh station were used. It was assumed that the spatial variability of rainfall is negligible due to the limited observation data and the scope of the study is mainly in the city center. The same rainfall input at Ha Tinh station, located at the city center, was applied for the whole city. Daily potential evapotranspiration (ETo) was estimated using the standard Penman-Monteith Food and Agriculture Organization (FAO) 56 method (Allen et al. 1998). All data used for model setup are summarized in Table A1 (Appendix A).

Model validation data
The inundation depths observed used for model calibration and validation are only available for the main streets of the city center where the traffic is seriously affected by flooding ( Figure A2, Appendix A). The other reference sources including photos, information reported by the media, and surveyed interview with local inhabitants, were reviewed as additional information. The calibration was based on two historical flood events, October 9, 2017 and July 16, 2018, while the validation was undertaken for the other two flood events, April 23, 2015 and September 16, 2015.

IDF curves and design storms
IDF curves of extreme rainfall intensity, presenting relationships between mean rainfall intensity, storm duration, and frequency of occurrence, are widely used in hydrological applications. In order to obtain IDF relationships, the type of distribution is chosen based on extreme value analyses. Independent rainfall intensity extremes were extracted from the time series as peakover-threshold (POT) values by the moving-average technique. A minimum time interval of 12 hours was selected as independent criterion. The distribution parameters were then calibrated for different aggregation-levels in the range of 1 h to 15 days using weighted regression in quantile-quantile plots, following the procedure proposed by Willems (2000;Willems, Beirlant, and Vi 2007). The historical hourly rainfall data for 1975-2018 at Ha Tinh station were applied for that.
Because the Ha Tinh urban drainage system has response times to rainfall smaller than 1 hour, downscaling of the IDF relationships down to 10 minutes was conducted using the scaling properties of rainfall (e.g. Olsson and Niemczynowicz 1994;Willems 2000). The rainfall intensity at Ha Tinh has multiscaling properties: different scaling for aggregation-levels smaller and larger than 7 hours ( Figure A3, Appendix A). Based on these IDF relationships, corresponding design storms were derived up to a total duration of 6 hours.

Data used (climate model runs and historical data)
Large ensembles of GCM and RCM simulation runs were considered (Table A2, Appendix A) for climate model grid cells covering the meteorological station at Ha Tinh. A total number of 170 GCM runs, available from CMIP5, were considered covering the greenhouse gas concentration scenarios RCP 2.6, RCP 4.5, RCP 6.0, and RCP 8.5. For the RCMs, a total number of 20 runs were considered, covering the RCP 4.5 and RCP 8.5 scenarios. A 30-year period covering wet, dry, warm, and cool years is considered sufficiently long to determine the climate for a given location. The period 1961-1990 recommended by the IPCC is considered as the control or baseline period and the period 2071-2100 as the future climate scenario period. The climate change impacts presented in this study, therefore, correspond with the changes from the 30-year control period compared to the 30-year scenario period. Local historical daily rainfall intensity at 15 stations were considered for the period 1961-2018 (Table A3, Appendix A).

Quantile-perturbation based downscaling approach
Perturbation-based downscaling approaches have been widely applied to downscale rainfall intensities by many researchers (Hundecha et al. 2016;Ntegeka et al. 2014;Sunyer et al. 2015;Teutschbein and Seibert 2012;Van Uytven and Willems 2018;Willems 2013). The method is based on the assumption that climate model bias remains unchanged in future climate conditions. It obtains climate change signals in the form of the relative changes from the control to the scenario period, also called change factors or perturbation factors, which are then applied to the historical rainfall data to obtain future rainfall data.
In the quantile-perturbation method (QPM), the perturbation factors are obtained as a function of the return period (Ntegeka et al. 2014;Willems and Vrac 2011). They were studied and applied to the historical rainfall series, for the changes in mean daily rainfall intensity and wet day frequency. A wet day was defined as a day with daily rainfall value greater than 1 mm (Stephens et al. 2010). The relative changes in wet day frequency were obtained as the ratios of the number of wet days in a given month in the future period (2071-2100) over the number of wet days during the corresponding month in the control period . The relative changes in daily rainfall value were obtained in a similar way, but depend on the return period: where φðT i Þ is the relative change of quantile for a given return period T i , R f ðT i Þ is the future (2071-2100) rainfall quantile, and R c ðT i Þ is the current (1961-1900) rainfall quantile, N is the data length (in this study, it is 30 years for both control and future periods), and i is the quantile rank (1 for the highest). The changes are considered for the rainfall intensity quantiles corresponding to the highest 300 daily rainfall intensities in the 30-year period, thus for return periods >0.1 year.

Model calibration and validation
The model was calibrated with a set of parameters for the hydrological and 1D hydrodynamic pipe flow models (Table 1) and for the 2D overland flow model. For the sensitivity analysis of 2D model parameters (Ks, n, and imperviousness), the reader is referred to the supplementary material of this paper. The statistical model evaluation measures, including the Model Efficiency coefficient (EF), derived by Nash and Sutcliffe (1970), and the Mean Error (ME) are summarized in Table A4 (Appendix A). Overall, the statistics indicate that the urban flood model performs well for the study catchment. The inundation depth ME values are less than 4 cm, and the EF values are higher than 0.6 for both calibration and validation periods. Hamouz, Møller-Pedersen, and Muthanna (2020), Moriasi et al. (2007) suggested that in general model performance can be judged as satisfactory with EF > 0.50. The scatterplots of inundation depths between the observations and model results in Figure A4 (Appendix A) show a fairly good performance, without major deviations. Taking the context of data scarce regions into account, the model performance is considered adequate.  return periods for the historical climate conditions . The inundation depths tend to be limited (less than 15 cm) for the less intense events (i.e. lower return periods), while there is an increase of the flooded area with inundation depths of over 30 cm for the more intense events (i.e. higher return periods). The flood extents for various return periods are shown in Figure  2(b). The difference in flood extent between the scenarios is more obvious for the peri-urban area. Figure 2(c) shows the maximum inundation depths and flooded area for different return periods. It shows that most of the city center does flood even for the lowest return period of 0.5 year (i.e. flood event occurs twice per year). These impact results were confirmed by the field interviews with local inhabitants.

Analysis of changes in rainfall intensities
Prior to the analysis of the GCM and RCM projections for rainfall intensity, a comparison of the climate model simulation results with local observations was performed for the control period . Figure 3(a) shows this comparison for the mean monthly rainfall depths for all available GCM and RCM control runs. The GCM-based monthly rainfall depths are systematically lower than the observed depths in the rainy season. For the RCM runs, the differences are smaller. In the months with the highest rainfall intensity, October and November, the RCM runs do show slightly lower intensities compared to the local observations, likely due to the coarse resolutions used in the current RCMs and the anomalous orographic precipitation occurring in the central coast of Vietnam. Figure 3(b) shows the comparison for the daily rainfall intensities at various return periods during the rainy months. The climate-model-based rainfall intensity quantiles are systematically lower compared to the observed intensities for both the GCM and RCM runs. However, some RCM runs fall within the range of observations. In previous studies, Katzfey et al. (2016), Nguyen, Katzfey, and McGregor (2014) and Schubert et al. (2017) used RCMs for dynamical downscaling and projecting rainfall intensity across Vietnam. In line with our findings, these studies also found that the RCMs do underestimate the rainfall amounts during the rainy months along the north central coast of Vietnam including Ha Tinh province. In this study, by applying a perturbation-based downscaling approach, the effect of bias is minimized in the future projections of rainfall intensity (Ntegeka et al. 2014;Willems and Vrac 2011). The relative changes, as applied by the quantileperturbation-based statistical downscaling, were investigated for the relative changes of daily rainfall intensity, wet day frequency, and daily rainfall intensity for different return periods. To investigate the sensitivity of the RCP scenarios and hence the uncertainty related to the future greenhouse gas concentration conditions, the ensemble size, and the climate model resolutions, the relative changes obtained from the RCM ensembles (20 runs) were compared to those derived from the GCM large ensembles (170 runs) and from the subset of GCM runs for all RCP scenarios combined and for the individual RCP 4.5 and RCP 8.5 scenarios. The subset of GCM runs consists of the 17 GCM runs in which RCMs are nested. The changes were calculated for seasonal and monthly time scales. Figure 4 shows the relative changes for daily rainfall intensity and wet day frequency at seasonal scale and in the rainy season, for all RCPs combined and for the individual RCPs. The daily rainfall intensity is projected to increase for all climate models in the rainy season with the median change ranging between 5 and 20%. This result is generally consistent with the previous study conducted by Tran et al. (2016). Under the National Programme of Climate Change Adaptation and Mitigation in Vietnam, dynamical downscaling on five GCMs and RCMs for RCP 4.5 and RCP 8.5 was applied to project the future climate for the whole of Vietnam. That research reported that the rainfall depths in the rainy season in Ha Tinh province are projected to increase between 4 and 30% (median of 17.6%) for the future period 2080-2099. Comparison of the changes for the GCM-subset with the full GCM-ensemble shows that the GCM-subset considered for the RCMs is not fully representative of the large ensemble. The former does slightly overestimate the daily rainfall intensity by approximately 10% compared to the latter. The comparison moreover indicates that the finer scale RCM runs project larger changes in comparison with the GCM runs, especially for the RCP 8.5 scenario for which the average difference is around 20%. By comparing the runs in individual RCPs, as shown from previous researches (e.g. Donat et al. 2017;Toreti et al. 2013;Hosseinzadehtalaei, Tabari, and Willems 2018), the relative changes (φ) obviously increase for higher RCP scenarios (φ RCP8:5 >φ RCP4:5 ).
For wet day frequency in the rainy season, the differences in relative changes due to differences in climate model resolutions, number of considered runs and RCP scenarios for the large ensemble of GCM runs, the subset of GCM runs and the RCMs, are minor. The mean changes are negative, which means that, on average, the wet day frequency is projected to decrease. However, the uncertainty in these changes is large: the zero change values are inside the interquartile uncertainty ranges of the ensembles, as presented by the box-plots.
The relative changes for daily rainfall intensity and wet day frequency are presented in Figure 5 for all RCPs combined and for each month of the year. For daily rainfall intensity, the RCM results show smaller uncertainty in the changes compared to the GCM results, also for the subset of GCM runs. For most of the months, the subset of GCM runs shows slightly higher changes compared to the large ensemble of GCM runs. The RCM runs moreover show higher changes for most of the months compared to the GCM runs. In terms of mean changes, smaller changes are identified from March to July, while there are larger changes for the other months including the rainy season months.
For all months, the wet day frequency decreases for the mean changes. The uncertainty in these changes is, however, large. This uncertainty is smaller for the months from April to September, while it is larger for the remaining months. The zero change is inside the interquartile ranges of the wet day frequency change for most of the months.
Relative changes in daily rainfall intensity calculated for different return periods for all RCPs combined are shown in Figure 6(a). The mean changes are similar for the different return periods, but the uncertainty in the changes for the higher return periods is larger, which is trivial as a result of the smaller sample size on which the higher return period calculation is based on. By contrast, Hosseinzadehtalaei, Tabari, and Willems (2018) studied extreme rainfall intensity for central Belgium based on an ensemble of 88 RCMs. That study reported that the changes in rainfall intensity do depend on the return period, with higher changes in rainfall for higher return periods. When the change ranges are compared between the RCP scenarios ( Figure 6(b)), the changes are   higher for the RCP 8.5 scenario than for the RCP 4.5 scenario, for all return periods. The changes for the subset of GCM runs tend to be slightly overestimated compared to the large ensemble of GCM runs. The changes for the finer scale RCMs tend to be higher than the GCMs.
The relative changes derived from the higher resolution of CORDEX RCMs for the 95th percentile were selected as a high climate scenario (Tabari, Teferi, and Willems 2015) for the impact analysis on urban flooding in Ha Tinh. Future impacts are expected to be situated with high likelihood between that high climate scenario and the current climate conditions.

Analysis of changes in IDF relationships and design storms
To examine the changes in rainfall intensity, a 44-year hourly historical rainfall intensity series  of Ha Tinh station was perturbed by the quantile-perturbation approach for each of 20 CORDEX RCM control-scenario run combinations. This involves random wet day frequency changes and rainfall intensity changes as a function of the return period of the storms in the rainfall time series, according to the relative changes identified in the previous section and the high impact scenario selected for 2071-2100. It results in a perturbed rainfall series, which is a modified version of the historical series, aimed to be representative of the selected future climate condition. Details about the quantile-perturbation method can be found in Willems and Vrac (2011) and Willems (2013).
Given the focus to study climate change impact on the urban flood hazard, IDF relationships and design storms were derived from the perturbed rainfall series using the same method as for the current climate conditions. Figure 7(a) shows an example of a comparison of historical and future IDF curves under the high impact scenario of RCM RCP 4.5 for different return periods. It can be seen that for the high impact scenario the IDF curve for a 2-year return period comes close to the IDF curve for a 20-year return period in the current climate conditions. In other words, under the high impact scenario of RCM RCP 4.5, a 20-year rainfall intensity would become a 2-year intensity in the future condition. Similarly, a return period in the range of 2-5 years or 50-100 years (depending on the aggregation levels) would become about 1 year and 5 years, respectively, under the RCM RCP 4.5 scenario. These results are of direct use in urban drainage design, management, and planning applications. The analyses show that flooding might occur more frequently and be more extreme in Ha Tinh. It suggests that the design storms have to be adjusted for use when designing new systems and climate adaptation plans implemented to keep future flood safety levels to the current levels. It should be noted that for these analyses, it is assumed that land use, water management, and urban planning remain unchanged. Figure 7(b) presents the changes in design storms from the current climate and the high climate change scenarios of RCM RCP 4.5 and RCP 8.5 for a 2-year return period. Consistent with the relative changes identified, the design storms for the future climate conditions are significantly larger in comparison with those for the current conditions, and the intensity changes are higher for the RCP 8.5 scenario than for the RCP 4.5 scenario.

Changes in urban flood hazard
The maximum flood extents for different return periods for the CORDEX RCM RCP 4.5 and RCP 8.5 high scenarios are shown in Figure A5(a) while the maximum inundation depths and flooded areas of these two high scenarios for different return periods (0.5 year and 20 years) are compared in Figure A5(b) (Appendix A). Further analysis of the results, as shown in Figure 8, Table A6 and Table A7 (Appendix A) and, indicates that the flood hazard in Ha Tinh city is projected to increase significantly up to 30-40%. The simulated flooded area for a 20-year return period is projected to increase from 3.79 km 2 for the current climate to 5.75 km 2 and 6.34 km 2 corresponding to the RCM RCP 4.5 and RCM RCP 8.5 scenarios, respectively. In a similar study on the impact of climate change and urbanization on urban flooding in Can Tho city, located in the south of Vietnam, Huong and Pathirana (2013) assessed increasing inundation areas by 50% in the year 2100. These changes are of the same order of magnitude as in this study.

Conclusions and recommendations
This paper offers some insights into the urban flood hazard for Ha Tinh city in North Central Vietnam under present and future climate conditions. A 1D/2D hydrodynamic drainage and surface inundation model was developed for the city. The flood hazard for recent historical climate conditions was assessed using a 44-year hourly rainfall time series  at Ha Tinh station. Extrapolation of rainfall intensity to time intervals of less than 10 minutes was performed based on rainfall scaling properties. IDF curves and design storms were then derived for various aggregation-levels from 10 minutes to 15 days. Flood depth and flood extent maps were obtained and compared for different return periods. The results show that the city center was mostly flooded even for a 0.5-year return period, but with low inundation depths of less than 15 cm.
Climate change impact on rainfall intensity in Ha Tinh was assessed based on a quantile-perturbation approach applied as a statistical downscaling method to different ensemble sets of global and regional climate model runs. The results show that daily rainfall intensity is projected to increase, meanwhile, on average the wet day frequency is projected to decrease, but with some uncertainty. The daily rainfall intensity was found to be sensitive to the ensemble size and spatial resolution of the climate models considered. The subset of GCM runs considered for the RCM runs (dynamic downscaling approach) is not fully representative of the larger ensemble of available GCM runs. The former may overestimate the relative changes in comparison with the latter. The finer scale of the RCMs leads to larger changes compared to the coarser scale GCMs. For the wet day frequency, the climate change impact results are less sensitive to the GCM/RCM ensemble size and spatial scale.
It was also found that the relative changes in rainfall properties strongly depend on the time scale, i.e. seasonal versus monthly. The shorter the time scale, the higher the uncertainty in the impact results. In addition, the results show that the relative changes in rainfall intensity are independent of the return period, but with larger uncertainty for the higher return periods.
For the 95% quantile upper limit of the RCM-based changes, called high impact scenario, corresponding changes in IDF relationships and design storms were obtained and applied to assess the impact of climate change on the urban flood hazard. The impact results show that flooding may occur more frequently and become more extreme by the end of the century. The current sewer system designed for a 20-year return period may be only able to accommodate a 2-year return period by the end of the century under the high impact scenario for RCP 4.5. The warning message might contribute to better future urban planning for Ha Tinh city since the combined sewer system currently covers only 57% of the city area. It is, therefore, suggested that when a new development area is proposed, the revised and future design storms are applied and that flexible adaptation measures are designed to take into account the impact of climate change and urban growth, together with the uncertainties on these future trends.
Sustainable development strategies for future urban planning would be highly recommended: these should include both innovative adaptations such as storm water retention and infiltration in open and/or green spaces in the city (blue-green water integration), rainwater tanks, and traditional end-of-pipe solutions, such as upgrading or installing new drainage systems and pumping stations. This paper only considers the impact of climate change on rainfall intensity; the other elements which also have impacts on urban flooding, such as urbanization due to the obvious trend of rapid urbanization in the region, should be considered for future research.