Towards a high-resolution modelling scheme for local-scale urban flood risk assessment based on digital aerial photogrammetry

Abstract

With the rapid development of cities and the impact of climate change, cities located near rivers are facing an increasingly serious flood threat. Urban flood risk prediction management is pressing. The peak discharge and duration are important factors in urban flood management as well as the important characteristics of flood hygrograph. Mostly, hydrological model is used to obtain the upstream flood hygrograph to drive inundation model. However, lack of information on reservoirs, barrage structures and land use make it difficult to construct high-precision hydrological models, especially in upstream cities where data is lacking. In this study, therefore, we propose an approach to urban flood management based on measured flood data from urban hydrological stations in close proximity to derive flood hydrography for different return periods, and establish a high-precision urban-scale river flood risk management method by combining with Unmanned Aerial Vehicle (UAV) survey data. This method can allow information on river construction and underlying surface change to be introduced into the flood hygrograph implicitly, thus avoiding the difficulty of establishing hydrological and hydrodynamic coupling of the whole basin. The applicability and accuracy of the methodology are explored in this article with reference to Chenxi City in the upper reaches of Yuanshui River. Comparison the flood process of the 10-year with the actual flood in 2016 indicates that the extent of inundation is strongly dependent on the instantaneous discharge, with the peak flow largely determining the maximum inundation extent and risk level. This illustrates the feasibility of risk assessment of urban flooding based on flood hygrograph with different period levels derived from measured flood processes at adjacent sites. Subsequently, the different return periods scenario is simulated and analyzed. This approach provides technical guidance for flood risk assessment in areas where data is lacking (e.g., upstream mountainous cities).

KEYWORDS:

• Urban fluvial flood
• UAV
• flood hydrography
• risk assessment

1 Introduction

Urbanization is a global long-term process, accompanied by population and economy regionalization. Industry and economy are concentrated in cities accompanied by a considerable percentage of GDP and assets (high value property and infrastructure) (Bates et al., 2005; Olbert et al., 2017). Most cities are located along rivers and subject to a great threat of flooding owing to the destructive, pervasive, and frequency of flooding (Sanyal & Lu, 2004). Over the past few decades, the frequency of urban floods has increased (Atta-ur-Rahman et al., 2016; Ziegler, 2012), and floods cause more than a third of global economic damage according to the EM-DAT international disaster database (Environment Agency, 2018). In the UK, approximately 40% of damage and associated economic loss in cities result from flooding (Douglas et al., 2010). In 2017, 55.15 million people in China were affected by floods, with 214.3 billion yuan a direct economic loss (MEEPRC, 2017). Furthermore, recent studies indicate that rainfall will have a further potential to intensify in the near future (owing to, for example, the impact of El Nino) (Chandra et al., 2015; Corringham & Cayan, 2019; Xu & Luo, 2015). Hence, cities confront the dual pressures of human activities and climate change. In order to cope with flood risk events, flood risk classification (Wu et al., 2015) and residential evacuation planning (Baba et al., 2017) become increasingly important. Obviously, precise modelling is required of inundation extent, water depths, discharge zoning, and flow velocity in urbanized flood-prone areas. As the key parameters for flood assessment, an appropriate combination of water depth and flow velocity represents the potential impact of flooding on a city, such as the direct potential impact assessment on structures and buildings (Bermúdez & Zischg, 2018; Milanesi et al., 2018), people (Milanesi et al., 2016; Xia et al., 2014) and vehicles (Martínez-Gomariz et al., 2016). As a consequence, accurate modelling and forecasting of flooding play a major role in the better management of flood risk.

With the development of numerical methods and computational techniques, numerical modelling has become the most popular method for providing detailed information on flood risk management and urban planning (Teng et al., 2017). Comprehensiveness and accuracy includes 1-D (Brunner, 2016; Papaioannou et al., 2016), 2-D (Dazzi et al., 2019; Sanders & Schubert, 2019), 1-D/2-D schematizations, and even 3D models (Prakash et al., 2014; Rong et al., 2020; Vacondio et al., 2012), and experimental numerical simulation research on flow diverted at a bifurcating road crossing (Wang et al., 2022). In general, numerical simulation is mainly used for full-scale basin floods, including hydrological basin models, river networks and inundation models. Hydrological models are usually used to provide upstream boundaries to force river networks and inundation models (Collischonn & Pontes, 2016 ; Nguyen et al., 2016). Kim et al. (2012) developed a spatially-distributed, fully-coupled hydrologic and hydrodynamic model, namely TRIBS-OFM. Nguyen et al. (2016) coupled hydrologic model (HL-RDHM) with a hydraulic model (BreZo) for flash flood modelling in Oklahoma. Li et al. (2019) employed the hydrological model TOPMODEL to obtain a flood hydrograph and the 1D-2D hydrodynamic model MIKEFLOOD to simulate floodplain inundation in the Pajiang River in South China. Although such models can consider the flooding process from a river basin perspective, the simulation of a hydrological curve based on the Digital Elevation Model (DEM) and underlying surface data has certain errors owing to the existence of complex terrain and a river network, and hydrological models can’t completely describe the dynamic process of flood wave generation and propagation. In addition, a 1D-2D model is limited by ignoring momentum exchange and backflow between channels within a floodplain, which lacks the integrity of river and urban inundation models (Papaioannou et al., 2016). The great achievement of Xia et al. (2019) was to establish a high-precision two-dimensional hydrodynamics framework (HIPIMS) to simulate the entire flooding process from rainfall to inundation. However, the detailed representation of catchment topographic features such as a high-quality DEM, a land-use map and geological information for the whole basin are challenging, especially for developing countries. In addition, compared to the whole basin, there are more concerns about the flood risk in vulnerable urban areas. Therefore, how to reduce the modelling difficulty and improve the accuracy of urban flood assessment (including watershed flood characteristics and urban high precision) is the key problem of urban flood analysis, especially for rivers in mountainous regions with limited data availability and difficulties in establishing detailed and accurate hydrological models.

Accurate simulation of urban flood requires reasonable flood hygrographs, and more importantly, high precision elevation data that can represent complex urban terrain. The rational design and evaluation of measures to improve the resilience to urban flooding should be based on the analysis of a range of scenarios, including past floods and flood hazard maps designed to provide spatial visualization of potentially hazardous areas (Macchione et al., 2019; Sy et al., 2019). Flood Frequency Analysis (FFA) is a fundamental and effective method for producing flood hygrographs, and is divided into two approaches: rainfall-runoff simulation of flood peak discharge (the rainfall-based approach) and FFA of historical discharge data (the discharge-based approach) (Tanaka et al., 2017). Rainfall-runoff models are a common method of obtaining upstream flow boundaries and driving inundation models (Hasan et al., 2019; Collischonn & Pontes, 2016; Nguyen et al., 2016). Generally speaking, extreme rainfall is the main cause of urban flooding (Huang et al., 2020; Xu et al., 2021), but the local climate effect caused by climate change and strong human activities intensifies the heterogeneity of regional spatial and temporal distribution of rainfall, and changes in the spatial and temporal distribution characteristics of rainfall inevitably lead to changes in flood characteristics such as peak discharge and flood ephemeris, increasing the uncertainty of flood hazards (Sanches et al., 2019; Zhao et al., 2020). Simultaneously, for floods of different return periods, the traditional assumption based on the same frequency of rainfall and flooding (i.e. n-year rainfall causes n-year floods) is subject to great uncertainty owing to heterogeneity (Zhou et al., 2017). Zhu et al. (2018) pointed out that, under the influence of spatial and temporal heterogeneity of rainfall, a relatively small amount of precipitation (such as a 20-year event) can also lead to a flood of a larger magnitude (such as a 500-year event). Meanwhile, the construction of dams and sluices, as well as the accuracy of the underlying surface data, can also contribute to errors in rainfall-runoff models, especially for mountainous region river cities where data are scarce and flooding duration is short. However, the discharge-based approach, e.g. Bulletin 17 C (England et al., 2015), based on the data of gauging stations, can effectively reflect the local hydrological character (such as peak magnitude, flood volume and duration). The different return period levels of floods obtained from measured flood process and historical data based on the site can effectively reduce the rainfall complexity, hydrological variability and the influence of human factors. On the other hand, urban flooding depends on surface runoff under the action of gravity, and its flow characteristics strongly depend on surface elevation. Therefore, detailed topography and structures play an important role in flood evolution. High-precision horizontal terrain resolution that can represent the complex structure of a city is basically 1 m, or even 0.1–0.4 m (Tanaka et al., 2020). Unmanned Aerial Vehicles (UAVs) and ground-based Laser Radar (LiDAR) technology have made it possible to obtain high precision urban terrain data owing to their low cost and ease of operation (Leitão, Moy De Vitry et al., 2016; Leitão, Prodanović et al., 2016). Rong et al. (2020) constructed a coastal urban flood simulation system based on UAV and building information modeling (BIM) measured data and quantified the flood risk of coastal cities. Trepekli et al. (2022) showed that UAV LiDAR effectively improved urban flood modelling and assessment.

In this article, an urban catchment integrated river−urban flood evolution model is established with high-precision DEM data through UAV survey technology. Based on flooding process data measured at the site, the flooding process under different return periods (10-year, 20-year, 50-year, and 100-year) is considered to analyse the risk of flood inundation in Chenxi city. As a new and simple urban flood assessment modelling method, the purpose is firstly to reduce the challenge of modelling data by focusing on the key areas of people’s concern without constructing the overall watershed model. Secondly, based on the measured flood process of the gauging stations, which represent the ultimate result of rainfall-to-runoff processes and, to some extent, incorporate actual influencing factors that may be challenging to capture in hydrological models such as the peak reduction due to reservoir storage, the local hydrological characteristics can be effectively included. Thirdly, the process of flood inundation can be effectively simulated based on the high precision aerial survey data of UAV. This article aims to check the applicability of the method, and provide ideas and design basis for urban flood control management.

2 Study area

Chenxi city, located in the west of Hunan Province in China, the upper and middle reaches of Yuanjiang River (see Figure 1). It belongs to the subtropical monsoon climate, with an annual rainfall of about 1328.4 mm. The rainfall is concentrated from March to July and is characterized by concentration and high intensity. As a result, it’s easy to generate storm runoff causing flood. In recent years, the region has experienced frequent flooding, such as the 2016 and 2017 Flood (Figure 2). In addition, at the confluence of the Yuanjiang river and Chenshui river (the first tributary of Yuanjiang River), the water level is easy to increase by the jacking action of current.

Figure 1. Location of study area and river information.

Figure 2. The flood inundation picture of: (a) the right bank of Chenshui River in the 2016 flood; and (b) the right bank of Yuanjiang River in the 2017 flood.

3 Data

3.1 UAV DEM

Highly accurate topographic data is the first element in the construction of an urban flood model. In this article, the UVA aero-photographic system based on multi-camera oblique photography technology was used for data acquisition. The flight measurement is implemented by the DJ Phantom 4 RT UAV with the 3-axis (pitch, roll, yaw) stabilization system, and the maximum operating area of a single flight is about 1 km², which meets the accuracy requirements of 1:500 topographic map aerial photogrammetry industry specifications in GB/T7930-2008. Meanwhile, a real-time kinematic (RTK) system with 35 photo-control-points was adopted, as shown in Figure 3, and more detailed information is shown in Table 1, resulting in the horizontal and vertical accuracy of the data reached centimetre level. In this survey, the altitude of the flight was 400 m and the measured area was 11.2 km². In order to ensure the quality of collected data, the route adopts a Z-shape with graph overlap rate of 60%.

Figure 3. The distribution of photo-control points.

Table 1. Photo-control points coordinate information.

Point	X	Y	Elevation (m)	Point	X	Y	Elevation (m)
P1	421,920.641	3,101,014.617	132.146	P19	418,296.173	3,099,751.656	154.775
P2	421,632.265	3,100,719.882	149.512	P20	419,101.351	3,099,099.741	116.099
P3	421,426.949	3,100,243.802	128.864	P21	419,022.251	3,098,581.837	127.660
P4	421,073.997	3,100,486.876	125.393	P22	419,133.542	3,097,842.916	121.083
P5	420,701.648	3,100,302.289	118.646	P23	418,626.604	3,097,474.097	122.535
P6	420,044.821	3,099,831.009	123.841	P24	419,396.113	3,097,300.198	124.539
P7	422,561.311	3,098,488.927	126.214	P25	419,043.129	3,097,013.562	123.159
P8	420,137.689	3,099,887.767	121.218	P26	419,529.930	3,096,899.778	124.339
P9	422,007.586	3,098,587.697	129.763	P27	420,044.877	3,096,913.103	125.598
P10	421,686.629	3,098,819.467	143.620	P28	419,311.133	3,098,419.899	120.449
P11	421,142.989	3,098,878.605	127.817	P29	419,760.264	3,097,645.390	125.028
P12	420,843.894	3,099,017.328	128.253	P30	420,128.439	3,098,519.875	127.430
P13	420,415.214	3,099,125.308	128.096	P31	420,803.306	3,098,304.498	129.658
P14	419,922.153	3,099,228.413	125.068	P32	420,299.887	3,098,160.626	135.720
P15	419,565.863	3,099,298.128	124.662	P33	421,529.159	3,098,053.009	131.465
P16	419,102.100	3,099,624.560	117.701	P34	422,227.581	3,097,845.816	135.720
P17	418,772.148	3,099,952.336	116.053	P35	420,530.652	3,099,979.520	123.823
P18	418,356.655	3,100,281.293	121.509

Figure 4 shows the 3D city model obtained by UAV. Although the UAV data contains the surface elevation in most urban areas, there are many non-reality elevation information (e.g. trees, bridges and billboard), which will cause large errors to simulation results. Moreover, UAV cannot measure the elevation of the river bottom. Therefore, the original data need further processing, including deleting the water data and unifying the canopy elevation to the ground elevation.

Figure 4. The 3D digital city model.

Finally, urban DEM data with a horizontal accuracy of 1m is obtained, which contains 2, 590, 3917 scatter data. To demonstrate the high-resolution of DEM data, we compare the DEM cloud map with UAV aerial survey image (Figure 5). It’s clearly show that high-resolution DEM can effectively distinguish complex landforms (e.g. streets, buildings, roads) in urban areas. Besides, the river bottom topography is based on ship survey data came from the 1:65,000 topographic map of Yutan Junction project measured in 2015 year, and the elevation datum was unified with DEM data.

Figure 5. Terrain extracted from the digital city model geometry.

3.2 Flood hygrograph

Flood management policies and designs are generally based on an estimate of flood frequency, namely the return period (Gilroy & McCuen, 2012). The flood peak discharge is the main concern in flood estimation (Rosbjerg et al., 2013), and it’s the major factor in the magnitude of flood damage. Furthermore, for the prevention of flood disasters and the design of hydraulic structures, it is also important to understand the volume of flooding and the shape of flood hydrograph (Mediero et al., 2010). Different flood hydrograph (e.g. flood volume and duration) will lead to differences in cost of hydraulic structures, and affect flood control policies and flood management strategies (Yue et al., 2002). Similarly, the characteristic quantities (e.g. flood duration and flood hydrograph shapes) of flood process has great influence on urban submergence, which is related to local engineering construction, such as dam and river dikes (Tanaka et al., 2017).

In this article, the 2017 flood events measured at Pushi station (See Figure 6) is taken as a typical flood hydrograph, which is the largest flood in recent years with a peak value of about one in 50 years for the Yuanshui River, and has the function of connecting the upper and lower floods. The parabolic shape hydrograph consists of a steep increase to peak discharge and a gentle subsequent decline (See Figure 6(a)). According to the flood frequency analysis data of Dafutan Water Conservancy Project upstream of the city (8.5 km) (Location is shown in Figure 7), the 100, 50, 20, and 10-year return period flood events, where the peak discharge is 27900, 24900, 20900, and 17700 m³/s, respectively (see Table 2), are adopted for flood research scenarios. The flood hydrographs at different frequencies are obtained by using the same frequency amplification principle and the corresponding formula as follows: (1) $Q_f=Q_{obs}\times \frac{Pea{k}_{max}}{Q_{max}}$(1) where $Pea{k}_{max}$ is the peak discharge of certain frequency, $Q_{max}$ is the observed peak discharge, $Q_{obs}$ and $Q_f$ corresponding to the flood discharge process of measured and certain frequency. This generates the hydrographs as shown in Figure 6(b).

Figure 6. (a) Hydrograph of the flood that occurred in 2017. (b) Hydrographs of different frequencies.

Figure 7. The study area and schematic diagram of the local grid.

Table 2. Discharge for different return periods.

Return period	100 years	50 years	20 years	10 years
Discharge (m^3/s)	27,900	24,900	20,900	17,700

4 Model description and validation

4.1 Model description and setup

The two dimensions (2D) hydrodynamic models based on the shallow water equation are widely used in flood process simulation, such as HiPIMS based on the GPU framework (Xia et al., 2019), ANUGA (Mungkasi & Roberts, 2013), TUFLOW-FV2 (BMT-WBM, 2016), TUFLOW-HPC (BMT-WBM, 2018). In this research, the model is based on the solution of 3D Reynolds averaged NS equations, subject to the assumptions of hydrostatic pressure. The equations can be expressed as. (2) $\begin{array}{rl}& \frac{\mathrm{\partial }U}{\mathrm{\partial }t}+\frac{\mathrm{\partial }F}{\mathrm{\partial }x}+\frac{\mathrm{\partial }G}{\mathrm{\partial }y}=S+\frac{\mathrm{\partial }R}{\mathrm{\partial }x}+\frac{\mathrm{\partial }Q}{\mathrm{\partial }y}\end{array}$(2) (3) $\begin{array}{rl}& U=\left(\begin{array}{c}\eta \\ hu\\ hv\end{array}\right),F=\left(\begin{array}{c}hu\\ h{u}^2+g{h}^2/2\\ huv\end{array}\right),\\ & G=\left(\begin{array}{c}hv\\ huv\\ h{u}^2+g{h}^2/2\end{array}\right),\end{array}$(3) (4) $\begin{array}{rl}& S=\left(\begin{array}{c}0\\ -gh(S_{bx}-S_{fx})\\ -gh(S_{by}-S_{fy})\end{array}\right),R=\left(\begin{array}{c}0\\ \nu {\displaystyle \frac{\mathrm{\partial }}{\mathrm{\partial }x}}\left(hu\right)\\ \nu {\displaystyle \frac{\mathrm{\partial }}{\mathrm{\partial }x}}\left(hv\right)\end{array}\right),\\ & Q=\left(\begin{array}{c}0\\ \nu {\displaystyle \frac{\mathrm{\partial }}{\mathrm{\partial }y}}\left(hu\right)\\ \nu {\displaystyle \frac{\mathrm{\partial }}{\mathrm{\partial }y}}\left(hv\right)\end{array}\right)\end{array}$(4)

In which: t(s) is time; x(m), y(m) are the horizontal coordinates; h(m) is the water depth $u$ (m/s); $v$ (m/s) is the depth-averaged flow velocities in x-directions and y-directions; g (m/s²) is the gravitational acceleration. $\eta$(m) is the water level. F, G are the convective flows in the x and y directions, and R Q are the diffusion fluxes in the x and y directions. $S_{bx}$, $S_{by}$ is bottom slope, $S_{bx}=\mathrm{\partial }{z}_b/\mathrm{\partial }x,S_{by}=\mathrm{\partial }{z}_b/\mathrm{\partial }y$, $z_b$ is the bottom elevation. $S_{fx}$, $S_{fy}$ are the bottom friction in x -directions and y -directions, which is modelled using Manning’s resistance law. $\nu$ is the coefficient of kinematic viscosity. The governing equations in (2) are solved using the finite-volume method on non-overlapping irregular triangular meshes. In order to capture discontinuities, an approximate Riemann solver (Roe’s scheme) is used to calculate the convective fluxes at the interface (Harten et al., 1983; Roe, 1981). Model employing a linear gradient-reconstruction technique to achieve the Second-order spatial accuracy, and the average gradients are estimated using the method by Jawahar and Kamath (2000). Meanwhile, the second order TVD slope limiter is used to avoid numerical oscillations (Darwish & Moukalled, 2003; Hirsch, 1990). A third-order Runge–Kutta scheme is integrated in time steps, and wetting and drying are also implemented. This approach has been successfully applied to many realistic problems (Tang, Chien et al., 2013; Tang, Kraatz et al., 2013). In this simulation, the upstream boundary conditions are characterized by the discharge, while the downstream boundary condition is controlled by water level derived from the water level-discharge relationship curve. As shown in Figure 7. The model includes Yuanjiang river, Chenshui river and Chenxi city, with Pushi hydrological station as the boundary in the downstream and Dabutan Water Conservancy Project as the boundary in upstream.

4.2 Model validation

The 2017-year fluvial flood event is simulated in this work to demonstrate the performance of the hydrodynamic model. In this flood, the peak discharge was 26524 m³/s, accompanied with the maximum river water level in the urban section at 127.56 m, which is only 0.44 m lower than the historical maximum water level. The vast majority of urban area was severely inundated, with the maximum water depth of 10.06 m and the flood levee submerged water depth of 1.1 m. The river backed up into the city and the city drainage system failed, so the drainage function of the pipe net is neglected. Meanwhile, as the overbank flood, rainfall accumulation in urban areas haven’t been considered.

The urban flooding assessment requires a scientifically credible and computationally feasible model, so model mesh resolution needs to consider a compromise balance between describing phenomena in complex environments and computational feasibility. In this article, the mesh precision of 4m, 8m and 12m is considered. The results are shown in Figure 8. The inundation range between the mesh resolution of 4m and the 8m remains basically the same, with some of the streets in the Chenxi City (red dashed box B in Figure 8(a)) having a greater inundation range for the 4m mesh resolution model, but which is also reflected in the 8m mesh resolution. However, the flood inundation area at 12m mesh resolution shows an obvious underestimation, The flood evolution in the urban area to the right of the Yuanshui River (see red dashed box A in Figure 8(a)) is not represented, mainly because the 12m mesh resolution is unable to identify flood transmission channels. Similarly, the right side of the Chenshui River and the left side of the Yuanshui River are underestimated to varying degrees. Based on the above results, therefore, considering the model accuracy and computational efficiency, this manuscript uses 8m mesh resolution for computational analysis.

Figure 8. Comparison of the maximum inundation area of the flood in 2017 with mesh accuracies of 4, 8, and 12 m.

In this study, the Manning coefficient of the models has been checked. Considering that Manning coefficient has spatial variability with land cover types, different Manning coefficients are respectively assigned to the river and other areas (e.g. urban areas and crop zones, etc.) according to the Giovanni Forzieri et al. (2012). Different values are tested, and finally the value (0.031 m^1/3/s for river and 0.027–0.04 m^1/3/s for the land area) produces reasonable results, which is adopted to support the simulations. In order to evaluate the validity of model, two evaluation metrics are used, including the root mean squared error (RMSE) and the Nash-Sutcliffe model efficiency coefficient (NSE). The RMSE is calculated using. (5) $RMSE\text{ = }\sqrt{\frac{{\sum }_1^N{(l_m^{\text{n}}-l_o^n)}^2}{N}}$(5) in which N is the total number of observe data time steps, $l_o^{\text{n}}$ is the observed water elevation at time step n and the $l_m^{\text{n}}$ is the corresponding model water elevation. The NSE is estimated as (6) $NSE=1-\frac{{\sum }_1^N{(l_m^n-l_o^n)}^2}{{\sum }_1^N{(l_m^n-\overline{l_o})}^2}$(6) as $\overline{l_o}$ is the mean observed data. NSE ranges from $-\mathrm{\infty }$ to 1, NSE = 1 indicates the model prediction agreement with observation. NSE = 0 suggest that the mean observation value has been predicted by the model.

As shown in the Figure 9(b), the numerical result is good agreement with the observations at Chenxi Hydrology Station (see Figure 7). The water level gradually rises to peak and then decreases at similar rate, forming a triangular shape of typical flood hygrograph. Moreover, to provide quantitative assessment of the simulation results, RMSE and NSE are calculated, with 0.39 and 0.97, respectively. Although there’s only Chenxi station, it basically represents the water level change of flood in the urban river section as the special geographical location. Of course, there are still some errors between simulation and observation, mainly in that the simulated value is greater than the measured value, which is caused by the fact that the river bottom topography cannot be updated in time. The river bottom topography data is 2015 year, and the recent river sand mining activities have caused the river bottom elevation to decline. As the result, there are some differences between the model river bottom elevation and the actual river topography. Overall, the numerical simulation is consistent with the observations, and the RMSE and NSE are acceptable, which indicate well performance of the model.

Figure 9. Model validation with: (a) comparison of inundation area with risk maps developed by the Hydrographic Board; and (b) comparison of the water level at Chenxi Hydrology Station.

In addition, we also compared the inundation area with the 2004 edition of the flood risk map compiled by the local Hydrology Bureau, as shown in Figure 9(a). In this image, the thick solid red line shows the 50-year flood risk range, while the bathymetric map represents the flood inundation area of 2017 obtained by numerical simulation. The peak discharge of the 2017-year flood is closer to that of the 50-year return period. Form the Figure 9(a), the location of flooded regions are generally consistent between the risk map and simulation except for the northern part of Yuanjiang and Chenshui left bank area. The main reason is the construction of breakwater on the right bank of Yuanjiang River (see Figure 10) and the change of urban development, such as roads and buildings. Unfortunately, the evaluation index was not calculated because the digital data of reported flood area was unviable. As a summary, the model successfully reproduces the flooding process and inundation extent of the Chenxi city.

Figure 10. The flood dike on the right-hand side of Yuanjiang River.

5 Results

The applicability of urban flood research depends on the risk assessment and hazard maps, which are also the development strategies and guidelines for local governments to manage urban construction, risk and emergency management personnel, and the common people to prepare action plans. In this article, therefore, a method based on flood parameters (D,V, and the combination of both D × V) is used to establish the evaluation of flood disaster index to describe the vulnerability and risk of flood disaster in urban environment at different flood frequencies. These results are presented and discussed in the following sections.

5.1 Flood hydrographs accuracy assessment

This article proposes an approach for urban flood management based on the measured flood data from urban proximity hydrological stations to derive flood hydrography for different return periods, and establish a high-precision urban-scale river flood risk management method by combining with UAV aerial survey technology. In order to verify the rationality of this method, we compare the flood hydrograph in 2016 with the 10-year return period, and peak flood discharges of 180,00 and 17,700 m³/s, respectively. As shown in Figure 11, before the flood peak, the predicted flood hydrograph is higher than the measured floods, mainly owing to the staggered flood peaks of Chenshui River and Yuanjiang River, but remain consistent during high flows (>12,000 m³/s). For the flood recession, the rest are consistent except for the observed short-term reduction. In general, the correlation coefficient between them was 0.89, and the predicted flood process is basically consistent with the actual measurements.

Figure 11. Comparison of the observed flood process in 2016 with the 10-year flood hydrograph at Pushi Station.

In order to compare the inundation of the two scenarios further, this manuscript simulates the 10-year flood process derived based on the 2017 measured flood and the actual flood process in 2016. Subsequently, a comparison of the maximum inundation areas between the two is shown in Figure 12. As can be seen from the figure, the maximum flooded areas for the two boundary conditions are almost identical, except for a slight difference between the two banks of Chenshui River (the dashed circles). The results illustrate the feasibility of deriving flood process simulation approaches for different return periods from risk assessment based on measured flood hydrographs.

Figure 12. Maximum inundation extent of the 10-year flood scenario (peak discharge: 17,700 m³/s) and the 2016 flood (peak discharge: 18,000 m³/s).

Figure 13 represents the temporal variation of the inundated area in the 36 h before and after the flood peak, as the flow rate of the 10-year flood process is greater than the measured flood process in 2016 (see Figure 11); therefore, the corresponding inundated area shows the same phenomenon. Obviously, the 10-year flood and the 2016 flood have almost identical maximum flood inundation areas. For upstream river channels without lakes or ocean topographic effects, the relationship between flow and water level is relatively homogeneous. The water level in the river channel remains consistent with the flood discharge. Therefore, peak flow is the key factor influencing the maximum extent of flood inundation and largely determines the risk level of a flood. Further, this reveals that assessing urban flood inundation risk based on peak flow has a theoretical foundation and practical value.

Figure 13. Curves of the flood inundated area for the 10-year flood scenario and the 2016 flood.

5.2 Evolution of flood inundation under different return periods

In order to show the dynamic process of flood evolution, the time evolution of inundation areas for all scenarios is presented. Figure 14 illustrate the process of 10-year flood inundation, including the flood depth maps for Chenxi City at t = −24, −12, 0, 12, and 24 h taking the flood peak time as the origin of the time axis, the negative values representing before the flood. As shown in Figure 14 at t = −24 h, certain low-lying areas along Chenshui River have been flooded although there have been no serious bank overflows. With the further increase of the upstream discharge, the floodplain occurred at the confluence of Chenshui and Yuanjiang Rivers and in the area to the left of Chenshui River at t = −12 h. Since the inundation area is connected to Chenshui River, this definitely shows that the inundation is mostly caused by overbank flow from Chenshui River. As the simulation time increases, the inundation area expands further and reaches its maximum at t = 0 h. By comparing the flood maps at t = −12 and t = 0 h, it can be seen that the flood mainly spreads far along low-lying areas, and there are also differences in the timing of floods in different areas, which can provide initial locations and routes for evacuating people and property. After the peak passed, the flood water in the floodplain started to retreat gradually, as shown by the shrinking inundation maps in Figure 14 for t = 12 and 24 h. On the whole, the submergence scope of the10-year scenario is mainly concentrated on the left bank of Chenshui and its junction, which is mostly agricultural land with low terrain. Additionally, the Chenshui floodplain is the main area of flood, hence it is of great significance to strengthen the construction of the Chenshui embankment.

Figure 14. Water depth and extent of inundation at different times for the 10-year return period.

Similarly, the evolution of the flood inundation area at 20 years, 50 years, and 100 years is presented in Figure 15. However, considering that flood losses are mainly concentrated before the arrival of flood peaks, here we only focus on the temporal and spatial variability of the flood rise process. Meanwhile, the time series variation curve of flood inundated areas at different frequencies is shown in Figure 16. In terms of the temporal change trend, the flood plain at each frequency occurs within 48 h before the flood peak, with the characteristics of rapid rising and slow retreating (see Figure 16(a)). In terms of spatial variation, similar to the case of a 10-year flood, the inundation area first occurred on the left bank of Chenshui River, as shown in Figure 15(a), 15(c), and 15(i). However, the submerged area of the urban area on the right bank of Yuanshui River was increased in the 50-year and 100-year floods, which is mainly distributed on both sides of the stream in the main urban area.

Figure 15. Water depth and inundated area at different times at 20 years (a)−(d), 50 years (e)−(h), and 100 years (i)−(l), while t = −24, t = −12, t = −6, and t = 0 correspond to (a, e, i), (b, f, j), (c, g, k), and (d, h, l), respectively.

Figure 16. (a) Time curves of submerged area under different return periods. (b) Net inundation area under different return periods.

As the flood crest approaches, the submerged area and water depth increase further. Under the 20-year return period, the floodplain phenomenon occurs in the main urban area on the right-hand side of Yuanjiang River at t = −12 (see Figures 15(b), 15(f), and 15(j)), which is similar to the 50-year (Figure 15(c)) and 100-year (Figure 15(i)) scenarios. At the same time, the right bank of Chenshui River has an inundation area in the 50-year and 100-year return periods, but not for the 20-year return period. Significantly, the area was not submerged in the whole process of the one flood in 10 years. More importantly, the floodplain occurred on both banks of Yuanshui River, especially in the main urban area on the right of the river, at t = −6 of the 50-year and 100-year floods. Meanwhile, under the 20-year flood scenario, an overbank phenomenon occurred on the right bank of Chenshui River (see Figures 15(c), 15(g), and 15(k)).

At the peak moment, the flood inundation range and depth reached the maximum as shown in Figures 15(d), 15(h), and 15(l), and after the flood peak, the flood gradually subsided, accompanied by a gradually decreasing water depth. It took almost 72 h for the flood to dissipate, as shown in Figure 16(a). The net inundated area at different frequencies (excluding the original water area of the river, for which Q = 8100 m³/s) is shown in Figure 16(b). As expected, the inundation extents and depths gradually increased with increasing recurrence intervals.

To summarize, there are differences in the evolution of the inundation area under different frequencies. The main inundation area of the 10-year flood occurrence is concentrated on the left bank of Chenshui River, at the intersection with Yuanjiang River. For the 20-year return period, there are new inundation areas on the right-hand side of Yuanjiang River and Chenshui River compared to the 10-year return period. Similarly, compared to 20-year flood, the inundation area of Yuanshui river along the both banks were increased under the conditions of 50-year and 100-year scenario. Consequently, the temporal and spatial characteristics of flood evolution can provide better decision-making for urban flood control management.

5.3 Maximum inundation area

In this section, the flood extents at various return periods were examined to understand different characteristics of inundation situations. The Figure 17 show maximum inundation area over the study area for fluvial flooding at different return periods ranging from 10 years, 20 years, 50 years to 100 years. Consistent with the Section 5.1 analysis, fluvial flooding occurs on the left bank of Chenshui River and the confluence of Chenshui and Yuanjiang at the 10-year return period, owing to the dykes in these places being low or even missing. The floodplain phenomenon appears in the main urban area under the 20-year, 50-year, and 100-year return periods, which mainly occurs around the inner city part of the river, so it is important to strengthen the water level monitoring and control. From Figure 17, a similar inundation pattern is observed for both simulations (50-year and 100-year) in the study area, but with different flood depths. This can largely be attributed to the topographic confinement of the river floodplain. In general, Figure 17 shows the flood inundation areas for different return periods and has a certain guiding value in disaster estimation and personnel deployment.

Figure 17. Maximum flood inundation for 10-year, 20-year, 50-year, and 100-year return periods.

The flow velocity is another index involved in flood hazard assessment. A similar analysis was performed but only the peak flow velocity was considered because velocity tends to occur only during water level rise and at the end of simulation when water is no longer flowing. As shown in Figure 18, in most areas the velocity is small (<0.6 m/s). The regions with velocities greater than 1 m/s are mainly concentrated on the right bank of Chenshui (10-year), the left bank of Chenshui (20-year, 50-year, and 100-year), and the two banks of Yuanjiang (50-year and 100-year). Obviously, the value and distribution range of velocity increase with increasing return period, but the spatial distribution of speed is consistent. As we all know, momentum is the most direct form of speed destruction, so greater attention needs to be paid to vehicle and personnel safety in these areas.

Figure 18. Flow velocity distributions for 10-year, 20-year, 50-year, and 100-year return periods.

5.4 Flood hazard risk map

Section 5.2 analysed the distribution of maximum inundation range and maximum velocity from the perspective of a single variable at each frequency. However, flood disasters in waterlogged urban areas, affecting the safety of pedestrians and vehicles may occur with some combination of flow velocity and depth, rather than simply at the peak of flow velocity or water depth. Therefore, a combination of $D\times V$ based on the flood disaster curve established by Smith et al. (2014) and recommended by AIDR (2017) has been discussed and compared to other methods in a recent review on vehicles stability during floods (Bocanegra et al., 2020), and is regarded as an index to quantify flood disasters. In this method, a hazard classification ranging from H1 to H5 is established for each cell as reported in Costabile et al. (2020), as shown in Figure 19 and Table 3. This classification is helpful for identifying the variability of the spatial resolution of flood disaster parameters. Based on this information, the flood-hazard maps developed for Chenxi City under 10-year, 20-year, 50-year, and 100-year return periods are presented in Figure 20. Simultaneously, Figure 21 illustrates statistical data concerning the inundated area probabilities at each level of flood hazard classification for each recurrence frequency.

Figure 19. Flood hazard curves according to the AIDR (2017) and Costabile et al. (2020) approaches.

Table 3. Flood hazard index classification.

Hazard risk index	Hazard class, $D\times V$ (m^2/s)	Inundation depth (m)	Flow velocity (m/s)	Description
H1	≤0.3	0 ≤ D ≤ 0.3	≤1	Generally safe for people, vehicles and buildings.
H2	≤0.6	0.3 ≤ D ≤ 1.2	≤0.5	Unsafe for vehicles, children and the elderly.
H3	≤1	1.2 ≤ D ≤ 2	≤0.5	Unsafe for people and vehicles.
H4	≤4	2 ≤ D ≤ 4	≤1	Unsafe for people and vehicles. All buildings vulnerable to structural damage.
H5	>4	D > 4	>1	Unsafe for people and vehicles. All building types considered vulnerable to failure.

As shown in the Figure 20, most of the flooded areas were at risk level H1 in each of the return periods (10-year, 20-year, 50-year, and 100-year), and correspondingly, the percentages of the total submerged areas were 75%, 66%, 62% and 61% (see Figure 21), which were distributed in areas far from river channels. H1 indicates relatively safe flood areas, while the other risk levels pose threats to people, vehicles, and buildings. To be more precise, in the 10-year scenario, the risk levels of H2, H3, and H4 are mainly distributed on the left-hand side of Chenshui River, whereas H3 and H4 are mainly distributed along roads in a long narrow shape. The results for the 20-year scenario are similar to those for the 10-year scenario. Although the right bank of Chenshui River is submerged, its risk level is relatively low. The risk distributions for the 50-year and 100-year scenarios are consistent (see Figure 20). But the flood zones in the H2 category and above appear on Chenshui River and Yuanjiang River banks, and especially H4 are relatively large on the left bank of Chenshui River, which means that the impact of flooding on developments in this area cannot be ignored.

In general, the inundation area of each risk category increases with increasing return period. However, the proportion of inundation shows different trends. As shown in Figure 21, the proportion of H1 decreases with the return period increases, while H2 increases first and then decreases, subsequently reaching its peak in the 20-year scenario. This is highly likely because the step distribution of topographic elevation affects the flood evolution process.

The curves presented in Figure 22 show the changes of cell number for each flood hazard classification over time at different frequencies. H1 increases gradually from zero to maximum within the time step 2500–6000, and then decreases. Although the slope is disparate in different return periods, the trend is basically the same, and the process of flooding and retreating is basically symmetric under certain return periods. The law for H2, H3, and H4 is similar to that for H1, but there are local fluctuations in the process of flooding for H2 and H3, even retreating for H2. The main reason is that the flow velocity at certain places will change with the water level in the process of flood evolution, thus the risk level will change accordingly.

6 Discussion

Urban hydrology is sensitive to various natural and man-made factors on multiple spatial scales (Torabi et al., 2020; Pirnia, Darabi et al., 2019; Pirnia, Golshan et al., 2019). Therefore, urban flood risk management requires a detailed understanding of flood evolution (in urban sub-catchments) and hydrological process interactions between urban sub-catchments and adjacent rivers in urban areas (Vercruysse et al., 2019), and coupled hydrological and hydrodynamic models are often used for this purpose. In this process, the accuracy of hydrodynamic boundary conditions, namely flood hydrograph and urban model precision, are key factors in flood risk assessment. Nevertheless, the complexity of catchment geographic information (such as underlying changes and hydraulic engineering) is challenging for the generation of accurate flood hydrographs. In addition, the spatial heterogeneity of rainfall likewise affects the flood characteristics of rivers (Sanches et al., 2019; Zhao et al., 2020; Zhou et al., 2017; Zhu et al., 2018), especially those where tributaries exist, and this is the subject of subsequent work continuing our research on the effects of different regional rainfall and intensity levels on urban flooding. This study has deduced flood process curves under different return periods based on the measured typical flooding process in 2017. On the one hand, it implicitly included the changes in flood hydrographs caused by various factors in the upper reaches of the city, and on the other hand, it effectively reduced the data difficulty in constructing a whole basin model. Comparing the flood inundation of the 10-year flood process derived based on the measured 2017 flood with the actual flood process in 2016 (both with the same peak discharge) shows that the extent of inundation is strongly dependent on the instantaneous flow (see Figure 13), with the peak flow largely determining the maximum inundation extent and risk level. This indicates the effectiveness of using peak flow rate as the main parameter and deriving flood processes based on the observed flood processes at neighbouring stations for risk assessment. As an overland flow driven by gravity, the accuracy of urban flood prediction is strongly dependent on the quality (e.g. accuracy and resolution) of the input data (Leitão & de Sousa, 2018). In this study, high-precision terrain data (1 m × 1 m) are obtained by UAV, which effectively characterize the terrain surface features, such as roads, buildings, and other man-made features.

There are, of course, certain limitations associated with this approach. In this article, the same ratio method is adopted to calculate the flood hydrograph under different return periods. This method only considers the flood peak discharge and ignores other hydrograph characteristics such as hydrograph volume and shape (Mediero et al., 2010), which provides only a limited description of a flood event (see Figure 6(b)). Furthermore, there are many other ways to deduce flood hydrographs, for instance based on runoff observation data and the combination of flood variable estimates and flood shape results in a Synthetic Design Hydrograph (SDH) (Brunner et al., 2017), a multi-characteristic Synthesis Index Method (MSI) (Xiao et al., 2009) as well as the construction of joint flood distributions based on the Copula function (Klein et al., 2010). These methods can effectively reflect the differences in flood hydrographs, which is what we are interested in, and consider the impact of hydrograph shape on flood processes to improve further the method of flood risk forecasting. Additionally, in consideration of computational efficiency, the article failed to consider the three-dimensional effect of water flow, although three-dimensional models are more accurate in complex flow fields (Munoz & Constantinescu, 2018; Rong et al., 2020), or the numerical effects of vegetation on river resistance (Xiao et al., 2022). But these are only minor questions, and the study aimed to provide ideas for flood simulation at the city scale.

Figure 20. Distribution of the risk index for combinations of flow velocity and depth for each return period.

Figure 21. Area statistics for each risk level under different return periods.

Figure 22. Number of cells for each risk level over time.

7 Conclusion

To address the problem of urban-scale fluvial flooding, particularly for rivers in mountainous areas where data is limited, establishing detailed and accurate hydrological models can be challenging. This article proposes an approach to urban flood management based on measured flood data from urban hydrological stations in close proximity to derive flood hydrography for different return periods, and establish a high-precision urban-scale river flood risk management method by combining with UAV aerial survey technology. Taking Chenxi City in the upper reaches of Yuanshui River as an example, the applicability and accuracy of this method are discussed. Comparing the flood inundation of the 10-year flood hydrograph derived based on 2017 flood measurements with the actual flood process in 2016 (both with the same peak discharge, 18,000 m³/s for the 2016 flood process and 17,700 m³/s for the 10-year recurrence period, respectively) shows that the extent of inundation is strongly dependent on the instantaneous discharge, with the peak flow largely determining the maximum inundation extent and risk level, making the different recurrence period level flood flows derived based on historical river observations important for the assessment of urban flooding. This demonstrates the feasibility of risk assessment based on measured flood processes at adjacent sites.

Subsequently, the flood evolution and corresponding risk levels under 10-year, 20-year, 50-year, and 100-year scenarios were analysed. The results reveal that the left bank of Chenshui River and the confluence of Chenshui and Yuanjiang Rivers are flood-prone areas, as well as floodplains appearing on the right banks of Chenshui River and Yuanjiang River in the 20-year, 50-year, and 100-year scenarios, while floodplains on the left bank of Yuanjiang River only occurred in the 50-year and 100-year scenarios. Floodplain area increases with increasing return period, with 4.59, 5.25, 6.11, and 6.57 km² corresponding to the 10-year, 20-year, 50-year, and 100-year scenarios. Meanwhile, combined velocity and depth indexes were used to evaluate the flood process at different frequencies. The results showed that flood risk level H3 (unsafe for people and vehicles) appeared on the left bank of Chenshui River, the right bank of Chenshui River (in the 20-year, 50-year, and 100-year scenarios), and the two banks of Yuanjiang River (in the 50-year and 100-year scenarios).

For rivers having limited availability of fundamental data on such factors as rainfall, evaporation, and DEM, as well as reservoirs and channelization projects, constructing high-precision hydrological models poses significant challenges. Therefore, this study proposes using field measurements of flood processes to derive flood events for different return periods and combining them with unmanned aerial vehicle surveying techniques to establish a high-precision framework for urban-scale flood risk assessment. Compared with whole catchment models, this approach first reduces the difficulty of modelling (whole catchment models require a wide range of high-precision data such as surface data and data on reservoirs, hydraulic structures, etc.). Secondly, typical flood hydrograph measurements are used as the basic data, which retain some process-based information on flood events (such as the influence of man-made buildings and land type changes on flood hydrographs in the upper reaches of the city). This approach provides technical guidance for flood risk assessment in areas where data is lacking (e.g. upstream of cities in mountainous areas).

Acknowledgements

Thanks are due to the four anonymous reviewers for their valuable comments.

Declaration on competing interests

The authors declare that they have no known competing financial interests or personal relationships that could appear to influence the work reported in this article.

Additional information

Funding

This work is sponsored by the National Natural Science Foundation of China [#51839002 and #51879015]; the Natural Science Foundation of Hunan Province, China [#2021JJ20043].