A Framework for Refined Hydrodynamic Model Based on High Resolution Urban Hydrological Unit

Abstract

With the accelerating pace of urbanization, cities are increasingly affected by rainstorm and flood disasters, which pose severe threats to the safety of residents’ lives and property. Existing models are increasingly inadequate in meeting the accuracy requirements for flood simulation in highly urbanized regions. Thus, it is urgent to develop a new method for flood inundation simulation based on high-resolution urban hydrological units. The novelty of the model lies in the novel structure of the high-resolution Urban Hydrological Units model (HRGM), which replaces coarse sub-catchments with a fine-grained network of urban hydrological units. The primary innovation is the node-based coupling strategy, in which the HRGM provides precise overflow hydrographs at drainage inlets as point sources for LISFLOOD-FP, rather than relying on diffuse runoff inputs from larger areas. In this paper, a high-resolution hydraulic model (HRGM) based on urban hydrological units coupled with a 2D hydrodynamic model (LISFLOOD-FP) was constructed and successfully applied in the Chebeichong watershed. Results show that the model’s simulations align well with observed data, achieving a Nash efficiency coefficient above 0.8 under typical rainfall events. Compared with the SWMM model, the simulation results of HRGM were significantly improved and more consistent with measured results. Taking the rainstorm event on 10 August 2021 as an example, the Nash coefficient increased from 0.7 to 0.85, while the peak flow error decreased markedly from 15.8% to 3.1%. It should be emphasized that urban waterlogging distribution is not continuous but appears as patchy, discontinuous, and fragmented patterns due to the segmentation and blocking effects of roads and buildings in urban areas. The framework presented in this study shows potential for application in other regions requiring flood risk assessment at urban agglomeration scales, offering a valuable reference for advancing flood prediction methodologies and disaster mitigation strategies.

1. Introduction

Over the past decade, floods have affected more than two billion people globally and, together with storms and droughts, account for 80–90% of natural disasters worldwide [1]. Currently, one of the major threats facing cities is the escalating risk of flooding, driven by climate change and accelerated urbanization.

Global temperatures have continued to rise under the influence of climate change and human activities. This warming has triggered significant changes in the global hydrological cycle [2,3], including shifts in the spatial and temporal distribution of precipitation [4], an increase in the frequency and intensity of extreme rainfall events [5], and greater variability in river runoff [6,7]. These changes pose severe challenges to urban flood control and drainage infrastructure. Milly et al. [8] noted that, under a changing global hydrological system, traditional methods of estimating hydrological event probabilities—based on stationarity assumptions—have become increasingly inadequate for accurately capturing the long-term evolution of water resources and hydrological processes in non-stationary environments. Therefore, there is a growing urgency to improve the accuracy and predictability of meteorological and flood forecasts to support the construction of drainage infrastructure and the design of emergency management measures. Consequently, there is a pressing need for advanced modeling tools that can accurately simulate flood processes, identify critical risk areas, and ultimately inform decision-making for infrastructure reinforcement, spatial planning, and emergency response.

With the rapid advancement of computer technology, accurate flood modeling and forecasting have become essential for effective flood risk management. In the early stages of urban flood simulation, cities were often treated as single watersheds for runoff calculation, yielding a “point” output—such as a hydrograph at the outlet of a catchment or sub-catchment [9]. However, as urbanization continues and urban flood disasters become more severe, requirements for simulation have shifted from “point” to “surface” —focusing more on the spatiotemporal variation of inundation extent and depth. Traditional hydrological models focused on “point” outputs can no longer meet these demands. As a result, flood forecasting based on hydrodynamic models, or coupled hydrological–hydrodynamic models, has become a major research direction [10,11,12].

Many hydrodynamic-based models are already available for urban flood simulation, such as the MIKE series by the Danish Hydraulic Institute, the STORM model by the U.S. Army Corps of Engineers, the Storm Water Management Model (SWMM) (Version 5.1) by the U.S. Environmental Protection Agency [13], as well as LISFLOOD-FP [14] (Version 7.1) and FloodMap [15] (Version 2011). To leverage the strengths of different models—such as SWMM in pipe network flow routing and LISFLOOD-FP in surface inundation simulation—some researchers have attempted to couple multiple models to improve simulation performance [16]. For example, Zeng and Wu coupled SWMM with LISFLOOD-FP by enabling real-time data exchange to simulate both underground pipe flow and surface flooding in Dongguan. Their results demonstrated that the coupled model can effectively simulate urban flood processes and shows promise for real-time forecasting [5].

In recent years, machine learning (ML) and deep learning (DL) have emerged as powerful, data-driven alternatives to traditional process-based hydrological and hydrodynamic models for flood forecasting and inundation mapping. Unlike physics-based models that rely on explicit numerical representations of hydrological processes, ML models learn complex, non-linear relationships between input variables (e.g., rainfall, topography, land use) and target outputs (e.g., streamflow, inundation depth) directly from historical data [17]. This data-driven paradigm offers several advantages, including rapid computation for real-time forecasting, applicability in data-scarce regions, and often comparable or superior accuracy to conventional methods.

However, many challenges remain in urban flood forecasting and early warning systems. These include the scarcity of distributed flood models suitable for large watersheds with dense river networks and engineering controls; the need for high resolution, multi-source data inputs; and the difficulty of coupling multiple physical processes. Large-scale, high-intensity urbanization has altered regional flow patterns, increasing the risk and uncertainty of flood evolution. The impermeabilization of surfaces, combined with the interplay between underground drainage systems and dense tidal river networks, adds to the complexity of flood dynamics, making accurate prediction extremely difficult.

Improving model accuracy typically requires higher spatial resolution and more detailed parameterization. Parameter calibration must be performed individually for each urban hydrological unit—especially in real-time forecasting, where time-varying parameter adjustment is needed. Moreover, most current urban flood models focus either on underground drainage or coupled surface–subsurface flow [5,18,19]. These models often overlook key factors, such as the regulatory effect of sluice gates in densely river-networked, highly urbanized areas, making them inadequate for flood simulation in urban agglomerations.

Conventional runoff yield and flow concentration models for urban areas mostly adopt an approach that divides the watershed into several sub-catchments. Within these sub-catchments, the underlying surfaces are classified into broad categories such as water surface, pervious areas, and impervious areas [10]. The runoff yield within each sub-catchment and the surface flow concentration processes are then calculated separately. However, this classification method is relatively coarse, and the spatial resolution of the sub-catchments is often low. This makes it difficult to accurately characterize the complex underlying surface features of urban areas, and the approach struggles to adapt to the detailed flood simulation required in highly urbanized regions with complex underlying surfaces [20,21]. In particular, most existing flood models use sub-catchments for surface runoff calculation. These sub-catchments often have low spatial resolution and are delineated somewhat arbitrarily. It is widely recognized that higher spatial resolution yields more accurate results, provided the modeling framework can support such detail.

In summary, for highly urbanized areas with complex underlying surfaces, there is a critical need to develop an integrated modeling framework that couples surface runoff generation, subsurface pipe flow, and two-dimensional hydrodynamic processes. To overcome the limitations of low spatial resolution in current models and improve simulation accuracy, it is urgent to develop new flood inundation methods based on high-resolution urban hydrological units.

Therefore, the objectives of this study are to:

(a)

construct a high-resolution Urban Hydrological Units model (HRGM) based on urban hydrological units and SWMM, coupled with the 2D hydrodynamic model LISFLOOD-FP, to simulate complete flood inundation processes;

(b)

calibrate and validate the coupled model using monitored data, then apply it to investigate simulation performance under typical storm events;

(c)

assess and compare the performance and accuracy of the two model frameworks—HRGM/LISFLOOD-FP and SWMM/LISFLOOD-FP—under different scenarios.

This study is expected to provide a potential approach for improving the accuracy of numerical flood simulation in highly urbanized areas.

2. Study Area and Data

2.1. Study Area

This study selected the Chebeichong Watershed in Guangzhou, China, as a case study to validate the effectiveness of the hydrodynamic model within the high-resolution urban hydrological unit modeling framework. Covering an area of approximately 73.6 km² (Figure 1), the watershed is characterized by a subtropical maritime monsoon climate, with an average annual temperature ranging from 20 to 22 °C and an average annual precipitation of 1720 mm. The area features low-lying, highly urbanized terrain with dense building coverage. Its high susceptibility to waterlogging is attributed to a high percentage of impervious surfaces (accounting for 70% of the area), low-lying topography, and a relatively low design standard of the drainage system. Although the drainage network in the Chebeichong Watershed is well-developed, major junctions are predominantly located in low-lying areas, making them prone to water accumulation and frequently leading to paralysis of the road network. The frequent occurrence of urban inundation in this area already poses a significant challenge to city management, highlighting the urgency of evaluating the role of the drainage system in this region.

2.2. Monitoring Data

The coupled hydrodynamic model was constructed using a 2 m resolution digital elevation model (DEM) obtained from the Guangzhou Land Resources and Planning Commission. Additional data comprised land-use information provided by the National Geographic Center of China and drainage network data sourced from the Guangzhou Water Affairs Bureau. While large watersheds typically employ permanent measuring stations to record rainfall and runoff, flow and rainfall measurements for this urban sewer basin were acquired from an automatic observation station to characterize its hydrological regime. These rainfall data acted as the hydrological forcing input for the High-Resolution Grid Model (HRGM), whereas the flow measurements were utilized for model calibration and evaluation. The rainfall data covering the period from May to August 2021 are presented in Figure 2.

2.3. Data for Model Establishment

Based on high-resolution DEM and underlying surface data, the study area was meshed. In order to improve the computational efficiency of the model, the Chebeichong watershed was discretized into 7813 grid units (Urban Hydrological Units, UHUs) with a spatial resolution of 100 m (Figure 3c). Figure 3a shows the distribution of the underlying surface in the study area, including water area, urban area, forest and grass, farmland, and urban land. In the SWMM model, the study area was divided into 182 sub-catchments (Figure 3b). The original underground drainage network in the Chebeichong Watershed is complex and dense. If it is directly applied to flood model, it will seriously complicate and reduce the computational speed of the model. Based on the Strahler ordering method [22], the original 95,219 pipes and 95,318 nodes were simplified to 604 and 606 (Figure 3b,c). Both the HRGM and SWMM models utilized identical parameters for their respective drainage pipe and river sub-models.

3. Methodology

3.1. Model Framework

The numerical models of urban surface runoff based on the hydrodynamic method include one-dimensional (1D) Saint-Venant equations, two-dimensional (2D) shallow water equations (SWEs), 1D–2D shallow water equations (SWEs), Boussinesq equations, kinematic wave equations, diffusion wave equations, and the porosity model.

Table 1. Summary and comparison of types of numerical simulation models.

Subprogram	Typical Application	Available Numerical Model	Remark
Surface runoff unit	Type of catchment	Sub-catchment (Wu et al. [22]; Wang et al. [23]; Wu et al. [24]) UHUs (Li et al. [25], Present Work)	• Sub-catchment: Random and simple • UHUs: High resolution and easy to use
Surface runoff	Rainfall-runoff yield model	SCS-CN method (Li et al. [25]), Infiltration curve method (Wang et al. [23]; Wu et al. [24]), Runoff coefficient method (Present Work)	• Impervious area: Runoff coefficient method; SCS-CN method • Permeable areas: Runoff coefficient method • Suitable for areas with limited data
	Surface confluence model	Reasoning formula method, (Present Work) Non-linear reservoir method (Wang et al. [23]; Wu et al. [24], Present Work)	• Higher precision: Nonlinear reservoir method
Drainage system	Pipe flow	1D Saint-Venart equations (Li et al. [25], Wang et al. [23]; Wu et al. [24]), Attenuation coefficient method (Present Work)	• Widely used pipe flow calculation model: 1D Saint-Venant equations, Preissmann model
Regional flow interaction	Exchange flow	Orifice equation, Weir equation (Li et al. [25], Wang et al. [23]; Wu et al. [24], Present Work)	• Surface—pipe free exchange flow: Weir equation • Surface—pipe submerged flow: Orifice equation
2D-Surface runoff	Hydrodynamic method	2D SWEs (Li et al. [25]), Kinematic wave (Wang et al. [23]; Wu et al. [24], Present Work)	• 2D SWEs: More complex calculation • Kinematic wave: Has good application prospects

lists the application of these numerical models. The spatial resolution of the urban hydrological units within the study area was not arbitrarily chosen, but rather a deliberate choice based on a balance between computational feasibility and physical representativeness. The Chebeichong watershed was discretized into 7813 grid units (Urban Hydrological Units, UHUs) with a spatial resolution of 100 m (Figure 3c). The model framework, as illustrated in Figure 7, integrates surface runoff generation, 1D channel flow, and 2D floodplain dynamics. The core computational steps are outlined below.

3.2. Watershed Discretization and Flow Path Determination

The watershed was conceptualized through a standard DEM preprocessing procedure:

Depression Filling: All depressions in the Digital Elevation Model (DEM) were filled using the leveling and filling method [26] to ensure continuous flow paths.

For depressions with elevation values lower than the surrounding area, the processing method is to elevate the elevation value of the grid where the depression is located equal to the average elevation value of the adjacent grids around it.

If $Z_i<Z_j$, the elevation can be adjusted using Formula (1).

$$Z_i=mean\left(Z_{j=1}^n\right)$$

(1)

where $Z_i$ is the elevation of Urban Hydrological Unit (UHU) numbered i, m; $Z_j$ is the elevation of the urban hydrological units around i, m. Repeat the above process, iterate through all the entire watershed until the total outlet of the watershed was found (Figure 4).

Flow Direction: The D8 algorithm was employed to determine the single flow direction for each UHU based on the steepest descent principle (Equation (2)). The slope value (S) can be adjusted using Formula (2).

$$S=\arctan {\displaystyle \frac{\left(Z_i-Z_j\right)}{\left|L\right|}}$$

(2)

where $Z_i$ is the elevation of Urban Hydrological Unit (UHU) numbered i, m; $Z_j$ is the elevation of the urban hydrological units around i, m; L is the distance between UHU i and UHU j, m.

Flow Accumulation and Routing Order: The flow accumulation matrix was calculated based on the flow directions to identify upstream–downstream relationships and define the routing order for hydrological computations (Figure 5).

The basic idea is that each grid cell of DEM has a unit of water volume, and the water volume flowing through each grid unit (the total number of all grid units in the upstream) is calculated. The confluence value of grid units in the basin reflects the strength of flow convergence capacity. The larger the value, the more water flows to the grid, and the easier the basin is to generate surface runoff.

3.3. Runoff Generation and Confluence

3.3.1. Infiltration and Runoff Yield

The Horton infiltration model was selected to calculate the infiltration loss for each pervious UHU (Equation (3)).

$$f_i=\left(f_{0_i}-f_{c_i}\right)e^{-kt}+f_{c_i}$$

(3)

where $f_i$ is the infiltration of Urban Hydrological Unit (UHU) numbered i, mm/s; $f_0$ is the initial infiltration rate of Urban Hydrological Unit (UHU) numbered i, mm/s; $f_c$ is the stable infiltration rate of Urban Hydrological Unit (UHU) numbered i, mm/s; k is a constant.

The flow of the unit was calculated using Formula (4).

$$r_i=\left(I_i-f_{c_i}\right)\cdot M_i\cdot \mathrm{\Delta }t\cdot {\lambda }_i$$

(4)

where $r_i$ is the flow of Urban Hydrological Unit (UHU) numbered i; $I_i$ is the rainfall of Urban Hydrological Unit (UHU) numbered i mm/s; ${\lambda }_i$ is the coefficient of imperviousness numbered i, $M_i$ is the area of Urban Hydrological Unit (UHU) numbered i, m².

The flow of the grid unit numbered can be calculated by Formula (5).

$$Q_{i in}^{t+\mathrm{\Delta }t}=Q_j{\epsilon }_j+R_i$$

(5)

where $Q_j$ is the flow from grid unit numbered j to grid unit numbered i, m³/s; ${\epsilon }_j$ is the flow attenuation coefficient of grid unit numbered j, $R_i$ is the yield of grid unit numbered i within the calculation step, m³/s and $Q_{i in}^{t+\mathrm{\Delta }t}$ is the flow of grid unit numbered i at time of t + Δt.

3.3.2. Overland Flow Routing

The overland flow from one UHU to its downstream neighbor was routed using a kinematic wave approach. The flow from grid unit numbered i to grid unit numbered i + 1 downstream can be calculated by Formulas (6)–(10).

$$t_i^{slope}=1.45{\left({\displaystyle \frac{{\alpha }_i{L}_i}{\sqrt{S_i}}}\right)}^{0.467}$$

(6)

$$Q_{in}^{t+\mathrm{\Delta }t}={\displaystyle \frac{1}{n_i}}{W}_i\cdot d_1\cdot R_i^{\frac{2}{3}}\cdot S_i^{\frac{1}{2}}$$

(7)

$${\displaystyle \frac{d_2-d_1}{\mathrm{\Delta }t}}=\frac{(i_i^t-f_{c_i})+(i_i^{t+\mathrm{\Delta }t}-f_{c_i})}{2}-{\displaystyle \frac{1.49W_i}{A_i\cdot n_i}}{S}_i^{\frac{1}{2}}{\left[{\displaystyle \frac{d_2+d_1}{2}}-d_{p_i}\right]}^{\frac{5}{3}}$$

(8)

$$d={\displaystyle \frac{d_1+d_2}{2}}$$

(9)

$$Q_{i,out}^{t+\mathrm{\Delta }t}={\displaystyle \frac{1}{n_i}}{W}_i\cdot d\cdot R_i^{\frac{2}{3}}\cdot S_i^{\frac{1}{2}}$$

(10)

where $W_i$ is the grid width of the grid unit numbered i, m; $R_i$ is the hydraulic radius of grid unit numbered i, m; $S_i$ is the slope of grid unit numbered i, $n_i$ is the manning coefficient of grid unit numbered i, $f_{c_i}$ is the stable infiltration rate of grid unit numbered i, $i_i^t$ is the rainfall intensity at time t of grid unit numbered i, $i_i^{t+\mathrm{\Delta }t}$ is the rainfall intensity at time t + Δt of grid unit numbered i, and $d_{p_i}$ is the depression depth of grid unit numbered i, m; $d$ is the water depth of grid unit numbered i, m; which is an intermediate variable.

3.4. Surface-Drainage System Coupling

This model considers street inlets (grate inlets) as the primary linkage between the surface and the underground drainage network, rather than manholes.

Surface to Pipe Flow: When the water surface elevation in a UHU exceeds the elevation of the inlet, water flows into the drainage network. The actual inflow rate Q_inlet is the lesser of the hydraulic capacity calculated by a weir equation (Equation (11)) and the inlet’s maximum intake capacity Q_max (Equation (12)).

$$q_i=\omega A\left(Z_{g_i}-Z_p\right)$$

(11)

where $Z_{g_i}$ is the water depth elevation of grid unit numbered i, m; $Z_p$ is the water surface elevation of the rainwater well, m; $\omega$ is the flow contraction coefficient of rainwater well, and $A$ is the area of rainwater well, m²; $q_i$ is the flow of grid unit numbered i, m³/s.

$$q_{\max }=A{H}_t$$

(12)

where $H_t$ is the depth of the rainwater well, m and $q_{\max }$ is the maximum collection flow of the rainwater well, m³/s.

If $q_i<q_{\max }$, the flow of the grid unit will all flow into the rainwater well, so the actual flow of the grid unit into the rainwater well can be calculated by Formula (13).

$$q_k=q_i$$

(13)

If $q_i\ge q_{\max }$, the flow part of the grid cell flows into the rainwater well, and the rest flows to the next grid cell according to the confluence trajectory. The flow into the rainwater well can be calculated by Formula (14).

$$q_k=q_{\max }$$

(14)

Pipe to Surface Flow (Surcharge): When the water level in the drainage pipe (at the inlet) exceeds the ground surface elevation, it discharges back to the surface (Figure 6b). This overflow Q_overflow is also calculated using a weir equation (Equation (15)).

$$q_l=A\left(Z_{p_i}-Z_g\right)$$

(15)

where $q_l$ is the overflow flow of the rainwater well, m³/s; $Z_{g_i}$ is the water depth elevation numbered i as the grid unit, m and $Z_p$ is the water surface elevation of the rainwater well, m.

3.5. Drainage Network Hydraulics

A simplified hydraulic routing method was adopted for efficient computation within the pipe network. The flow in a pipe segment is calculated using a joint solution of Formulas (16) and (17) that considers the inflow, the pipe’s characteristics, and a friction loss coefficient.

In the initial calculation, the flow in the pipeline is the flow from the grid unit to the rainwater well plus the flow existing in the pipeline in the last calculation step with the gradual evolution of the calculation.

$$P_1=\mu B\sqrt{2g{H}_1}$$

(16)

$$P_2=\theta P_1$$

(17)

where $P_1$ is the initial flow of the pipeline, m³/s; $\mu$ is the flow coefficient of the pipeline, $B$ is the cross-sectional area of the pipeline, m²; $H_1$ is the water depth of the rainwater well, m; $P_2$ is the discharge of the pipeline to the downstream sump, m³/s; $\theta$ is the loss coefficient of water flow in the pipeline.

3.6. Model Coupling: HRGM and LISFLOOD-FP

The model couples a 1D drainage network model (HRGM) with the 2D hydrodynamic model LISFLOOD-FP [12] to simulate integrated urban flooding.

1D Channel Flow (Pipes/Channels): Solved using the 1D Saint-Venant equations (Equations (18) and (19)).

$${\displaystyle \frac{\partial Q}{\partial x}}+{\displaystyle \frac{\partial A}{\partial t}}=q$$

(18)

$$S_0-{\displaystyle \frac{n^2{P}^{\raisebox{1ex}{$4$}\!\left/ \!\raisebox{-1ex}{$3$}\right.}{Q}^2}{A^{\raisebox{1ex}{$10$}\!\left/ \!\raisebox{-1ex}{$3$}\right.}}}-\left[{\displaystyle \frac{\partial h}{\partial x}}\right]=0$$

(19)

where $Q$ is the volumetric flow rate in the channel, m³/s; $A$ is the cross sectional area of the flow, m²; $q$ is the flow into the channel from other sources (i.e., from the floodplain or possibly tributary channels), m³/s; $S_0$ is the down-slope of the bed, $n$ is the Manning’s coefficient of friction, $P$ is the wetted perimeter of the flow, m, and $h$ is the flow depth, m. For problems with no channels present, this function can simply be switched off.

2D Floodplain Flow: Solved using a simplified form of the 2D shallow water equations on a raster grid. The flow between two surface grid cells is a function of the free surface height difference (Equations (20) and (21)). This can be switched on or off in the model to enable both kinematic and diffusion wave approximations to be tested. Equations (18) and (19) are discretized using finite differences and a fully implicit scheme for the time dependence.

$${\displaystyle \frac{d{h}^{i,j}}{dt}}={\displaystyle \frac{Q_x^{i-1,j}-Q_x^{i,j}+Q_y^{i-1,j}-Q_y^{i,j}}{\mathrm{\Delta }x\mathrm{\Delta }y}}$$

(20)

$$Q_x^{i,j}={\displaystyle \frac{h_{flow}^{\raisebox{1ex}{$5$}\!\left/ \!\raisebox{-1ex}{$3$}\right.}}{n}}\left({\displaystyle \frac{h^{i-1,j}-h^{i,j}}{\mathrm{\Delta }x}}\right)\mathrm{\Delta }y$$

(21)

where $h^{i,j}$ is the water free surface height at the node (i, j), m; $\mathrm{\Delta }x$ and $\mathrm{\Delta }y$ are the cell dimensions, $n$ is the effective grid scale Manning’s friction coefficient for the floodplain, and $Q_x$ and $Q_y$ describe the volumetric flow rates between floodplain cells, m³/s. $Q_y$ is defined analogously to $Q_x$. is the flow depth, $h_{flow}$ represents the depth through which water can flow between two cells, m³/s, and is defined as the difference between the highest water free surface in the two cells and the highest bed elevation (this definition has been found to give sensible results for both wetting cells and for flows linking floodplain and channel cells).

The HRGM model specializes in simulating the hydrodynamics in rainwater piping systems (1D). It is a model of open-source management, and secondary development can be conducted in this model. Therefore, this function was also employed to couple the HRGM with the LISFLOOD-FP model.

It is capable of yielding good predictions of maximum inundation extent for flooding problems [27,28,29]. The main advantage of LISFLOOD-FP is that the model can be integrated with high resolution DEMs that are increasingly available for floodplain areas recently. It is recognized as a simpler 2D model in comparison to most other 2D models [14].

The models were connected by exchanging data at shared nodes. A timing synchronization technique ensured simultaneous discharge information transfer between them. At each manhole, the interacting discharge was calculated using either the Weir or orifice equation, based on the hydraulic heads within the manhole and the adjacent aboveground water surface. The coupling procedure implemented on the C# platform consisted of the following steps: (a) extracting overflow data from the HRGM result file (.out) to generate a boundary file (.bdy); (b) creating the overflow point coordinate file (.bci) and the Digital Elevation Model file (.asc); (c) preparing a two-dimensional simulation run file (*.par) that specified parameters such as simulation duration, time step, wet-dry boundaries, water surface slope threshold, and default values; and (d) executing both models to produce the output results. A flowchart of the coupling process is provided in Figure 7 of the Appendix A. Further details regarding the model coupling methodology are available in prior publications [27,28,29].

4. Results

4.1. Model Calibration and Validation Results

We used the rainstorm events of 29 May 2021 and 2 June 2021 to calibrate the developed model (HRGM), and the 10 August 2021 rainstorm event to verify the results (Figure 2). The optimized parameter values obtained from each calibration exercise are summarized in

Table 2. Results of parameter calibration for HRGM/LISFLOOD-FP model.

Parameter		Description	Type of UHU	Recommended Perturbation Range	Value
Runoff generation	$\lambda$	The coefficient of imperviousness [-]	Water/River	0	0
			Grass/Woodland	0–1	0.2
			Urban land	0–1	0.75
			Farmland	0–1	0.35
	$f_c$	Stable infiltration rate [mm·s−1]	Water/River	0–0.001	0.001
			Grass/Woodland	0.0002–0.089	0.015
			Urban land	0.0001–0.004	0.001
			Farmland	0.006–0.09	0.02
Runoff concentration	$\alpha$	Surface roughness value [-]	Water/River	0–1	1
			Grass/Woodland	0–1	0.66
			Urban land	0–1	0.90
			Farmland	0–1	0.75
	$n$	Manning’s n value [-]	Water/River	0.01–0.017	0.01
			Grass/Woodland	0.03–0.05	0.03
			Urban land	0.01–0.025	0.01
			Farmland	0.018–0.025	0.02
	$d_p$	Water storage value [$mm$]	Water/River	0	0
			Grass/Woodland	10–20	15
			Urban land	10–50	20
			Farmland	10–50	10
	$\epsilon$	Flow attenuation value [-]	Water/River	0–1	1
			Grass/Woodland	0–1	0.65
			Urban land	0–1	0.88
			Farmland	0–1	0.72
Flow routing	$\mu$	The coefficient of pipe flow [-]	/	/	0.8
	$\theta$	Pipe roughness [-]	/	/	0.92
	$\omega$	The coefficient of junction flow [-]	/	/	0.95
	$H_t$	Maximum depth in junction [m]	/	0.5–2.5	2

. For the convenience of subsequent comparison between these two models, the parameter calibration results of SWMM are summarized in

Table 3. Results of parameter calibration for SWMM/LISFLOOD-FP model.

Parameter	Description	Recommended Perturbation Range	Value
Rough-R	Roughness of River	0.010–0.14	0.014
Rough-P	Roughness of Pipes	0.010–0.14	0.014
N-Imperv	Manning’ value of impermeable area	0.005–0.05	0.015
N-Perv	Manning’ value of permeable area	0.05–0.5	0.2
S-Imperv/mm	Storage capacity value impermeable area	1–20	2
S-Perv/mm	Storage capacity value impermeable area	1–50	5
MaxRate/(mm/h)	Maximum infiltration rate	80–150	90
MinRate/(mm/h)	Minimum infiltration rate	1–50	7.3
Decay	Coefficient of attenuation	1–10	5.3
Kwidth	The characteristic width	0.2–5	4.8

.

Through parameter calibration, the simulation results at the observation sections are presented in Figure 8. The simulated hydrograph shows relatively good agreement with the observed flow process, effectively capturing the overall trend, including the onset and recession phases of the flood event.

To evaluate the accuracy of the developed model, the Nash–Sutcliffe efficiency coefficient and peak flow error and Root Mean Square Error were calculated using Equations (22)–(24), respectively. The calibration was performed using a manual trial-and-error approach. This choice was made due to the complex interactions between model parameters and the need for iterative evaluation against the multiple objectives mentioned above. The parameter sets for both models were derived through this parallel, iterative process until a satisfactory and comparable fit was achieved for the three objective functions across the calibration events.

$$E_{NS}=1-{\displaystyle \frac{{\Sigma }_{t=1}^N(Q_{t, obs}-Q_{t, sim}{)}^2}{{\Sigma }_{t=1}^N(Q_{t, obs}-\overline{Q_{obs}}{)}^2}}$$

(22)

$$E_{PR}={\displaystyle \frac{\left|Q_{p, obs}-Q_{p, sim}\right|}{Q_{p, obs}}}\times 100\%$$

(23)

$$RMSE=\sqrt{{\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n(Q_{t, obs}-Q_{t, sim}{)}^2}}$$

(24)

On the whole, the coupled model developed for the study area yielded reasonably accurate results. During the calibration period, the Nash–Sutcliffe efficiency coefficients were 0.812 and 0.831, and the peak flow error remained below 10% (

Table 4. Error statistics for simulation results.

	Rainstorm Event	Time (h)	Rainfall (mm)	NSE (-)	PRE (%)	RMSE (-)
Calibration	29 May 2021	24	60	0.812	7.3	2.14
	2 June 2021	28	52	0.831	5.4	1.50
Verification	10 August 2021	28	55	0.858	3.1	2.21

Note: NSE is the Nash–Sutcliffe coefficient; PRE is the relative peak flow error; RMSE is Root Mean Square Error.

). In the validation period, the model achieved a Nash–Sutcliffe coefficient of 0.858 with a peak flow error of 3.1%. It is worth noting that during calibration, the simulated peak flow was slightly lower than the observed value. This discrepancy may be attributed to blockages in the drainage networks during actual operation, leading to a systematic deviation in the simulation results.

Due to the lack of accurate flood measurement data and limitations in the quality of available flood records, we relied on historical waterlogging point data provided by the municipal water department [11,24]. The coupled model demonstrated a satisfactory ability to simulate the waterlogging caused by this rainstorm event. As shown in Figure 9, most of the historical waterlogging points released by the Guangzhou Water Affairs Bureau (Figure 8a) were accurately situated within the simulated inundated areas. Therefore, the coupled model developed for the study area can be considered reliable and reasonable and was subsequently applied in the following analysis.

4.2. Overall Simulation Results for HRGM/LISFLOOD-FP Model Under Typical Rainstorm Event

To further evaluate the model performance, the Peak Time Error (PTE) and Relative Runoff Error (RRE) were selected as assessment indicators. As summarized in

Table 5. Summary of flow rate results for HRGM/LISFLOOD-FP model.

Rainstorm Event	${\mathit{Q}}_{\mathit{p}\mathit{s}}$ (m3/s)	${\mathit{Q}}_{\mathit{p}\mathit{o}}$ (m3/s)	${\mathit{Q}}_{\mathit{t}\mathit{s}}$ (m3)	${\mathit{Q}}_{\mathit{t}\mathit{o}}$ (m3)	PTE (min)	RRE (%)
29 May 2021	26.31	31.61	149.94	154.68	0	−3.0
2 June 2021	41.54	43.93	201.87	198.10	0	1.9
10 August 2021	41.23	42.53	298.14	275.10	0	8.4

Note: $Q_{ps}$ is the simulated peak flow rate; $Q_{po}$ is the observed peak flow rate; $Q_{ts}$ is the simulated total flow volume; $Q_{to}$ is the observed total flow volume; PTE is the peak time error between simulated and observed data; RRE is the relative flow volume error.

, the simulated peak flow consistently coincided with the measured peak flow in terms of timing. In addition, the relative runoff error (RRE) was below 10% for all events, further confirming the model’s good accuracy.

In general, different inundation depths have varying impacts on urban transportation and daily activities. Inundation depths exceeding 1 m tend to cause severe disruption in urban areas. Based on the analysis of inundation areas across different water depths (Figure 8 and

Table 6. Summary of simulation inundation results for HRGM/LISFLOOD-FP model.

Rainstorm Events	Area of Index Depth (km2)
	<0.5 (m)	0.5–1.0 (m)	1.0–1.5 (m)	>1.5 (m)	Total
29 May 2021	0.77	1.01	0.80	0.26	2.84
2 June 2021	1.21	0.91	0.32	0.03	2.47
10 August 2021	0.88	1.11	0.47	0.05	2.51

), most of the flooded areas exhibited water levels below 1 m. With the exception of the event on 29 May 2021, the proportion of the inundated area with water depth over 1 m accounted for less than 20% of the total flooded area. Moreover, the spatial distribution of flooding (Figure 8) shows that submergence was primarily concentrated in the downstream portion of the basin—a highly urbanized region.

4.3. Overall Simulation Results Comparison Under Typical Rainstorm Event

As shown in Figure 10, the differences in results between the HRGM and SWMM models primarily stem from their distinct approaches to runoff generation and concentration. The SWMM model uses delineated sub-catchments to calculate surface runoff, whereas the HRGM model employs urban hydrological units to simulate the surface runoff process. Evidently, the runoff process simulated by the HRGM model aligns more closely with measured values under typical rainstorm events (Figure 10), and it accurately captures the rising and falling limbs of the hydrograph.

Table 7. Summary of flow rate results under typical rainstorm event.

Rainstorm Event	NSE (-)		PRE (%)		PTE (min)		RMSE (-)
	HRGM	SWMM	HRGM	SWMM	HRGM	SWMM	HRGM	SWMM
29 May 2021	0.81	0.73	7.3	13.6	0	0	2.14	2.18
2 June 2021	0.83	0.78	5.4	14.7	0	−5	1.50	1.94
10 August 2021	0.85	0.70	3.1	15.8	0	0	2.21	3.03

Note: NSE is the Nash–Sutcliffe coefficient; PRE is the relative peak flow error; PTE is the peak flow time error between simulated and observed data; RMSE is Root Mean Square Error.

summarizes the simulation results of the HRGM and SWMM models for the typical rainstorm events on 20210529, 20210602, and 20210810, including the NSE, PRE, and PTE values. The NSE values of the two models are 0.81 and 0.73 for event 20210529, 0.83 and 0.78 for event 20210602, and 0.85 and 0.70 for event 20210810. Regarding PRE, the results obtained from HRGM are lower than those from the SWMM model. Notably, HRGM consistently shows higher accuracy than SWMM in terms of peak time error. In all rainfall events, the peak flow error of the HRGM model was less than 10%. Taking the rainstorm event on 10 August 2021 as an example, the Nash coefficient increased from 0.7 to 0.85, while the peak flow error rate decreased significantly from 15.8% to 3.1%. This indicates that, compared with SWMM, the HRGM model demonstrates superior accuracy in flood prediction.

Figure 11 presents the simulated inundation depths from the HRGM/LISFLOOD-FP and SWMM/LISFLOOD-FP models during typical rainstorm events, while

Table 8. Comparison of inundation area between HRGM/LISFLOOD-FP and SWMM/LISFLOOD-FP model under different rainstorm events.

Rainstorm Events	Inundation Area by HRGM/LISFLOOD-FP (km2)					Inundation Area by SWMM/LISFLOOD-FP (km2)
	<0.5 (m)	0.5–1.0 (m)	1.0–1.5 (m)	>1.5 (m)	Total (km2)	<0.5 (m)	0.5–1.0 (m)	1.0–1.5 (m)	>1.5 (m)	Total (km2)
29 May 2021	0.77	1.01	0.80	0.26	2.84	5.23	4.61	2.0	0.2	12.04
2 June 2021	1.21	0.91	0.32	0.03	2.47	5.73	4.38	1.34	0.15	11.6
10 August 2021	0.88	1.11	0.47	0.05	2.51	6.51	4.51	1.68	0.19	12.89

summarizes the corresponding inundation areas across different water depth ranges. As shown in Table 8, most inundated areas have water depths below 0.5 m, with regions exceeding 1.5 m accounting for only a small proportion of the total flooded area. Both models consistently reflect this distribution pattern. It should be noted that although the predicted waterlogging areas from both models generally correspond with the recorded waterlogging points, the HRGM/LISFLOOD-FP model yields more precise predictions. The simulated inundation extents show nearly perfect alignment with the observed waterlogging locations, as evidenced in Figure 11a–c.

To further evaluate the performance advantages of the two models in predicting waterlogging depth, we analyzed the correlation between their predicted waterlogging points. Figure 12 compares the inundation depths simulated by the HRGM/LISFLOOD-FP and SWMM/LISFLOOD-FP models for each unit during the events of 20210529, 20210602, and 20210810. The correlation coefficients (R²) between the two models for these events were 0.15, 0.21, and 0.12, respectively. The correlation plots indicate that the predictions of the two models show relatively high consistency in areas where the simulated inundation depth is below 0.5 m. However, in regions with water depths exceeding 1.5 m, the HRGM/LISFLOOD-FP model demonstrates better performance.

The inundation depth and extent maps generated in this study can be directly used to update the city’s flood risk maps and optimize emergency response plans. For instance, the simulations identified specific waterlogging points with depths surpassing 1.5 m, which are likely to cause traffic disruptions and safety hazards. Such locations should be designated as high-priority control sites, where measures such as pre-positioning mobile pumps and adjusting evacuation routes can be implemented to allocate resources effectively to the highest-risk areas.

The waterlogging points predicted by the HRGM/LISFLOOD-FP model are more dispersed, fragmented, and discontinuous, which aligns closely with the actual distribution patterns of urban waterlogging. This fragmented pattern arises due to the obstruction effects of buildings and the segmentation by road networks, resulting in a decentralized and scattered distribution of waterlogged areas.

5. Discussion

Existing flood models, such as those utilizing runoff generation and flow algorithms like SCS and SWMM, primarily depend on meticulously delineated sub-catchments [25]. The method of sub-catchment delineation and their scale significantly influence model precision. Furthermore, within the current research scope, sub-catchment division has not yet been standardized, often relying heavily on researchers’ personal experience, which introduces randomness and uncertainty into the division process. The developed flood inundation model demonstrates robust performance in simulating both the spatial extent and temporal dynamics of flooding events under varying hydrological conditions. However, in highly urbanized areas, this coarse and arbitrary division method is clearly inadequate to meet the accuracy standards required for reliable modeling.

Based on urban hydrological units, parameters for different unit types were calibrated to develop a model suitable for highly urbanized areas, which was subsequently validated under typical storm events. Compared with observed flood maps from historical events, the model achieves satisfactory accuracy (NSE > 0.8, PRE < 10%), though some discrepancies are noted in urban areas with dense infrastructure. These errors likely stem from simplified representations of drainage systems and neglected micro-topographic features in the DEM data—a limitation also reported by Wu et al. [22] in similar contexts.

The evaluation results of this study align with existing research conclusions and the actual conditions of the Chebeichong basin. For example, during the event of 10 June 2021, at locations such as Longdong West Street, No. 381 Changxing Road, the intersection of Xingtang Street and Daguan Road, Chebei North Road, Dongpu Second Road, and the Chebei Olympic Sports Vehicle Tunnel in the lower reaches of Chebeichong, a comparison with waterlogging points identified by the Guangzhou municipal water authority shows that concentrated overflow points correspond to actual waterlogging sites. These areas were correctly identified as high-risk zones in this study.

Accuracy analysis of the two models indicates that model results are highly dependent on spatial resolution, consistent with the findings of Hsu et al. [30], who emphasized the necessity of high-resolution topographic data in urban flood modeling. It is worth noting that when spatial resolution is improved, model performance increases significantly (Figure 11), demonstrating its application potential in urban flood simulation and suggesting suitability for operational flood early warning systems.

When constructing an efficient and high-resolution urban flood model, the complexity of urban terrain makes high-quality terrain data pivotal for ensuring simulation accuracy. Low-resolution terrain data fail to accurately capture topographic variations and key structural features such as buildings, inevitably leading to deviations between simulated results and actual conditions. For instance, in a low-resolution Digital Elevation Model (DEM), features like walls that block water flow or minor diversion channels may be overlooked. Hsu et al. [30] explored the impact of DEM resolution on flood inundation simulations using the Sanyi watershed in Tainan City as a case study, finding that assessments of inundation areas tend to increase when coarser DEMs are used. Wang et al. [24] employed 1 m resolution DEM and Digital Surface Model (DSM) data (from 2016) provided by the UK Environment Agency, with a vertical accuracy of ±0.15 m, to develop an urban inundation model. Evidently, spatial resolution is crucial for generating reliable flood simulation results.

In high-density urban areas, accurately characterizing the guiding and blocking effects of roads and building complexes on surface flow using a 100 m grid remains challenging. Research has shown that while high-resolution data can enhance validation effectiveness, it is not a panacea [24]. Model performance is also influenced by data limitations and structural uncertainties. The primary constraint of limited validation data and its impact on the certainty of our spatial predictions. The potential effects of UHU generalization on masking fine-scale drainage features. The accuracy of the model heavily depends on the precision of input data, such as digital elevation models and land-use data. For example, DEM resampling operations can significantly affect flood depth predictions.

Although this study has validated the model in typical storm events by introducing urban hydrological response units and coupling with a two-dimensional hydrodynamic module—showing good accuracy and adaptability—some limitations remain. First, the model assumes that all drainage networks are fully unobstructed and does not account for potential blockages or damage, which diverges from real-world conditions. Second, due to the lack of accurately measured inundation depth data, some parameters in the two-dimensional hydrodynamic module could not be quantitatively studied or precisely calibrated. The primary constraint on a fully quantitative, pixel-by-pixel skill-score analysis is the nature of our validation data—namely, the historical waterlogging points which provide only point-location information, rather than continuous observed inundation depth or extent maps. Consequently, while our comparison of total inundated area and spatial patterns between HRGM and SWMM is informative, it remains necessarily descriptive and qualitative at this stage. Obtaining high-resolution, spatially continuous flood observations (e.g., from remote sensing or detailed post-event surveys) to enable rigorous quantitative map-based validation is a critical and high-priority objective for our future research.

Despite these limitations, the model provides valuable insights for flood risk management, especially offering a reliable basis for refined waterlogging simulation in highly urbanized regions. With the rapid development of machine learning technology, the application of deep learning methods in urban waterlogging simulation shows broad prospects.

6. Conclusions

To accurately simulate waterlogging in highly urbanized areas, a high-resolution urban hydrological model (HRGM) was developed by integrating the concept of Urban Hydrological Units with the Storm Water Management Model (SWMM) and coupling it with the 2D hydrodynamic model LISFLOOD-FP. Both coupled models were validated in the Chebeichong watershed. Their simulations showed good agreement with measured data, particularly in replicating flood hydrographs, confirming the models’ accuracy and reliability for urban inundation scenarios. The main findings of this study are as follows:

(a)

The developed high-resolution Urban Hydrological Units model (HRGM) was successfully applied in the Chebeichong watershed. Its simulation results align well with observations, achieving Nash–Sutcliffe efficiency coefficients above 0.8 for typical rainfall events.

(b)

In terms of accuracy, HRGM showed significant improvement over the standalone SWMM model and agreed better with measurements. For example, during the storm event on 10 August 2021, the Nash–Sutcliffe efficiency coefficient increased from 0.7 to 0.85, and the peak flow error decreased markedly from 15.8% to 3.1%.

(c)

The waterlogging patches simulated by the coupled HRGM/LISFLOOD-FP framework are more dispersed, fragmented, and discontinuous. This pattern aligns well with the actual characteristics of urban waterlogging, which often exhibits a discrete and scattered distribution due to the obstruction of buildings and the segmentation by roads.

Furthermore, collecting more comprehensive field-measured water depth data during flood events would greatly enhance the calibration of the model’s key parameters, leading to more accurate and reliable simulations.