Flood simulation using the hydrological model and the hydrological–hydrodynamic coupling model in a small watershed in semi-arid and sub-humid region, North China

With global warming in recent decades, the frequency and intensity of rainstorms have increased significantly worldwide, causing a large number of flood disasters. The southeastern Australian states of New South Wales and Queensland experienced their worst flooding in 50 years in late March 2021. From the end of March to the beginning of April 2021, Indonesia suffered from flash floods and mudslides due to continuous heavy rainfall, at least 86 people were killed and 71 people were missing. The extreme rainstorm on 20 July 2021, in Zhengzhou City, China, caused huge casualties and economic losses (Alfieri et al. 2017; Davenport et al. 2021; Yu et al. 2022). The need to manage available water resources is felt more than ever to reduce harm to humanity (Ngakan et al. 2022). Flood simulation is always an important means of flood control and disaster mitigation, which plays a very important role in flood forecasting, early warning, and flood risk assessment (Walega & Ksiazek 2016).

Although the main flood generation mechanisms have been recognized through field observation and some hydrological models have been proposed, the low flood simulation accuracy is still the main problem at present (Teng et al. 2017; Khosravi et al. 2018). Especially in small mountainous watersheds, flood simulation is hard work due to its complex topography and geomorphological conditions. The flood hydrograph has the characteristics of short duration, high peak discharge, and sudden rise and fall. Once a flood event occurs, it will bring serious consequences. Therefore, it is of great significance to deeply explore the flood formation mechanism in small mountainous watersheds, and then improve the flood simulation accuracy (Kusumastuti et al. 2016; Janabi et al. 2021). The existing runoff prediction models can be divided into two categories: process-driven and data-driven. Compared with the process-driven model, the data-driven model is highly operable. It does not need to consider the physical mechanism of runoff generation, only needs to conduct mathematical analysis on the time series, and establish the functional relationship between the input variables and the output variables. Due to factors such as human activities and climate change, runoff sequences exhibit complex nonlinearity and nonstationary characteristics, which increases the difficulty of runoff forecasting. In order to solve this problem, nonlinear data-driven models have been proposed one after another. Liu et al. (2023) constructed the hybrid forecasting model by combining ESMD and NNBR. The results showed that this method can improve the performance of the forecasting model. With the rapid development of computer technology, the long short-term memory neural network (LSTM) based on deep learning theory has received extensive research. Sun et al. (2022) proposed a variational mode decomposition (VMD) and sparrow search algorithm (SSA) coupled with LSTM to improve the accuracy of monthly runoff prediction and established a monthly runoff prediction model (VMD-SSA-LSTM). In recent years, machine learning has made significant progress in the hydrological field due to its powerful learning ability, and has been widely applied in runoff prediction work. Adnan et al. (2021) proposed the ELM-PSOGWO method for predicting monthly runoff, which provides more accurate results than the standalone ELM, a hybrid of ELM-PSO, and binary hybrid PSOGSA (hybrid of PSO with gravitational search algorithm) methods. Subsequently, Adnan et al. (2022) also proposed a GBO-ANFIS mixture model and compared its performance with that of the independent ANFIS model. The results showed that the method successfully improved the accuracy of runoff prediction. Guo et al. (2022) proposed a stepwise decomposition ensemble (VMD-SVM-KDE) model combining the variable mode decomposition method, the support vector machine model, and the kernel density estimation method, which performs both point prediction and interval prediction. Wei (2021) proposed a computational fluid dynamics-based dynamic numerical simulation technique for competitive aerobics to improve the accuracy of numerical simulation. The experimental results indicated that the methods have more significant practical application value.

The process-driven model based on the hydrological model has also been widely used. The advantage of this method is that it can describe the key mechanism of the hydrological process, and restore the whole process of rainfall transforming into watershed outlet runoff to a certain extent. According to the structure, the hydrological model can be divided into lumped, semi-distributed, and distributed hydrological models (Singh & Woolhiser 2002; Wu & Liu 2002; Kour et al. 2016). Among them, the semi-distributed HEC-HMS hydrological model developed by the U.S. Army Hydrological Engineering Center (USACE 2015) has a relatively simple model structure and less measured data requirements (Liu et al. 2013). It has been widely used in different climate regions at a global scale. Kazezylmaz-Alhan et al. (2020) used the HEC-HMS model to simulate the flood event that occurred in the Ayamama River Watershed in Istanbul, Turkey. The simulated flood hydrographs by the HEC-HMS model are in good agreement with the results obtained by the rational method. Cao et al. (2020) constructed the HEC-HMS hydrological model in the Xihe River Watershed in Liaoning Province, China. In order to explore the applicability of different runoff generation and flow concentration methods, two different calculation schemes were used to simulate floods in the watershed. The results showed that the infiltration excess runoff generation scheme is more applicable in this watershed. Cheng et al. (2018) used the HEC-HMS model to simulate floods in the watershed in the upper reaches of the Beichuan River in Shanxi Province, China. The results showed that the HEC-HMS model had good applicability in the hilly areas of semi-arid regions. However, the distributed hydrological model has strict requirements for hydrological, topographic, geomorphological and other data, and involves a large number of differential equation solving, which makes the model complex and inefficient. In the actual flood forecasting system, the lack of detailed DEM, rainfall, and other data often leads to the reduction of model calculation accuracy. The difficulty in obtaining and integrating data also directly limits the development of distributed hydrological models. Simplifying the structure of the distributed hydrological model and reducing the data requirements of the hydrological–hydrodynamic coupling model become the way to solve this problem.

For a watershed, the runoff generation and concentration is a complex process, there may be some limitations in using a single hydrological model or hydrodynamic model, and sometimes the key hydrological elements in the watershed cannot be reflected realistically (Rui 1997). The hydrological–hydrodynamic coupling model can link the hydrological elements with the hydrodynamic elements and can solve the more complex river flood simulation problems (Felder et al. 2017). Carter et al. (2005) coupled the hydrological model LSPC and the hydrodynamic model EFDC to simulate the hydrological process of the Sacramento Watershed in the United States. The coupled model simulated the outlet flow processes of the 17 sub-watersheds simulated by the LSPC model as the input of the EFDC model, and the simulation achieved ideal results. Jiang et al. (2020) used the BTOPMC hydrological model to couple a hydrodynamic model. The mountain runoff generation calculated by the hydrological model was input into the hydrodynamic model to simulate the urban inundation process. The simulation result was in good agreement with the reality. Wang et al. (2021) used the Xin'anjiang model to couple the IFMS hydrodynamic model in the Puyang River Watershed in the Pearl River Delta region. The upstream runoff calculated by the Xin'anjiang model was input into the downstream two-dimensional hydrodynamic model to simulate floods. The results showed that the hydrological–hydrodynamic coupling model could reflect the hydrological and hydrodynamic characteristics of the watershed. In addition, some mature models such as the MIKE and the HEC series models are often used in coupling models (Thompson et al. 2004; Doulgeris et al. 2012). The coupling of the MIKE 11 NAM hydrological model and the MIKE 11 HD one-dimensional hydrodynamic model has been widely used and the model accuracy has been verified. However, the use of this software module is limited, and the two cannot be freely integrated into the self-developed flood forecasting and scheduling system, making it difficult to carry out secondary development and utilization. Therefore, it is crucial to accurately select a suitable and relatively simple hydrological and hydrodynamic model for coupling on the premise of fully understanding the runoff generation and flood routing laws of different watersheds. Both the HEC-HMS hydrological model and the HEC-RAS hydrodynamic model were developed by the U.S. Army Engineering and Hydrology Center. The coupling process between the two models is simple and has good connectivity. Knebl et al. (2005) coupled the HEC-HMS hydrological model with the HEC-RAS hydrodynamic model and conducted flood simulations in the San Antonio Watershed in the United States. The runoff yield of 12 sub-watersheds simulated by the HEC-HMS model was input into the HEC-RAS model for river flood simulation. The results showed that the coupled model achieved ideal results. For more hydrology–hydrodynamic coupling models and similar studies, please refer to the related literature (Betrie et al. 2011; Paiva et al. 2011; Choi & Mantilla 2015; Rouya et al. 2018).

At present, the hydrological–hydrodynamic coupling models have been widely applied, but there are still some problems that need to be solved. Firstly, the prerequisite for establishing a coupled model is to have accurate and complete river network geographic information data, but obtaining relevant data poses certain difficulties; secondly, there are many existing hydrological models and hydrodynamic models with increasingly rich functions. When dealing with the hydrological process of complex watersheds, it is necessary to ensure that the models have sufficient accuracy and strong adaptability, but the adaptability of different models to different watersheds needs further research. The main contribution and innovation of this article lie in (1) obtaining high-precision terrain data of the research area using drone tilt photography technology; the HEC-HMS hydrological model and the HEC-HMS and HEC-RAS coupling model were used to simulate the historical flood events of a small watershed. The simulation results of the two models were compared and analyzed. (2) The measured river section data was used for model construction to simulate the evolution process of floods in the river, and the functions of three small dams in the mainstream were also considered. This research will answer whether it is necessary to measure the topography of major rivers in small watersheds and use a hydrological–hydrodynamic coupling model to improve the accuracy of flood simulation.

As far as current research is concerned, it is necessary to find a watershed that is prone to flood disasters, with a climate type preferably drought. The runoff generation meets the two characteristics of runoff yield under saturated storage and runoff yield under excess infiltration, so as to test the application of the hydrological model in areas with complex runoff generation modes, and evaluate the fitting effect of the model on floods with different characteristics. In addition, in order to obtain high-precision topographic data, the area of the study area should not be too large.

Liulin Watershed is located in Neiqiu County, Hebei Province, China. The main stream is the Xiaoma River within the Ziya River System. The watershed has a drainage area of 57.4 km², an average slope of 30.9‰, and a width of 4.35 km. The elevation of the watershed gradually decreases from west to east, with an elevation range of 150–1,141 m. About 80% of the watershed is agricultural land with main crops of wheat, corn, cotton, and peanuts. The watershed belongs to a semi-arid and sub-humid climate zone, with a temperate continental monsoon climate. In spring, the rainfall is small, the temperature difference between day and night is large, and the surface evaporation is large; in summer, it is hot and rainy, with concentrated precipitation; autumn is dry and sunny, with high altitude and cool air, and less rain; winter is cold, with less rain and snow, and is prone to northerly winds. The annual average temperature is 11.7–13.6 °C, and the sunshine hours are 2,600 h. The average annual precipitation is 594.5 mm, and the average annual runoff depth is 80.7 mm. The distribution of precipitation within the year is extremely uneven. Floods mostly occur from June to September, and most of them occur in late July and early August. The observed maximum peak flow in historical records was 570 m³/s in 1996.

The HEC-HMS hydrological model was used to simulate the flood processes in the watershed. The HEC-HMS model consists of four modules, namely the watershed module, the meteorological module, the control operation module, and the data storage module. Among them, the HEC-GeoHMS module is used to preprocess the original DEM, including the steps of depression filling, flow direction analysis, and confluence calculation. Compared with the traditional hydrological model, the HEC-HMS model has the advantages of a wide application range, comprehensive functions, and strong adaptability to different climatic environments, and has been widely used in different regions in different countries and regions.

According to previous research in other watersheds in semi-arid and sub-humid climate areas, the Soil Conservation Service (SCS) curve number method was selected for the runoff calculation, the Snyder unit hydrograph method for the direct runoff calculation, the recession method for the base flow calculation, and the Muskingum method for the river confluence calculation. The details of each method can be found in the literature (Gumindoga et al. 2017; Ren 2017; Zhang et al. 2021; Liang et al. 2022).

The simulation results are evaluated by four indicators: Nash–Sutcliffe efficiency (NSE), the relative error of flood peak flow, the relative error of total flood volume, and peak flow occurrence time error. NSE not only reflects the degree of consistency between simulated flood processes and observed flood processes, but also matches the peak weighted root mean square objective function used in the model, making the model more efficient in parameter optimization, while other indicators cannot meet these characteristics well.

The HEC-RAS hydrodynamic model was developed by the American Hydrological Engineering Center and includes four hydrodynamic modules: constant flow surface line calculation, one-dimensional, two-dimensional unsteady flow numerical simulation, sediment transport simulation, and temperature or other components (USACE 2010). Compared with other hydrodynamic models, the HEC-RAS model has a very clear operating interface and modeling steps. Users can clearly know the location and physical meaning of their data input at each step and can set various boundary conditions during operation. It can not only calculate hydraulic elements such as water surface profile, velocity, and depth of river flow but also simulate various hydraulic structures on the river, which can be used to evaluate the impact on water flow. Since there is almost no constant flow in natural channels, the elements of water flow such as velocity, flow, and water level will change with time. Therefore, the hydrodynamic module uses one-dimensional unsteady flow to simulate the flow in the river, and its theoretical basis is the energy equation and the momentum equation (Fang et al. 2011; Shayannejad et al. 2022).

According to the data transmission method between models, the coupling methods of different models are mainly divided into external coupling and internal coupling. Although the internal coupling method has higher accuracy, due to the relatively complex model, external coupling is often used for research in practice.

The HEC-HMS model and the HEC-RAS model are coupled by the external coupling method (Inoue et al. 2008). That is, the simulated flow at each sub-watershed outlet by the HEC-HMS hydrological model is used as the input of the HEC-RAS hydrodynamic model, to simulate the flood process at the watershed outlet. The coupled model calculates the hydrological process independently from the beginning to the end and then calculates the hydrodynamic process independently. The flow at the outlet of the sub-watershed obtained from the hydrological calculation is used as the boundary condition of the hydrodynamic calculation. The hydrological process affects the hydrodynamic process, and the hydrodynamic process has no impact on the hydrological process.

• (1)

  River course generalization

Upstream boundary conditions: The W120 and W140 sub-watershed flow processes are used as the upper boundary conditions of tributary 1, and the W100 and W110 sub-watershed flow processes are used as the upper boundary conditions of tributary 2. The flow processes generated in the remaining sub-watersheds all flow into the channel section in the form of lateral inflow. According to the area weight, the flow process of the sub-watershed where the hydrodynamic area is located is injected into the corresponding section in the form of lateral inflow. Table 1 shows the distribution of flow processes in each sub-watershed after calculation.

Downstream boundary condition: The channel gradient of 1% at the watershed outlet is the downstream boundary condition (USACE 2010).

• (2)

  Roughness setting

The riverbed in the downstream reaches of the Xiaoma River is composed of masonry, sand pebbles, and sandy loam. The upstream riverbed is sandy, and the riverbed surface is relatively flat. After trial-and-error tests, the roughness of the upstream tributaries 1 and 2 is taken as 0.024, and the roughness of the main river channel in the downstream is taken as 0.026.

First, the HEC-HMS model is used to calculate the flooding process of each sub-watershed, and it is input into the HEC-RAS one-dimensional hydrodynamic model as the upper boundary condition or the inner boundary condition. The coupling method is external coupling, and finally, the flood process at the Liulin Hydrological Station cross-section is obtained.

The two parameters that need to be determined by the Snyder unit hydrograph method are the peaking lag time (TP) and the peaking coefficient (CP). They reflect the time of peak occurrence and the magnitude of flood peak flow. The three parameters that need to be determined in the regression method are the initial base flow, decay constant, and ratio of peak. The initial base flow is the average flow of the river channel on the day before the rainfall event. The change rate between the base flow at a certain time on the base flow hydrograph and the base flow of the previous day is the attenuation coefficient. The Muskingum method is used for the river flow confluence calculation, and the flow specific gravity factor X and the slope K of the tank storage curve need to be calibrated. Among them, the storage constant K represents the propagation time of the flood in the river channel, and the initial value can be determined based on the characteristics of the river section. It is represented by the ratio of the length of the calculated river section to the flood wave velocity. The flow specific gravity factor X represents the slope of the river, ranging from 0 to 0.5. The larger the value of X, the steeper the slope of the river, and the greater the discharge. When X approaches 0, it indicates that the river channel is in the maximum attenuation state.

Sixteen floods from 1982 to 1996 were selected for model calibration. The key parameters are calibrated using the univariate gradient optimization method combined with manual adjustment. As an important parameter in the flow generation process, in order to obtain a suitable CN value for this model, the initial value calculated by Equation (2) is first inputted into the model for simulation, and then manually calibrated based on the simulation results to obtain the optimal value. The remaining parameters can be automatically iteratively calculated by the model to provide the optimal values. The values of calibrated parameters are shown in Tables 2 and 3. The simulation results are shown in Table 4.

From the flood simulation results in Table 4, it can be seen that the simulated runoff depth of the 26 floods is generally larger than the observed, but the simulated flood peak flows are generally lower. The reason for the larger runoff depth is that the model does not consider the influence of evapotranspiration, interception, and other factors in the runoff generation calculation. In addition, the model performance for flood events of different sizes is quite different. The specific performance for small floods is poor, for medium floods it is moderate, and for large floods it is good. It is also confirmed that the influence of interception and evapotranspiration is not negligible for small rainfall. When the rainfall is large, the proportion of interception and evapotranspiration decreases, which reduces the influence on the simulation results. From Figure 8, the simulated flood hydrographs are basically consistent with the observed flood hydrographs, and the simulation effect of the model on single-peak flood is better than that of multi-peak flood. Among them, the No. 950816 flood event had four flood peaks, but the simulated hydrograph only had two flood peaks.

In general, the simulated flood hydrographs are basically consistent with the observed flood hydrographs. Floods with simulation errors exceeding 20% are mainly small flood events with runoff depths of less than 5 mm. For the No. 960804 large flood, the relative errors of the runoff depth and peak flow are both within 20%, and the NSE is 0.80. For the selected 10 flood events during the validation period, only two floods, 960820 and 160807, have poor simulation results. The relative errors of the runoff depth and peak flow of the No. 960820 flood event exceeded 50%, and the errors of the runoff depth and peak flow of the remaining floods were all within 20%. For the No. 160720 large flood with a peak flow of 368 m³/s, the relative error of the simulated runoff depth was 11.01%, the relative error of the peak flow was −10.68%, and the NSE was 0.87, indicating a good performance of the HEC-HMS hydrological model.

It can be seen from Table 5 that the simulated runoff depth of the 26 flood events is overall relatively large. The specific reasons have been analyzed above. However, the simulation results of flood peak flow are generally smaller. This is because the three dams in the coupled model had a greater impact on the medium and small flood simulations.

It can be seen from Figure 9 that the simulated flood hydrographs are basically consistent with the observed flood hydrographs. The coupled model simulates unimodal floods better than multimodal floods. Especially for the No. 160812 flood event, the simulated flood process is almost the same as the measured flood process.

The average value of the NSE is 0.64 for the 26 flood events simulated by the coupled model. There are 5 flood events with runoff depth errors exceeding 20%, and 10 flood events with relative error of the peak flow exceeding 20%. Among them, the No. 960820 flood event has the largest error, reaching −62.76%. However, for a flood with a large peak flow such as the No. 160720 flood event with a peak flow of 368 m³/s, the relative error was −14.96%.

From the flood simulation of the two models, it can be seen that the floods simulated by the hydrological model with a relative error of runoff depth less than 20% accounts for 73.1% out of the total number of flood events. The relative error of the flood peak flow less than 20% accounts for 73.1% out of the total number of flood events. The peak flow occurrence time error is all within 3 h, and the average value of the NSE is 0.72.

Obviously, the structure of the model itself has brought certain errors to the simulation results, but subjective factors have also caused certain errors during model construction. From the perspective of the model, both models have a calculation step size of 1 h. In order to output a flow sequence with a step size of 1 h, certain interpolation processing was performed on the collected rainfall and runoff data. From this characteristic, it can be seen that during a period of the same flow rate, the shorter the calculation time step, the more accurate the output peak flow rate. An increase in the calculation time step will cause the peak flow rate to become uniform during this period, resulting in a smaller peak flow rate. When there is short-term heavy rainfall, the simulated flood process will experience a ‘flattening’ phenomenon. From the analysis of the coupling mode between the hydrological model and the hydrodynamic model, because the scope of the flood inundation area changes with time during the rainstorm process, the location of the flow boundary points of the hydrodynamic model also changes with time. The boundary points of the sub-watersheds confluence into the inundation area are continuously distributed in space rather than a finite number of scattered points, and the external coupling cannot accurately determine the location of these flow boundary points.

This study used a UAV equipped with five-lens tilt photography technology to obtain high-precision terrain of the watershed and DEM with a resolution of 7 cm × 7 cm. The channel cross-sections were extracted and 1 m × 1 m resolution DEM was resampled for hydrological and hydrodynamic modeling. On this basis, the hydrological model and the hydrological–hydrodynamic coupling model were constructed to simulate floods in the Liulin Watershed and achieved good simulation results. However, there are still some deficiencies in this paper to be further studied.

The rainfall-runoff data are from 1982 to 2016, with a time span of 35 years. During this period, human activities may have had a greater impact on the underlying surface of the watershed. The terrain data used in the model was measured in March 2021, especially the river course may have changed during these years, which resulted in some inaccurate flood simulations. There are three small dams on the mainstream, and the storage capacity cannot be negligible. However, the detailed construction year cannot be surveyed. When we consider the three dams in the hydrodynamic model, the flood simulation before the construction year must be affected. In addition, we did not consider the land surface change during the simulation period, which might result in some discrepancies. Land use/land cover change would have influenced flood formation and the characteristics. This impact has been confirmed by many studies. Rukundo & Dogan (2016) studied the impact of land use change on the flood process in the Nyabugogo Watershed in Rwanda, and inserted the land use data into the hydrological model to simulate flood peak flow. The results showed that land use change affected the hydrological processes by increasing surface runoff, which in turn led to a significant increase in flooding. Hu et al. (2016) established a distributed hydrological model to study the impact of land use change and small upstream water conservancy projects on watershed runoff. The results showed that the flood simulation accuracy can be significantly improved after considering land use change and water conservancy engineering. Ma (2019) used the TVGM model to study the impact of changes in the underlying surface on floods in the Xihe River Watershed in Liaoning Province. The conclusion showed that the change of the underlying surface conditions can reduce the peak flow and total flood volume, and the smaller the flood events, the greater the attenuation effect.

The rainfall spatial distribution in the Liulin Watershed is not uniform. There are only five rain gauges in the watershed and the distance between them is far. When the rainfall of some floods is relatively concentrated, five rain gauges may not reflect the real temporal and spatial distributions of the rainfall, resulting in some errors in the conversion of point rainfall to area rainfall in the model, which affects the flood simulation results. In the Liulin Watershed, an X-band precipitation radar was installed in 2021, with a detection distance of 150 km, and an overall scan is completed every 6 min. In the future, the radar monitoring data can be combined with the rain gauge observed data to obtain a more precise temporal and spatial distribution of rainfall, which plays a crucial role in reducing the uncertainty in the simulation of hydrological processes. This technology has been adopted by some scholars. Sun et al. (2000) integrated rain gauge and radar data in a 1,060 km² watershed with only three rain gauges in Australia. The results showed that the fusion data can show great potential for hydrological applications even when the density of rain gauges is very small. Wijayarathne et al. (2021) integrated the S-band and C-band radar QPEs into event-based hydrological models to improve the calibration of model parameters for streamflow simulation and flood mapping in an urban setting. The results showed that the bias correction of radar QPEs can enhance the hydrological model calibration. Reichel et al. (2009) illustrated that the fusion of radar and rainfall data will have good potential for hydrological modeling and hydrological forecasting.

In the construction of the HEC-HMS hydrological model in the Liulin Watershed, the SCS curve number method, the Snyder unit hydrograph method, the recession method, and the Muskingum method were used to simulate the flooding process. In fact, the runoff generation process of the watershed is affected by many factors, such as soil type, vegetation coverage type, soil moisture content, and underlying surface conditions. The SCS curve method sums up many influencing factors into a parameter CN, but some influencing factors are not in the same step in time or space, resulting in large errors between the simulated runoff and the measured runoff of some floods. In the HEC-HMS model, each module has a variety of calculation methods for runoff generation and flow concentration. In particular, the calculation methods of runoff generation reflecting different runoff mechanisms have a great influence on the flood simulation. Therefore, it is an important direction to improve the flood simulation accuracy by using different model structures to simulate the flood process in the Liulin Watershed, and evaluating the uncertainty of the model structure on the flood simulation. Wang et al. (2018) selected five common hydrological models to carry out flood simulations in 14 small watersheds in typical hilly areas, to compare the applicability of different models. The results showed that the HEC-HMS hydrological model has the strongest applicability in small sub-humid and semi-arid watersheds, but in some small watersheds where both saturation excess and infiltration excess runoff generation coexist, it is also difficult with the HEC-HMS hydrological model to meet the requirements of flood forecasting. Furthermore, in order to explore the difference between the characteristics of the HEC-HMS model and the TOPMODEL model and their simulation effects, Li et al. (2019) selected 10 representative floods in the Dongzhuang Watershed of Shanxi Province for simulation. The results showed that the sensitive parameters of the two models were very different. The HEC-HMS model has a high sensitivity to the parameter CN. Since the TOPMODEL model considers evapotranspiration in the runoff calculation, which has a greater impact on humid regions than in arid regions, it has better simulation results in humid regions.

At present, improving the basic theory of rainfall-runoff generation is an urgent problem to be solved in developing and improving the hydrological model in sub-humid and semi-arid small watersheds in hilly areas. In practical applications, combined with local climatic conditions and watershed characteristics, more advanced monitoring methods can be used to obtain better simulation results. As for the Liulin Watershed, the causes of flood disasters can be divided into meteorological factors and topographic factors. As the main factor, meteorological factors have more unpredictability, and more observation methods need to be added to the existing foundation to explore the occurrence mechanism of flood disasters. In 2021, on the basis of the original 5 rain gauges, another 5 meteorological stations have been installed, and 15 TDR soil moisture monitoring stations have been set up in the Liulin Watershed. In the future, combined with high-density and high-precision rainfall, soil water content, and flow observation in the watershed, we can analyze the runoff generation mechanism with different rainfall patterns and previous soil water content conditions, and explore the applicability of different model structures to rainfall patterns.

The high-resolution DEM data of the Liulin Watershed was obtained by using the UAV equipped with five-lens tilt photography technology. The channel cross-sections and dams were extracted, and the DEM was resampled to obtain a DEM with a resolution of 1 m × 1 m, which has higher accuracy compared to DEM obtained from geospatial data cloud. On this basis, the HEC-HMS hydrological model with high calculation efficiency is used to couple with the HEC-RAS hydrodynamic model with a clearer physical mechanism. Twenty-six historical flood events were selected to calibrate and verify the HEC-HMS hydrological model. The simulated flood discharge by the HEC-HMS model of each sub-watershed is imported into the HEC-RAS hydrodynamic model by external coupling to simulate the flood hydrograph at the outlet of the Liulin Watershed. The main conclusions can be summarized as follows:

• (1)

  The HEC-HMS hydrological model was constructed according to the DEM with a resolution of 1 m × 1 m in the watershed; not only does it truly reflect the terrain and geomorphological characteristics of the watershed, but it also makes the process of runoff generation and concentration more accurate. The parameters of the model were calibrated and verified using 26 observed rainfall and runoff events. The floods with a relative error of peak flow less than 20% accounted for 73.1% of the total flood events, and the floods with a relative error of runoff depth less than 20% accounted for 73.1% of the total flood events. The average value of the NSE during the calibration period is 0.71, and during the validation period it is 0.72. The model exhibits stable performance and reliable results, which can be applied to flood simulation and prediction in the research area.

• (2)

  The hydrological–hydrodynamic coupling model takes into account the evolution process of floods in river channels, which can better simulate water flow movement. From the simulation results of the coupled model, it can be seen that the model has significant errors in simulating small floods, due to the consideration of the water storage effect of three dams in the hydrodynamic module, resulting in a smaller simulation value. For large floods such as No. 160812, the simulation results are closer to the measured values. Overall, although the construction of hydraulic structures in river channels and changes in underlying surfaces make it difficult for coupled models to simulate the production and concentration processes of small floods, they are still applicable to large floods, especially those that are relatively recent. However, in this study case, due to the unsatisfactory simulation results of certain flood events, there is still room for further optimization of model parameters. The external coupling method cannot accurately reflect the internal interaction relationship between models. In future research, it is necessary to explore how to further optimize the coupling method of the model, such as adopting a tightly coupled approach to make the model more efficient. The hydrodynamic module adopts the one-dimensional channel hydrodynamic calculation method, which only simulates the flood routing process in the river channel, without considering the situation of the flow outflow channels, and can build a two-dimensional channel hydrodynamic model for simulation.

J.L. proposed the main idea for the hydrological–hydrodynamic coupling model; Z.W. established the hydrological model as well as the hydrological–hydrodynamic coupling model; and T.Z. dealt with the digital elevation model.

We are grateful to Xingtai Hydrology Survey and Research Center, Hebei Province for providing the rainfall and flood data.

All relevant data are included in the paper or its Supplementary Information.

The authors declare there is no conflict.