Hydrologic models are used to solve a range of specific problems in the management and development of water and land resources, including flood simulation and prediction, aquifer recharge management, runoff estimation, and drainage network design (e.g., [1
]). The uncertainties inherent in these models can be reduced with the availability of accurate input and watershed data [4
]. Hydrologic models include lumped models, semi-distributed models, or fully distributed models [5
]. These models can also be classified based on the model formulation as empirical or physically-based models with conceptual formulations typically used in lumped models and physical equations in the other two types [6
]. The parameters of the lumped models are not directly related to the watershed’s physical characteristics, which is an important shortcoming of these models [7
]. Generally, the application of lumped models is limited to gauged watersheds not undergoing significant change in their conditions, and they have to be calibrated using significant observational data [7
In a semi-distributed model, the basin to be simulated is divided into smaller sub-basins to capture the spatial variability of the inputs and to represent the heterogeneity within the basin [8
]. The data required for a semi-distributed model is typically less than that needed for a fully distributed model. In addition, the spatial representation and mathematical formulations of the hydrologic processes are simpler [9
]. Although a semi-distributed model requires a limited number of parameters compared to a fully distributed model, it requires more calibration data [10
]. In general, the natural spatial and temporal variations of the hydrological properties such as soil type, land use, and rainfall are not accurately represented in the semi-distributed hydrologic models [11
The fully distributed models try to simulate the watershed conditions accurately while the physical processes and/or the spatial heterogeneity are mathematically treated at the grid cell scale [12
]. The distributed models can be either empirically distributed models or physically-based distributed models. The difference between them is the transformation of the surface runoff, which is empirically simulated in empirical distributed models and through physical formulations in physically-based distributed models [14
]. The model grid size is selected based on the required level of detail of runoff simulations and the spatial resolution of inputs and watershed data [15
Several studies have compared different types of hydrologic models. Paudel et al. [17
] compared the performances of lumped and distributed models and concluded that the simulation of land use changes using distributed models is significantly better than using lumped models. Sith & Nadaoka [18
] found that the physically-based, fully distributes model GSSHA (Gridded Surface/Subsurface Hydrologic Analysis) performed slightly better than the semi-distributed, semi-physically-based model SWAT (the Soil and Water Assessment Tool) for short-term streamflow simulations, while, for long-term simulations, both models had acceptable performance. Hui-lan [19
] concluded that the hydrographs predicted by a distributed model compared with field observations were much better than those predicted by a semi-distributed model. El-Nasr et al. [20
] found that the semi-distributed SWAT model and the fully distributed MIKE SHE model performed well when used to simulate the hydrology of a large catchment, but the MIKE SHE model was slightly better. Meselhe et al. [21
] investigated the impact of spatial and temporal sampling of rainfall on runoff predictions using a conceptual and physically-based hydrologic model over a small watershed. They found that the physically-based hydrologic model was more sensitive to both the temporal and spatial samplings of rainfall than the conceptual model. Other researchers evaluated the performance of HEC-HMS (the Hydrologic Engineering Center-Hydrologic Modelling System) and GSSHA models and reported a much better performance by GSSHA (e.g., [22
Several hydrologic studies have been conducted in the Arab Peninsula using different modeling approaches. Al Abdouli et al. [24
] used HEC-HMS to quantify coastal runoff and examine its temporal and spatial distribution in several watersheds in the United Arab Emirates (UAE) to assess the potential for recharge of depleted aquifers and flood mitigation in urban areas. Sharif et al. [25
] simulated an extreme flood event in Hafr Al Batin city, Saudi Arabia using the physically-based, fully-Distributed Hydrologic model (GSSHA) driven by Global Precipitation Measurement (GPM) satellite rainfall data. Their results demonstrated that the urban portion, which is 6.8% of the watershed area, produced about 85% of the generated runoff. Embaby et al. [26
] developed an integrated approach combining geotechnical investigation and hydrologic modeling to generate hazard zones for Tabuk city, Saudi Arabia. The approach was successful in identifying areas prone to flooding. Al-Zahrani et al. [6
] developed a flood hazard map for Hafr Al-Batin city by integrating the results of three models (hydrologic, hydraulic, and flood delineation models). They recommended that their result can be used for designing flood control structures and other planning purposes. Fathy et al. [27
] evaluated hydrologic models over Wadi Sudr, Sinia, Egypt, and found that the runoff hydrograph can be estimated more accurately using the distributed model. To the authors’ knowledge, no study tried to investigate the differences between conceptual semi-distributed and physically-based fully distributed models in simulating extreme flood events in the region.
In recent years, the Makkah region in western Saudi Arabia has been suffering from an increase in the frequency of flash floods due to its natural complex topography, which is dominated by mountains with steep slopes and very shallow soil layers. These events are triggered by the occurrence of infrequent high convective short-lived rainfall events. A few hydrologic studies have been performed in the Makkah region. However, all studies used lumped or semi-distributed models and were mostly focused on simulating peak discharge. Abdelkarim & Gaber [28
] assessed the impact of flash flooding in the Makkah area by constructing a flood risk map for a watershed in Makkah City using HEC-HMS and HEC-RAS. Their findings suggested a flood control policy based on using the existing hydraulic facilities and a range of new structures that can help protect lives and the urban infrastructure. Elfeki et al. [29
] evaluated the flood hazard on another watershed in Makkah, also using HEC-HMS and HEC-RAS, and recommended some structural measures to mitigate floods. Bastawesy et al. [30
] assessed the potential of flash flooding in a third watershed in Makkah using GIS and recommended a range of mitigation measures to contain the 50-year design flood. Dawod et al. [31
] applied simple hydrologic models and four national regression models to estimate the flood peak discharge in several ungauged small watersheds in Makkah. They concluded that the Curve Number (CN) approach is the best methodology for flood estimation in case of the availability of soil type, land use, and metrological data. Al-ghamdi et al. [32
] used the Curve Number method to estimate flood hazards in Makkah City for the period 1990 and 2010. They found that the expansion of the residential areas was behind the significant rise in significant flood hazards. Al Saud [33
] analyzed the morphometry of Wadi Aurnah, Makkah, using GIS and introduced a simple approach for rapid assessment of the flooding potential.
The main goal of this study is to simulate recent floods in the Makkah region using a physically-based fully distributed hydrologic model and compare its performance to that of the semi-distributed HEC-HMS model. Two flood events that occurred in 2010 and 2018 were used in this study. One event was used for model calibration and the other for validation. The difference in model output was discussed and was further investigated by comparing the different implementations of HEC-HMS to explore the impact of watershed heterogeneity on the model results.
2. Study Area
Makkah province is one of thirteen provinces in Saudi Arabia and is considered one of the most important regions due to its location in the historic Hejaz region and its extended coastline on the Red Sea. The total area of Makkah region is approximately 153,200 km2
occupied by around 8,557,000 people [34
], making it the third-largest and most populous province. The city of Makkah (Makkah Al-Mukarramah), the Makkah Province’s capital, is located 70 km from Jeddah on the Red Sea with an average elevation of 277 m above sea level. The region is characterized by a dry climate with sparse, inconsistent rainfall events [35
]. Plains and mountains are the main geomorphological units of the study area. In the past few decades, the city of Makkah has witnessed unusually rapid urbanization, making it the third most populated metropolitan area in Saudi Arabia. The increase in the urban areas in Makkah province between 1992 and 2013 was threefold, as reported by Alahmadi & Atkinson [36
]. This growth is mainly due to the significant increase in the area covered by Makkah City from about 12% of the province in 1992 to 22% in 2016 [37
]. Recently, Abdelkarim & Gaber [28
] reported that the urban areas exposed to flooding have increased by approximately 25 fold between 1988 and 2019. In addition, a high percentage of Makkah’s road network (more than 50%) is subjected to inundation during major flood events [36
]. Because of the rapid urban growth, municipal authorities now require flood impact studies being conducted before any land development activity within the city limits. The Makkah watershed used in this study was delineated using a high-resolution digital elevation model (DEM) of 10 m obtained from King Abdulaziz City of Science and Technology (KACST). The delineated watershed includes most of Makkah city with a drainage area of 1725 km2
3. Rainfall Events
The calibration and validation run of the two models were driven by two storm events that occurred recently in the Makkah watershed. Due to the very sparse rain gauge network in the study area, capturing the features of the extreme precipitation through ground observations was not possible. Moreover, the very limited rain gauge data are available only at a daily time scale. Therefore, the two hydrologic models were forced using the satellite rainfall data, which would have the ability to capture the spatial and temporal variability of the storm events at much higher resolutions.
A previous study by the authors using the Integrated Multi-satellitE Retrievals for GPM (IMERG) satellite rainfall products showed reasonable performance of the IMERG Early run product in the Makkah region compared to the other IMERG products. Accordingly, the two simulated models were forced by the IMERG Early run product in this study. The IMERG Early run data with spatial and temporal resolutions of 0.1° × 0.1° and 30 min were downloaded using FileZilla software from the Precipitation Measurement Missions (PMM) website (http://pmm.nasa.gov/data-access/downloads/gpm
(accessed on 7 March 2020)). The data were converted from HDF5 into a gridded ASCII format using an R script developed by the authors and was then reformatted as input to the hydrologic model. The latest version of the IMERG product, Version 6 (IMERG-V06), was used. The latest product includes significant updates such as a higher maximum rainfall threshold from 50 to 200 mm/h, full inter-calibration to the GPM combined instrument dataset, and the use of an updated rain retrieval algorithm [38
The first storm that occurred on 13 February 2010 with an average watershed total of 54.7 mm based on the Early IMERG product (Figure 2
) was used for calibration. The IMERG Early product showed that the intense rainfall was concentrated over the northern portion for the 13 February 2010 event and the southeastern part of the watershed for the 3 November 2018 event. The second storm that occurred on 3 November 2018 was used for validation. Three rain gauge stations recorded the total rainfall accumulation for this event: Makkah (J114), Arafah (9004), and Muntasaf-Huda (J205), with rainfall totals of 7.8, 22.0, and 20.0 mm, respectively. The watershed-averaged total rainfall observed by the three gauges was 18.36 mm based on the Inverse Distance Squared Weighting method, while it was 42 mm based on Version 6 (Figure 2
). The Early IMERG product significantly overestimated the rainfall amount for the 3 November 2018 storm event compared to observations by the ground rain gages. However, the average rainfall estimated by the Early IMERGV06 is closer to what can be inferred from a study that has been done on Wadi Al-Nu’man watershed in Makkah [29
Rainfall frequency analysis was also used to estimate the rainfall for different return periods to assess the models’ performance. The recorded data from 2006 to 2018 was used to estimate the return period for maximum daily rainfall over the study area using Log-Pearson Type III distribution (LPT III) [39
]. Figure 3
shows the frequency distribution as fitted to the LPT III distribution. Table 1
shows the maximum 24-h rainfall values for the return periods of 2, 5, 10, 25, 50, and 100 years.
5. Watershed Data
The primary input data required for the GSSHA and HEC-HMS models include rainfall data, digital elevation models (DEM), soil types, and land use/cover, which need some processing and preparation in the right input format. The DEM data, at 10 m resolution, were obtained from King Abdulaziz City of Science and Technology (KACST). The soil type data were downloaded from the SoilGridsTM
global digital soil mapping system (www.SoilGrids.org
(accessed on 20 March 2020)) as mass fractions of clay, silt and sand in percentages. The final soil type map (Figure 6
) was developed for the Makkah watershed using ArcGIS10.6 [49
], and the soil texture classification based on the clay, silt, and sand contents and then adjusted after comparison with aerial photographs and field information. The land use/cover data (Figure 6
) were obtained from the OpenLandMap data portal (www.openlandmap.org
(accessed on 25 March 2020)), and adjustments were made after comparison with aerial photographs. The hydrologic soil group data were obtained from the HYSOGs250m product ([50
(accessed on 26 March 2020)) at 250 m resolution as shown in Figure 7
a. This data were used to estimate the CN values (Figure 7
b) needed for the HEC-HMS model.
The methodology adopted in this study is described through the flowchart shown in Figure 8
. ArcGIS 10.6 was used to prepare the raster data for WMS [51
] to be processed and exported to HEC-HMS. The process in WMS includes delineation of the watershed, flow accumulations, flow directions, and computing CN values based on the hydrologic groups in Figure 7
. The watershed was subdivided into 55 sub-basins for HEC-HMS simulations, as shown in Figure 9
a. The SCS Curve Number and SCS dimensionless transformation methods were applied to compute the runoff volumes and to model direct runoff, respectively, while the Muskingum method was used for channel routing. The inverse distance method was used to generate the rainfall hyetograph for the selected storm events before being used as input to the HEC-HMS model.
The channel roughness, channel cross-sections, and other parameters such as hydraulic conductivity were required for the GSSHA model. ArcGIS and WMS were also used to prepare the raster data for the GSSHA model. All streams were smoothed to avoid the effect of the adverse and flat slopes resulting from inaccuracies in the DEM. The watershed was divided into grid cells at 150 × 150 m resolution (Figure 9
b). The land use/cover and soil types were imported by WMS to prepare the grid index maps which then were used to estimate the required parameters such as infiltration parameters, roughness, initial moisture, overland retention depth, and impervious fraction. The WMS was also used to extract channel cross-sections from the 10 m DEM (Figure 10
). The no-flow condition was adopted for all first-order streams boundary conditions in GSSHA as there is no baseflow in the basin [47
In this study, a physically-based, fully distributed model (GSSHA) and a conceptual semi-distributed model (HEC-HMS) were used to simulate flood events over an arid watershed in Makkah, Saudi Arabia. The steep slopes and large fraction of impervious surfaces of the study area increase drainages efficiency and contribute to the occasional flash floods in the city. Hydrographs simulated by the two models were compared to observed peak discharge for one flood event using satellite rainfall as input. One more event was used for model validation. The two models were prepared with the minimum required data so that the comparison between the two models would be meaningful. The performance of the models was measured by the peaks’ discharge. The effort required to build the two models was comparable since the same input data were used, and a similar number of parameters were calibrated. However, additional effort was needed to define channel cross-sections for the GSSHA model. Moreover, cell-to-cell surface runoff computations in GSSHA required more time than did the lower resolution sub-basin runoff computations of HEC-HMS.
Although the GSSHA model performed better than HEC-HMS, which significantly overestimated the peak discharge before calibration, the two models’ performance after calibration was comparable. However, the runoff volumes produced by HEC-HMS were much smaller than GSSHA for all calibration and validation events. The failure of HEC-HMS to capture details of the complex terrain and land surface heterogeneity was likely behind this discrepancy. The cell-to-cell routing of flow that was only possible with GSSHA. The GSSHA model can make use of higher-resolution DEMs and rainfall products that are becoming more available and are more suitable in urban areas, where the natural landscape is significantly modified.
The difference in the output of GSSHA and HEC-HMS was investigated further by comparing GSSHA hydrograph to those produced using three implementations of HEC-HMS (semi-distributed (SCS-UH), semi-distributed (C-UH), and ModClark). A uniform design storm was used as input. The two semi-distributed (SCS-UH and C-UH) models were implemented with the same input data, losses method, and routing. The results showed that the semi-distributed (C-UH) model produced less peak discharge than the semi-distributed (SCS-UH) one. This could be related to the effect of the storage coefficient in the C-UH model, which is geomorphological data and unique for each watershed. The HEC-HMS ModClark model hydrograph was comparable to that of GSSHA, indicating that it was able to capture the effect of the complex topography of the watershed. The comparison results were intended only for situations when a quick and simple hydrologic model simulation is to be performed. All models could have been built more deliberately by incorporating more of the watershed features, such as the detention/retention structure in residential areas, if observed runoff data were available. This can be investigated in more detail in future studies.
A major advantage of the GSSHA model is that the detailed information on high infiltration areas produced by the model can help identify the locations where the excess runoff can be utilized in aquifer recharge. This can help reduce flooding without the need for costly structural measures. Moreover, the detailed information on the runoff depths produced by the GSSHA can be used for designing flood control structures and other planning purposes.