This study investigates an extreme weather event that impacted the
United Arab Emirates (UAE) in March 2016, using the Weather Research and
Forecasting (WRF) model version 3.7.1 coupled with its hydrological modeling
extension package (WRF-Hydro). Six-hourly forecasted forcing records at
0.5
Changes in rainfall patterns directly impact hydrological processes overall and impact particularly the timing and magnitude of floods. In order to produce reliable flash flood warnings, accurate predictions of precipitation timings and amounts, along with their impact on resulting runoff, are needed. However, discrepancies are largely reported when forecasting precipitation using numerical weather prediction (NWP) models. The inaccurate predictions are then magnified in flood forecasts when simulated rainfall is used to drive a hydrological model (Wang and Seaman, 1997; Nielsen-Gammon et al., 2005; Yousef and Ouarda, 2015). Standard hydrological models are often driven by precipitation products inferred from radar, rain gauge, and remote sensing observations, or a combination of them. The lack of dense radar and rain gauge networks in areas like the Arabian peninsula makes the reliance on precipitation remote sensing more attractive. However, such products come with coarse spatial resolution and fail to capture rainfall structures forced by mesoscale orography and land surface interactions, which are prevalent across the Arabian peninsula, including the UAE (Mandoos, 2006). A number of studies reported higher inaccuracies of precipitation products in arid regions (Fekete et al., 2004; Milewski et al., 2015; Wehbe et al., 2017, 2018), which suggests the interest in enhancing mesoscale modeling of weather processes in such regions to generate reliable precipitation products, and therefore accurate predictions of extreme hydrometeorological events. The Weather Research and Forecasting (WRF) model is a mesoscale NWP system created for the dual purpose of assisting with the needs of operational forecasting and facilitating atmospheric research. WRF is designed to be a next-generation mesoscale forecast model and data assimilation system for the purpose of advancing the understanding and prediction of mesoscale weather, and accelerating the transfer of research advances into operations (Skamarock et al., 2005). Powers et al. (2017) give a detailed overview of the initial phases of the WRF project.
A limited number of studies utilizing the WRF model focused on areas in the Middle East region. Awad et al. (2007) summarize the achievements of the UAE Air Force and Air Defense in the use of the WRF model, where a local operational suite was developed for predicting weather numerically over the Middle East region generally, and specifically over the Arabian peninsula and UAE areas. Recently, Chaouch et al. (2017) studied the consistency of WRF simulations with seven different planetary boundary layer (PBL) schemes and showed that a better performance is obtained with the quasi-normal scale elimination (QNSE-EDMF) scheme while the remaining schemes showed a comparable performance. El Afandi et al. (2013) simulated heavy rainfall events in the Sinai Peninsula using WRF for the purpose of exploring how early warnings could be issued for flash flood risk mitigation. They found WRF simulations of a flash flood event on 18 January 2010 consistent with measurements recorded at rainfall gauges for different parts of the peninsula with RMSE values below 5 %.
The enhanced WRF Hydrological modeling extension package (WRF-Hydro) has shown improvement in prediction capabilities of hydrometeorological forecasts using numerical prediction tools when tested over other regions with climate conditions different from those observed in the Middle East (Parodi et al., 2013). It has been used for flash flood prediction, land–atmosphere coupling studies, regional hydroclimate impacts assessment, and seasonal forecasting of water resources (Gochis et al., 2013). Flash flood predictions using the WRF-Hydro model have been applied across the United States of America (Unal, 2015; Gochis et al., 2015; Read, 2015) and various parts of the world. Fiori et al. (2014) analyzed a convective system responsible for an extreme flash flood event that occurred in Genoa, Italy, on 4 November 2011, using WRF coupled with WRF-Hydro. The study outlined the effectiveness of the model in predicting quantitative precipitation for the purpose of flash flood forecasting. Streamflow forecasting from the fully coupled WRF-Hydro modeling system was evaluated through comparisons with both observations and uncoupled hydrological model results. The results of the study highlighted the need to consider multiple factors and sources of error for the prediction of hydrometeorological events, and it presented optimal configurations of WRF-Hydro for future extreme flash flood events in the Mediterranean region. A study conducted for the western Black Sea region of northern Turkey by Yucel et al. (2015) also analyzed the potential of the WRF-Hydro modeling system for flash flood predictions. The study explored the potential of improving runoff and streamflow simulations through the application of data assimilation for precipitation prediction in NWP models. The analyses involved concentrated on assessing the capabilities of the WRF and WRF-Hydro models in forecasting the responses of floods for the vast range of hydrological settings associated with 10 different events that occurred in the study area. Precipitation inputs for the forcing of the WRF-Hydro model were derived from the WRF model and the European Organisation for the Exploitation of Meteorological Satellites (EUMETSAT) multi-sensor precipitation estimates (MPEs). The use of a data assimilation scheme with the WRF model provided significant improvements in simulations of streamflow, while the MPE product led to less accurate streamflow simulations. The optimum results in error reduction were obtained when both WRF model data assimilation and hydrological model calibration were employed.
In a recent study, Givati et al. (2016) calibrated and evaluated the coupled WRF–WRF-Hydro modeling system with the standalone WRF model for flood forecasting of Wadi Musrara, Israel, for two major winter storm events in January and December 2013. Higher correlations (0.89 and 0.85) with the station observations were recorded by the coupled WRF-Hydro model than those of the standalone WRF model (0.85 and 0.80) for both events. Lower RMSE values were also obtained by the coupling (12.2 and 24 compared to 16 and 30). The findings of the study showed that the coupled WRF–WRF-Hydro modeling system resulted in improved precipitation and hydrological simulations when compared with the results of the standalone WRF simulations. Therefore, the authors proposed the employment of atmospheric–hydrological coupling due to its potential to produce improved precipitation predictions, which translates to better hydrological forecasts for early flood hazard mitigation. The application of WRF-Hydro for flood forecasting in arid environments has also been explored by Silver et al. (2017). They simulated six storm events over seven basins in arid and semi-arid regions of Israel and Jordan using WRF-Hydro while incorporating field-based soil characterization data into the land surface model initialization. Nash–Sutcliffe efficiency coefficient (NSE) values of up to 0.415 were recorded between the observed streamflow records and WRF-Hydro simulated streamflow. In desert regions, high soil porosity and hydraulic conductivity of the prevailing sandy soil implies rapid infiltration and runoff drainage. This suggests that the impact on latent heat and, therefore, on the surface radiation budget would be minimal. Alternatively, precipitation largely contributes to soil physical crust formation in desert environments (Fang et al., 2007). Precipitation compacts fine particulate and fills the pores of the top soil layer, forming a hard shell. Dust is also washed out of the atmosphere by precipitation over desert environments which increases amounts of finer particulate at the surface layer to further accelerate the crust formation process. This translates to less vertical infiltration and more lateral flow processes. These mechanisms are specific to arid regions and corroborate the importance of accounting for lateral flow and surface feedback in the coupled WRF–WRF-Hydro model to correctly capture the atmospheric and hydrological process.
The present study expands the ongoing research addressing the prediction of extreme hydrometeorological events in arid regions and investigates the potential of coupling atmospheric and hydrological processes in short-term prediction. There is interest in determining the potential improvement in the simulation that the online land–atmosphere coupling could bring in the case of an extreme hydrometeorological event. To our knowledge, such coupling has never been assessed in hyper-arid environments like the one observed in the UAE where hydrological and atmospheric processes are specific and different from other study domains where similar coupling was evaluated. We focus in this study on an extreme hydrometeorological event recorded on 9 March 2016 over the UAE. The standalone WRF and fully coupled WRF–WRF-Hydro simulations are assessed against (i) weather station surface observations of rainfall, 2 m surface temperature and global radiation, and (ii) satellite remote sensing retrievals of Global Precipitation Measurement (GPM) rainfall, the Advanced Microwave Scanning Radiometer 2 (AMSR2)-derived soil moisture, and the Moderate Resolution Imaging Spectroradiometer Atmosphere product (MODATM) cloud fraction. In the absence of ground-based streamflow gauges in the study domain, satellite data were valuable to understand the dynamics of the event and changes in precipitation, and soil moisture distributions.
WRF model nested domains at 9, 3 (d02), and 1 km (d03)
horizontal resolutions
This study consists of three nested domains (see Fig. 1), with the parent
domain covering the Arabian peninsula, Iran, Iraq, Afghanistan, and parts of
Pakistan, Syria, and Ethiopia (9 to 37
Figure 1 also shows the high-resolution (100 m) WRF-Hydro terrain model domain over the UAE, along with the associated topography map derived from the 30 m resolution Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) digital elevation model (DEM) (Toutin, 2008). The Al Hajar Mountains, located along the Gulf and Oman coasts, foster local convergence zones that trigger small-scale convection initiation (Chaouch et al., 2017). Consequently, the northeastern part of the UAE receives more rainfall compared to the country's 100 mm annual average (Ghebreyesus et al., 2016a, b; Ouarda et al., 2014; Wehbe et al., 2017). The inland city of Al Ain is of close proximity to the Al Hajar Mountains on the southerly downside, designating it as a susceptible area to flash floods triggered by accumulated upstream runoff from the northern highlands. Farming and agriculture is concentrated in this area to benefit from the oasis effects of high rainfall rates and fresh groundwater, compared to other parts of the country.
On 9 March 2016, a low pressure system passing from the UAE and Oman to
southeastern Iran produced thunderstorms, fierce winds, large hail, and
severe flooding. According to the UAE National Center of Meteorology (NCM),
over 240 mm of rain was recorded in Dubai while winds of up to
126 km h
Satellite and ground-based observations were used for the verification of WRF and WRF-Hydro simulations. In the case of the former, necessary data re-gridding were done to match WRF outputs spatially and temporally. The following datasets were used to quantitatively and qualitatively assess the performance of the simulations.
The simulated precipitation was compared to GPM satellite retrievals. The GPM
mission, launched in February 2014, provides high temporal (30 min) and
spatial (0.1
Soil moisture estimates from the AMSR2 mission were used to verify the simulated soil moisture and its simulated spatial extent. In this study we rely on AMSR2 datasets for the verification of simulated soil moisture as products from other datasets were masked out over the study area during the investigated event. AMSR2 was launched on 18 May 2012 on-board the Global Change Observation Mission 1st-Water (GCOM-W) platform, with the capability of measuring passive microwave emissions from the surface and atmosphere. The window channels on-board the sensor are capable of retrieving key surface parameters like soil moisture. AMSR2 L3 datasets provide daytime and nighttime soil moisture measurements with near-global coverage over 3 d (Wentz et al., 2014). Moreover, AMSR2 is equipped with frequencies higher than the L band that are more suitable for soil moisture retrievals. Such frequencies, ranging from 6.93 to 89 GHz, are currently available on-board Soil Moisture and Ocean Salinity (SMOS) and Soil Moisture Active Passive (SMAP) missions.
Radiosonde skew-T profiles retrieved at Abu Dhabi Airport on 9 March 2016 at 00:00 Z
The MODATM dataset from Platnick et al. (2015) contains measurements in 36 spectral bands (250 to 1000 m resolution in nadir) for a combination of key atmospheric parameters at daily and 5 or 10 km resolutions (parameter dependent), including aerosol properties, water vapor profiles, and cloud properties, starting 13 October 2003 and are currently ongoing. The cloud mask is derived as a probabilistic variable from multispectral testing techniques proposed by Ackerman et al. (1998). The cloud fraction (10 km resolution), used in this study, is obtained from the Terra platform's infrared retrievals during both day and night.
Rainfall, surface temperature (2 m), and average global radiation records were obtained from a network of 57 automatic weather stations (see Fig. 1) operated by the NCM across the UAE. Rainfall observations were recorded at 15 min intervals during the event, while hourly observations of temperature and global radiation were made available after quality control.
Two simulations were carried out in this study: one with the standalone WRF
version and another with the online coupled WRF–WRF-Hydro version. In both
simulations, the initial and lateral boundary conditions of the parent domain
were defined from the Global Operational Analysis and Forecast Products of
the National Center for Environmental Prediction (NCEP-GFS) at 0.5
The existing Noah–multiparameterization land surface model (Noah-MP LSM) in the standalone WRF considers single vertical columns (one-dimensional) of terrain properties at each overland grid cell (Niu et al., 2011; Ek et al., 2003). It fails to account for horizontal interactions between adjacent grid cells to calculate soil moisture, temperature profiles, runoff, and water and energy fluxes at the land surface. Therefore, the runoff–infiltration partitioning in the standalone WRF simulation is described as a purely isolated vertical process with no intake from (or discharge to) neighboring grid cells, as dictated by topography. On the other hand, the WRF-Hydro model utilizes the Noah-MP LSM 1-D representations and attempts to improve the simulation of terrestrial hydrologic processes at high spatial and temporal resolutions by including lateral redistribution of overland and saturated subsurface flows for runoff prediction (Gochis et al., 2013). More importantly for this study, WRF-Hydro is run in coupled mode with WRF to permit the feedback of land surface fluxes of energy and moisture to the atmosphere, which impacts the simulated precipitation fields (Arnault et al., 2016; Larsen et al., 2016; Senatore et al., 2015; Koster et al., 2004). Hence, the coupled WRF–WRF-Hydro configuration differs from the standalone WRF configuration by its (i) lateral distribution of surface runoff and (ii) feedback of surface fluxes to the atmosphere.
The model configuration and
Based on the comparable performance of planetary boundary layer (PBL) schemes over the UAE (Chaouch et al., 2017) and given the unstable conditions during the event, the non-local Yonsei University (YSU) scheme is selected as the most favorable scheme for maintaining an entrainment flux proportional to the surface flux (Hariprasad et al., 2014; Hu et al., 2010; Shin and Hong, 2011). The selection of a microphysics scheme has been shown to be of least importance compared to PBL and cumulus schemes (Argüeso et al., 2011).
In light of the findings of Givati et al. (2011), the physics options chosen for this study included the following: the Noah-MP land surface scheme (Niu, 2011), Monin–Obukhov surface layer scheme (Monin and Obukhov, 1954), the rapid radiative transfer model for general circulation models (RRTMG) longwave (Mlawer et al., 1997) and shortwave radiation schemes (Iacono et al., 2008), and the Morrison double-moment cloud microphysics scheme (Morrison et al., 2009) with improved strati-formed cloud persistence compared to the single-moment scheme. The WRF and WRF-Hydro configurations are listed in Table 1.
WRF and WRF-Hydro model configurations.
The WRF Hydrological modeling extension package (WRF-Hydro) from Gochis et al. (2013) was developed through research collaborations between the National Center for Atmospheric Research (NCAR) and its partners, the National Aeronautics and Space Administration (NASA), and the National Oceanic and Atmospheric Administration (NOAA). It can be used as a standalone model and has an architecture to facilitate coupling between hydrological models and atmospheric models. It provides the capability to spatially relate meteorological variables to physical and terrestrial attributes (elevation, soil, land use, etc.) and their associated feedbacks. Several hydrologic routing physics options exist in the version 3.0 of WRF-Hydro. Here, a fully distributed three-dimensional overland surface flow model configuration was employed. Soil moisture initialization was provided from the WPS along with the predefined soil hydraulic parameters for the Noah-MP LSM on the basis of USGS soil classifications (Ek et al., 2003).
The geographic information system (GIS) Python-based preprocessing approach from Gochis and Sampson (2015) was followed to derive the WRF-Hydro LSM and routing grids. Inputs of the static terrain properties from the WPS GEOGRID file and the high-resolution (30 m) ASTER DEM were used. A re-gridding factor of 10 was applied to reach the 100 m LSM resolution from the 1 km GEOGRID resolution. The minimum basin size was defined by a threshold of 20 pixels per stream, following the analytical method for stream network delineation proposed by Tarboton et al. (1991). Overall, a dominant parallel pattern of stream networks is obtained, draining water westward from the mountainous region in eastern UAE towards the western side of the Arabian Gulf (Fig. 1). The northern part of the UAE shows a large number of streams flowing in the opposite direction towards the eastern side of the country. The sandy nature of the soil in the desert in the southwestern side of the country favors rapid infiltration of runoff and its drainage towards the Gulf as groundwater, before it makes surface around the western coastal regions (Fig. 1). This contributes to the existence of salt flats around those regions. To our knowledge, a dense drainage network which may allow us to identify the main wadis in the western region does not exist in the UAE and this study proposes a first version of such a network. The challenge faced when studying hydrological processes in hyper-arid regions, like the UAE, is the absence of gauged watersheds and lack of in situ data to calibrate and verify hydrological models. In this study, the performance of the hydrological simulations was evaluated using satellite soil moisture data and the analysis of the changes in its lateral extent.
The RMSE and rBIAS methods of evaluation were implemented, with the value observed representing the station measurement, or the satellite
retrieval in the absence of a station observation (Eqs. 1 and 2). The RMSE
and rBIAS reflect the average error and degree of over- or underestimation of
the model output fields.
The Pearson correlation coefficient (PCC) (Benesty et al., 2009) was used in
relating the observed values and model output (Eq. 3). The Pearson
correlation reflects the statistical association between variables and can
range between
Figure 3 shows the cumulative event rainfall over the UAE in d03 between 06:00 Z on 8 March 2016 and 06:00 Z on 10 March 2016. This includes values recorded by the rain gauges (a), values retrieved from the GPM product (b), and values simulated from the WRF and WRF–WRF-Hydro runs (c and d, respectively). The highest cumulative rainfall of 288 mm was recorded by the rain gauge at Al Shiweb weather station (see Fig. 3a) in the northeastern part of the UAE. Lower rainfall rates between 30 to 70 mm were recorded along the coast and western parts of the country. The minimum, maximum, and mean from each rainfall source are listed in Table 2, in addition to the correlation, RMSE, and rBIAS values of the simulated and GPM-retrieved rainfalls against the station records over the UAE.
Accumulated storm rainfall over the UAE (d03) from station
observations
Total storm rainfall statistical measures from collocated GPM retrievals, station observations, and simulations (WRF and WRF–WRF-Hydro).
GPM retrievals recorded a maximum cumulative rainfall of 189 mm over the eastern coast of Oman which is lower than the maximum obtained from the station data in the UAE (readings in Oman are not available). This implies that GPM could have underestimated rainfall records as it did not capture the maximum cumulative reading reported at Al Shiweb weather station (288 mm). This could be explained by the lower temporal coverage of GPM (30 min) compared to the stations (15 min), and this could have led to the former missing the peak of rainfall records. Another reason that could explain the underestimation is the coarse spatial resolution of GPM data (10 km), which senses rainfall over a footprint larger than the one represented by the local rain gauge readings. On the other hand, GPM overestimates cumulative rainfall values ranging between 50 to 90 mm in the northeastern part of the UAE and highlands (see Fig. 3b). Minimum rainfalls between 10 to 40 mm were recorded by stations along the western coast lines and central inlands, while no significant rainfall was recorded over the western areas. In a study that validated a number of precipitation products over UAE for a 10-year period (Wehbe et al., 2017), it was shown that the analyzed remote sensing products perform better at higher elevations (> 250 m), which is the case of Al Shiweb weather station (306 m). In line with the previous study, the lowland stations in the western region showed higher biases by GPM and retrievals over areas not receiving any rainfall during the event. More events are needed to accurately assess the performance of GPM in the region and demonstrate any potential improvement over its predecessor, Tropical Rainfall Measurement Mission (TRMM) (Huffman et al., 2007), among others.
The standalone WRF and fully coupled WRF–WRF-Hydro simulated rainfall exhibited spatial rainfall patterns similar to those obtained from GPM and station data, with the highest rainfall simulated over the coast of Oman and eastern UAE (see Fig. 3c and d). However, the coupled WRF–WRF-Hydro model resulted in a maximum cumulative rainfall of 156 mm over the northeastern border with Oman, which is closer to the 189 mm obtained from GPM readings compared to the 122 mm from the standalone WRF system. Also, more rainfall was reproduced over the western quarter by the standalone WRF than by the coupled WRF–WRF-Hydro where GPM and stations do not report significant readings. Overall, more of the rainfall retrieved by GPM or measured by the stations was reproduced by the coupled WRF–WRF-Hydro model, and a better agreement in the spatial pattern of rainfall and its magnitude is obtained.
More importantly, the lowest rBIAS (0.21) was obtained from the WRF–WRF-Hydro setup, followed by the WRF setup (0.34), and the GPM retrievals (0.71). The higher bias associated with the GPM rainfall retrievals is thought to be related to the difference in spatial scales, ice-scattering microwave retrieval mechanism, and locality of orographic rainfall events, among other factors. The improved performance of the coupled WRF–WRF-Hydro system compared to the standalone WRF is evident across all measures in Table 2 – closer to the observed 48 mm mean and 288 mm maximum (39 vs. 32 mm and 156 vs. 122 mm), higher correlation (0.82 vs. 0.76), and lower RMSE (10.89 vs. 14.24) and rBIAS (0.21 vs. 0.34). The resulting coupling improvements translate into a 24 % and a 13 % decrease in RMSE and rBIAS, respectively. The enhanced precipitation forecast will directly impact land surface processes. The degree of improvement achieved here is in line with the findings of Givati et al. (2016) over the Ayalon basin in Israel during a 1-month simulation of WRF versus WRF–WRF-Hydro, in which they recorded a 21 % decrease in the coupled precipitation RMSE. Figure 4 shows the observed, simulated (WRF and WRF–WRF-Hydro), and GPM-retrieved cumulative rainfall during the event at four selected stations, namely, the coastal Abu Dhabi station (Fig. 4a), inland Al Ain station (Fig. 4b), and mountainous Jabal Hafeet (Fig. 4c) and the Jabal Mebreh (Fig. 4d) stations situated on the northeastern highlands at altitudes of 1080 and 1433 m (see Fig. 1), respectively. The stations were selected to capture the spatial variability of the rainfall across the study domain as well as the potential impact of the topography with stations located at flat open terrain and others at higher altitude. In Fig. 4, time intervals of 15 min were used to derive the cumulative plots, while 30 min intervals were available for the GPM product.
Comparison of accumulated rainfall at Abu Dhabi
RMSE (rBIAS) of simulated and retrieved rainfall at Abu Dhabi, Al
Ain, Jabal Hafeet, and Jabal Mebreh stations for 30 min intervals (
Table 3 summarizes the error measures at each station. In all four cases,
the simulated WRF rainfall recorded the poorest performance and highest
deviations (underestimation and overestimation) from the observed rainfall
pattern with rBIAS values of 0.84,
Comparison of 2 m temperature observations at Abu Dhabi
Figure 5 shows the hourly observed and simulated WRF and WRF–WRF-Hydro 2 m temperature at the stations of Abu Dhabi (Fig. 5a), Al Ain (Fig. 5b), Jabal Hafeet (Fig. 5c), and Jabal Mebreh (Fig. 5d). Both WRF and WRF–WRF-Hydro show a strong warm bias across all four stations during the morning and day hours, with the higher biases occurring during the first day of the simulation (8 March 2016). Also, both models produced smaller cold biases during the night hours. The stations recorded a sharp decrease in temperature between 07:00 Z and 10:00 Z on 9 March 2016 which is temporally consistent with the associated rainfall initiation at each station in Fig. 4. Both simulations with WRF–WRF-Hydro and WRF reproduced the decline in temperature especially at Abu Dhabi (Fig. 5a) and Al Ain stations (Fig. 5b). The decrease in temperature between 07:00 Z and 10:00 Z on 9 March 2016 at the mountain stations, namely, Jabal Hafeet (Fig. 5c) and Jabal Mebreh (Fig. 5d) was not significant. Overall, the difference between WRF–WRF-Hydro and WRF temperatures was more significant at Al Ain and Abu Dhabi stations than at the mountain stations. The decrease in temperature simulated by WRF–WRF-Hydro and WRF between 07:00 Z and 10:00 Z on 9 March 2016 preceded the decline in the observed temperatures, which is in line with the lag between simulated and observed rainfall, especially at the Abu Dhabi station, with simulated rainfall initiated earlier than the observed.
Scatterplot of 2 m temperature from WRF and WRF–WRF-Hydro simulations versus station observations across the UAE.
Figure 6 shows the scatterplot of the hourly 2 m temperature from the WRF and WRF–WRF-Hydro output fields against the temperature observations from all 57 stations in the network. The warm biases appear to increase with the rise in temperature, confirming the previously noted stronger daytime biases and its consistency at all stations.
The 2 m temperature statistical measures from collocated stations and simulation outputs.
Table 4 lists the statistical measures obtained from both simulations
compared to the observed records. Slight improvement (below 1
The increase in rainfall should have increased soil moisture and therefore led to an increase of the latent heat over the sensible heat, causing a decline in air temperature. The improvement of temperature simulation with WRF-Hydro over WRF could be attributed to the capability of WRF-Hydro to simulate soil moisture spatial distribution more reliably than WRF. Nevertheless, both models still show a warm bias. The heat exchange between land and atmosphere which controls the change in air temperature is site specific. It also depends on land cover conditions which defines the roughness length for heat and local topography (macro roughness and surface geometry) which defines the roughness length for momentum (Yang et al., 2008). This is also in combination with the soil moisture effect which impacts surface emissivity, surface temperature, and therefore the simulated air temperature. This explains the different behaviors of temperature at the different sites.
Comparison of global radiation observations at Abu Dhabi
Figure 7 shows the hourly observed and simulated WRF and WRF–WRF-Hydro global radiation at the stations of Abu Dhabi (Fig. 7a), Al Ain (Fig. 7b), Jabal Hafeet (Fig. 7c), and Jabal Mebreh (Fig. 7d). Both WRF and WRF–WRF-Hydro show overestimations, especially for the Jabal Mebreh station, over the first day of the simulation (8 March 2016). Similar to the 2 m temperature warm bias evolution, the radiation overestimations are reduced in the second day of the simulation (9 March 2016). Also, the lower radiation readings during the second day are temporally consistent with the 2 m temperature depressions (Fig. 5) and rainfall initiation (Fig. 4) at each station. It was observed that the WRF–WRF-Hydro simulation better matches the observed temporal variation and magnitudes than the WRF simulation.
Scatterplot of global radiation from WRF and WRF–WRF-Hydro simulations versus station observations across the UAE.
Figure 8 shows the scatterplot of the global radiation from the WRF and WRF–WRF-Hydro simulations fields against the station observations. The WRF–WRF-Hydro radiation variability demonstrates much less deviation from the station observations compared to that of the WRF model.
Global radiation statistical measures from the stations and
simulations at 15 min intervals (
Table 5 lists the statistical measures obtained from both simulations
compared to the observed records. Higher agreement with the observed mean (101.96 W m
Zempila et al. (2016) assessed the performance of different shortwave radiation schemes, namely the Dudhia, updated Goddard and the Goddard Fluid Dynamics Laboratory (GFDL), and RRTMG (used here). They compared the simulated global radiation to a set of hourly measurements at 12 stations over Greece. Overestimations between 40 % and 70 % were recorded for all schemes, while a better agreement was achieved during clear (cloudless) sky conditions. In the present study, improved bias results of 33 % and 21 % biases were obtained from WRF and WRF–WRF-Hydro, respectively (see Table 5).
Cumulative cloud fraction retrieved from MODATM and simulated by WRF and WRF–WRF-Hydro for 9 March 2016
Figure 9 shows the cumulative cloud fraction from the MODATM level-2 retrievals (10 km) and both the WRF and WRF–WRF-Hydro simulations (1 km) for 9 March 2016. The cloud fraction ranges from 0 (cloud-free) to 1 (complete cloud cover). The cloud base altitude was found to be in the range of 5 km from both MODATM, standalone WRF, and WRF–WRF-Hydro, which is also in line with the profiles of Fig. 2. Overall, the simulated cloud cover was reproduced to a much smaller extent, especially overland in the western coastal areas. Díaz et al. (2015) assessed the capability of the WRF model to reproduce clouds over the African region with varying configurations and physics options and their comparisons with satellite observations. They concluded an overall underestimation of cloud cover from their nine WRF simulations, particularly in the case of marine-boundary layer clouds over coastal areas. Their simulations resolved a high number of thick clouds and too few clouds with lower optical thickness. The net result was an underestimation of low cloud cover, which is the case of the present study. Otkin and Greenwald (2008) also examined the ability of WRF to reproduce cloud properties during an extratropical cyclone over the North Atlantic Ocean. Similar to the present study, they relied on MODATM retrievals for model verification with different combinations of cloud microphysics and PBL schemes, and found consistent underestimation. They attributed the underestimation to the utilization of radiance and reflectance data on a 1 km grid by MODATM, and therefore its ability to capture small cumulus clouds, whereas the WRF model horizontal resolution (4 km) failed to explicitly resolve all processes. This is not the case with the present study, given the matching 1 km horizontal resolution used for d03. Hence, model resolution is shown to be of less significance for cloud resolving, while focus should be placed on the inclusion of a cumulus parameterization (not explored here) which may improve the model simulations through a better representation of the subgrid-scale cumulus clouds within this region. It is also possible that MODATM overestimated cloud coverage over the study domain. It is known that in the presence of high reflective surfaces (high albedo) in the background, such as snow, ice, or desert, cloud products become less accurate (Kotarba, 2010).
As a key factor in land–atmosphere interactions, clouds directly impact the radiation balance in terms of the amount reflected, absorbed, and emitted, depending on various cloud physical properties. Therefore, the underestimation of cloud fraction in both model configurations explains the observed overestimation in simulated global radiation, with less reflectance and more radiation reaching the surface. Also, the underestimation of precipitation during the first 12 h (see Fig. 4) is primarily attributed to the models' spin-up time and GFS bias during initialization. This could be supported with Fig. 9 which shows an underestimation of cloud extent in the cases of both WRF and WRF–WRF-Hydro.
Comparison of simulated WRF-Hydro and retrieved AMSR2 soil moisture during (9 March 2016, 00:00:00) and after (10 March 2016, 00:45:00) the event.
Figure 10 shows the soil moisture retrievals from ASMR2 (10 km), the standalone WRF (1-D Noah-MP – 1 km), and the simulated WRF–WRF-Hydro (100 m) during 9 March 2016 at 00:00:00 and after the event at 00:45:00 on 10 March 2016. The increase in soil moisture from both model simulations along the coast and western desert areas after the event on 10 March 2016 at 00:45:00 verifies the anticipated soil exfiltration and runoff drainage direction from the northeastern mountains toward the western lowlands (see Fig. 1). The higher soil moisture values resulting from the WRF–WRF-Hydro simulation are primarily attributed to the increase in simulated precipitation compared to standalone WRF (see Fig. 3).
Scatterplots of AMSR2 and WRF-Hydro soil moisture estimates across the UAE.
Figure 11 shows scatter plots of the soil moisture values from ASMR2 and WRF–WRF-Hydro from both timings after re-gridding at 10 km through least-squares interpolation. The spatial comparison of WRF–WRF-Hydro and AMSR2 soil moisture estimates revealed three increasing soil moisture classes: R1, R2, and R3 (delineated in Fig. 10) with an overall RMSE and rBIAS of 0.07 and 0.08 (8 %), respectively.
The region over the western part of the country (R1) received negligible
rainfall (see Fig. 3), and consequently, recorded the lowest soil moisture
value class ranging from 0 to 0.15 m
The Soil Moisture Operational Products System (SMOPS), provided by the
National Oceanic and Atmospheric Administration (NOAA), merges soil moisture
retrievals from multi-satellites or sensors to generate a global product at
higher spatial and temporal coverage (Liu et al., 2016). Relevant to the
current study period, SMOPS now incorporates near real-time SMAP data and
includes soil moisture retrievals from the GPM Microwave Imager (GMI). The
6-hourly product mapped at 0.25
A comparison of soil moisture evolution at the upstream and downstream of a
wadi within the study domain is expected to verify whether soil moisture
transport occurs over the storm timescale. A wadi within the coverage of the
Saih Al Salem station (24
Time series of simulated soil moisture from WRF–WRF-Hydro at wadi upstream and downstream locations, along with collocated SMOPS retrievals. The hyetograph recorded at the Saih Al Salem station is shown on top.
An increase in water content of the top soil layer decreases both the surface albedo and the Bowen ratio. A lower surface albedo dictates more absorbance of net radiation, while lower Bowen ratios are a result of higher water vapor content in the boundary layer and more downwards flux of terrestrial radiation at the surface due to the water vapor greenhouse effect. This dual effect amounts to a larger total flux of heat from the surface into the boundary layer (Eltahir, 1998). Furthermore, the cooling of surface temperature accompanied by the moisture should be associated with a reduced sensible heat flux and a smaller PBL height. Figure 13 shows the PBL heights from both simulations with larger collapses resolved from the coupled model. According to Seidel et al. (2010), PBL heights can be inferred from radiosonde data (Fig. 2), particularly based on determining maximum or minimum vertical gradients of relative humidity or specific humidity. Such methods yield better agreement compared to those relying on locations of elevated temperature inversions or mixing height. Hence, using the Abu Dhabi radiosonde profile of Fig. 2, and basing on the gradient approach, the PBL height can be estimated to be in the range of 90–200 m at 12:00 Z, which is closer to that simulated from the coupled WRF–WRF-Hydro (190 m) compared to standalone WRF (750 m).
Planetary boundary layer (PBL) heights from WRF and WRF–WRF-Hydro simulations at each of the four stations.
Similar to the present study, Xiang et al. (2017) used the coupled WRF-Hydro system for short-term (72 h) simulations of storm events to discretize the effects of higher soil moisture conditions on precipitation generation, using the framework proposed by Eltahir (1998) to diagnose mechanisms of positive soil moisture–precipitation feedback. They captured an increase of up to 26 mm in WRF–WRF-Hydro precipitation over 48 h, which is in line with the present study with a maximum increase of 23 mm.
Koster et al. (2004) identified regional hot spots, including the Arabian peninsula, and particularly the UAE, where a global initialization of soil moisture may enhance precipitation prediction skill during summer in the Northern Hemisphere. Assuming predominantly local soil moisture impacts, the hot spots indicate where regular monitoring of soil moisture using in situ and satellite observation may lead to an enhancement in boreal summer seasonal forecasting. Their study also referred to a main challenge related to the dependency of the models on soil moisture computational estimates, and especially that a long spin up period might be required to reproduce reliable soil moisture values for seasonal forecasting. Senatore et al. (2015) showed that simulations for one month in a watershed in Italy required a 2-month spin up period. This fosters the importance of deploying dense soil moisture monitoring networks in the region (AlJassar et al., 2015; Temimi et al., 2014; Fares et al., 2013) which should contribute to a better understanding and characterization of soil moisture variability and hydrological processes in desert and hyper-arid environments. The description of routing and lateral flow by WRF-Hydro improved the quality of the simulated atmospheric processes in this case study. This promises improved seasonal precipitation forecasts, as well as short-term predictions assessed here.
In this study, we simulated an extreme weather event in March 2016 over the UAE, a country within the Arabian peninsula of particular interest for hydrometeorological research and monitoring. The event was simulated from both standalone WRF and fully coupled WRF–WRF-Hydro model configurations and compared to station observations and ongoing satellite products. The main objective of the study was to investigate the added value of coupled land surface–atmospheric modeling for precipitation forecasts over the hyper-arid environment of the UAE, while employing current modeling tools to aid in operational forecasting efforts in the region.
Results showed reductions of 24 % and 13 % in RMSE and rBIAS measures, respectively, for precipitation forecasts from the coupled model configuration. Furthermore, the coupled WRF–WRF-Hydro system was found to outperform GPM rainfall retrievals at some stations (e.g., Abu Dhabi and Jabal Mebreh). The demonstrated improvement in coupled precipitation simulation, at the local scale, greatly enhances the accuracy of hydrologic forecasts and flash flood guidance systems. Senatore et al. (2015) explained the higher precipitation with their WRF–WRF-Hydro simulation by the differences in surface temperatures as a warmer surface boundary condition may lead to more convection and therefore higher energy and rainfall. The same interpretation could be also adopted in this study. The lateral boundary conditions are expected to severely restrict our model during these short 48 h simulations, however internal model variability is a direct consequence of the non-linear dynamical and physical internal processes being active and detectable (Christensen et al., 2001). However, despite the more skillful forecasts of the coupled system, the bias remains high (21 %), which dictates the need for ongoing hydrometeorological forecast enhancement.
The coupled system also showed improvements in global radiation forecasts (45 % and 12 % for RMSE and rBIAS, respectively), while less significant enhancements were observed in the case of surface temperature (3.1 % and 1 %). Both parameters were subject to high positive biases during the morning and daytime. The warm temperature biases were attributed to dry biases in the NCEP-GFS boundary conditions observed by Chaouch et al. (2017) and Yang et al. (2011) and the uncaptured cooling mechanisms of aerosols and sea breezes, while the underestimation of cloud cover explained the overestimations in global radiation. The diurnal temperature signal was not captured, even in the observed values, due to the extreme event. Also, the discrepancies in temperature simulations could be caused by soil moisture simulation and its spatial organization within the study domain which impacts the latent heat, the heat exchange, and therefore the temperature difference. More importantly, higher spin up times (6 h used here) can add to the model accuracy in terms of both atmospheric dynamics and hydrological processes (Lo et al., 2008). Soil moisture validation – a challenging application over arid regions – showed varying response classes across the UAE and were consistent with the expected surface flow directions. The difference of spatial scales between the retrieved and simulated soil moisture, and the impact of high reflectance from desert land cover on the AMSR2 microwave retrieval algorithms may have contributed to the observed discrepancies (Wehbe et al., 2018).
The fully coupled model configuration captures the complete dynamics of the water and energy cycles, starting from the upper atmosphere to the unsaturated and saturated zones on the land surface and back. Land surface–atmospheric interactions are primarily governed by two key hydrological parameters, namely rainfall and soil moisture. Hence, future work with in situ soil moisture data assimilation is expected to enhance the model accuracy, both overland and in the atmosphere, through the captured feedbacks. This case study focused on a regional event triggered by a large-scale system. Hence, the impact of accounting for hydrological processes through the online coupling is not expected to be significant. Nevertheless, an improvement in the simulation of precipitation was obtained with the coupled WRF–WRF-Hydro model. To further discretize the added value of the coupling demonstrated in this study, ensemble approaches should be the focus of future work in order to assess the robustness of the potential improvements with WRF-Hydro. In the present study, the exact contribution of lateral flow versus internal atmospheric variability on the captured improvement remains an open question that is subject to further research in this area, particularly for arid environments that have not been receiving much attention.
The GPM data are provided by the NASA's Goddard Space Flight Center and archived at the Goddard Earth Sciences Data and Information Services Center (GES DISC)
(
MT and YW conceived and designed the case study. YW performed the simulations, postprocessing, and data analyses. YW and MT analyzed the results and wrote the article. MW, NC, OB, and TS advised on compiling and configuring the model in coupled mode. VW reviewed the results and revised the article. XZ and JL provided the NOAA SMOPS data and reviewed results. AA ensured quality control of the weather station observations.
The authors declare that they have no conflict of interest.
The authors acknowledge Khalifa University of Science and Technology (KUST) for the provided support. The authors would also like to thank the UAE National Center of Meteorology (NCM) for the quality-controlled weather station observations used in this study. The authors thank the NCM executive director, Abdulla Al Mandous, and program director, Alya Al Mazroui, for their continued support.
This research has been supported by the Khalifa University of Science and Technology (grant no. SSG2015-000060). This work was also conducted within the framework of a project funded by the NCM under the UAE Research Program for Rain Enhancement Science (UAEREP).
This paper was edited by Gregor C. Leckebusch and reviewed by Nir Krakauer and one anonymous referee.