Evaluation of Nine Operational Models in Forecasting Different Types of Synoptic Dust Events in the Middle East

Abstract

This study investigates four types of synoptic dust events in the Middle East region, including cyclonic, pre-frontal, post-frontal and Shamal dust storms. For each of these types, three intense and pervasive dust events are analyzed from a synoptic meteorological and numerical simulation perspective. The performance of 9 operational dust models in forecasting these dust events in the Middle East is qualitatively and quantitatively evaluated against Terra-MODIS observations and AERONET measurements during the dust events. The comparison of model AOD outputs with Terra-MODIS retrievals reveals that despite the significant discrepancies, all models have a relatively acceptable performance in forecasting the AOD patterns in the Middle East. The models enable to represent the high AODs along the dust plumes, although they underestimate them, especially for cyclonic dust storms. In general, the outputs of the NASA-GEOS and DREAM8-MACC models present greater similarity with the satellite and AERONET observations in most of the cases, also exhibiting the highest correlation coefficient, although it is difficult to introduce a single model as the best for all cases. Model AOD predictions over the AERONET stations showed that DREAM8-MACC exhibited the highest R² of 0.78, followed by NASA_GEOS model (R² = 0.74), which both initially use MODIS data assimilation. Although the outputs of all models correspond to valid time more than 24 h after the initial time, the effect of data assimilation on increasing the accuracy is important. The different dust emission schemes, soil and vegetation mapping, initial and boundary meteorological conditions and spatial resolution between the models, are the main factors influencing the differences in forecasting the dust AODs in the Middle East.

1. Introduction

Dust storms are considered environmental hazards due to their high impact on climate change, widespread environmental degradation, financial and human losses [1,2,3]. The impacts of dust aerosols on radiative forcing, atmospheric circulation and climate, especially in the Middle East, have been well recognized in previous studies (e.g., [4,5,6]). Air pollution arising from dust storms poses an extra threat to public health with serious respiratory and heart diseases [7,8,9]. The relationship between mortality or morbidity and the occurrence of dust storms has been extensively studied even in areas outside the dust belt such as Korea [10], Taiwan [11] and Greece [12] among many others [13]. Decreased horizontal visibility due to massive dust storms, causes traffic accidents and increases the death rate every year [14,15,16,17]. Furthermore, land degradation by wind erosion affects many parts of the world [18,19,20,21,22], and therefore, the occurrence of dust storms has destructive effects on crop growth, livestock production, agricultural infrastructure and ecosystem health [23,24].

According to global data from satellite observations, ground stations and model simulations, the amount of dust that enters the atmosphere reaches several hundred to several thousand tons each year [25]. For example, Zender et al. [26] calculated the annual average of global dust emissions in the 1990s to be 1490 ± 160 (Tg yr⁻¹), while Ginoux et al. [27] estimated the total annual dust emissions between 1604 and 1960 Tg yr⁻¹ in a 5-year simulation study. Huneeus et al. [28], based on global model ensemble simulations, estimated the global dust emissions to about 500–4000 (Tg yr⁻¹), of which 400–2200 (Tg yr⁻¹) is related to North Africa and 26–526 (Tg yr⁻¹) is emitted from the Middle East. Many studies have been conducted to identify dust sources in the Middle East, mainly across the Arabian Peninsula, Syria, Egypt, Iraq and Iran [29,30,31,32]. Moridnejad et al. [33] used MODIS images of 70 dust storms that occurred between 2001 and 2012 to develop a new map of the main dust sources in the Middle East. Their results showed that 247 different points were involved in the dust production in the Middle East; among them, Iraq and Syria have the most dust sources in the region.

Synoptic analysis of meteorological patterns and weather clustering techniques during dust storms in the Middle East has been studied by many researchers [34,35,36,37]. Alamodi [38] studied a severe dust storm in Saudi Arabia on 13 May 2005. He reported that the low-pressure movement from Europe to the southeast, over Turkey, along with Sudan low-pressure shift to the northeast toward central Saudi Arabia, caused strong pressure and temperature gradients and high-speed winds triggering a severe dust storm over Saudi Arabia. Hamidi et al. [39] examined 60 dust storms over a period of 9 years (2003 to 2011) in the Middle East and 12 of them were analyzed from a synoptic perspective. They concluded that dust storms in the Middle East can be divided into two main categories i.e., Shamal and frontal dust storms [39]. Other studies have divided dust activity in the Middle East into three main types (i) Shamal, (ii) frontal and (iii) convective dust storms [40,41,42,43], while many of them have focused on Shamal dust storms (e.g., [44,45,46]). On the other hand, one of the main factors for the formation of dust storms, especially in winter and spring, is the presence of atmospheric fronts. Strong winds accompanied by cold or warm fronts and low humidity levels facilitate intense dust emissions. Hamzeh et al. [47] investigated six frontal dust storms in the Middle East. They concluded that the presence of a deep trough over Turkey, Syria and northern Iraq plays a key role in the formation and transport of frontal dust. Karami et al. [48] studied the frontal dust storm on 19 January 2018 over Syria, Iraq and Iran. In that dust event, the presence of a cold front over the region caused high dust emissions from the deserts in Iraq and Syria and transport of dust over long distances. On the other hand, some dust events are due to formation of a dynamic low-pressure system and are not necessarily associated with atmospheric fronts (e.g., [35,43,49]). Al-Jumaily and Ibrahim [50] analyzed two severe dust storms on 22 February 2010 and 10 December 2011 in Iraq, indicating the presence of a deep low-pressure system that was formed in the area during these dust storms. This type of synoptic dust storm is mentioned here as cyclonic dust. More recently, Mohammadpour et al. [36,37] examined the synoptic meteorological conditions and the dominant weather clusters that are associated with dust accumulation over specific parts of Iran, with more emphasis on the western parts that are affected by dust storms of various types from the Iraqi plains.

Due to the importance of forecasting dust storms in order to mitigate their negative impacts, many studies have been conducted to evaluate the performance of numerical dust forecasting models in the Middle East [47,51,52,53]. Beegum et al. [34] investigated the predictability of seven dust events in the Arabian Peninsula using the CHIMERE model coupled with the WRF weather forecast model. Five of those dust events were associated with synoptic-scale meteorological dynamics, such as frontal systems and summer Shamal winds, while two of them were meso-scale convective storms. Kontos et al. [54] parameterized the dust emissions in the Middle East, while Karami et al. [55] compared the outputs of DREAM-MACC, CAMS and WRF-Chem models during two dust storms originated from the desiccated Aral Sea and Hamoon lakes. That study showed that the CAMS model represented better the AOD values, likely related to MODIS data assimilation, although all models underestimated the high AODs compared to Terra-MODIS retrievals, especially near the dust source. There are some difficulties in calculating AOD by dust forecasting models, with the uncertainties mostly related to the complex refractive index (scattering and absorbing properties of dust), as well as to the size distribution (relative partitioning of the fine and coarse modes) [56]. So, for satisfactory predictions of the dust-AOD, scattering, absorption processes and particle size distribution should be accurately parameterized in the model, as they have different responses to AOD values. Ukhov et al. [56] revealed this challenging issue and the effect of fine particles in dust AOD calculations for the WRF-Chem model. Huneeus et al. [28] compared the outputs of 15 global aerosol models in the AeroCom project. The compared outputs included different parameters such as AOD and dust deposition. Since AOD is highly dependent on the particle load, refractive index and size distribution, their results were so different in various areas, since some models overestimated the AOD values in some areas and underestimated them in others. Flaounas et al. [57] showed that the model AOD outputs were more realistic when dust emissions were over-estimated and more accurate dust concentrations were obtained when the AOD values were less than the observational ones.

In this study, the performance of nine operational dust forecasting models included in the WMO Sand and Dust Storm Warning Advisory and Assessment System (WMO SDS-WAS; https://sds-was.aemet.es/, accessed on 12–14 April 2021) is examined over the Middle East during several dust events (3 in each case, 12 in total) of different types (mechanisms) of dust outbreaks such as (i) cyclonic dust, (ii) post-frontal dust, (iii) pre-frontal dust and (iv) Shamal dust. The performance of the model forecasting is examined against Terra-MODIS AOD observations and via statistical indicators and Taylor diagrams. Furthermore, ground measurements of AODs at AERONET stations over the region affected by these dust storms were also used for the evaluation process. Selected dust events of each type are also analyzed from the meteorological perspective in order to reveal the different mechanisms that triggered the dust emission and transport. This work provides a comprehensive evaluation of the dust forecasts over the Middle East, aiming to reveal the challenges and biases of each model. An important aspect of this study is the model evaluation during four different types of dust events in the Middle East, aiming to identify which model is more reliable for a certain dust type.

2. Study Area, Materials and Methods

The Middle East includes vast deserts and is strongly affected by dust throughout the year. Syrian Desert, Al-Nafud, Al-Dahna, Rub Al-Khali, Dasht-e Kavir and Dasht-e Lut are the most important desert areas in the Middle East. The prevailing wind regime in the Middle East region is north and northwesterly, especially during the summer season, which in local language is referred to as “Shamal wind”. Shamal is a strong low-level northwesterly wind blowing over Iraq, Saudi Arabia and Kuwait, often strong during the day, but decreasing at night. Sometimes, Shamal blows also in winter and is associated with a cold front over the eastern Mediterranean and western parts of the Middle East [47,58], facilitating severe dust storms that may affect many countries [59,60,61]. Abdi Vishkaee et al. [44] analyzed a severe dust storm over Iraq and northwest Iran in 22–23 February 2010 which resulted from a strong winter Shamal after the passage of a cold front.

In general, dust storms are divided into two main categories based on the factors affecting their formation and their spatial extent, namely (i) synoptic and (ii) mesoscale [62]. In the Middle East, one of these effective factors is Shamal wind, which triggers intense dust storms [46,63]. In summer, a low-pressure system develops over northwestern India and Pakistan, extending into Afghanistan, south Iran and the eastern Arabian Peninsula [64], which in conjunction with the high-pressure system over the eastern Mediterranean [36,65], causes a strong pressure gradient and northwesterly winds across the Iraqi plains [66]. Furthermore, in summer, a dominant high-pressure system is usually observed over the central-eastern part of the Mediterranean Basin, which leads to a prevalence of dry conditions over the region [65]. The summer Shamal wind is strengthened by the Zagros Mountains in southwest Iran, via the channeling effect of the northwesterly currents [46]. In winter, the presence of atmospheric fronts causes strong surface winds, which in the absence of sufficient moisture, may cause massive dust emissions. In these cases, a dust storm may occur in front of or behind the frontal system, characterized as pre-frontal or post-frontal dust, respectively [47,67]. Sometimes a dynamic low-pressure center that is not necessarily associated with a front may also cause dust emissions [43,68]. In this case, the presence of high relative vorticity, convergence and upward motions at the surface initiate the dust outbreak [69].

In this study, four types of synoptic dust events in the Middle East (cyclonic, pre-frontal, post-frontal and Shamal dust) are investigated. The first mechanism for the dust emission is due to the presence of a surface low-pressure associated with intense near-surface winds [43,68,69]. In these cases, the rotation of dust particles is observed around the low-pressure closed center, and due to the upward movements in the center of the cyclone, the dust particles are transported to higher altitudes. The main difference from this dust type against Shamal dust storms is that cyclonic dust is driven only by strong winds around a cyclone (mostly thermal) over the desert terrain and not by large-scale pressure gradients that trigger the Shamal wind, which covers a large area in the Middle East. In post-frontal and pre-frontal dust storms, dust is emitted from the surface due to strong winds behind and in front of the frontal areas, respectively, and the dust mass is mainly observed as a strip parallel to the atmospheric fronts [47]. Because of the greater instability in the cold front, dust is usually seen in front or back of the cold front. If the position of the front is such that the dust-producing areas are located behind it, the dust occurs behind the front and if it is in front of it, the dust occurs before the front. The prevailing wind direction in the post-frontal areas is northwesterly and in the pre-frontal areas is south or southwesterly. The Shamal dust storms, as discussed above, mainly occur during the summer season, as a large part of Iraq and Saudi Arabia is under the influence of the strong northwesterly Shamal wind [46,52,63]. This wind is the result of the low-pressure tongue from the Arabian Sea into the northern regions of Iraq, the return branch of which causes the northwesterly winds to blow.

Three cases of intense and pervasive dust events of each type were identified and analyzed from a synoptic perspective, in total 12 dust events, summarized in

Table 1. Case studies of the synoptic dust events examined here.

Type/Mechanism	Date
Cyclonic	1 September 2015; 29 September 2016; 3 December 2016
Pre-Frontal	13 April 2017; 30 March 2018; 14 March 2020
Post-Frontal	17–18 February 2017; 2–3 November 2017; 8 January 2020
Shamal	5 July 2017; 29 June 2018; 26 July 2018

. We chose these synoptic dust events with criteria for the different mechanisms (4 types), the spatial extend, focusing on dust events that occurred on a large scale and lasted more than one day, and the availability of model predictions. Furthermore, the model AOD forecasting for these dust events is compared against the MODIS spatial AOD distribution in order to evaluate the performance of each model.

In order to determine the dust-affected areas, true-color images and AOD retrievals from Terra-MODIS observations over the Middle East are examined. Global Forecast Grids (GFS) re-analysis data (https://rda.ucar.edu/datasets/ds084.1/, accessed on 17 April 2021) with 0.5° × 0.5° horizontal resolution are used to prepare synoptic maps of the mean sea level pressure (MSLP), temperature and geopotential heights at 850 and 500 hPa levels during the dust storms. Due to the impact of various factors (dust schemes, initial and boundary conditions, soil characteristics, dynamic processes, spatial resolution) that affect the forecasting and simulations of dust events from numerical models [43,51,53,70], the outputs of 9 operational dust forecasting models included in https://sds-was.aemet.es (accessed on 12–14 April 2021) are examined and inter-compared during the selected dust storms in the Middle East. The AOD outputs of the models related to their forecasting (initial time 24 h before the valid time) were extracted from the archives of their data. The models that are examined in this study are BSC-DREAM8b-V2 [71,72], DREAM8-MACC [25], DREAMABOL [73], EMA-RegCM4 [74], MACC-ECMWF [75], NASA-GEOS [76], NCEP-NGAC [77], NMMB-BSC [78] and NOA-WRF-Chem [79]. The main characteristics and numerical equations of these models (dust schemes, initial, boundary conditions, etc.) are briefly described in Appendix A. To evaluate the performance of these models in forecasting the examined dust storms, the AOD outputs over a selected spatial domain in the Middle East are qualitatively and quantitatively evaluated against the Terra-MODIS satellite data with 1° × 1° spatial resolution (level 3), retrieved by combined Dark Target and Deep Blue algorithms. MODIS AOD₅₅₀ from Collection 6.1 and level 3 [80] are derived from the Giovanni tool for data access and visualization (https://giovanni.gsfc.nasa.gov, accessed on 15 April 2021). Since the resolution of all model outputs is finer than that of the observational data, the satellite pixels are considered as the basis, and the AOD values are compared pixel by pixel, after rescaling the model outputs. There are 2386 pixels in the study domain, but some of them may not be valid on certain dust events due to cloud contamination. Taylor diagrams, including statistical indicators of Pearson correlation coefficient (r) and standard deviation, are used to evaluate the performance of each model for the different types of dust storms. Furthermore, ground-based AOD data from 3 AERONET stations mostly affected by these dust storms, i.e., Kuwait University (29.32° N, 47.97° E), KAUST-campus in Saudi Arabia (22.305° N, 39.103° E) and Masdar Institute (24.442° N, 54.617° E) in the United Arab Emirates, were used during the dust-event days, collocated in time with satellite and model-forecasting outputs, in order to evaluate the model’s performance.

3. Results and Discussion

The results section is separated into three sub-sections including satellite observations, synoptic analysis of the selected dust storms and evaluation of the numerical model outputs for dust forecasting.

3.1. Satellite Observations

Figure 1 shows the Terra-MODIS true color imagery during selected dust storms of each type, while the rest cases are included in Figure S1 (in the Supplementary Materials). In the first case of a cyclonic dust storm (Figure 1a: 2 December 2016), intense dust plumes existed over the eastern regions of Syria and north Iraq. More or less similar dust-covered areas in north Iraq are observed in the other cyclonic-dust days (Figure S1), while common characteristics of these dust storms are their relative narrow expansion and the oval shape due to strong rotation of wind around a deep low-pressure center [43]. The post-frontal dust storm (Figure 1b) was mainly observed over the mid-south part of the Iraqi plains and affected south Iraq, southwest parts of Iran and northeast Saudi Arabia. In the pre-frontal dust storm of 30 March 2018 (Figure 1c), dust covered a large area from west Arabia to south Iraq and west Iran, lying in front of or below cloud masses that were associated with a frontal system. These dust storms were caused by cold fronts that usually stretched from southeast Turkey and northern Iraq to northeast Saudi Arabia [44]. The Shamal dust storm during 5 July 2017 (Figure 1d), stretched across the Iraqi-plains to southwest Iran and Kuwait, also affecting the eastern part of Saudi Arabia and the northern Persian Gulf. The other Shamal dust storms (Figure S1) affected similar areas in the Middle East and the Arabian Peninsula.

Figure 2 shows the spatial distribution of the Terra-MODIS AODs during four dust events (rest are included in Figure S2), indicating a remarkable spatial heterogeneity of the AOD values and the contribution of various sources in dust emissions, which are mostly detected in the Iraqi plains (Tigris-Euphrates alluvial plain), along the eastern part of the Arabian Peninsula, Rub-Al-Khali and Oman Deserts. The cyclonic dust storm on 2 December 2016 showed a dust hotspot over northern Iraq due to cyclonic conditions that will be analyzed in the following, while the spatial AOD distribution in this case shows increased dust loading over Kuwait and parts of the Persian Gulf and coastal Iran as well (Figure 2a), attributed to other local dust plumes.

On 18 February 2017 (Figure 2b), the maximum AOD is observed in southwest Iran and southern Iraq, associated with a frontal system over this area, which is analyzed in the next section. On this day, increased AOD values were also observed in southern Saudi Arabia, Oman and Yemen.

A characteristic pre-frontal dust storm occurred on 30 March 2018 (Figure 2c), when high AODs were observed from Sudan to west Iran, indicating the presence of a huge and long-expanded frontal dust storm.

The examined Shamal dust storm (Figure 2d) exhibited increased dust AODs along the corridor from the Iraqi plains to the east Saudi Arabia, which is the area affected by the summer Shamal wind [46,63], and over parts of the southern Arabian Peninsula and the south Red Sea.

3.2. Synoptic Meteorology

In this section, the synoptic meteorological patterns of selected dust events of each type are analyzed, while the rest of the cases are included in Figures S3–Figure S6. Figure 3 shows the MSLP patterns superimposed with 10-m vector winds, the geopotential heights and temperature at 850 hPa level and geopotential heights at 500 hPa level for the selected dust events.

In the case of the cyclonic dust storm on 2 December 2016 (00:00 UTC), the MSLP pattern shows a low-pressure center in eastern Turkey and northern Syria, while its southern expansion affected central Iraq (Figure 3a). This low-pressure center facilitated cyclonic winds around it and dust advection over the western part of Syria and northern part of Iraq (Figure 1a and Figure 2a). Deep low-pressure centers are responsible for massive dust storms of this type over Syria and Iraq, as was the case in September 2015 [43]. At 850 hPa level (Figure 3b), a low-altitude closed center (cut-off low) was formed over northern Syria, southeast Turkey, with a 1340 gpm contour at the center, while its trough reached southeast Iraq. A strong temperature gradient was observed over central Iraq, while at 500 hPa (Figure 3c), a relatively deep trough was observed over the eastern Mediterranean and western part of the Middle East, with surface low-pressure in front of its axis in eastern Turkey and northern Syria. The upward motions in front of the trough axis caused surface low-pressure reinforcement in this area.

Synoptic analysis of the post-frontal dust storm on 17 February 2017 (12:00 UTC) shows that a low-pressure system with two closed centers at the surface was formed in the central regions of Iran (Figure 3d). The cold front of this system was stretched from central Iraq to eastern Saudi Arabia, associated with extensive cloudiness in its northern part (Figure 1b). High-pressure conditions dominated behind the frontal system over Syria and northwest Iraq, associated with strong northwesterlies over the area, which were shifted to southwesterlies in front of the frontal system over central Iran and southeastern Arabian Peninsula. This is a typical MSLP pattern for post-frontal dust storms in the Middle East, mostly affecting the Tigris-Euphrates Basin and the northern Persian Gulf [47]. At 850 hPa geopotential height (Figure 3e), a low-altitude system (cut-off low) was observed over the northern parts of Iraq (Kurdish cut-off low), which was characteristic for dust events affecting southwest Iran in late winter and spring [37,81]. This “cut-off” low was further intensified by the deep trough at 500 hPa level (Figure 3f) extended from the Caspian Sea and northwest Iran towards central Iraq, where the low surface pressure was located in front of its axis and was strengthened by the upward motions, thus triggering the frontal dust storm [47,82].

At 00 UTC on 30 March 2018, the MSLP pattern demonstrated a narrow low-pressure center formed over southern Iraq, accompanied by a cold front extended from Iraq to the west of Saudi Arabia (strong pressure gradient at 850 hPa level) (Figure 3g,h). The pre-frontal dust storm revealed large AODs just ahead of the frontal area covering a large part of the Middle East with a southwest to northeast axis (Figures S1 and S2, in the Supplementary Materials). At 850 hPa level, a low-altitude center was formed just above the surface low-pressure over southern Iraq, whose trough penetrated as far as the Persian Gulf. Baroclinicity and cold advection were observed over central Iraq. At 500 hPa (Figure 3i), the trough of a low-altitude center over eastern Turkey reached the northern parts of Saudi Arabia, while advection of warm air occurred ahead, from southeast Arabia to central Iran. Mid-troposphere troughs over the northwest parts of the Middle East (eastern Turkey, northern Syria) are a major dynamic force for pre-frontal dust storms over southern Iraqi plains [47].

Shamal dust storms usually occur in summer [46,63,83], when the Indian thermal low expanded over southern Iran and the Arabian Peninsula dominates [34,37]. This was also the case on 5 July 2017 (Figure 3j), since the Iraqi plains and the eastern part of the Arabian Peninsula were under the influence of an extended low-pressure system associated with the Indian summer monsoon. Nearly the whole Middle East land exhibited very high temperatures favoring the presence of thermal lows in the surface and the lower troposphere (850 hPa), while over the eastern Mediterranean and North Africa higher pressure conditions dominated (Figure 3k), which caused a strong pressure gradient and low-troposphere winds over the eastern Syria and Iraq. The 10-m wind field showed strong northerly and northwesterly winds over the southern half of Iraq and a large part of eastern Saudi Arabia, associated with enhanced dust AODs over these areas (Figure 2d). At 500 hPa geopotential height map (Figure 3l), two high-altitude centers were seen. One was located in the northern part of Iran and the Caspian Sea and the other was extended from the west of the Red Sea to Saudi Arabia, which is a characteristic 500 hPa atmospheric pattern in summer [84]. A low-altitude center (cut-off low) was located between these high-pressure systems over southern Iraq, which was a rather rare situation, as the Red Sea/Iranian trough is usually continuous [36,37,84].

3.3. Model Simulations of AOD during Dust Events

This section presents the model forecasting of the AOD spatial distribution for selected events of each dust type, aiming to qualitatively compare the outputs from the 9 models. For this evaluation, one case was selected for each dust type, same as those in Figure 1, Figure 2 and Figure 3, while the model results for the remaining case studies are included in the Supplementary Materials. The model AOD distributions are presented at 7:00 UTC (10:30 LST) close to Terra-MODIS overpass time.

3.3.1. Cyclonic Dust Storms

On 2 December 2016 (Figure 4), all models except the EMA-RegCM4 predicted much lower AODs than those observed by MODIS over the Middle East and especially over the dust-source regions that were located in the northern part of Iraq (cyclonic dust storm) and in a vast area in southern Iraq, Kuwait and northeast Saudi Arabia (Figure 2a). In general terms, all models, despite the differences between them, forecasted enhanced AODs over these regions, indicating a satisfactory representation at least for the spatial distribution of the AOD. However, DREAMABOL and NMMB-BSC models did not forecast well the enhanced AOD over north Iraq and showed dust presence only over the surrounding areas of the northern Persian Gulf.

On the other hand, MACC-ECMWF, NOA-WRF-Chem, BSC_DREAM8b and DREAM8-MACC models exhibited greater similarity with MODIS regarding the spatial distribution of AOD. Among these models, the BSC_DREAM8b_V2 underestimated the AOD values especially over Saudi Arabia but showed a closed center of dust mass in northern Iraq, which was related to surface low pressure and originated from arid regions around Tharthar Lake [32]. The failure of some models to identify this closed dust center could be likely attributed to underestimation of wind speed and dynamic processes around the low-pressure center that triggered the cyclonic dust storm. Since the wind speed and threshold wind speed significantly affect the surface dust emissions, and dust flux is usually proportional to the third power of wind speed, errors in the forecasting of the wind regime may lead to large uncertainties in predicting dust emissions and concentrations [85,86]. Tegen and Miller [87] compared the surface wind speed of two models and resulted that the differences in dust emissions between the models were also dependent on the model resolution and parametrizations of the atmospheric boundary layer. In general, many studies have indicated that the accuracy of the atmospheric/meteorological model plays an important role in forecasting/simulating dust events (e.g., [88,89]). Global models are known to underestimate the dust-lifting power of the wind [90,91], thus predicting generally lower dust emissions and related AODs. In our case, the presence of smaller dust sources in the northern part of Iraq [92] compared to the vast deserts in Saudi Arabia is another factor that models produce larger discrepancies in forecasting dust in north Iraq and east Syria. In some dust emission schemes such as that in EMA-RegCM4 model, larger dust emissions and AODs are considered over the arid regions dominated by sand, such as the Arabian Desert, than areas covered with fine clay and silt particles in central Iraq [47,54]. Based on the Food and Agriculture Organization (FAO) soil map, which is widely used in dust forecast models, there is more clay fraction in vast parts of Iran and north half of Iraq but there is more sand fraction in the whole Saudi Arabia and Turkmenistan [47]. The model forecasts of the AOD distributions during the other two cyclonic dust storms are included in Figures S7 and S8 (in the Supplementary Materials), indicating similar results about the models’ performance in forecasting the dust storms, also exhibiting an AOD overestimation from EMA-RegCM4 compared to other models.

3.3.2. Post-Frontal Dust Storms

The AOD outputs from the 9 models during the post-frontal dust storm on 18 February 2017 (Figure 5) showed that all models underestimated the MODIS AODs (Figure 2b) in southwest Iran, south Iraq, Kuwait and north part of the Persian Gulf, the areas that were most impacted by the post-front dust storm and exhibited the maximum AODs (Figure 2b). However, most models predicted the highest AODs over southern Saudi Arabia (Rub-Al-Khali Desert), in general agreement with MODIS, although the differences in the AOD values and spatial distribution between the models are especially high, indicating large discrepancies. Therefore, the models present some difficulties in forecasting of this post-frontal dust storm over southern Iraq and southwest Iran, as also noted by Hamzeh et al. [47]. BSC_DREAM8b_V2 and DREAM8-MACC models represent satisfactorily the post-frontal dust storm, whereas EMA-RegCM4 highly overestimated the dust emissions from Saudi Arabian Deserts. Furthermore, NMMB-BSC predicted very high AODs, corresponding to intense dust storm transported from Rub-Al-Khali to southern Iran. NOA-WRF-Chem model also forecasted significant emissions from the Rub-Al-Khali Desert but not from the southern Iraqi plains. The surface dust flux in many models is calculated by Ginoux et al. [27] dust source function, which simulates enhanced emissions from the most erodible regions [26,93]. So, it’s necessary to use modified source functions to get more agreement with the observational data [94]. The use of dust source functions based on topographic depressions [27] or remote sensing dust load [95] improves the forecasting of the dust life cycle [26,96], but the use of these semi-empirical source functions cause missing of dust particle physics. Other models do not consider the availability of sediment and soil properties in the calculation of dust flux [28]. On the other hand, discrepancies in the detection of the position and strength of the frontal system is a main source of error in frontal-dust prediction [47]. The results for the other two post-frontal dust storms are included in Figures S9 and S10 (in the Supplementary Materials). Some models highly underestimated the MODIS AODs, while significant discrepancies and differences in dust forecasting were observed between the models.

3.3.3. Pre-Frontal Dust Storms

All dust models revealed similar outputs regarding the spatial distribution of the AOD during the pre-frontal dust storm on 30 March 2018 (Figure 6). The models generally forecasted high AODs in the zone covered by the extended dust plume on that day, starting from Sudan and the Red Sea till northwest Iran and the Caspian Sea, although cloud cover along the dust plume prevented satellite retrievals (Figure 2c). However, the large expansion of this intense dust plume was not represented equally by all models, since several revealed much fewer dust AODs over the central parts of the Arabian Peninsula. Furthermore, the dust AODs may be underestimated by some models (i.e., BSC_DREAM8b_V2, NASA-GEOS, NCEP-NGAC) in the central part of Iran. In general, qualitative comparison with the satellite observations revealed that the dust AOD patterns obtained by the MACC-ECMWF and NOA-WRF-Chem models exhibited good performance. One of the effective factors in forecasting and simulating atmospheric dust is the surface wind speed. Most models consider that the amount of dust flux is proportional to the higher powers of surface wind speed or friction velocity. Because of the rapid change in wind speed and direction during the passage of frontal systems, many models are incapable or face difficulties in representing such sudden changes, leading to large biases in dust forecast [97]. The Zagros mountains in the west and southwest Iran and the rough topography of the Middle East are other factors that lead to uncertainties in the simulations of dust life cycle and concentrations due to their impact on the wind regime [51,98,99]. Model forecasts for the rest pre-frontal dust storms are included in the Supplement (Figures S11 and S12).

3.3.4. Shamal Dust Storms

As stated above, the Shamal dust storms mainly occur in summer, when higher AODs dominate over the southern Arabian Peninsula and the Arabian Sea due to increased dust activity as a result of low-pressure conditions and strong winds [83,100].

On 5 July 2017 (Figure 7), the comparison of all model outputs revealed very high variability and differences in the AOD spatial distribution. All models, except DREAMABOL, underestimated the AOD values, especially over Iraq, in the source area of the Shamal dust storm. The AOD outputs of the NASA-GEOS model were mostly similar to the observational data except over Oman. Other models showed increased AODs of various intensity over specific regions such as Iraqi plains, Rub-Al-Khali Desert, Nubian Desert in Sudan and in Oman Desert, while MODIS observations also revealed enhanced AODs due to dust plumes over these regions (Figure 2d). There are several factors that affect critical parameters for dust emissions such as surface heat flux [101], land use [102], topography [103] and surface viscosity properties [104]. On the other hand, pressure gradient [63], orography [51,105] and sea-breeze circulation [106] are the main factors controlling the variability of Shamal wind, which can lead to considerable errors in predicting dust emissions from Shamal dust storms, in case of inaccurate representation [42,107]. The model forecasts during this Shamal dust storm, as well as those shown in the rest cases (Figures S13 and S14, in the Supplementary Materials), indicated that accurate representation of temperature, pressure gradients, wind speed and direction is the most important and challenging issue in numerical predictions of Shamal dust storms.

3.4. Model Evaluation

This section evaluates the models’ performance in forecasting the spatial AOD distributions against Terra-MODIS observations, including all 12 dust events over the Middle East (main text and Supplementary Materials). In this respect, the Taylor diagrams of Pearson correlation coefficient (r) and standard deviation, along with root mean square error (RMSE) values are used (Equation (1)):

$$\mathrm{RMSE}=\sqrt{\frac{{\sum }_{\mathrm{i}=1}^{\mathrm{N}}{\left({\mathrm{I}}_{\mathrm{m}}-{\mathrm{I}}_{\mathrm{c}}\right)}^2}{{\mathrm{N}}_{\mathrm{t}}}},$$

(1)

where I_m and I_c are the measured (observed) and forecasted AODs, respectively and N_t is the total number of valid pixels in the studied model domain. Taylor diagram is used to determine the similarity between the modeled and observational data in the form of three statistical quantities: the Pearson correlation coefficient, the standard deviation and the root-mean-square error (RMSE). The Pearson correlation coefficient is associated with the azimuthal angle in the diagram; the standard deviation is related to the radial distance from the origin point and the RMSE is proportional to the distance from the observed point on the x-axis.

Figure 8 shows the scatter diagrams of AODs over the studied domain between model predictions and Terra-MODIS retrievals for the three cases of cyclonic dust storms. The results show that all models underestimated the AODs, and this difference increased with increasing AOD values, indicating less model performance near the dust source. Among the models, EMA_RegCM4 presents a high fraction of overestimated AODs compared to MODIS, while NMMB_BSC model systematically underestimated the MODIS-AODs. The DREAM8_MACC and NASA_GEOS models exhibited the largest r values (0.718, 0.703), while NOA-WRF-Chem exhibited the lowest RMSE value (

Table 2. Statistical results (r, p-value and RMSE) from the evaluation of the forecasted AODs of each model against MODIS-AODs over the Middle East during dust events of various types.

Model	Cyclonic			Post-Front			Pre-Front			Shamal
	Correlation	p-Value	RMSE	Correlation	p-Value	RMSE	Correlation	p-Value	RMSE	Correlation	p-Value	RMSE
EMA_RegCM4	0.439	<0.001	0.265	0.281	<0.001	0.058	0.186	<0.001	0.216	0.340	<0.001	0.249
DREAM8_MACC	0.718	<0.001	0.206	0.553	<0.001	0.283	0.348	<0.001	0.283	0.447	<0.001	0.404
MACC_ECMWF	0.70	<0.001	0.220	0.433	<0.001	0.323	0.586	<0.001	0.386	0.529	<0.001	0.424
NMMB_BSC	0.572	<0.001	0.307	0.301	<0.001	0.298	0.478	<0.001	0.315	0.552	<0.001	0.461
DREAM8b_V2	0.596	<0.001	0.270	0.472	<0.001	0.286	0.449	<0.001	0.298	0.510	<0.001	0.448
DREAMABOL	0.541	<0.001	0.232	0.442	<0.001	0.351	0.398	<0.001	0.220	0.497	<0.001	0.453
NASA_GEOS	0.703	<0.001	0.254	0.555	<0.001	0.399	0.623	<0.001	0.323	0.579	<0.001	0.485
NCEP_NGAC	0.655	<0.001	0.263	0.452	<0.001	0.320	0.155	<0.001	0.182	0.412	<0.001	0.330
NOA-WRF-Chem	0.492	<0.001	0.194	0.433	<0.001	0.263	0.557	<0.001	0.282	0.593	<0.001	0.469

; Figure 8). According to the Taylor diagram (Figure 8), the points related to these three models as well as MACC_ECMWF are closer to the observation point, indicating satisfactory performance in forecasting the examined cyclonic dust storms. The MACC and NASA-GEOS models assimilate MODIS AOD and generally provide satisfactory performance compared to other models. The closer the points to the diameter of the circle, the more accurate the model outputs. The point of the EMA_RegCM4 model is far beyond those of the other models, indicating a very different behavior in dust simulations, as shown in the AOD spatial distribution patterns (Figure 4).

Figure 9 shows the respective scatter plots and Taylor diagram for the model forecasts of post-frontal dust storms in the Middle East. The results show a rather similar model performance as in the previous cases, since EMA_RegCM4 exhibited a large possibility of AOD overestimation for the lower values, while NMMB_BSC highly underestimated the high AODs. Among the studied models, DREAM8_MACC and NASA_GEOS exhibited the highest r values (0.55), while NOA-WRF-Chem presented low RMSE (Table 2; Figure 9). The Taylor diagram shows that among the studied models, the DREAM8_MACC, NCEP_NGAC, MACC_ECMWF and NASA_GEOS outputs were closer to the MODIS observations.

The models’ evaluation for the pre-frontal dust storms (Figure 10) also shows that most of the models underestimated the MODIS AODs. The AOD outputs of DREAM8_MACC, NASA_GEOS and MACC_ECMWF models were more similar to MODIS, exhibiting higher r values due to MODIS data assimilation as mentioned above, while the NOA-WRF-Chem model performed also satisfactorily (Table 2; Figure 10). Based on the Taylor diagram, the outputs of all models, except EMA-RegCM4, were similar to each other, indicating a similar capability of the models for forecasting pre-frontal dust storms in the Middle East.

In the case of Shamal dust storms (Figure 11), the models exhibit different performances and a general underestimation for the high AODs. However, there is evidence that models are capable to forecast the synoptic conditions and pressure gradients that trigger the Shamal dust storms over the Middle East [45,107,108] than the frontal systems, which are fast-moving and the changing winds and temperatures are more difficult to be accurately represented [47]. The outputs of the NCEP_NGAC model were more uncertain compared to the other models and NMMB_BSC exhibited good performance (Figure 11). NASA_GEOS and NOA-WRF-Chem exhibit the highest r values (close to some other models), while NCEP_NGAC exhibited low RMSE (Table 2, Figure 11). The Taylor diagram for Shamal dust storms shows that NOA-WRF-Chem, NASA-GEOS and MACC-ECMWF outputs were closer to MODIS observations.

According to the combined results summarized in Table 2 and Figure 12, in general, the highest correlation coefficient is related mostly to NASA_GEOS and DREAM8_MACC models, both of which used MODIS data assimilation schemes (Appendix A). These results indicate that even though the model forecasts correspond to valid time more than 24 h after the initial time, the effect of data assimilation increases the accuracy of the forecasts. These two models also have a relatively high resolution compared to other examined models. NOA-WRF-Chem also presented satisfactory results with lower RMSE values compared to other models. Although the NCEP_NGAC model has finer resolution than MACC_ECMWF model, its outputs were less accurate. The various models exhibited higher correlation coefficients and lower RMSE values (Table 2, Figure 12) for the cyclonic dust storms compared to frontal dust storms. The larger inaccuracies in the predictions of the dust AODs during pre- and post-frontal dust storms may be explained by difficulties in accurate representation/forecasting of the frontal systems and the change of wind in the frontal areas, which triggers the dust storms. On the other hand, high RMSE values between the models were also found for the Shamal dust storms, which are likely attributed to the large coverage of the high dust-AODs during Shamal dust storms (Figure 2 and Figure S2). On the contrary, during cyclonic dust storms, the dust covers a relatively small area around the cyclonic center and its representation by the models was generally satisfactory, while a large part of the study domain was usually free of dust.

In general, the model forecasts for several dust events of different types and meteorological conditions over the Middle East revealed specific difficulties in representing the AOD over the dust source regions, which may be related to topography, surface and dynamic meteorological parameters [51,53]. Therefore, the source of the errors in dust forecasting models is very diverse including surface parameters such as soil texture, vegetation, land use and soil moisture, as well as atmospheric parameters such as wind, precipitation, and the instability associated with boundary layer related quantities. Furthermore, the relative coarse spatial resolution of the model AODs could explain the generally reduced level of verified performance of these numerical models, since especially over the complex topography of the Middle East, the finer model resolution increased the accuracy of the dust predictions [51,53,55]. On the other hand, since wind regime is an important factor for dust emissions and transportation, the accuracy in simulations of wind speed and direction in each model plays a major role in model performance [107,108,109]. Overall, the results show that there is not a single model that can be characterized as “better” or “more accurate” in all dust cases.

Finally, ground-based AOD data from 3 AERONET stations (Kuwait University, KAUST-campus and Masdar Institute), during the 12 examined dust storms revealed a satisfactory agreement with MODIS data (R² = 0.75) collocated in time and space over these stations (

Table 3. Statistical results (R²) from the evaluation of the forecasted AODs of each model against AODs at 3 AERONET stations in the Middle East (Kuwait University, KAUST-campus and Masdar Institute) during the 12 dust storms.

	BSC_DREAM8b_V2	DREAMABOL	EMA-RegCM4	NASA-GEOS	NCEP-NGAC	DREAM8-MACC	MACC-ECMWF	NMMB-BSC	NOA-WRF-CHEM	TERRA
R2	0.48	0.61	0.04	0.74	0.62	0.78	0.73	0.35	0.21	0.75

). Model AOD forecasts over the AERONET stations, showed that DREAM8-MACC exhibited the highest R² of 0.78, followed by NASA_GEOS model (R² = 0.74) and MACC-ECMWF (R² = 0.73), considering available data during the 12 dust events. These results agree with the model’s performance against MODIS observations, indicating the positive feedback of data assimilation in the forecasting process. On the other hand, the AOD forecasts from other models such as EMA-RegCM4 and NMMB-BSC highly departed from the AERONET measurements (Table 3), justifying the complexity of various parameters such as size distribution, in model AOD estimates [56].

4. Conclusions

This study analyzed four types of synoptic dust events over the Middle East region, including cyclonic, pre-frontal, post-frontal and Shamal dust storms. For each of these types, three intense and pervasive dust events were examined in view of meteorological dynamics and the capability of numerical models to forecast the spatial distribution of dust. True-color imagery and AOD retrievals from Terra-MODIS observations were analyzed over the region, in order to determine the dust sources and affected areas in each case study. Since various factors affect the outputs of the dust forecasting models, the performance of 9 operational dust models was qualitatively and quantitatively evaluated during these dust events in the Middle East.

The synoptic meteorological analysis showed the presence of low-pressure systems over parts of Iraq and Syria, in all three cases of cyclonic dust storms, along with cyclonic winds around them. The upward motion at the center of the low pressure caused dust emission. In pre-frontal and post-frontal dust storms, low-pressure systems associated with warm and cold fronts triggered the dust events. The change of wind direction and the rotation of isobars were signs of the presence of fronts in the area. The intersection of geopotential height contours and isotherms well showed the baroclinic area and thermal advection, which helped in the determination of the exact location of atmospheric fronts. In the three Shamal dust events, the thermal-low troughs entered parts of Iraq and Syria from the southeast, and associated with ridges of high-pressure systems dominated over the Mediterranean Sea, caused strong pressure gradients and lower-troposphere winds that initiated the dust storms.

The comparison of the models’ AOD outputs with Terra-MODIS AODs showed that all models had a relatively acceptable performance in forecasting the AOD patterns over the region. The studied models were able to forecast the higher AODs along the dust-affected areas, although most of them underestimated the values compared to MODIS, especially near the dust source areas with the maximum AODs. Regarding the frontal dust storms, it can be concluded that all models were able to represent the position of atmospheric fronts to some extent, but in most cases, they underestimated the large AODs around the fronts. This bias can be likely attributed to discrepancies in representing the wind intensity in the area or the amount of atmospheric moisture and wind profiles. By comparing the model outputs in all studied cases, it can be concluded that the models performed satisfactorily in forecasting the cyclonic dust storms. Examination of the Taylor diagrams showed that the performance of the models in forecasting the frontal dust storms was similar to each other except for the EMA-RegCM model, which exhibited large uncertainties. The highest correlation coefficients between the model outputs and the satellite AODs referred to cyclonic dust storms and BSC-DREAM8b-V2, DREAM8-MACC, MACC-ECMWF, NASA-GEOS, NCEP-NGAC and NMMB- BSC models, with r values of about 0.6–0.7. Regarding the other dust types, in most cases, the correlation coefficient was between 0.4 to 0.6, which can be considered acceptable. Although it is not possible to introduce a single model as the best for all the dust storms, it can be said that the outputs of the MACC and NASA-GEOS models that used MODIS data assimilation showed greater similarity with MODIS observations and AERONET measurements in most of the cases.

Accurate forecasting of dust phenomena, which is one of the main preconditions for the development of early warning systems, can highly reduce dust impacts. However, there are relatively large uncertainties between the outputs of different dust forecasting models that could be due to differences in their inputs, including statistical data related to the land surface such as topography, vegetation cover and soil texture, atmospheric data, initial and boundary conditions, as well as the use of data assimilation. Overall, this study revealed that various models are capable in forecasting dust storms of various types in the Middle East, although accurate representation of dust AODs remains a challenge. Using more accurate land surface data, better parameterization schemes, data assimilation and ensemble forecasting are some of the solutions to improve the outputs of dust forecasting models.