Remote Sensing Applications: Society and Environment 23 (2021) 100532

Contents lists available at ScienceDirect

Remote Sensing Applications: Society and Environment

journal homepage: [Link]/locate/rsase

Spatiotemporal variations and long term trends analysis of aerosol optical

depth over the United Arab Emirates
Abdelgadir Abuelgasim a, Muhammad Bilal b, *, Ibrahim Abdalla Alfaki c
a
Department of Geography and Urban Sustainability, National Water and Energy Center, College of Humanities and Social Sciences, United Arab Emirates University, Al
Ain, Abu Dhabi, 15551, United Arab Emirates
b
School of Marine Sciences, Nanjing University of Information Science & Technology, Nanjing, 210044, China
c
Department of Analytics in the Digital Era, College of Business and Economics, United Arab Emirates University, Al Ain, Abu Dhabi, 15551, United Arab Emirates

A R T I C L E I N F O A B S T R A C T

Keywords: This research provides a comprehensive long term time series analysis of the spatiotemporal variability of aerosol
Aerosol optical depth optical depth (AOD) over the United Arab Emirates (UAE) using MODIS (Moderate-resolution imaging spec­
Spatial variability troradiometer) MAIAC (Multi-Angle Implementation of Atmospheric Correction) AOD (aerosol optical depth)
Temporal variability
product for the period 2003–2018. The UAE is located in one of the driest places on earth, the Arabian Peninsula,
Aerosols
Arid regions
with two large extended desert regions that form major atmospheric dust sources in the world. The atmospheric
dust emitted by these vast deserts is transported locally and regionally and it is the primary major contributor to
the overall aerosol optical depth in this area. The investigation of the spatiotemporal trends of AOD over the UAE
showed a significant obvious annual seasonal variability. Results showed a significant increase in AOD during
summer due to high wind speed, resulting in scattered events of small to large scale haboobs within the study
area, high temperature, and high relative humidity. The increased humidity levels are due to increased tem­
peratures in the summer season leading to significant evaporation from the neighbouring water bodies (mainly
the Arabian Gulf and the Sea of Oman) leading to a significant increase in water vapour levels in the atmosphere.
Summar AOD values during the study period ranged between 0.42 – 0.73. The second highest AOD was observed
in spring followed by winter. Spring AOD values ranged between 0.33-0.58 and winter values ranged between
0.25 – 0.41. The spatial distributions of AOD showed higher values over the coastal and desert areas. Coastal
areas portrayed higher values as well likely due to the high levels of evaporation and increased anthropogenic
activities such as coastal urban developments and automobile pollution. Time series analysis of the multi­
temporal dataset confirmed the SARIMA forecast model as the best fit which can be used to provide an accurate
and reliable monthly forecast for the AOD concentrations in the UAE.

1. Introduction In addition to their impacts on climate and weather, large concen­

trations of aerosols in the atmosphere at local scales have significant
Natural and anthropogenic atmospheric aerosols (fine solid or liquid impacts on the ambient air quality and have adverse effects on human
particles suspended in the air) play a major role in the earth’s climate health (Atkinson et al., 2013; Pope et al., 2011). Studying the formation,
system, weather, and regional and global climate change, by influencing and spatiotemporal distribution of aerosols in the atmosphere is a
the earth’s radiation budget and affecting the regional hydrological fundamental part of atmospheric science. It is generally believed that the
cycle (Ramanthan et al., 2001). The earth radiation budget is signifi­ concentrations of aerosols in the atmosphere are strongly influenced by
cantly affected by the interactions, both scattering and absorption, be­ the underlying local land cover, local land-use practices, and the
tween the incoming solar radiation and the presence of atmospheric meteorological conditions of the atmosphere in an area leading to great
aerosols in the atmosphere. Depending upon their size, type, and loca­ variability in their spatiotemporal distributions (Abuelgasim and Far­
tion, aerosols can either cool the surface or warm it. They can help ahat 2020a). Identifying the spatiotemporal variability of aerosol con­
clouds to form, or they can inhibit cloud formation. centrations requires regular measurements of their optical properties

* Corresponding author.
E-mail addresses: [Link]@[Link] (A. Abuelgasim), [Link]@[Link] (M. Bilal), [Link]@[Link] (I.A. Alfaki).

[Link]
Received 24 August 2020; Received in revised form 9 January 2021; Accepted 28 April 2021
Available online 11 May 2021
2352-9385/© 2021 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license
([Link]
A. Abuelgasim et al. Remote Sensing Applications: Society and Environment 23 (2021) 100532

Fig. 1. Geographic location of the UAE

such as aerosol optical depth (AOD). AOD is defined as the extinction data using an appropriate seasonal time series model to predict the
coefficient of solar light in the atmosphere due to aerosols. It represents evolution of AOD in the UAE and further forecast (out-of-sample) future
the sum of both absorption and scattering coefficients. annual AOD scores.
Satellite remote sensing is the only viable method for estimating
AOD over a large spatial extent on a continuous and temporal basis. 2. Study area and dataset
Advancement in satellite remote-sensing techniques has opened new
corridors for the monitoring and mapping of air pollution over large 2.1. Study area
regions. Currently, several satellites in orbit have instruments suited for
AOD estimation such as the Moderate Resolution Imaging Spectro- The UAE is located in the Arabian Peninsula at the south-eastern part
radiometer (MODIS) carried on both Aqua and Terra platforms (Levy between latitudes 22◦ 50′ - 26◦ 4′ N and longitudes 51◦ 5′ - 56◦ 25′ E
et al., 2013), Sea-viewing Wide of view Sensor (SeaWiFS) (Sayer et al., (Fig. 1). It is bordered by Saudi Arabia to the west, Oman to the east and
2012) and the Advanced Very High-Resolution Radiometer (AVHRR) south, and the Arabian Gulf waters to the north. It is a relatively small
(Riffler et al., 2010). Satellites use visible and near-infrared spectra to country with approximately a total area of about 83,600 km2. The UAE
derive accurate estimates of aerosol information overland (Omari, et al., is a federal union made up of seven emirates with the emirate of Abu
2019). In recent years, there has been progressing in estimating AOD Dhabi being the largest in terms of land size, population, and largest
using only one spectral band (Bilal et al., 2013). Section 2.2 discusses economic development and hydrocarbon industry. In spite of its rela­
some other methods for AOD retrieval based on relative transfer theory tively small land size, several landscape types and climate systems exist
from optical remote sensing imagery. including coastal, mountain, and desert climate along with precipitation
Numerous previous studies have analyzed the spatiotemporal vari­ scarcity all year round except with sporadic low precipitation rainfall in
ations of AOD concentrations over the Arabian Peninsula at large and the winter (Böer 1997). The climate is predominantly arid, dry, and hot
Saudi Arabia in particular (Ali et al., 2017; Ali and Assiri, 2019; Alharbi, with local coastal and mountain climates leading to a variety of envi­
et al., 2013; Farahat, 2019; Almazroui, 2019; Kumar et al., 2018; Jin ronments such as Gulf islands, coastal and desert areas, and major dif­
et al., 2018; Maghrabai and Alotaibi, 2018). Very little research is done ferences in relative humidity and therefore differences of aerosol optical
with a focus on the United Arab Emirates (UAE). The UAE is a small properties.
country with little information about the spatiotemporal variability of Deserts cover a significant portion (about 80 percent) of the total
AOD (Albaloushi et al., 2016). Its landscape is predominantly composed UAE area and are predominantly occupied with different types and
of desert surfaces frequented by occasional dust storms that significantly shapes of dunes. Dunes are particularly common in the southern and
contributes to the country’s overall AOD concentrations. In addition, the western parts of the country making them a dominant landscape feature.
major urban expansions, industries, and population growth add further Several mountain ranges stretching from the Strait of Hormuz are found
anthropogenic aerosols to the total country’s AOD. along the border with Oman. The desert environment of the UAE suffers
The primary objective of this study is to analyze the long-term trends from frequent sand storms and haboobs leading to the transport of dust
of AOD over the UAE and identify its spatiotemporal variations using particles, industrial and hydrocarbon pollutants all over the country.
MODIS MAIAC AOD products. The work in this research uses the The frequent sand storms lead to the rise of dust particles into the air and
advanced MAIAC algorithm with a much finer spatial resolution AOD thus significantly contributing to the overall country’s AOD.
product in comparison to previous studies in the Arabian Peninsula. Vegetation agricultural cover is scarce in the UAE except for some
Furthermore, an additional objective of the study is to model the AOD mountainous areas in the northeast and ancient oasis areas along the

2
A. Abuelgasim et al. Remote Sensing Applications: Society and Environment 23 (2021) 100532

Omani border. The dominant agricultural activity is limited to palm and shown a large uncertainty over urban areas (Nichol and Bilal, 2016; Bilal
a few scattered vegetable plantations. As desert is the dominant land et al., 2019).
cover it is continuously used for animal raising (such as camels) and
care. In the past four decades, the UAE has implemented massive urban 3. Research methodology
planning projects resulting in the development of many major cities and
urban population centers. 3.1. Image processing

2.2. MODIS data Image processing of MODIS data consisted of mosaicking the
different titles covering the UAE and further cropping areas outside the
There are several methods developed for AOD retrieval based on boundaries of the country. The average daily AOD concentrations were
radiative transfer theory from optical remote sensing imagery (Kaufman calculated for each day for the study period with other statistical pa­
et al, 1994, 1997; Levy et al., 2007). The most popular method is known rameters such as standard deviation, maximum, and minimum AOD. To
as the Dense Dark Vegetation (DDV) (Kaufman et al, 1994, 1997; Levy speed the process of mosaicking and statistical calculations a short
et al., 2007). This method was successfully used to retrieve AOD over model code was written to perform the image processing in ArcGIS
areas with dense vegetation canopies (Kaufman, et al., 1997; Levy et al., 10.6.1. MODIS image data were further used to calculate the average
2007). In this method, the contribution of surface reflectance is removed annual AOD concentrations per season and per year to analyze its
by taking advantage of the correlations between the shortwave infrared spatiotemporal variability. The output statistical information for each
(2100 nm) and the red (660 nm) and the blue (470 nm) and the negli­ day was statistically analyzed to identify multi-temporal trends daily,
gible effect of aerosols in the shortwave infrared (SWIR) band. This weekly, monthly, and annually and served as inputs for the time series
approach has become the operational method for AOD estimation for the analysis models. The R statistical package was used throughout the
Moderate Resolution Imaging Spectroradiometer (MODIS) carried on study for statistical analysis.
both Aqua and Terra platforms (Levy et al., 2013), Sea-viewing Wide of
view Sensor (SeaWiFS) (Sayer et al., 2012), and the Advanced Very 3.2. Temporal modeling and forecasting
High-Resolution Radiometer (AVHRR) (Riffler et al., 2010). However,
this approach showed great limitations over bright targets such as in arid In the temporal dimension, the objective of the study is to model
and semi-arid regions, and over urban areas. AOD data using an appropriate seasonal time series model to predict the
This limitation is overcome by the Multi-angle Implementation of evolution of AOD in the UAE and further forecast (out-of-sample) future
Atmospheric Correction (MAIAC) algorithm which is designed and annual AOD scores. The literature on time series analysis extensively
developed to retrieve simultaneously surface reflectance and atmo­ covers several well-known traditional modeling and forecasting tech­
spheric products at 1 km spatial resolution over dark vegetated and niques that are often used to forecast time series data with at most one
bright surfaces using time-series observations from MODIS and clima­ seasonal component. Among the most popular and frequently used
tology data from Aerosol Robotic Network (AERONET) (Lyapustin et al., methods, for univariate time series, are the ARIMA (autoregressive in­
2010). The surface parameterization is derived based on the spectral tegrated moving average), the exponential smoothing models, and their
regression coefficients that explain the surface bidirectional reflectance subsequent modifications discussed recently in Hyndman et al. (2008).
distribution function (BRDF) in the MODIS bands of Blue (470 nm),
green (550 nm), and shortwave infrared (2130 nm). This MAIAC algo­ 3.3. Seasonal autoregressive integrated moving average (SARIMA) model
rithm uses seven distinct regional aerosol models for the distinct regions
over the globe, and for the dust-related aerosols, either the dust or The general Box-Jenkins model of order (p, P, q, Q), denoted
background models have used alike the DT algorithm (Levy et al., 2013). as SARIMA(p, d, q)(P, D, Q)m , is outlined by Bowerman et al. (2005) as
Moreover, a smoke test is applied to distinguish between smoke and
clouds particles (Lyapustin et al., 2012). This test depends on the cor­ φp (B)φP (Bm )zt = δ + θq (B)θQ (Bm )εt (1)
responding rise in aerosol absorption at the MODIS band of 412 nm as
compared to bands 470 nm and 670 nm because of various scattering zt is the stationary transformation of the original time series yt ,
and absorption by organic carbon, released during biomass burning expressed in terms of nonseasonal ∇ , seasonal ∇m operators and a
activities. The series of MAIAC-based atmospheric products covers predifferencing transformation y*t of the original time series yt as
aerosol optical depth (AOD) at 470- and 550 nm, cloud mask, and col­
umn water vapour (WV) at 940 nm from MODIS NIR bands at 1 km zt = ∇Dm ∇d y*t = (1 − Bm )D (1 − B)d y*t (2)
resolution. The expected error (EE) of MAIAC over land is ± (0.05 +
where d and D are the degrees of nonseasonal and seasonal differencing,
10%‒15%) (Lyapustin et al., 2018).
respectively, and B and Bm are the backshift operators, shifting the
In this study, daily MODIS MAIAC AOD product was used for the
subscripts of a time series observation backward in time by one period
period 2003–2018 for the UAE which was acquired from the Level-1 and
and m periods, respectively.
Atmosphere Archive & Distribution System (LAADS) Distributed Active
φp (B) and θq (B) are denoting the non-seasonal autoregressive and
Archive Center (DAAC). MODIS is a principle instrument onboard the
Terra (18 December 1999) and Aqua (4 May 2002) satellites to provide moving average operators of order p and q, respectively. The φP (Bm ) and
geophysical data products with a temporal resolution of 1–2 days at θQ (Bm ) are, respectively, the seasonal autoregressive and moving
three spatial resolutions (e.g., 250 m, 500 m, and 1 km) using 36 spectral average operators of order P and Q. The constant term δ is given as δ =
bands ranging from visible to infrared wavelengths (i.e., 0.14–14.4 μm). φp (B)φP (Bm )μ, where μ denotes the true mean of the stationary time
MODIS explicitly uses three different inversion algorithms such as Dark series, Bμ = μ. The εt s are independent random shocks with mean zero
Target (DT) land algorithm, (DT) ocean algorithm (Levy et al., 2013), and constant variance σ2 , εt ∼ N(0, σ 2 ).
and Deep Blue (DB) land algorithm (Hsu et al., 2013) to retrieve aerosol
optical depth (AOD) over the bright-reflecting surfaces, urban and 3.4. The innovations state space exponential smoothing models (ETS)
vegetation surfaces (Bilal et al., 2019). MODIS DT and DB level 2 aerosol
products provide daily observation at 10 km coarse resolution and un­ Holt-Winter time series forecasting approach (Holt, 1957; Winter
able to provide detailed information about spatial distributions of 1960) is one variant of a class of exponential smoothing (ES) methods
aerosols over a small region such as the UAE. MODIS DT algorithm also that allow modeling of trend and univariate seasonality of a time series.
provides AOD at 3 km resolution (Levy et al., 2013) but this product has One disadvantage of the ES methods is the disability to admit maximum

3
A. Abuelgasim et al. Remote Sensing Applications: Society and Environment 23 (2021) 100532

Fig. 2. Daily average AOD over UAE from 2003 to 2018.

likelihood calculations and derivation of prediction intervals. Therefore, ℓt = (ℓt− 1 + bt− 1 )(1 − σεt ,
Hyndman et al. (2008) proposed the ES methods within the framework
of the innovations state-space models. The developed models are bt = bt− 1 + ​ β(ℓt− 1 + bt− 1 )εt ,
referred to as ETS, where E stands for the errors, T for the trend, and S for
the seasonal components. Each of these components can be character­ or
ized as “additive (A)”, “multiplicative (M)”, or “none “(N)”. Accordingly,
the ETS framework provides over 30 possible options for choosing the yt = [11]xt− 1 (1 + εt ),
best model. For instance, the ETS(A, ​ A, ​ N) model refers to Holt’s ad­ [ ] [ ]
1 1 α
ditive trend, with additive errors and no seasonality pattern. The xt = xt− 1 + [11]xt− 1 ε,
0 1 α t
ETS(M, ​ Ad, ​ M) model refers to Holt-Winter damped additive trend,
with multiplicative errors, and multiplicative seasonality pattern. For Hyndman et al. (2008) expressed all the 30 variations of the state
full details of alternative models please consult Hyndman et al. (2008). space exponential smoothing models as a general model that involves
The state-space Holt’s linear trend exponential smoothing model the state vector xt = (ℓt , ​ bt , ​ st , ​ st− 1 , …, ​ st− m+1 )T ,
with additive errors, ETS(A, ​ A, ​ N), for example, can be derived as
yt = w(xt− 1 ) + r(xt− 1 )εt ,
follows, Hyndman et al. (2008): Let μt = ̂ y t = ℓt− 1 + bt− 1 denote the
one-step ahead forecast of yt , assuming all parameters are known. Let
xt = f (xt− 1 ) + g(xt− 1 )εt ,
εt = yt − μt be the one-step ahead forecast error at time t. Then from the
⃒
⃒
forecast equation, ̂ y t+h ⃒⃒t = t + bt h, where {εt } is a Gaussian white noise process with variance σ 2 , and μt =
w(xt− 1 ). A model with additive errors would have r(xt− 1 ) = 1, hence yt =
yt = ℓt− 1 + bt− 1 + εt μt + εt . Similarly, a multiplicative errors model would have r(xt− 1 ) = μt
and yt = μt (1 + εt ). Thus, εt = (yt − μt )/μt is the relative error for the
from the level and growth equations (ℓt = αyt + (1 − α)(ℓt− 1 +bt− 1 ) and multiplicative model.
bt = β* (ℓt − ℓt− 1 ) + (1 − β* )bt− 1 , respectively) we can write
4. Results and discussions
ℓt = ℓt− 1 + bt− 1 + αεt
4.1. Exploring AOD data
bt = bt− 1 + β∗ (ℓt − ℓt− 1 − bt− 1 ) = bt− 1 + αβ∗ εt = bt− 1 + ​ βεt , ⋅ for ⋅ β = αβ∗
The error term εt is assumed to be independent and identically 4.1.1. Average daily AOD data
distributed with mean zero and variance σ2 , εt ∼ NID(0, σ2 ). The method Daily, monthly and annual average AOD retrievals were calculated
can be expressed in the state space form by defining the state vector xt = using 5844 MODIS daily observations spanning the period from January
1, 2003, to December 31, 2018.
[ℓt , bt ]T and re-expressing the three above equations as
Fig. 2 showed lower values of AOD at the beginning of each year
yt = [11]xt− 1 + εt , during winter and starts rising by spring to peak during the summer
[ ] [ ] season. That is quite the norm as dust storm season starts in April and
xt =
1 1
xt− 1 +
α
ε. peaks by May–July. Results also depict a stable, possibly stationary, time
0 1 α t series with no obvious trend. The elevated AOD spikes in 2009, 2011,
Similarly, the state space Holt’s linear trend exponential smoothing 2015, and 2018 were most likely associated with the massive dust
model with multiplicative errors, ETS(M, ​ A, ​ N), can be derived by first storms that occurred in the UAE leading to the significant rise of AOD.
setting εt = (yt − μt )/μt , so that εt is the relative error. The following To gain further insight into the behavior of the daily average AOD in
the UAE. First, to reduce the influence of outlying AOD scores, outliers in
similar derivation as for the additive errors model above, the following
the data were subdued using the “tsclean()” function implemented in the
equations are obtained
R “forecast” package. Second, to further understand the various patterns
yt = (ℓt− 1 + bt− 1 )(1 + εt ), exhibited in the daily average AOD data, it is helpful decomposing the

4
A. Abuelgasim et al. Remote Sensing Applications: Society and Environment 23 (2021) 100532

behavior that tends to fluctuate from one week to another, confirming

inconsistent pattern surrounding the weekly AOD cycle.

4.2. Monthly average AOD data

The apparent inconsistent weekly pattern discussed earlier, led to the

conclusion that the AOD data was better recognized as an annual one-
season time series. Accordingly, the whole daily data were aggregated
to a monthly level and the average AOD concentration per month was
calculated. Fig. 4 shows a graphical time-series representation of the
monthly AOD averages.
To further understand the behavior of the monthly average AOD, the
STL decomposition (Cleveland et al., 1990) approach was again utilized.
Results depicted in Fig. 5 below confirmed the presence of a strong
one-season annual pattern in the monthly data.
Close examination of the average monthly data in Fig. 6 shows that
high scores of AOD are evident during the summer period, June, July,
and August. There is a clear decline in concentration levels during the
winter or low-temperature months, mainly from October to February.
This is generally consistent with the overall temporal fluctuations of
AOD over the UAE.

Fig. 3. The AOD data (a) (top) and its four additive components (b–e) obtained 4.3. Average annual AOD
from a robust MSTL (multiple seasonal) decompositions with flexible trend-
cycle and fixed seasonality. Table 1 shows the average annual AOD over the UAE during the
study period together with the days per year where the AOD level ex­
series into its, possibly, natural components, each represents the un­ ceeds 1.0, 1.5, 2. The annual AOD values ranged from a minimum of
derlying categories of patterns. Three types of time series patterns are 0.32 in 2004 to a maximum of 0.47 in 2018. Notably, the year 2015,
commonly discussed: namely, trend, seasonality, and cycles. The trend heralds a continuing increase in the level of AOD. In addition, the table
and cycle components are commonly combined in one component shows the number of days where AOD have exceeded 1.5, 2, 2.5, 3 and
denoted simply as a trend (Hyndman, 2018). Several methods are 3.5. It is not possible to conclude from this table that extreme AOD
employed to extract these components from a time series. For this study, values are becoming much more frequent.
a graphical decomposition, using the Seasonal and Trend decomposition
using Loess (STL) decomposition approach (Cleveland et al., 1990), was 4.4. Spatial distribution of AOD in the UAE
utilized to decompose the daily time series into seasonal, trend and
remainder (random) components. This provides for choosing an 4.4.1. Average annual trend
appropriate forecasting model that most closely describes the pattern Fig. 7 shows the spatial distribution of annual average MODIS AOD
generated by the data. over the UAE from 2003 to 2018.
Overall, the structure of the decomposed data depicted in Fig. 3 The coastal areas along the Arabian Gulf are portraying the highest
revealed a nonlinear trend (Fig. 3(b)) with a slight increase in the concentrations during the study period. The concentrations are gener­
average AOD levels after the year 2015. The seasonal behavior of the ally higher than the desert areas. This pattern is likely due to three
data (Fig. 3(d)) confirmed the presence of annual seasonality in the primary reasons. First, the naturally high levels of water vapor present in
behavior of the average AOD. the atmosphere due to significant evaporation from the nearby Arabian
Weekly seasonality or cycles (Fig. 3(c)) are, however, not evident Gulf waters (Srivastava et al., 2008) in addition to the sea salt aerosol
from the plot. This component shows irregular and unpredictable particles. Secondly, urban pollution from the coastal cities in the UAE
where this part of the country contains the three largest cities in the

Fig. 4. Monthly average AOD in UAE, 2003–2018.

5
A. Abuelgasim et al. Remote Sensing Applications: Society and Environment 23 (2021) 100532

Fig. 5. The monthly average AOD data (top) and its three additive components were obtained from a robust MSTL decomposition with flexible trend-cycle and fixed
seasonality.

Fig. 6. The Polar and the Seasonal subseries plots depicting the underlying monthly pattern and changes. The horizontal bars in the seasonal subseries plot show the
average monthly AOD values over the study period, 2003–2018.

country with large population centers, namely, Abu Dhabi, Dubai, and AOD were observed during this year compared to the other years
Sharjah. The higher levels could be due to errors of AOD retrievals due to (Fig. 7). By the spring, AOD starts to rise due to the increased wind
sea haze. activity and reaches the highest concentrations during the summer
The north and northeast part of the country along the border with season. It comes as a surprise to note that the autumn season is showing
Oman near the Strait of Hormuz is mostly showing lower concentrations lower levels of AOD in comparison to the winter season. This is generally
(shown as a black circle in the 2003 image in Fig. 7). The dominant land due to the high levels of fog concentrations along with coastal areas in
cover in this part of the country is the large mountains of Oman the winter, and also in different areas, the fog extends deep into the
extending from the strait inland and thus providing much lower sand desert regions. By the spring, AOD starts to rise much further due to the
and dust. Similarly, are the regions along the Sea of Oman are portraying increased wind activity and reaches the highest concentrations during
lower AOD concentrations. the summer season. By the autumn, concentration levels start to drop
reaching their lowest.
4.4.2. Average Seasonal trend Fig. 8(b), below, shows the land cover distributions in the UAE. The
Table 2 shows the seasonal variations of AOD over the UAE during spatial distribution of AOD concentrations in the different land cover
the study period. The summer season is showing the highest levels of types is not very different from the annual pattern, however, with
AOD, followed by the spring and winter season for years investigated. coastal areas showing the highest levels in all seasons due to both nat­
Seasonal AOD distributions (Fig. 8(a)) were analyzed for the year ural and anthropogenic effects. In the spring and autumn seasons, some
2018 as the highest AOD (Table 1) as well as significant variations in locations, other than coastal areas, are showing high AOD. These are

6
A. Abuelgasim et al. Remote Sensing Applications: Society and Environment 23 (2021) 100532

Table 1
Average annual AOD in the UAE, 2003 to 2018.
Year Mean Std Max Min Days per year with AOD exceeding

1.0 1.5 2.0 2.5 3.0 3.5

2003 0.4 0.27 1.56 0.02 16 1 0 0 0 0

2004 0.32 0.16 1.05 0.06 2 0 0 0 0 0
2005 0.34 0.22 2.23 0.05 2 1 1 0 0 0
2006 0.38 0.26 2.25 0.02 5 4 1 0 0 0
2007 0.36 0.18 1.01 0.07 1 0 0 0 0 0
2008 0.42 0.29 1.87 0.06 13 4 0 0 0 0
2009 0.44 0.34 2.73 0.06 13 2 3 1 0 0
2010 0.37 0.22 1.12 0.07 5 0 0 0 0 0
2011 0.43 0.32 3.22 0.08 14 1 0 1 1 0
2012 0.44 0.28 1.87 0.06 14 3 0 0 0 0
2013 0.39 0.23 1.56 0.08 7 1 0 0 0 0
2014 0.36 0.18 1.12 0.08 2 0 0 0 0 0
2015 0.4 0.27 3.3 0.02 9 0 0 0 1 0
2016 0.4 0.25 2.03 0.06 4 2 1 0 0 0
2017 0.41 0.23 1.29 0.013 15 0 0 0 0 0
2018 0.47 0.38 3.82 0.09 8 2 1 0 2 1

Fig. 7. Spatial Distribution of the Average Annual AOD (the black circle in 2003 shows mountainous regions).

associated with the calcareous desert surface and sand dunes field. and Farahat, 2020b). Most of these studies are, however, focused on the
Furthermore, in the summer, areas of the country along the sea of Oman Arabian Peninsula at large (Kumar et al., 2018) or Saudi Arabia in
shows higher than average AOD concentration levels. particular except for Abuelgasim and Farahat (2020b). The major
Previous studies of AOD concentrations and distributions are findings of studies over the Arabian Peninsula are showing a general
showing on average similar patterns (Ali et al., 2017; Ali and Assiri, increase of AOD levels across the Arabian Peninsula in both spring and
2019; Alharbi, et al., 2013; Farahat, 2019; Almazroui, 2019; Abuelgasim summer and comparison lower levels in the autumn and winter. Some

7
A. Abuelgasim et al. Remote Sensing Applications: Society and Environment 23 (2021) 100532

Table 2
Seasonal variation of average AOD over the UAE.
Year Year Year Year Year Year Year Year

Season 2003 2004 2005 2006 2007 2008 2009 2010

Winter 0.32 0.25 0.30 0.27 0.32 0.33 0.40 0.27

Spring 0.49 0.37 0.33 0.40 0.43 0.56 0.47 0.45
Summer 0.56 0.42 0.50 0.58 0.42 0.56 0.62 0.52
Autumn 0.23 0.23 0.23 0.29 0.25 0.21 0.26 0.24
Year Year Year Year Year Year Year Year
Season 2011 2012 2013 2014 2015 2016 2017 2018

Winter 0.37 0.41 0.28 0.31 0.30 0.31 0.38 0.29

Spring 0.55 0.54 0.49 0.40 0.52 0.43 0.43 0.58
Summer 0.56 0.55 0.55 0.48 0.50 0.60 0.59 0.73
Autumn 0.23 0.25 0.24 0.24 0.27 0.27 0.26 0.26

Fig. 8. (a): Average seasonal AOD concentrations in 2018 (b): Land cover Map of UAE.

differences are observed in the levels of AOD concentrations between Merged Dark Target Deep Blue Algorithm. As described earlier the
the findings in this study and the study of Ali and Assiri (2019) over the MAIAC algorithm yields much more accurate estimates of AOD in
UAE. Ali and Assiri (2019) showed higher levels over the UAE in the comparison to the other algorithms, and as such differences in the
autumn and lower in the winter in contrast to this study. This is likely estimated AOD are expected.
due to the different algorithms used to derived the MODIS AOD product.
This study uses the MAIAC algorithm while the other studies use the

8
A. Abuelgasim et al. Remote Sensing Applications: Society and Environment 23 (2021) 100532

Fig. 10. One year ahead forecast, out-of-sample, of AOD based on the SARIMA
Model. The dark and light grey shaded area shows the 80% and 95% predic­
tion intervals.

function in the forecast package of the R software, the best model based
on the AIC criterion is the ETS(M, N, M] model, which means a multi­
plicative error, no trend and multiplicative seasonality. Similarly, the
ETS[M, N, M] also seems to provide a good fit to the data. Fig. 9(b)
shows the model forecast over the test sample.
Table 3 below compares the performance of the derived SARIMA and
Fig. 9. Forecasts from (a) a SARIMA and (b) an ETS model fitted to the UAE
ETS models in terms of fitting accuracy over the training and test sets.
AOD Data, Training (2003–2016), and Test (2017–2018) sets. The dark and
light grey shaded areas show 80% and 95% prediction intervals, respectively.
The SARIMA model is slightly accurate on the test set when using the
accuracy measures MAE, ​ MAPE, ​ and ​ MASE, compared to the ETS
model. However, the ETS model is better based on RMSE.
Table 3 Both the SARIMA and the ETS models are known for short-term
ARIMA(0,0,1)(2,1,)[12] and ETS[M,N,M] performance, AOD data. forecasting horizons. Thus, one year ahead (12 months), out-of-
Data set RMSE MAE MAPE MASE sample, forecast of AOD based on the SARIMA model is shown in Fig. 10.
SARIMA Training 0.07415945 0.05348690 14.27841 0.7150817
Test 0.08078897 0.05730816 12.34778 0.7661692 5. Conclusions
ETS Training 0.06389007 0.04823168 12.67347 0.6448231
Test 0.07707349 0.05786737 12.79281 0.7736454 This study investigated the long spatiotemporal trends of AOD in the
UAE. The investigation of the spatiotemporal trends of AOD showed
significant seasonal variability. A significant increase in AOD during
4.5. Empirical results: model fit and out-of-sample forecast
summer was observed due to high wind speed, resulting in scattered
events of small to large scale haboobs within the study area, high tem­
To evaluate the derived models’ fit, the AOD data were partitioned
perature, and high relative humidity. The second highest AOD was
into a training or estimation set, covering the period from January 2003
observed in the spring followed by the winter season. The spatial dis­
to December 2016, and a test set covers the 24 months of the last two
tributions of AOD showed higher values over the coastal and desert
years 2017 and 2018. All analyses and potential models were, therefore,
areas. Coastal areas portrayed higher values are likely due to the high
implemented using the training set and model performance was later
levels of evaporation, sea-salt aerosols, and increased anthropogenic
evaluated using the data provided by the test set. For that reason, and to
activities such as coastal urban developments and automobile pollution.
estimate forecasting accuracy, four metrics were utilized; namely, the
The time-series analysis showed that using the monthly average AOD
Root Mean Squared Error (RMSE), the Mean Absolute Error (MAE), the
data, two univariate time series models were fitted to the data; namely,
Mean Absolute Percentage Error (MAPE), and the Mean Absolute Square
the SARIMA and the ETS models. Both models seemed to have achieved
Error (MASE) (De Liverta et al., 2011).
parity in performance. The forecast produced by the SARIMA model,
however, seemed to outperform the ETS model forecast in three out of
4.6. SARIMA model four utilized accuracy measures, MAE, MAPE, and MASE, except in
RMSE. The results suggest that the two univariate time series models can
The training set, covering the period from January 2003 to December provide an accurate and reliable monthly forecast for the AOD con­
2016 was used to fit the seasonal ARIMA model. The function “auto. centrations in the UAE.
arima()” of the “forecast” package in R software was employed in this
process to provide a combination of model parameters that produce the Ethical statement
lowest AIC (Akaike Information Criterion). An ARIMA(0,0,1)〖(2,1,0)〗
_12 with an AIC score of -339.54 was selected as the best fit for the AOD I declare that all ethical practices have been followed in relation to
training data. The model seems to do well, capturing the dynamics in the the development, writing, and publication of the article.
AOD data. The model residuals seem to resemble white noise, with a
Ljung-Box test statistic resulting in a p-value = 0.2367, confirming that Credit author statement
the model does not show a lack of fit. Fig. 9(a) displays the model
forecast over the AOD test set. Abdelgadir Abuelgasim: Conceptualization, Data curation, Fund­
ing, Investigation, Methodology, Software, Writing - original draft.
4.7. ETS model Muhammad Bilal: Writing - review & editing, Visualization, Supervi­
sion, Funding. Ibrahim Abdalla Alfaki: Conceptualization, Data cura­
Similarly, fitting the ETS model to the training data using the “ets()” tion, Formal analysis, Investigation, Methodology, Software, Writing -

9
A. Abuelgasim et al. Remote Sensing Applications: Society and Environment 23 (2021) 100532

original draft. Farahat, A., 2019. Comparative analysis of MODIS, MISR, and AERONET climatology
over the Middle East and north africa. Ann. Geophys. 37, 49–64. [Link]
10.5194/angeo-37-49-2019, 2019.
Declaration of competing interest Holt, C.C., 1957. Forecasting Trends and Seasonals by Exponentially Weighted Averages,
O.N.R. Memorandum 52/1957, vol. 15. Carnegie Institute of Technology, p. 44 [5,
The authors declare that they have no known competing financial 14.
Hsu, N.C., Jeong, M.-J., Bettenhausen, C., Sayer, A.M., Hansell, R., Seftor, C.S., et al.,
interests or personal relationships that could have appeared to influence 2013. Enhanced deep blue aerosol retrieval algorithm: the second generation.
the work reported in this paper. J. Geophys. Res.: Atmosphere 118 (16), 9296–9315. [Link]
50712.
Hyndman, R.J., Koehler, A.B., Ord, J.K., Snyder, R.D., 2008. Forecasting with
Acknowledgment exponential smoothing: the state space approach. In: Library of Congress Control
Number: 2008924784. Springer-Verlag Berlin Heidelberg, 978-3-540-71916-8, e-
Dr. Abdelgadir Abuelgasim and Dr. Ibrahim Alfaki would like to ISBN 978-3-540-71918-2.
Hyndman, R.J.G.A., 2018. Forecasting: Principles and Practice, second ed. [Link].
acknowledge the continuous support from the United Arab Emirates [Link]
National Water and Energy Center under grant number 31R155- Jin, Q., Wei, J., Pu, B., Yang, Z.-L., Parajuli, S.P., 2018. High summertime aerosol
Research Center- NWC -3-2017 and the College of Humanities and Social loadings over the Arabian Sea and their transport pathways. J. Geophys. Res.:
Atmospheres 123 (10), 590. [Link] 568– 10.
Sciences. This work was supported by the Special Project of Jiangsu Kaufman, Y.J., Gitelson, A., Karnieli, A., Ganor, E., Fraser, R.S., Nakajima, T., Mattoo, S.,
Distinguished Professor (1421061801003 & 1421061901001). Holben, B.N., 1994. Size distribution and scattering phase function of aerosol
particles retrieved from sky brightness measurements. J. Geophys. Res. 99 (D5),
10341–10356. [Link]
Appendix A. Supplementary data
Kaufman, Y.J., Tanr, D., Remer, L.A., Vermote, E.F., Chu, A., 1997. Operational remote
sensing of tropospheric aerosol over land from EOS moderate resolution imaging
Supplementary data to this article can be found online at [Link] spectroradiometer. J. Geophys. Res. 102 (96), 51–67.
org/10.1016/[Link].2021.100532. Kumar, K.R., Attada, R., Dasari, H.P., Vellore, R.K., Langodan, S., Abualnaja, Y.O.,
Hoteit, I., 2018. Aerosol optical depth variability over the arabian Peninsula as
inferred from satellite measurements. Atmos. Environ. 187, 346–357. [Link]
References org/10.1016/[Link].2018.06.011.
Levy, R.C., Remer, L.A., Dubovik, O., 2007. Global aerosol optical properties and
Abuelgasim, A., Farahat, A., 2020a. Effect of dust loadings, meteorological conditions, application to Moderate Resolution Imaging Spectroradiometer aerosol retrieval
and local emissions on aerosol mixing and loading variability over highly urbanized over land. J. Geophys. Res. 112, D13210. [Link]
semiarid countries: United Arab Emirates case study. J. Atmos. Sol. Terr. Phys. 199, Levy, R.C., Mattoo, S., Munchak, L.A., Remer, L.A., Sayer, A.M., Patadia, F., Hsu, N.C.,
105215. 2013. The Collection 6 MODIS aerosol products over land and ocean. Atmos. Meas.
Abuelgasim, A., Farahat, A., 2020b. Investigations on PM10, PM2.5, and their ratio over Tech. 6, 2989–3034. [Link]
the emirate of Abu Dhabi, United Arab Emirates. Earth Syst. Environ. [Link] Lyapustin, A., Wang, Y., Laszlo, I., Kahn, R., Korkin, S., Remer, L., Levy, R., Reid, J.S.,
org/10.1007/s41748-020-00186-2, 2020. 2010. Multiangle implementation of atmospheric correction (MAIAC): 2. Aerosol
Albaloushi, R., Alghafri, A., Ghazal, S., Aljaberi, A., Abuelgasim, A., 2016. Investigations algorithm. J. Geophys. Res. 116, D03211. [Link]
on the seasonal and inter-annual variations of the atmospheric aerosol optical depth 2011.
in the United Arab Emirates using MODIS satellite data. In: 37th Asian Conference Lyapustin, A., Korkin, S., Wang, Y., Quayle, B., Laszlo, I., 2012. Discrimination of
on Remote Sensing: Fostering Resilient Growth in Asia, ACRS 2016. Colombo, Sri biomass burning smoke and clouds in MAIAC algorithm. Atmos. Chem. Phys. 12,
Lanka October 16-21, 2016. 9679–9686. [Link]
Alharbi, B.H., Maghrabi, A., Tapper, N., 2013. The March 2009 dust event in Saudi Lyapustin, A., Wang, Y., 2018. MCD19A2 MODIS/Terra+Aqua Land Aerosol Optical
Arabia: precursor and supportive environment. Bull. Am. Meteorol. Soc. 94, Depth Daily L2G Global 1km SIN Grid V006 [Data Set], NASA EOSDIS Land
515–528. [Link] Processes DAAC. [Link]
Ali, M.A., Assiri, M., Dambul, R., 2017. Seasonal aerosol optical depth (AOD) variability Maghrabi, A.H., Alotaibi, R.N., 2018. Long-term variations of AOD from an AERONET
using satellite data and its comparison over Saudi Arabia for the period 2002‒2013. station in the central Arabian Peninsula. Theor. Appl. Climatol. 134, 1015–1026.
Aerosol Air Qual. Res. 17, 1267–1280. [Link] [Link]
Ali, M.A., Assiri, M., 2019. Analysis of AOD from MODIS-merged DT–DB products over Nichol, J., Bilal, M., 2016. Validation of MODIS 3 km resolution aerosol optical depth
the arabian Peninsula. Earth Syst. Environ. 3, 625–636. [Link] retrievals over asia. Rem. Sens. 8, 328.
s41748-019-00108-x. Omari, K., Abuelgasim, A., Alhebsi, K., 2019. Aerosol optical depth retrieval over the city
Almazroui, M., 2019. A comparison study between AOD data from MODIS deep blue of Abu Dhabi, United Arab Emirates (UAE) using Landsat-8 OLI images. Atmospheric
collections 51 and 06 and from AERONET over Saudi Arabia. Atmos. Res. 225, Pollut. Res. 10, 1075–1083.
88–95. [Link] ISSN 0169-8095. Pope, C., Burnett, R., Turner, M., Cohen, A., Krewski, D., Jerrett, M., Gapstur, S.,
Atkinson, R., Carey, I., Kent, A., Van Staa, T., Anderson, H., Cook, D., 2013. Long-term Thun, M., 2011. Lung cancer and cardiovascular disease mortality associated with
exposure to outdoor air pollution and incidence of cardiovascular diseases. ambient air pollution and cigarette smoke: shape of the exposure–response
Epidemiology 24 (1), 44–53. [Link] relationships. Environ. Health Perspect. 119, 1616–1621.
Bilal, M., Nichol, J.E., Bleiweiss, M.P., Dubois, D., Rse, J., 2013. A simplified high Ramanathan, V., Crutzen, P.J., Kiehl, T., Rosenfeld, D., 2001. Aerosols, climate and the
resolution modis aerosol retrieval algorithm (sara) for use over mixed surfaces. Rem. hydrologi-cal cycle. Science 294, 2119–2124.
Sens. Environ. 136, 135–145. Riffler, M., Popp, C., Hauser, A., Fontana, F., Wunderle, S., 2010. Validation of a
Bilal, M., Nazeer, M., Nichol, J., Qiu, Z., Wang, L., Bleiweiss, M.P., Shen, X., Campbell, J. modified AVHRR aerosol optical depth retrieval algorithm over Central Europe.
R., Lolli, S., 2019. Evaluation of terra-MODIS C6 and C6.1 aerosol products against Atmos. Meas. Tech. 3 (5), 1255–1270. [Link]
beijing, XiangHe, and xinglong AERONET sites in China during 2004− 2014. Rem. Sayer, A.M., Hsu, N.C., Bettenhausen, C., Ahmad, Z., Holben, B.N., Smirnov, A.,
Sens. 11, 486. [Link] Thomas, G.E., Zhang, J., 2012. SeaWiFS Ocean Aerosol Retrieval (SOAR): algorithm,
Böer, B., 1997. An introduction to the climate of the United Arab Emirates. J. Arid validation, and comparison with other data sets. J. Geophys. Res. 117, D03206.
Environ. 35, 3–16. [Link] [Link]
Bowerman, B., O’Connell. R. and Koehler. A (2005), Forecasting, time series, and Srivastava, A.K., Devara, P.C.S., Rao, Y.J., Bhavanikumar, Y., Rao, D.N., 2008. Aerosol
regression : an applied approach, 3rd Edition, Belmont, Calif. Duxbury Press, 2005. optical depth, ozone and water vapor measurements over gadanki, A tropical station
Cleveland, R.B., Cleveland, W.S., McRae, J.E., Terpenning, I.J., 1990. STL: a seasonal- in peninsular India. Aerosol Air Qual. Res. 8, 459–476. [Link]
trend decomposition procedure based on loess. J. Off. Stat. 6 (1), 3–33. [Link] aaqr.2008.05.0015.
ly/stl1990. Winters, P.R., 1960. Forecasting sales by exponentially weighted moving averages.
De Livera, A.M., Hyndman, R.J., Snyder, R.D., 2011. Forecasting time series with Manag. Sci. 6, 324–342 [5, 15, 46, 230, 231].
complex seasonal patterns using exponential smoothing. J. Am. Stat. Assoc. 106
(496), 1513–1527.

10