An Enhanced Satellite-Based Algorithm for Detecting and Tracking Dust Outbreaks by Means of SEVIRI Data

Abstract

Dust outbreaks are meteorological phenomena of great interest for scientists and authorities (because of their impact on the climate, environment, and human activities), which may be detected, monitored, and characterized from space using different methods and procedures. Among the recent dust detection algorithms, the RST_DUST multi-temporal technique has provided good results in different geographic areas (e.g., Mediterranean basin; Arabian Peninsula), exhibiting a better performance than traditional split window methods, in spite of some limitations. In this study, we present an optimized configuration of this technique, which better exploits data provided by Spinning Enhanced Visible and Infrared Imager (SEVIRI) aboard Meteosat Second Generation (MSG) satellites to address those issues (e.g., sensitivity reduction over arid and semi-arid regions; dependence on some meteorological clouds). Three massive dust events affecting Europe and the Mediterranean basin in May 2008/2010 are analysed in this work, using information provided by some independent and well-established aerosol products to assess the achieved results. The study shows that the proposed algorithm, christened eRST_DUST (i.e., enhanced RST_DUST), which provides qualitative information about dust outbreaks, is capable of increasing the trade-off between reliability and sensitivity. The results encourage further experimentations of this method in other periods of the year, also exploiting data provided by different satellite sensors, for better evaluating the advantages arising from the use of this dust detection technique in operational scenarios.

1. Introduction

Dust storms are complex processes determined by the interaction of wind speed with bared soils and dry air conditions. In recent years, climate changes and the poorly managed human activities have increased the number of dust events, amplifying their effects in arid and semi-arid regions [1]. The Sahara is the main source of desert dust aerosols on Earth, which are mainly composed of minerals such as quartz, feldspars, and clays [1,2,3]. Such aerosols, having a diameter ranging from 0.1–10 μm and a lifetime of hours to weeks, may be transported thousands of kilometres from the source [1,2,3,4]. Through scattering and absorbing the solar and terrestrial radiation, they affect the air temperature [5,6] and act as a condensation nuclei for meteorological clouds [7,8,9]. Dust particles influence the pH of rainwaters, as well hurricane and tropical cyclone dynamics [10,11], altering both sea and agriculture productivity [2,12]; moreover, they reduce the air quality by increasing the PM10 level in the ambient air, whose limit in Europe was established to be 50 μg/m³, causing serious diseases to populations [13,14,15]. In addition, dust outbreaks may destroy communication facilities, cause possible air traffic disruptions [16,17,18], and transport microorganisms, insects, and chemicals, having a strong toxicological effect on human health [19,20].

Hence, an efficient monitoring and forecasting of dust storms is crucial to mitigate the impact of these meteorological phenomena on both social and economic human activities.

Satellite remote sensing plays an important role in detecting, tracking, and characterizing (in terms of optical depth, altitude, and mean particle size) dust events (e.g., [21,22,23,24,25,26,27]). Among the space-based sensors, SEVIRI (Spinning Enhanced Visible and Infrared Imager), aboard MSG (Meteosat Second Generation) satellites, has been profitably used to study the diurnal cycle of dust emissions (e.g., [28,29]), thanks to the high frequency of observations (15 min) and the spatial resolution of about 3 km × 3 km at nadir. Moreover, this instrument is suited to map airborne dust over wide geographic regions, enabling the identification of dust source areas if proper methods are used (e.g., [30,31]).

In this work, we exploit SEVIRI data to study three massive Saharan dust events that have affected Europe and the Mediterranean basin in recent years, testing an optimized configuration of the RST_DUST (Robust Satellite Technique for Dust detection) multi-temporal technique, which performs a qualitative analysis of dust phenomena from space [32,33]. The algorithm presented here, christened eRST_DUST (enhanced RST_DUST), aims at improving the identification of dust outbreaks, particularly over arid and semi-arid regions, and at better discriminating dust from meteorological clouds, minimizing false detections.

2. Background

Since the 1970’s, satellite remote sensing has been used for studying and monitoring dust outbreaks. The first studies on this topic were performed using the MSS (Multi Spectral Scanner) and VHRR (Very High Resolution Radiometer) sensors, aboard ERTS-1 (Earth Resources Technology Satellites-1) and NOAA (National Oceanic and Atmospheric Administration) meteorological satellites (e.g., [34,35,36]). In these pioneering works, the authors analysed the signal acquired in the visible region of the electro-magnetic spectrum to retrieve the Aerosol Optical Depth (AOD) over oceanic regions, where the contribution of airborne dust to the upwelling visible radiance is stronger [37]. Almost concurrently, Shenk and Curran [38] investigated the brightness temperature difference (BTD) BT11–BT12, based on signals measured in the Thermal Infrared (TIR) band at around 11 μm and 12 μm wavelengths, to detect dust clouds over land surfaces. Indeed, the BTD generally decreases up to negative values in the presence of desert dust aerosols [39], owing to the reverse absorption effect of silicate-rich particles in comparison with water/ice droplets at the aforementioned wavelengths (e.g., [40]). Following this, a number of studies exploited TIR and/or MIR (Medium Infrared) radiances to investigate dust phenomena (e.g., [41]), also using data provided by sensors like AVHRR (Advanced Very High Resolution Radiometer), onboard NOAA (National Oceanic and Atmospheric Administration) satellites, and MODIS (Moderate Resolution Imaging Spectroradiometer), aboard NASA Terra and Aqua platforms.

Ackerman [42] developed a three-band technique to better distinguish dust from meteorological clouds, combining two brightness temperature differences: i.e., BT11–BT12 and BT8–BT11 (analysing the signal at 8.5 μm and 11 μm wavelengths), whose values depend on several factors such as particle size and concentration. Wald et al. [43] experimented with a similar approach to improve the identification of mineral dust over desert regions, implementing a bi-spectral threshold test on MODIS data. This method required an accurate characterization of background surface emissivity to provide reliable results [44].

The Saharan Dust Index (SDI) [45], developed for correcting sea surface temperatures retrieved by satellite, is another well-established method used for monitoring dust outbreaks. Specifically, the SDI uses four SEVIRI infrared spectral channels to detect airborne dust, including channel 4 (MIR) centred at a 3.9 μm wavelength. Because of the reflected solar component affecting this spectral channel, the SDI (which is less effective for satellite zenith angles greater than 60°) was only useful in night-time conditions [45]. Thereafter, a daytime SDI was developed, but with a degraded accuracy in comparison with the original algorithm configuration [46].

Recent dust detection methods have analysed SEVIRI imagery at two different time slots per day [47], employed a cloud screened BTD image [48], used dynamic reference brightness temperature differences [49], or integrated several fixed threshold tests on BTD and BTR (Brightness Temperature Ratio) signals [50]. Despite the advantages offered by these algorithms, as well as by other novel techniques (e.g., [51,52]) also combining information provided by different satellite sensors (e.g., [53]), an effective identification and mapping of airborne dust, regardless of background surfaces and of atmospheric/observational conditions, still represents a challenge for scientists. In particular, some critical scenarios remaining challenging for most of the algorithms analysing VIS/IR radiances developed so far include: airborne dust identification over bright surfaces (e.g., desert regions); the dependence of IR signals on dust plume features (e.g., plume height [54]); the sensitivity of BTD to variability in surface emissivity [48]; the impact of cirrus clouds on the BTD signal [48,55].

In this work, we address this topic proposing and testing an enhanced configuration of the RST_DUST multi-temporal algorithm [32], which is described in detail in the next section.

3. Methods

3.1. Standard RST_DUST Algorithm

The RST (Robust Satellite Techniques) multi-temporal approach considers each anomaly in the space-time domain as a deviation from an unperturbed state that may be preliminarily determined by processing plurennial time series of homogeneous (i.e., same calendar month, same overpass times) cloud-free satellite records. In more detail, RST uses a change detection index named ALICE (Absolutely Local Index of Change of Environment) to identify anomalous variations of the signal related to possible perturbing events [56]:

$${\otimes }_V(x,y,t)\equiv \frac{\lfloor V(x,y,t)-V_{ref}(x,y)\rfloor }{{\sigma }_V(x,y)}$$

(1)

In Equation (1), $V\left(x,y,t\right)$ is the satellite signal under investigation, $\left(x,y\right)$ are the geographic coordinates of the image pixel, and t is the acquisition time. $V_{ref}\left(x,y\right)$ is the expected value of the signal (e.g., minimum, maximum, average value), while ${\sigma }_V\left(x,y\right)$ stands for the temporal standard deviation. These terms are calculated after filtering out water/ice clouds by means of an adequate cloud detection method and using the iterative kσ clipping filter to remove signal outliers [56].

RST_DUST is a specific configuration of the RST approach; it uses the following local variation index, based on Equation (1), to identify dust clouds [28,29]:

$${\otimes }_{\mathsf{\Delta }TIR}(x,y,t)\equiv \frac{\lfloor \mathsf{\Delta }T(x,y,t)-{\mu }_{\mathsf{\Delta }T}(x,y)\rfloor }{{\sigma }_{\mathsf{\Delta }T}(x,y)}$$

(2)

In Equation (2), $\mathsf{\Delta }T\left(x,y,t\right)=B{T}_{11}\left(x,y,t\right)-B{T}_{12}\left(x,y,t\right)$ is the brightness temperature and difference at 11 μm and 12 μm wavelengths, whereas the terms ${\mu }_{\mathsf{\Delta }T}\left(x,y\right)$ and ${\sigma }_{\mathsf{\Delta }T}\left(x,y\right)$ are the relative temporal mean and standard deviation computed, for the same location (x,y), by means of the above-mentioned multi-temporal analysis. The ⊗_ΔTIR(x,y,t) index was originally used to detect ash plumes (e.g., [57]), since negative values of this index generally characterize ash/dust clouds dominated by particles having a diameter smaller than 10 μm [33]. Regarding the airborne dust, we assessed such an index behaviour studying an intense dust outbreak affecting Australia during September 2009 by means of infrared MTSAT-1R (Multifunctional Transport Satellites) data [32]. Afterwards, we demonstrated that RST_DUST is capable of performing better than traditional split window methods [33,58]. The algorithm, however, may be still affected by some meteorological clouds (e.g., high level clouds) and generally performs less well for arid and semi-arid regions [33]. These limitations suggested that we needed to optimize the algorithm performance, exploiting information provided by other SEVIRI spectral channels, as discussed in the next section.

3.2. Enhanced RST_DUST Algorithm (eRST_DUST)

In this study, we introduce two additional local variation indices, based on Equation (1), to increase the performance of RST_DUST:

$${\otimes }_{VIS}(x,y,t)\equiv \frac{\lfloor R_{VIS}(x,y,t)-{\mu }_{VIS}(x,y)\rfloor }{{\sigma }_{VIS}(x,y)}$$

(3)

$${\otimes }_{TIR}(x,y,t)\equiv \frac{\lfloor B{T}_{11}(x,y,t)-{\mu }_{BT11}(x,y)\rfloor }{{\sigma }_{BT11}(x,y)}$$

(4)

In Equation (3), $R_{VIS}\left(x,y,t\right)$ is the radiance measured in the visible channel of SEVIRI centred at around 0.6 μm, while ${\mu }_{VIS}\left(x,y\right)$ and ${\sigma }_{VIS}\left(x,y\right)$, which have the same meaning as before, in reference to the visible signal. Equation (4), $B{T}_{11}\left(x,y,t\right)$ is the TIR brightness temperature (at around 11 μm); ${\mu }_{BT11}\left(x,y\right)$ and ${\mathsf{\sigma }}_{BT11}\left(x,y\right)$ are the relative temporal mean and standard deviation.

The ⊗_VIS(x,y,t) index is integrated within the RST_DUST scheme to filter out pixels less reflective in the visible band, which are generally not affected by dust or by water/ice clouds (e.g., [59]). The ⊗_TIR(x,y,t) index is used for better distinguishing dust from meteorological clouds (e.g., high level clouds), exploiting their different spectral behaviour at TIR wavelengths (e.g., [60]).

Thus, eRST_DUST implements up to three local variation indices in combination to identify dusty pixels over land (L) and sea (S) areas, according to the following detection scheme:

Daytime:

⊗_VIS(x,y,t) > 0 AND ⊗_TIR(x,y,t) > −2 AND ⊗_ΔTIR(x,y,t) < 0, −1, −2, … (L)

(5)

⊗_VIS(x,y,t) > 1 AND ⊗_TIR(x,y,t) > −2 AND ⊗_ΔTIR(x,y,t) < 0, −1, −2, … (S)

(6)

Nighttime:

⊗_TIR(x,y,t) > −2 AND ⊗_ΔTIR(x,y,t) < −1, −2, −3 (L/S)

(7)

As can be seen from tests (5) and (6), two different cutting levels of the ⊗_VIS(x,y,t) index are used in the daytime to take into account the smaller contribution of airborne dust over land than sea areas to upwelling visible radiance measured by satellite (e.g., [37]). Obviously, during the night-time, the ⊗_VIS(x,y,t) index is unusable. Therefore, to guarantee a performance comparable to the daylight conditions, values of −1 ≤ ⊗_ΔTIR(x,y,t) < 0 are not investigated (see test (7)).

It is worth stating that eRST_DUST is a tuneable algorithm. Hence, the lower the ⊗_ΔTIR(x,y,t) index, which is the most sensitive to airborne dust, the higher the confidence level of detection.

Implementation on MSG-SEVIRI Data

To implement both the standard and the enhanced RST_DUST algorithm on MSG-SEVIRI data, we processed hundreds of satellite imagery, stratified according to the same month and time of acquisition, acquired at the School of Engineering (SI) of the University of Basilicata (UNIBAS) (Italy) since 2004 (e.g., data of 2004–2007 were processed for studying the dust episode of May 2008).

The first step of the pre-processing phase was the calibration of satellite imagery [61,62]. Afterwards, we extracted a Region of Interest (ROI) from the original full disk data, corresponding to a sub-scene of 725 × 533 pixels in size, covering North Africa and Europe (see Figure 1). Moreover, we integrated a specific cloud-screening procedure within the process to filter out clouds from the scenes (aerosol effects were removed by means of the iterative kσ clipping filter; see Section 3.1), making the spectral reference fields as representative as possible of unperturbed conditions (in the presence of residual outliers, an increase in the natural variability of the signal is expected). This procedure combines the outputs of the EUMETSAT (European Organisation for the Exploitation of Meteorological Satellites) cloud mask product, whose detailed description can be found in [63], with those of the One Channel Cloud Detection Approach (OCA) [64]. The latter is a RST-based method which identifies cloudy radiances, based on the signal divergence from an unperturbed “clear sky” condition, exploiting information provided by SEVIRI VIS (at 0.6 μm) and TIR (at 10.8 μm) channels (e.g., [64,65]). The joint use of EUMETSAT and OCA cloud mask products generally guarantees an efficient removal of cloudy pixels, both in night-time and daylight conditions.

In Figure 1, we show an example of temporal mean and standard deviation images which were generated, for the month of interest (i.e., May), in reference to the time slot of 06:00 UTC, after using the above-mentioned cloud mask procedure. As can be seen from the top panel (see Figure 1a), both arid/semi-arid regions of North Africa and ice/snow surfaces (e.g., Alps) show higher values of temporal mean and standard deviation than sea areas (see Mediterranean Sea) in the visible band. In general, the satellite signal tends to fluctuate less over sea than land regions, as is evident when looking at the right panels, referring to the temporal standard deviation images for the VIS (Figure 1a), TIR (Figure 1b), and BT11-BT12 (Figure 1c) signals.

4. Validation Products

In this work, we use the following aerosol products to assess the identification of dusty areas by satellite:

• NASA Terra and Aqua-MODIS AOD at 550 nm products used here (referring to same time and hour of processed SEVIRI data) belong to the MOD04_L2 and MYD04_L2 datasets of the Level-1 & Atmosphere Archive and Distribution System (LAADS). These aerosol products, with a spatial resolution of 10 km × 10 km (at nadir), are based on MODIS radiances analysed by means of two different algorithms: one for the deep ocean areas [66] and the other for dark targets (e.g., vegetated regions; dark soils; [67,68]). Independent analyses showed that the expected error in MODIS AOD at 550 nm is in the order of ± (0.05τ + 0.03) over oceans and ± (0.15τ + 0.05) over land regions [69], where τ is the aerosol optical depth. Since the traditional MODIS algorithm generally fails in retrieving the AOD over bright land areas [70], the MODIS Deep Blue AOD daily products are commonly used over arid and semi-arid regions (e.g., [71]). These products, which are generated by using the MODIS blue channel radiances, where the contribution from the surface is relatively low, provide information about dust aerosols residing close to the ground [72]. In this study, we exploit information provided by the MODIS Deep Blue AOD daily maps (generated from Giovanni MODIS data system) belonging to collection 5.1 (having the same spatial resolution as above) to assess the identification of airborne dust over North Africa. However, it is important to stress that we do not use the aforementioned aerosol products to perform a quantitative pixel-by pixel assessment of dust detections from space. On the other hand, we only investigate the possible spatial agreement between dusty areas identified by our algorithm and regions characterized by significant values of AOD (which does not require the resampling of SEVIRI data to fit with the lower spatial resolution of analysed MODIS aerosol products).
• CALIOP is a two-wavelength polarization lidar providing high-resolution vertical profiles of clouds and aerosols in the troposphere and lower stratosphere [73,74]. This lidar is the primary instrument of CALIPSO (Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observations), which is part of the NASA “A-train” constellation [73,74,75]. The Level 1 CALIOP data products include calibrated and geo-located profiles of the total backscatter signal at 532 nm and 1064 nm, as well as of the perpendicular component of the 532 nm backscatter return [72]. The CALIOP Level 2 products analyse aerosol and cloud layers, respectively, classifying them into different types and the ice/water phase [74,76]. CALIOP observations have been widely exploited to assess the presence of desert dust aerosols inferred by AOD products at 532 nm and 1064 nm (e.g., [51,77]).
• AERONET is a global network of ground-based sun photometers measuring the atmospheric aerosol properties and providing globally distributed observations of AOD (at 340, 380, 440, 500, 675, 870, 940, and 1020 nm) inversion products and precipitable water in different aerosol regimes [78]. AERONET observations were largely used to validate the satellite retrievals of desert dust (e.g., [26,68]). In this study, we analyse the temporal profiles of the Ångström exponent (α), made available online by the network, to assess the identification of airborne dust over specific locations of interest (see next section). The Ångström exponent gives qualitative information about the aerosol particle size (e.g., [79]). Specifically, values of α < 1 indicate a dominance of coarse mode aerosols (e.g., dust; sea salt particles), whereas values of α > 2 are associated with a size distribution dominated by fine mode aerosols (e.g., urban pollution and biomass-burning aerosols) (see [80] and reference herein).

5. Case Studies and Results

5.1. Saharan Dust Event of 18–23 May 2008

During 18–23 May 2008, a massive Saharan dust loading coming from Algeria and Tunisia moved towards Europe. On 18 May, the dust phenomenon mainly affected Italy and Central Europe. From 19 May, the dust plume dispersed over Eastern Europe, strongly affecting Greece (reaching an altitude of about 4–5 km above sea level) [81]. In the following hours, the dust concentration decreased, although desert dust aerosols remained in the troposphere for a long time [81].

Figure 2 shows the dust maps generated using the standard (Figure 2a) and enhanced (Figure 2b) RST_DUST algorithm, in reference to the SEVIRI data of 19 May 2008 at 09:15 UTC; the two colour codes indicate the regions characterized by a high (brown pixels) and medium-low (orange pixels) probability of dust presence in the atmosphere. The maps display the visible channel in the background for better emphasizing the cloudy areas (in bright tones). As can be seen from the figure, both RST algorithms flagged the presence of dust over the same portion of Mediterranean Sea. Nevertheless, according to the eRST_DUST map of Figure 2b, the dust phenomenon was more intense (see brown pixels) and spatially extended (especially over North Africa, Southern Italy, the Ionian Sea, and Eastern Europe) than that indicated by the standard RST_DUST technique (see Figure 2a).

To assess the identification of Saharan dust over the region of interest, Figure 3a displays the Terra-MODIS AOD at the 550 nm product of 19 May 2008 at 09:15 UTC. The latter shows that large AOD values (in red), indicating the presence of a high aerosol loading (e.g., [82]), affected the above-mentioned dusty region of the Mediterranean Sea. Moreover, values 0.4 < AOD < 0.6, compatible with the presence of a medium-dense dust layer (e.g., [83]), were independently retrieved over both the Puglia region (Southern Italy) and Ionian Sea (see region marked by the white circle in Figure 3a). Regarding the other dusty areas of Figure 2, a low aerosol loading affected southeastern Europe (e.g., Balkans; see Figure 3a), whereas the Terra-MODIS aerosol product did not provide any information over North Africa. It is worth mentioning that, according to a previous literature study, Saharan dust actually dispersed over eastern parts of Europe on 19 May 2008 [81]. Hence, it is realistic to assume that the dusty pixels flagged by eRST_DUST over the Balkans did not represent artefacts, but were rather associated with a thin dust layer.

To validate the dust detections over arid/semi-arid regions of North Africa, Figure 3b displays the Deep Blue MODIS AOD at the 550 nm daily product of 19 May 2008.

The map shows that medium-high values of AOD were retrieved over Libya, Tunisia, and Algeria, where eRST_DUST flagged several dusty pixels. Furthermore, where the algorithm considered the dust phenomenon as more intense (see region marked by the black ellipse in Figure 2b), the aerosol loading was particularly significant (see the high AOD values retrieved over same region in Figure 3b).

These investigations show that eRST_DUST is capable of performing better than the standard RST_DUST technique, particularly over land surfaces; including arid/semi-arid regions where we estimated (for the selected test case) an increase of sensitivity around 13%. Moreover, the algorithm proposed here seems to be more effective in discriminating dust from meteorological clouds (thanks to the use of ⊗_TIR(x,y,t) index; see Section 3.2), as is evident when looking at artefacts reduction over cloudy regions of North Africa, Italy, and Central Europe (see Figure 2b).

Figure 4 displays the dust map of May 19 at 12:30 UTC (based on SEVIRI data acquired about three hours after that of Figure 2) and the Aqua-MODIS AOD at the 550 nm product of the same day and hour, confirming the efficiency of eRST_DUST detections. In particular, Figure 4a shows the persistence of the dust phenomenon over the Mediterranean Sea and part of the Puglia region, in good agreement with that indicated by the MODIS aerosol product of Figure 4b (see high AOD values retrieved over those areas). However, some dusty regions of Figure 4a (e.g., Southern Greece and Aegean Sea; see white ellipse in Figure 4b) were not quantified in terms of AOD (they were partially out of the MODIS scene or considered as overcast). Thus, to verify the presence of a dust layer over specific locations of interest, we investigated the temporal profiles of the Ångström exponent, made available online by the AERONET stations of Lecce-University (Puglia region; [84]) and Athens (Greece; [85]), indicated by green crosses in Figure 4a. Those profiles, which are not reported here, show that values of α (in the range 440–870 nm) derived at the Lecce-University measurement station on 19 May 2008 ranged from 0.3 to 0.4 (see [84]). With regards to Athens, α was still small assuming values between 0.3 and 0.75 (see [85]). Hence, although we cannot assess the identification of airborne dust over Southern Greece (e.g., Peloponnese) by means of the MODIS aerosol product of Figure 4b, it is realistic to suppose that a dust layer actually affected that area (as well as the Puglia region).

To investigate the space-time evolution of the monitored dust phenomenon in a retrospective way, Figure 5 displays dust maps of 00:00 UTC, 06:00 UTC, and 12:00 UTC, generated from the processed SEVIRI data of 18 May–23 May 2008. The maps report dusty pixels identified by eRST_DUST (see orange/brown pixels) and cloudy areas (in black) flagged by the EUMETSAT Cloud Mask (CLM) product, in order to emphasize the impact of meteorological clouds on the results of dust detection. As can be inferred from the figure, on 18 May, the dust plume affected North Africa and part of the Mediterranean Sea (see region between 30–40N; 10–20E), moving towards E-NE. The day after the dust coverage extended to the Balkans, affecting Greece. During 20–21 May, airborne dust dispersed over Eastern Europe, whereas, starting from 22 May (apart from the early morning of same day, when it seems that a dust layer affected North Africa once again), the dust event become progressively less intense (e.g., see the reduction of dusty pixels, particularly those depicted in brown).

It is worth noting that a diffuse cloud coverage affected the ROI during the investigated period (e.g., see 18–20 May). Hence, some regions of the dust plume probably remained undetected by satellite in the presence of overlying meteorological clouds (e.g., cirrus clouds). The panel of 19 May at 00:00 UTC shows, for instance, that the dust event was clearly underestimated over the Mediterranean Sea because of cloud coverage (e.g., see the dust map of the same day at 06:00 UTC for comparison). Moreover, eRST_DUST flagged a lower number of dusty pixels in night-time scenes, when compared to daytime scenes (e.g., see North Africa on panels of 18–19 May at 00:00 UTC for a comparison with those of 06:00 UTC). In addition, some possible artefacts were generated in areas located far from the dust front (i.e., also close to cloudy regions of Central Europe) that did not fit with the observed dust trajectory (e.g., see dusty pixels flagged on 20–21 May panels).

Despite the above-mentioned limitations, the dust episode of 18–23 May 2008 was, in general, well tracked from space. Information provided, for instance, over Southern Greece, seems to be consistent with the temporal trend of the Ångström exponent independently investigated over the Athens AERONET site, showing very small values of α (even lower than 0.5) during 19–21 May, along with the increase of the same parameter in the following days [81], when according to eRST_DUST, the dust phenomenon weakened.

5.2. Saharan Dust Event of 27–30 May 2008

During 27–30 May 2008, a heavy dust loading coming from the North of Sahara region (North-Eastern Algeria, Tunisia, and Libya) moved towards Europe, diffusing up to Scandinavia and the Norwegian Sea. Although low-pressure surface systems over the Western Mediterranean and Central Europe caused heavy rainfalls, they did not prevent the transport of large amounts of dust through the Alps towards Central and Northern Europe [86].

Figure 6a displays the eRST_DUST product of 29 May at 00:00 UTC, showing the presence of a dust plume (see orange/brown pixels) over a region extending from part of the Mediterranean Sea to part of the Balkans. To assess the presence of airborne dust over this region (where the plume was probably only partially detected by the satellite because of cloud coverage), Figure 6b displays the Total (parallel + perpendicular) attenuated backscattered profile measured by CALIOP, for the same day of the analysed SEVIRI data, at 00:24:12–00:37:41 UTC along the orbit track of CALIPSO (depicted in purple in Figure 6a). The figure shows that an aerosol layer affected the B-C-D region (see black ellipse) where eRST_DUST flagged the presence of airborne dust, reaching an altitude of about 7 km above sea level (a.s.l).

In more detail, eRST_DUST identified a few dusty pixels over the region where the polluted dust was dominant (i.e., along the A-B orbit track of CALIPSO), as indicated by the analysis of the aerosol sub-type profile of Figure 6c. On the other hand, the number of dust detections drastically increased in the presence of desert dust aerosols, apart from the locations clearly affected by meteorological clouds (see black pixels in Figure 6a).

To assess the performance of eRST_DUST in tracking the second dust episode of May 2008, we show, in Figure 7, some dust maps generated through the processing of SEVIRI data (of the same hours of Figure 5) acquired at the end of month. As can be seen from the figure, the dust plume coming from North Africa dispersed first over the Mediterranean Sea and Italy, affecting Eastern Europe from 28 May. Afterwards, an evident intensity reduction characterised the monitored dust event (see dust maps of 30–31 May), as revealed by the lower number of dusty pixels depicted in brown. It is worth mentioning that during the first phase of the event, eRST_DUST also flagged the presence of airborne dust over some cloudy regions, which were in spatial contiguity with dusty areas of North Africa (e.g., see panel of 27 May at 06:00 UTC). The identification of dusty pixels over those regions seems to indicate that the dust layer was probably higher in altitude than meteorological clouds.

Independent EARLINET (European Aerosol Research Lidar Network) measurements performed over Italian and German regions during the period of interest [87], as well as OMI (Ozone Monitoring Instrument) observations reported and discussed by other authors in a previous study [88], seem to confirm the results of Figure 7. In particular, AI (Aerosol Index) maps of 12:00 UTC obtained from OMI data, describing regions affected by UV absorbing aerosols, revealed, on 27 May, the presence of a dust plume over North Africa and nearby southeastern France, dispersing over Italy and Eastern Europe the day after. On 29 May, a smaller dust coverage was observed over the Mediterranean Sea, while on 30 May, the AI map showed the movement of the dust front towards the north of Europe [88].

Information provided by eRST_DUST is also consistent with that forecasted by the BSC-DREAM8b model (predicting the atmospheric life cycle of eroded desert dust over North Africa, the Middle East, and Europe, by solving the Euler-type equation for dust mass continuity [89,90]), whose products were independently assessed in previous literature studies (e.g., [91]). In detail, a comparison of the dust maps of Figure 7 with dust forecast products of Figure 8 shows that detected and predicted dust plumes were similar in terms of the shape, spatial extent, and direction, with the main differences characterizing regions where eRST_DUST did not detect dust because of cloud coverage (e.g., see panels of 27 May at 12:00 UTC). This is evident when looking at the dust maps of 27–29 May 2008, revealing the identification of several dusty pixels over areas (e.g., see North Africa and Italy) where the BSC-DREAM8b model forecasted, for the same day and hour, the presence of a significant dust loading. Moreover, the intensity reduction of the monitored dust phenomenon observed by satellite during 30–31 May is compatible with that foreseen by the model for the period of interest (see Figure 8).

5.3. Saharan Dust Episode of May 2010

In mid-May 2010, a thick dust layer coming from the Sahara region affected the Mediterranean Sea and coasts of Libya. In the following days, airborne dust dispersed over the Peloponnese region [92]. During 13–14 May, a number of EARLINET stations observed the dust event in concomitance with volcanic ash emitted by the Eyjafjallajökull (Iceland) volcano [93,94].

Figure 9a displays the eRST_DUST map of 11 May at 12:30 UTC, showing the identification of a dust plume (depicted in brown) extending from Northeastern Libya to Sicily, dispersing over the region intersected by the orbit track of CALIPSO (in blue). In particular, the Total Attenuated Backscatter profile of Figure 9b, acquired on May 11 between 12:20 and 12:34 UTC, shows that an aerosol layer (reaching the altitude of ~5 km a.s.l.) affected the region marked by the black ellipse, where eRST_DUST flagged a significant number of dusty pixels (see B-C region in Figure 9a). In addition, the aerosol sub-type profile of Figure 9c classified that aerosol layer as desert dust, further corroborating the information provided by our algorithm.

It should be pointed out that only values of ⊗_VIS(x,y,t) > 0 (L), 1 (S) AND ⊗_TIR(x,y,t) > −2 AND ⊗_ΔTIR(x,y,t) < −1 were analysed in Figure 9a to identify dusty areas minimizing artefacts. Figure 10 displays the eRST_DUST products (with a two-hour time interval) of 11–13 May 2010 at 06:00–14:00 UTC, showing that when running the aforementioned daytime algorithm configuration, only a few pixels were erroneously flagged as dust (e.g., see artefacts generated over NE corner of image on panel of 12 May at 12:00 UTC). Figure 11 quantifies the total dusty area retrieved from Figure 10 (by counting the number of detected dusty pixels and assuming a mean SEVIRI pixel area of 15 km²). As can be seen from the plot, the dust coverage reached the maximum spatial extent in the morning of 11 May, when the Saharan dust dispersed over North Africa and the Mediterranean Sea. In the following days (i.e., when the plume moved towards SE affecting Greece), the total area affected by dust significantly decreased. It is worth noting, however, that despite the general decreasing trend characterizing the analysed parameter, a relative minimum in the total area affected by dust was systematically recorded at 12:00 UTC, indicating that eRST_DUST was probably less effective in detecting airborne dust on satellite imagery acquired at noon (i.e., at local early afternoon), regardless of the background surface.

Besides, whilst the use of the single (and most conservative) daytime algorithm arrangement minimized artefacts, it also reduced the algorithm performance over land areas.

This is evident when looking at Figure 12a, which shows that even pixels having values of ⊗_VIS(x,y,t) > 0 AND ⊗_TIR(x,y,t) > −2 AND ⊗_ΔTIR(x,y,t) < 0 had to be analysed to detect airborne dust over the coastal region of Tunisia for SEVIRI data of 12 May at 12:00 UTC. Over this region, a dust layer was presumably present, as indicated by the BSC-DREAM8b forecast product of Figure 12b.

The spatial profiles of eRST_DUST indices, retrieved along the A-B-C-D transect region (in yellow) of Figure 13a, better emphasize that discussed above. In detail, Figure 13b describes the spatial behavior of ⊗_ΔTIR(x,y,t) (i.e., ALICE BTD; in red) and of the BTD signal (in black) along the same transect, showing that both indicators assumed negative values over the A-B transect portion, owing to the presence of Saharan dust. However, along the contiguous B-C dusty region, the BTD increased, becoming positive, while values of −1 < ⊗_ΔTIR(x,y,t) < 0 were recorded. Regarding the C-D transect portion, intersecting part of the Mediterranean Sea (which was still affected by dust), the BTD decreased, remaining positive, whilst the ⊗_ΔTIR(x,y,t) index assumed values even lower than −2. Thus, eRST_DUST was capable of detecting the plume, even where the dust BTD signature was weakened (i.e., along the B-D transect region), and common fixed threshold tests applied to the simple BTD signal (e.g., BTD < 0) would have failed in detecting airborne dust.

As a matter of fact, Figure 13c, displaying the spatial trend of the ⊗_VIS(x,y,t) (i.e., ALICE VIS curve; in red) and ⊗_TIR(x,y,t) indices (i.e., ALICE TIR curve; in black), shows that eRST_DUST did not detect dust only in correspondence of a few land locations indicated by green arrows. Over these locations, the ⊗_VIS(x,y,t) index assumed negative values, preventing the possible identification of airborne dust. Since high values of the temporal mean typically characterize bright surfaces (e.g., desert regions, see Figure 1a), we can speculate that cloud shadows caused the abrupt reduction of the ⊗_VIS(x,y,t) index recorded in Figure 13c. This hypothesis seems to be corroborated by Figure 14, displaying the HRV (High Resolution Visible) SEVIRI image of 12 May 2010 at 12:00 UTC at a 1 km² spatial resolution.

It should also be noted that the ⊗_VIS(x,y,t) index strongly increased in magnitude (in the presence of desert dust aerosols) when passing from land to sea areas (i.e., from B-C to C-D transect region of Figure 13c), because of the changes in the values of the spectral reference fields previously discussed in Section 3. Moreover, the ⊗_VIS(x,y,t) index reached its peak over sea region where the lowest values of ⊗_ΔTIR(x,y,t) index were retrieved (see Figure 13b), indicating that a significant dust layer probably affected that area (see red pixels in Figure 12a). Finally, the increase of both the BTD signal (up to positive values) and the ⊗_ΔTIR(x,y,t) index recorded along the B-C transect region was possibly determined by differential changes in the surface emissivity, related to the presence of a different background (e.g., the transect of Figure 13a partially intersected the Chott el Jerid salt lake). Those emissivity variations affected the ⊗_ΔTIR(x,y,t) index (remaining negative) in a less significant way than the simple BTD signal (see Figure 13b), confirming the low impact of site effects on the performance of the algorithm presented and tested in this work.

6. Discussion

The RST_DUST method, which uses the single ⊗_ΔTIR(x,y,t) index to detect dust clouds, was investigated and produced good results for studying and monitoring dust outbreaks that have affected Australia, the Mediterranean basin, and the Arabian Peninsula in recent years, in spite of some limitations (e.g., reduction of sensitivity over desert regions, dependence on some meteorological clouds [28,58]). In this work, we have addressed these issues by assessing the performance of eRST_DUST, which is an optimized configuration of the above-mentioned multi-temporal technique. The algorithm proposed here uses up to three local variation indices in combination to identify Saharan dust from space, integrating information provided by single TIR and/or VIS channels of SEVIRI to that retrieved from the analysis of the BTD (i.e., BT11-BT12) signal.

The results presented in this study have shown that eRST_DUST is capable of performing better than the standard RST_DUST algorithm, increasing the trade-off between reliability and sensitivity; thanks to a more efficient identification of dusty pixels, especially over land surfaces (e.g., arid/semi-arid regions), coupled with an improved capacity to distinguish dust from meteorological clouds. Furthermore, eRST_DUST is capable of providing information even in regions where the dust BTD signature is weakened (e.g., in the presence of optically thin dust layers) and dust plumes are consequently more difficult to identify (see Section 5.3). Actually, eRST_DUST analysing relative rather than absolute signal variations is less affected (such as the standard RST_DUST algorithm) by site (e.g., seasonal variations of vegetation in semiarid regions causing changes in the surface emissivity) and atmospheric effects (e.g., rich-moisture environment; near-surface temperature inversion), generally affecting performance of satellite-based methods developed to detect airborne dust (e.g., [47,49,95]).

As the main limitations, eRST_DUST shows: (i) a reduction of sensitivity under specific illumination conditions; (ii) the underestimation of plume regions in the presence of meteorological clouds obscuring the BTD signal; and (iii) the generation of residual artefacts.

Sensitivity reductions particularly affect eRST_DUST in the night-time, when to guarantee a reliability level comparable to daylight conditions, values of −1 ≤ ⊗_ΔTIR(x,y,t) < 0 are not investigated (see Section 3), with the consequent reduction in the algorithm performance, mainly over land surfaces. On the other hand, possible reductions of eRST_DUST sensitivity may also occur in the daytime (although in a less significant way than night-time conditions) because of cloud shadows (causing negative values of ⊗_VIS(x,y,t) index; see Section 5.3) and probably due to the diurnal cycle of cloud cover [96]. This work has revealed a systematic reduction of dusty pixels on satellite imagery acquired at noon, owing to the increase of the cloudiness recorded over the ROI in the early afternoon, analysing the operational EUMETSAT CLM products. In addition, some investigations are currently in progress to assess the possible dependence of the ⊗_ΔTIR(x,y,t) index also on dust cloud features (e.g., plume height).

With regards to artefacts, a single eRST_DUST configuration may be used in the daytime for further minimize their occurrence in areas strongly affected by meteorological clouds (see Section 5.3); however, this arrangement would reduce the sensitivity of our method over bright surfaces, probably influencing its performance close to dust emission sources. A spatial noise filter, aiming at removing sparsely distributed pixels that are generally not ascribable to dust, could be then implemented within the process, without reducing the sensitivity. In addition, the multi-temporal analysis of the BT8.6–BT11 signal (which is also sensitive to desert dust aerosols, as discussed in Section 2), could enhance the identification of dust outbreaks under different illumination conditions.

It is important to stress that eRST_DUST requires the availability of a sufficiently populated dataset of multiyear satellite records to be correctly applied (i.e., the longer the time series, the higher the efficiency of dust detections). This offline analysis, devoted to generate the library of spectral reference fields, may require the use of advanced computational systems (e.g., Graphics Processing Unit—GPU) to reduce processing times over broad areas. Nevertheless, once the setup activity is accomplished, the algorithm runs fast, generating dust products a few minutes after the sensing time (e.g., only five minutes are required to generate the dust maps over the region of interest investigated in this work). This means that eRST_DUST running on SEVIRI data enables the monitoring of dust outbreaks in quasi real time. In addition, the algorithm may be potentially used in any geographic area and period of the year, allowing us to investigate dust events related to sources (e.g., Gobi desert, which is one of main sources of Asian dust [97]) different from the Sahara region. Obviously, to assess eRST_DUST performance in other geographic areas and/or in the presence of dust plumes of different features (e.g., for chemical/microphysical composition), specific investigations are required (e.g., exploiting data provided by sensors like MODIS).

It is important to remark that our algorithm only provides qualitative information about dust outbreaks, although dusty pixels associated with a high confidence level of detection affected regions with high AOD values (retrieved by independent aerosol products), while those indicating a less intense dust phenomenon were in good spatial agreement with areas characterized by a lower aerosol loading. Nonetheless, eRST_DUST detections could represent the starting point to perform retrieval analyses aiming at characterizing dust outbreaks from a quantitative point of view, whose accuracy and/or time of elaboration processes could benefit from a more efficient identification of regions affected by dust. Some investigations are currently in progress to assess, for instance, if indirect information about dust clouds may be inferred from the numerator of Equation (2) (i.e., the difference of the BTD signal and its temporal mean value approximating a clear sky condition), once dusty pixels are identified.

Finally, geostationary satellite platforms like GOES-R and Meteosat Third Generation (MTG) thanks to an increased spatial, temporal, and spectral resolution, should further increase eRST_DUST performance, allowing us to better detect and track dust outbreaks. In this direction, satellite images acquired during a 30-day period or belonging to contiguous time slots could be processed to run our algorithm, even in the absence of a plurennial time series of satellite records, increasing the statistics; even if the performance of eRST_DUST (e.g., in terms of reliability) needs to be carefully assessed under this condition.

7. Conclusions

Three massive dust events affecting the Mediterranean basin and Europe in May, which, along with July, represents the month most frequently affected by desert dust aerosols coming from the Sahara region [98], have been analysed in this work, using information provided by some independent and well-established aerosol products to assess eRST_DUST detections.

The results encourage further experimentations of the proposed method in different geographic contexts and periods of the year, even for a comparison with other recent dust detection techniques, exploiting information provided by different satellite sensors (i.e., polar/geostationary ones). These investigations will allow us to evaluate, in a more accurate way, the advantages of using eRST_DUST in operational contexts, assessing its potential in supporting early warning and continuous monitoring systems, devoted to mitigating the effects of dust outbreaks on human health and economic activities.