The sensitivity of the colour of dust in MSG-SEVIRI Desert Dust infrared composite imagery to surface and atmospheric conditions

. Infrared “Desert Dust” composite imagery taken by the Spinning Enhanced Visible and InfraRed Imager (SEVIRI), onboard the Meteosat Second Generation (MSG) series of satellites above the equatorial East Atlantic, has been widely used for more than a decade to identify and track the presence of dust storms from and over the Sahara Desert, the Middle East, and southern Africa. Dust is characterised by distinctive pink colours in the Desert Dust false-colour imagery, however the precise colour is 5 inﬂuenced by numerous environmental properties, such as the surface thermal emissivity and skin temperature, the atmospheric water vapour content, and the quantity and height of dust in the atmosphere. This paper is the follow-up to Banks et al. (2018), which analysed the sensitivity of the colour of the dust in the imagery to its infrared optical properties. The previous paper introduced a modelling system combining dust concentrations simulated by the aerosol transport model COSMO-MUSCAT (COSMO: COnsortium for Small-scale MOdelling; MUSCAT: MUltiScale Chemistry Aerosol Transport Model) with radia- 10 tive transfer simulations from the RTTOV (Radiative Transfer for TOVS) model, in order to simulate the SEVIRI infrared measurements and imagery. Investigating the sensitivity of the synthetic infrared imagery to the environmental properties over a six month summertime period from 2011 to 2013, it is conﬁrmed that water vapour is a major control on the apparent colour of dust, obscuring its presence when the moisture content is high. Of the three SEVIRI channels used in the imagery (8.7, 10.8, and 12.0 µ m), the channel at 10.8 µ m has the highest atmospheric transmittance

The three IR atmospheric window channels at 8.7, 10.8, and 12.0 µm are particularly useful for measurements and imaging of dust aerosol over desert regions, since clear-sky atmospheric transmission is high at these wavelengths: measurements at these wavelengths are therefore most representative of the apparent surface, whether that is land, ocean, cloud, or aerosol (e.g. dust). Meanwhile visible channels have greater difficulty in resolving dust aerosols over deserts due to the weak contrast in reflectance between the desert and dust aerosols. SEVIRI IR measurements are often interpreted in the form of brightness 5 temperatures T B , considering the IR radiance measurements to be representative of blackbody radiation; see Schmetz et al. (2002) and Lensky and Rosenfeld (2008) for more details.
'Desert Dust' RGB composite imagery from SEVIRI is defined in the following configuration, as specified by Lensky and Rosenfeld (2008): This formulation has been designed specifically to highlight the presence of dust aerosol in the imagery, making use of brightness temperatures at 10.8 µm in the blue beam and brightness temperature differences (BTDs) in the red and green beams.
RGB values have a range of 0 to 1, so for example if T B108 has a value of 300 K, which is greater than the maximum value of the model domain extends from 0 to 60 • N, and 30 • W to 35 • E. The horizontal grid spacing is 28 km, while vertically there are 40 sigma-p levels starting with a bottom layer 20 m thick. Simulation output is produced at a 3-hourly resolution.
The radiative transfer model RTTOV has been designed for fast radiative transfer simulations of satellite measurements (Saunders et al. (1999), Matricardi (2005), publicly available at https://www.nwpsaf.eu/site/software/rttov/, last accessed 15th November 2018). Taking surface, atmospheric, and aerosol properties as input, RTTOV calculates the radiances and brightness 5 temperatures that would be measured at TOA, for a defined satellite instrument and channel, for example the SEVIRI channels at 8.7, 10.8, and 12.0 µm. Associated with these radiances, RTTOV also provides the atmospheric transmittance as output, defined as the fraction of the radiance emitted by the surface which passes out to space. Surface properties include thermal emissivity and skin temperature, while atmospheric properties include pressure, temperature, and specific humidity. Solar and viewing zenith angles are also accounted for. Aerosol properties are defined by profiles of the absorption and scattering coeffi-10 cients (km −1 ) and of the back-scatter parameter, which is the integrated fraction of back-scattered energy (Matricardi, 2005).
COSMO-MUSCAT-RTTOV is the modelling system which combines COSMO-MUSCAT simulations with the capabilities of RTTOV in simulating the SEVIRI brightness temperatures (Banks et al., 2018), so as to produce synthetic Desert Dust imagery.
In order to simulate only dusty or clear scenes, RTTOV simulations are not performed for COSMO-MUSCAT grid cells with cloud fractions greater than 1%. Meanwhile in order to focus on desert regions and the desert margins, RTTOV simulations are 15 only performed in the latitude band between 12 and 36 • N, over land.

Optical properties of dust
Knowledge of the optical properties of dust is required both to calculate the aerosol optical depth (AOD), and to calculate the IR effects of dust on the SEVIRI TOA brightness temperatures as simulated by RTTOV. Optical properties are derived from the wavelength-dependent refractive indices of the dust type and the size and shape of the dust particles. Dust particles are 20 assumed to be spherical, see B2018 for more analysis of the influences of particle shape on the imagery.
Optical properties are calculated using Mie theory (e.g. Mishchenko et al., 2002), which requires as input the complex refractive index m(λ) and the scattering parameter 2πr eff / λ, dependent both on the wavelength λ and the particle effective radius r eff . These inputs are used to calculate the extinction, absorption, and scattering efficiencies, which are convolved over the SEVIRI channel filter response functions to calculate the SEVIRI channel efficiencies Q(r eff , ch). Extinction coefficients, 25 β(r eff , ch, t, x), for each channel, location x, and time t, are calculated with reference to the particle size and density ρ p (specified as that of quartz, 2.65 g cm −3 ), and to the mass concentration M in the grid box (e.g. Tegen et al., 2010).
The AOD is defined by the integration of the extinction coefficient over the entire atmospheric column. For consistency with satellite AOD retrieval products (e.g. Hsu et al., 2013), which commonly quote their AOD retrievals using visible wavelengths, COSMO-MUSCAT dust AODs (Helmert et al., 2007) are calculated at 550 nm using dust visible refractive indices presented 30 by Sinyuk et al. (2003). Validation of the North African COSMO-MUSCAT simulated AODs with respect to AOD retrievals from AERONET (Holben et al., 1998) has been performed and is described by B2018, Section 4.
In the IR, calculations of the SEVIRI channel extinction, absorption and scattering coefficients are carried out using refractive indices developed by Volz (1973), henceforth denoted VO73. Further optical properties (Sokolik and Toon, 1999;Hess et al., Di Biagio et al., 2017) databases were also tested by B2018, however the VO73 refractive indices appeared to give the closest match to the observed AODs and the composite image colour, and so in this paper only COSMO-MUSCAT-RTTOV simulations with spherical VO73 dust will be considered. Summarising the optical properties relevant to this paper, Table 1 specifies the size-resolved extinction efficiencies for the five COSMO-MUSCAT size bins, at 550 nm for the Sinyuk et al. (2003) dust, assumed to be spherical, and for the three SEVIRI IR channels of interest using VO73 dust. Specified also in Table   5 1 are the bin size ranges and effective radii.

The atmospheric and surface environment of North Africa
In Section 3.1 the meteorological background to the summertime North African environment is introduced, while Sections 3.2 to 3.5 consider the influences of the North African environment on the satellite IR measurements and imagery.
3.1 North African atmospheric dynamics 10 The summertime (June, July, August) Saharan climate is of particular research interest due to the complexity of the meteorological situation, giving rise to a number of different patterns of atmospheric dust generation and transport. During the summer, the atmospheric circulation over North Africa is predominantly characterised by the strength of the Harmattan north-easterly trade winds, the pulsating nature of the Saharan Heat Low (SHL, e.g. Engelstaedter et al. (2015)), and northward migration of monsoonal air (Schepanski et al., 2017). The southerly monsoon winds, which bring humid and cooler air masses from the Gulf 15 of Guinea into the continent, and the northerly Harmattan winds, which push dry and hot air masses southward, meet at the Inter-Tropical Discontinuity (ITD), a zone marked by a strong gradient in humidity and a jump in wind direction. Generally, dry and desert conditions are present over the North African continent north of the ITD, but inflows of cooler airmasses from the adjacent seas result in intermittent increases in humidity and reduced temperatures. Particular examples include Mediterranean cold air surges (Vizy and Cook, 2009) and the inflow of marine air masses over Western Sahara (Grams et al., 2010). 20 Dust entrained into the planetary boundary layer (PBL) is vertically mixed and eventually homogeneously distributed over the depth of the PBL (e.g. Schepanski et al., 2009b). Most dust source activations occur during the first half of the day (Schepanski et al., 2017), in particular after sunrise when the PBL is deepening and the turbulent mixing increases. However over some parts of the Sahara and the Sahel in summer more than half of the dust emission can be related to moist convection (haboobs) from late afternoon to early evening (Allen et al., 2013;Heinold et al., 2013). Whereas larger dust particles fall out 25 quickly within a couple of hours, smaller particles remain aloft for longer (days to weeks). These particles will be transported away from the source region eventually leaving the continent. At sunset, the turbulent eddies determining the depth and the mixing within the daytime PBL decay and a calm and shallower nocturnal boundary layer (NBL) grows from the surface into the former daytime PBL, referred to as the residual layer. Whereas dust suspended in the NBL settles down due to the lack of turbulent buoyancy, dust suspended in the residual layer may be transported efficiently across longer distances during the night 30 (e.g. Kalu, 1979;Schepanski et al., 2009a).

Surface thermal emissivity
The surface thermal emissivity is one of the controlling variables on the pristine-sky IR composite imagery. In this context, 'pristine-sky' refers to cloud-free and aerosol-free scenes. The emissivity is a wavelength-dependent property, by definition bounded within a range between 0 and 1, but in practice for the three SEVIRI channels of interest the values are always greater than 0.7 over North Africa (Figure 2). There is a much wider range of emissivity values in the 8.7 µm channel than in the 10.8 5 and 12.0 µm channels (panels (a) to (c)): for the latter two channels, the minimum values are greater than 0.9. A characteristic reduction in the emissivity occurs within the approximate range of 8-10 µm, due to the reststrahlen absorption band in quartz silicates (e.g. Wald et al., 1998;Seemann et al., 2008), and hence sandy soils tend to have a much lower emissivity at 8.7 µm than do rockier or more vegetated surfaces. As a result, the geographical and even the temporal variability in the 8.7 µm emissivity will be a strong influence on the variability of the pristine-sky green values in the SEVIRI dust composite imagery.

10
In this study we make use of surface emissivity values derived by Borbas and Ruston (2010) using MODIS data, following techniques developed by Seemann et al. (2008).
Figure 2(d) subdivides the Saharan domain within three ranges of the surface emissivity at 8.7 µm, betweeen 0.7-0.8, 0.8-0.9, and 0.9-1.0. There are differences in the climatological values for June and July, accounting for those grid cells which belong in two different zones ('low-medium' and 'medium-high') at different times. 'Low' emissivity zones generally correspond to 15 dune fields and sand seas, 'medium' zones to rockier desert hamadas and regs, while the 'high' emissivity zones encompass most other surface types including mountains and vegetated areas. It is not generally possible to distinguish between mountains and vegetated areas using the surface emissivity alone. To summarise this, Table 2 quantifies the percentages of the total data subset within each emissivity zone: ∼ 41% lie in each of the low and medium emissivity zones. 18% lie in the high emissivity zone, a consequence of the strict cloud-screening criteria implemented in subsampling the output model data, with clouds more 20 frequently occurring over highly emissive vegetated surfaces.
There is evidence from laboratory measurements that the thermal emissivity in the 8.2-9.2 µm range can increase by up to 16% with increased surface moisture (Mira et al., 2007), with respect to the dry soil case, especially for sandy soils. Li et al. (2012) argue that this results in a diurnal variability in emissivity, with moister soils and hence higher emissivity values at night. Within the seasonal context, it is therefore to be expected that Sahelian and southern Saharan soils in particular will be 25 moister with the onset of the monsoon in late June. This is very apparent for example in northern and eastern Mali between June and July: in the central Sahara in northern Mali many points flip from being in the low to being in the medium emissivity zones, with emissivity values hovering around 0.8 (Figure 2(d)).
Considering a priori the overall effect of the surface emissivity on the pristine-sky imagery colours and neglecting the differences in atmospheric transmission, it is apparent that the slightly higher emissivity values at 12.0 µm than at 10.8 µm would 30 give rise to very slightly positive brightness temperature differences (BTDs) between these channels and hence moderately high pristine-sky red values (Equation 1). Over low 8.7 µm emissivity surfaces the 10.8 µm emissivity values will be much higher, as will be the T B values, giving rise to high green values. For high 8.7 µm emissivity surfaces the contrast between the channels is much reduced, and so the green values will be correspondingly smaller. The combination of these is presented in Figure 2(e), which presents the surface-only contribution to the Desert Dust RGB composite imagery, calculated from these emissivity values and assuming a fixed skin temperature of 30 • C. There is thus a high degree of geographical diversity in the colours arising purely from the intrinsic surface properties.

Skin temperature and column moisture
In the infrared the primary control on the measured radiances is temperature. The radiative emission temperature at the land 5 surface is referred to as the skin temperature (T skin ). High skin temperatures are characteristic of central desert regions during the day, while the coolest skin temperatures occur most frequently on the desert margins and at night. The desert skin temperature has a strong amplitude in its diurnal cycle. Analogously to the emissivities, we define five T skin ranges with widths of 14 K, encompassing the total range of possible values in this data subset ( However, the influence of the skin temperature on the composite imagery colour can be masked by the variability in atmospheric water vapour (also referred to here as the total column water vapour, TCWV) and the consequent inter-channel differences in the atmospheric transmittance. Water vapour is a major absorber of IR radiation, and the differences in the channels' sensitivity to its presence govern the colour response of the imagery to variable water vapour content. As is to be expected, the simulated transmittances across all three window channels are strongly negatively correlated with increasing water vapour. 20 It has been noted before that the 12.0 µm channel has greater water vapour continuum absorption than either the 8.7 or the 10.8 µm channels, leading to lower atmospheric transmittance values in this channel (e.g. Brindley and Allan, 2003). This is confirmed by Table 3, which presents the mean COSMO-MUSCAT-RTTOV calculated channel transmittances, within the four column moisture ranges defined in Table 2. The transmittance is at its maximum in dry atmospheres in the 10.8 µm channel, while it is most reduced under wet conditions for the 12.0 µm channel. 25 The geographical relationship between the average column moisture and the average channel atmospheric transmittances as simulated by COSMO-MUSCAT-RTTOV is mapped in Figure 3. The ITD is apparent in a winding belt between ∼ 18 and 21 • N (panel (a)), south of which are the moister monsoonal atmospheres of southern West Africa, and north of which are the characteristically drier atmospheres of the central Sahara. Considering the overall distribution of column moistures (Table 2), points are relatively infrequent in the moister regimes, compared to the percentage of points in the two drier regimes: 83% have the greatest sensitivity to changes in skin temperature, while the brightness temperatures in the 12.0 µm channel will be the most responsive to changes in the column moisture, as has been noted before (Brindley et al., 2012;Banks et al., 2018).

Dust AOD and altitude
As a proxy for the quantity of dust in an atmospheric column, the AOD is an obvious property related to the colour response of the false-colour IR imagery in the presence of dust. It is calculated by COSMO-MUSCAT at a wavelength of 550 nm. B2018 5 showed that employing the VO73 dust refractive indices with typical COSMO-MUSCAT simulated dust size distributions leads to a peak in spectral extinction at 10.8 µm. For a given dust AOD the brightness temperatures at 10.8 µm will therefore be reduced more than those at 12.0 and 8.7 µm, behaviour also noted by Brindley et al. (2012) and exploited by the original Desert Dust imagery formulation, since following Equations 1 and 2 deeper red and weaker green colours occur with increasing AOD, triggering characteristic pink colours. Mean simulated dust AODs are mapped in Figure 4. In summer there are particularly 10 active dust source regions in the south and west of the Sahara, subject to convective systems protruding northwards from the wetter regions of southern West Africa. Moreover, dust has a tendency to linger within the SHL, and hence large areas of northern Mali and western Niger see mean simulated AODs of greater than 1 over the six months.
The atmospheric temperature of the dust layer has a close relationship with the altitude of the dust layer, and it is of particular relevance since the temperature contrast between the background surface and the dust layer can be used to discriminate dust in 15 IR measurements (e.g. Legrand et al., 2001;Brindley, 2007). Higher dust tends to be colder and to have the greatest contrast with the surface, although this is not necessarily the case at night or in colder seasons. It is also to be expected that differential atmospheric absorption between channels would also be an influence on the colour of the composite imagery for particularly elevated dust with respect to the background surface. To explore this impact of height, it is useful first to specify a representative height of IR optically active dust in an atmospheric column, considering also that there may be more than one layer. Therefore 20 considered here is an average height defined by the AOD at 10.8 µm: this is referred to as the dust-AOD median height, at which half of the AOD is below this height. It is defined above ground level, on the basis that it is the contrast between the atmospheric and the surface temperatures which is of most significance for the colour sensitivity.
Defined dust height ranges are specified in Table 2, with 'Layer 3' (2-3 km) being the layer with the most frequent dust heights, at 35%. In the case of the AODs, the lowest AOD regime ('near-pristine', AOD < 0.2) has the highest frequency of 25 points (44%), with higher AOD values becoming progressively less common. Over the whole dataset there is an anti-correlation between the AOD and the dust height, but it is weak (-0.30), an indicator of the diversity of the simulated dust storm situations.

Co-location of surface and atmospheric properties
Given the multitude of factors influencing the colour of dust in the imagery, a first step in disentangling their interlocking influences is to quantify the extent to which each of the properties co-vary with each other. Behaviours which may appear to 30 be correlated with one variable may in fact be caused by another variable co-located with it. One may expect for example, the sandy deserts characteristic of the low emissivity zone to be characterised also by particularly hot skin temperatures, dry atmospheres, and by high dust AODs. Meanwhile as noted before by, e.g., B2018, it can be difficult to perceive dust in the imagery over high emissivity surfaces (which may be more vegetated, for example), but is this due to the surface or due to a co-varying moist atmosphere? Figure 5 plots the co-location of the variables with reference to the AOD regimes, the key metric identifying the intensity of dust activity. The principle of this plot is to identify the distributions of the variables within the AOD regimes: within each plotted regime, the total is 100%. Panel (a) considers how the emissivity zones are distributed within the AOD regime subsets: 5 at the lowest AODs the distribution of emissivities is quite close to the overall 41/41/18% split in the emissivity zones (Table   2), indicating no preferred zone for pristine conditions. However as the AOD increases, the dust events become increasingly weighted away from the high emissivity zones and towards instead the medium emissivity zone. As indicated by Figures 2(d) and 4, many of the high AOD areas are in eastern Mali, southern Algeria, and western Niger, in the medium emissivity zone. proportion of dust in the lowest two height layers increases markedly with increasing AOD. Meanwhile dust only tends to be calculated to be within the higher altitude ranges when the simulated dust loading is light. Knowledge of the water vapour vertical distribution is also of significance for understanding how the dust signal at TOA may be obscured by water vapour (WV): defined here is a representative water vapour height above which the atmosphere is 'dry' (i.e. < 13 mm), in support of the analysis in Section 4.2, since it is the quantity of water vapour above the dust layer which is of most interest for its 25 influence on the dust signal. Near-pristine scenes are simulated to be relatively dry (panel (b)) and hence the WV heights tend to be correspondingly low (< 1 km). Thicker dust is often simulated to be accompanied by increased moisture content, and so the WV heights become more elevated, especially common within the 1-2 km layer. In comparison with panel (d) therefore, it is very often the case that the dust layer is suspended within a similar height range as the water vapour, which will have consequences for the apparent colour of dust in the synthetic imagery. This is a known challenge for automated dust detection (e.g. Ashpole and Washington, 2012). An example of how difficult it 5 can be to perceive dust at night even with IR imagery is apparent in Figure 1 However what is different is the colour of the surrounding background 10 environment, which is often light blue during the day but is a murky purple during the night. The contrast between the dust and the background environment is hence typically much clearer during the day than during the night, a result of the strong diurnal cycle in skin temperature.
The overall consequence of the surface cooling during the night on the simulated colours is displayed in Figure 6(e, f), (black symbols). The differences between channel brightness temperatures are critical, and explain the apparent colour patterns.
The diurnal cycle in near-pristine T B108 values is greater than the cycle in the other two channels, hence T B108 can be greater than T B120 during the day but lower than T B120 at night, giving rise to redder colours at night. This is a simple consequence of the atmospheric transmittances at 10.8 µm being greater than the transmittances in the other two channels under all but the wettest atmospheres (Section 3.3, Table 3), such that T B108 is more sensitive to the skin temperature than the other channels. 25 The diurnal cycle in the near-pristine brightness temperatures (not shown) has a greater amplitude than does the diurnal cycle in the 'thick' dusty brightness temperatures, such that on average the presence of dust during the day tends to cool the brightness temperatures with respect to the pristine case. During the night the opposite occurs, when the presence of dust acts to warm the brightness temperatures compared to the near-pristine case, something observed before by other authors (e.g. Legrand et al., 1988). In the simulations, 'thick' dust layers with AODs between 2 and 3 are most prevalent in the bottom 2013) which quantify the presence of dust using the assumption that the dust layer is cooler than the background surface, and which in consequence do not attempt to produce retrievals at night.
For thicker dust loadings, the colour patterns show much more similarity between day and night. The differences between the daytime and night-time T B120 -T B108 differences are small, such that the resultant differences in thick dust mean red colours between day and night are negligible: amongst the emissivity zones the mean red colours in the thick dust case (AODs between 5 2 and 3) are between 0.70 and 0.73 during the day, and between 0.71 and 0.72 at night. The T B108 -T B087 difference and the resultant green colours tell a similar story in the thick dust case. Given a sufficient quantity, dust appears very much the same in the composite imagery throughout the diurnal cycle.
As a result of this, the lengths of the colour tracks in Figure 6(e, f)) are much shorter at night than during the day, and so from a measurement perspective it is harder to resolve dust in the composite imagery at night. Longer colour tracks, as seen in 10 panel (e), give greater clarity between dust and non-dust scenes. It is still possible to perceive dust in the night-time imagery, especially when analysing consecutive images and thereby observing the subtle patterns of motion above the static background surface, but the overall clarity is reduced.

Dust colours simulated by COSMO-MUSCAT-RTTOV
Generalising the analysis introduced in Section 4.1, Figure 7 introduces the mean COSMO-MUSCAT-RTTOV simulated 15 colours categorised for all combinations of the three emissivity zones with the cool to very hot skin temperature regimes, and with the four moisture regimes, in an attempt to disentangle the influences of the various environmental factors and to provide a physical interpretation of the imagery colours. The rings represent the colours within specified AOD and height ranges: the innermost ring is the 'near-pristine' case, with AODs between 0.0 and 0.2 over all height ranges; the outer four rings display the output colours within the AOD range of 2-3 ('thick dust'), from the inner to the outermost rings subdivided 20 by the height ranges of 0-1, 1-2, 2-3, and 3-4 km. The entire 3-hourly diurnal cycle is represented, the distinction between day and night being implicit only in the skin temperature regimes. Within the thick dust AOD range there are no points within the 'cold' skin temperature regime, and so for clarity these are not represented here. The numbers for each segment are the mean red values, which may be considered as the most vital of the constituent colours for producing characteristic pink dust colours: with reference to Figure 6(e, f), pink colours become apparent between red values of ∼ 0.6 and 1, and between green values of 25 0 and ∼ 0.8. For the perception of dust it is more important that the red value is high than that the green value is low.
Considering first the 'near-pristine' case, mean red values are at their greatest over the low emissivity zone, the coolest skin temperature regimes, and under the driest atmospheres ('low-cool-dry'), the regime which has the highest mean red value the purple/pink colour domain. This description generalises the 'night-time effect', of redder background colours during the night, as being rather a 'cool skin temperature effect' which is applicable not just to the diurnal but also to the seasonal cycle.
Notable in the near-pristine case are the ranges of available mean red colours over the moisture sets depending on whether the skin temperature is cool or very hot. For example, in the low-cool set the maximum red value is 0.80 in the dry regime, and the minimum is 0.48 in the wet regime. In this low-cool set there is therefore a range of 0.32 in the mean red values. In 5 the medium-and high-cool sets, the corresponding red ranges are 0.34 and 0.29. However when the skin temperature is 'very hot', the corresponding red ranges are much larger: 0.52 (low emissivity, only from dry to moist), 0.59 (medium emissivity), and 0.53 (high). In near-pristine conditions, for a given amount of moisture, the atmosphere is less transparent in the 12.0 µm channel than in the 10.8 µm channel, and this difference between the channels increases with atmospheric moisture content.
The contrast is exacerbated when the surface skin temperature is hot because there is more surface emission in both channels 10 to interact with the overlying atmosphere, given the relatively flat surface emissivity between the two channels. This means that the near-pristine T B120 -T B108 difference, which is often negative, becomes more negative when the surface is 'hot' and the overlying atmosphere is 'wet', with the converse being true for 'cool' and 'dry' conditions.
Analysing the ranges from the perspective of the skin temperature, the dry sets have red ranges between 0.18 and 0.20 (depending on emissivity), while the wet T skin sets have ranges between 0.41 (cool to hot only) and 0.44. The near-pristine red 15 colour appears to be more sensitive to variations in the column moisture than to variations in the skin temperature, an obvious caveat to which is that the skin temperature informs the behaviour of the colour response to moisture. Meanwhile the red values are weakly sensitive to the emissivity zones, with the red ranges of the emissivity sets varying between just 0.025 and 0.121 amongst the 16 T skin /moisture combinations.
In contrast, the green colours are more strongly governed by the emissivity, hence the distinctly different shadings between 20 the emissivity zones. The green ranges of the emissivity sets vary between 0.17 and 0.30 (16 possible combinations), for the T skin sets the ranges are 0.07 to 0.15 (12), and for the moisture sets the ranges are 0.00 to 0.23 (12). There is less variability in the green behaviour due to T skin or moisture compared to that of the red, since the sensitivity of the transmittance in the 8.7 µm channel to water vapour is less than that of the 12.0 µm channel. The green beam is more a function of the surface properties alone. 25 Introducing dust into the analysis, generally the 'near-pristine' colours tend to appear bluer than the 'thick dust' colours, which are redder and hence pinker. This is always the case in the hot and very-hot regimes, with the most exceptions to this pattern occurring in the cool regime for dust in lower altitude ranges (inner dust rings). For thick dust in variably moist atmospheres, it is the case in almost all circumstances that dry dusty atmospheres are the reddest and hence pinkest within their respective emissivity-T skin -height regime. The corollary of this is that red/pink signals for a given amount of dust are at their 30 weakest in the wettest atmospheres, confirming that moisture acts to 'hide' the presence of dust in the IR imagery. Meanwhile the simulated red colours of thick dust storms are broadly independent of the surface emissivity.
It is always the case that higher altitude dust regimes (outermost rings in Figure 7) are redder than lower altitude dust, giving rise to pinker colours. The greatest mean red values of 1.0 can reliably be found for dry, high-altitude dust, a scenario which is insensitive to the emissivity or the skin temperature. This implies that higher altitude dust has higher values of T B120

35
13 Atmos. Chem. Phys. Discuss., https://doi.org /10.5194/acp-2018-1238 Manuscript under review for journal Atmos. Chem. Phys. Discussion started: 16 January 2019 c Author(s) 2019. CC BY 4.0 License. than of T B108 , compared to dust at lower levels. This difference can be explained by the optical path length between the dust layer and the TOA. Assuming the dust layer is optically thick and hence acts as the emitting 'surface', the optical path length (predominantly driven by the water vapour distribution) will be shorter for higher altitude dust, such that the impact of the variation in the 10.8 and 12.0 µm channel transmittances will be reduced ( Table 4). The atmosphere is very regularly 'dry' above dust in Layer 4 (3-4 km, as defined in Table 2), and is often drier than total atmospheric columns defined as 'dry'. Hence 5 examining the segments in Figure 7 radially outwards, it is always the case that the red values increase towards the outermost ring (highest dust altitudes); similarly by examining each moisture set clockwise from wet to dry, it is almost always the case that the red values increase in the dry direction. Consider dust in the medium-warm-wet regime at an altitude of 0-1 km: it has a red value of 0.52 and would be very hard to discriminate as dust in a SEVIRI image. If this atmosphere is replaced with a dry atmosphere then the red value is nudged up to 0.74, making it slightly easier to identify it as dust, but not convincingly. If instead 10 the dust in the wet atmosphere is moved up to between 3 and 4 km, then the red value is boosted to 0.81 and characteristic pink dust colours become apparent. Especially in the warm and cool regimes, dust in wet conditions may be clearly observed if it is at a sufficiently high altitude.
It is clear that dust height has a substantial impact on the apparent colour of dust in the IR imagery. If dust is in the bottom layer of the atmosphere then it is only likely to be detectable in the imagery if the skin temperature is particularly hot and 15 the atmospheric column is particularly dry. However as shown by Figure 5(c) it is rarely the case in the simulations that the column moisture is 'dry' within this AOD range (3.0 %), and so low-altitude thick dust will regularly be missed in the imagery.
High altitude thick dust (> 3 km) is always likely to be visible in the imagery, given the strong colour contrast between it and the background surface. Even under the wettest atmospheres high dust will be clearly noticeable, since despite possible purple dust colours there is a strong colour contrast with the dark blue background surface which will be clearly distinguishable in 20 successive imagery. However Figure 5(d) indicates that thick dust rarely has a defined height as high as 3-4 km. Figure 7 has provided a physical interpretation of the consequences of different environmental regimes for the apparent dust colour. It is worth considering how frequently such conditions occur, in order to identify those segments within Figure 7 which represent the most likely scenarios. Within the six-month June and July simulation period the most common height range for thick dust is calculated to be in Layer 2, between 1 and 2 km (57 %). Such dust is unlikely to be noticeable above cool surfaces, 25 although within this AOD and height range the warm skin temperature regime is the most common (71 %). In winter it is to be expected that the percentages would be weighted more towards the cooler regimes.
Subdividing further within the regime of thick dust at 1-2 km height, Table 5 shows that the medium emissivity, warm skin temperature, and moist column water vapour regime is the combination simulated to occur most frequently (26 % of the 48 possible combinations) during this period. Identifying this segment in Figure 7, it is apparent that this dust has a red colour of 30 0.70, displaying pink-purple colours, with a noticeable but not substantial contrast against the blue background surface, which has a red colour of 0.44. Table 5 indicates that most frequently attention should be focused on the dry-moist, moist, warm, and hot regimes. Of these, dust in the hot and moist regimes displays the most distinct simulated colour contrasts with the background surface. The more frequent warm points display murkier contrasts against the background surface, although the dust is still likely to be readily apparent in the imagery. B2018 showed that the simulated colours produced by COSMO-MUSCAT-RTTOV are likely to be insufficiently deep, when compared with SEVIRI observations and retrieved AODs. It is plausible that the near-pristine colours should be greener and less red, implying that the COSMO-MUSCAT simulated skin temperatures are likely to be too cool; meanwhile the thick dust colours should perhaps be redder and less green, implying that in the simulations the dust extinction properties are too weak, or that the atmosphere is simulated to be too moist. An implication of this is that dust in the real imagery may be visible at 5 lower AODs, and/or that it may be visible at lower altitudes than in the synthetic imagery.
In summation, cool skin temperatures boost the red beam when the dust loading is light, while hotter skin temperatures suppress it, a result of the high atmospheric transmittance at 10.8 µm. Moister atmospheres will display weaker red colours, a result of the higher atmospheric absorption at 12.0 µm due to water vapour compared to the other channels. Thicker dust loadings boost the red beam and reduce the green beam, a result of the greater IR absorption by dust at 10.8 µm. Dust at higher 10 altitudes has a shorter atmospheric column above it, and hence is less masked by water vapour, giving rise to redder and more vivid pink dust colours in the IR imagery.

Conclusions
This paper is the follow-up paper to Banks et al. (2018), which explored the sensitivity of the colour of dust in SEVIRI Desert Dust IR imagery to dust optical properties. Using the COSMO-MUSCAT-RTTOV modelling system, this paper has explored 15 the sensitivity of the colour of dust in SEVIRI Desert Dust IR composite imagery to various environmental properties, including the surface thermal emissivity, surface skin temperature, atmospheric column moisture and atmospheric transmittance, dust AOD, and dust height. These properties are often co-varying with each other. The relationships between all these variables are intricate, but display distinct patterns in the colour of the imagery, patterns which can be discriminated by considering numerous combinations of the environmental variables. 20 The surface thermal emissivity (at 8.7 µm) is a controlling variable on the colour, in particular the green beam, of the nearpristine imagery. It is less significant at higher AODs, although it still contributes to the intensity of the green beam. Skin temperatures are also an important factor governing the colour of near-pristine imagery, with hotter skin temperatures tending to give rise to bluer (i.e. less red) colours than cooler surfaces. Cooler surfaces are redder, explaining why night-time and winter Desert Dust imagery is characterised by a greater prevalence of pink and purple hues across the desert surface, while 25 hot summer and daytime surfaces are characterised by light blue colours. Pink atmospheric dust is much easier to discriminate against the light blue summer daytime surfaces.
The significance of the skin temperature for the colour of the imagery is related to the atmospheric water vapour content: water vapour is more absorbing at 12.0 µm than at 10.8 µm and hence the wetter the atmosphere the weaker the red colour.
The atmosphere is more transmissive at 10.8 µm and so the brightness temperature in this channel is more sensitive to the skin 30 temperature. A stronger diurnal cycle in T B108 than in T B120 gives rise to a diurnal cycle in the red values whereby the red values are greatest at night. Similarly, wet atmospheres also act to hide the presence of dust by reducing the apparent red colours and hence the distinctive 'pink dust' colour. The importance of the dust height in producing redder and pinker colours is clearly shown by Figure 7 (understanding colour with reference to dust height is a novel feature enabled by the use of the COSMO-MUSCAT-RTTOV system), but it may be an easy misconception to make that this is due to the larger temperature contrast between elevated dust and the background surface. Of more relevance is the difference in atmospheric transmittance above the dust layer between the 10.8 and 12.0 µm channels, the relative difference between which reduces for higher atmospheric dust layers. Some analogy can be made between 5 high dust and dry dust, in that the atmospheric absorption at 12.0 µm due to water vapour is reduced when the dust is elevated higher. Hence the most reliable red and pink colours arise from high-altitude dust in dry atmospheres. Meanwhile low altitude dust can often be hard to discriminate from the background surface, but this is not because the dust layer is warm, but instead because the dust is more readily hidden by atmospheric water vapour the lower in the atmosphere it is located.
In the Desert Dust RGB composite imagery, however thick the dust loading, dust is very likely to be invisible when it is very 10 low in the atmosphere (< 1 km), in an altitude range where it is very likely to be obscured by water vapour. In contrast, dust elevated to altitudes greater than 3 km will always be likely to be apparent in the imagery, and will have a lower AOD threshold for perception. It is usually difficult to distinguish dust within particularly wet atmospheres, but if the dust is high enough then the atmosphere above it will be drier and hence the dust itself will become more readily apparent.
A possible extension of the insights provided by this paper would be to refine further and optimise dynamically the Desert

15
Dust RGB composite imagery in order to highlight better the presence of atmospheric dust, if the background environmental properties are known or can be estimated. An example might be to reduce the range of the brightness temperature differences in the red beam when the skin temperatures are cool, requiring a mathematical relationship between the skin temperature and the red range to be determined. Such a technique could be performed with the aid of model simulations, or in future with the aid of atmospheric sounding information from the Infrared Sounders that will be included onboard the upcoming Meteosat 20 Third Generation (MTG) series of satellites. It is to be expected that the MTG satellites will have enhanced capabilities in the quantification of dust loading and properties, and there is potential for the COSMO-MUSCAT-RTTOV modelling system to be used to test the feasibility of future dust quantification techniques.
In this paper we have presented a thorough evaluation of the various environmental factors that can influence the appearance of the SEVIRI Desert Dust imagery product as simulated by the COSMO-MUSCAT-RTTOV modelling system. The results 25 should be taken in conjunction with those of our previous paper (B2018) which considered the role of variability in dust optical properties alone. We hope that users of the imagery will find both analyses helpful in terms of assessing the strengths and weaknesses of the imagery as a dust identification and tracking tool. In particular, we also believe that users should be aware of the environmental properties which affect the appearance of the dust imagery, to avoid erroneously attributing these effects to dust properties.     Table 3. COSMO-MUSCAT-RTTOV simulated mean pristine-sky channel transmittances within the specified total column water vapour ranges ('dry', 'dry-moist', 'moist', and 'wet'), averaged over the distribution defined in Table 2 Table 4. As with Table 3, COSMO-MUSCAT-RTTOV simulated mean atmosphere-only channel transmittances above the specified dust layers. These are averaged within the specified total column water vapour ranges (defined in Table 2 as 'dry', 'dry-moist', 'moist', and 'wet'), for points within the thick dust AOD range (from 2 to 3) for consistency with the outer rings of Figure 7. Included also are the mean water vapour column values above the specified dust layers.
Moisture ranges (mm) 8.7 µm 10.8 µm 12.0 µm WV above dust (mm)       Percentage within AOD range (d) 0 < dust height < 1 km 1 < dust height < 2 km 2 < dust height < 3 km 3 < dust height < 4 km 4 < dust height < 10 km dust height > 10 km 0.0-0.2 0.2-0.5 0.5-1.0 1.0-1.5 1.5-2. Percentage within AOD range (e) 0 WV height < 1 km 1 < WV height < 2 km 2 < WV height < 3 km 3 < WV height < 4 km (b) skin temperature (K); (c) column moisture (mm); (d) dust height (km); (e) water vapour (WV) height (km), defined here as the height above which there is less than 13 mm of water vapour. Within each cluster of bars for a specified AOD range, the sum of the bars is 100%.    Table 2. Values increase in the anti-clockwise direction, as indicated. The innermost ring is the 'near-pristine' case, with AODs < 0.2. The four remaining rings are the thick dust cases with AODs between 2 and 3, subdivided by height ranges. From the inner to the outer rings, the dust height ranges represented are 0-1 km, 1-2 km, 2-3 km, and 3-4 km (Layers 1-4). 0.2% of the points in this AOD range have a dust height of greater than 4 km, while no points in this range have a skin temperature of less than 286 K ('cold'). All 3-hourly timeslots from day and night are included. The numbers marked on each segment indicate the mean red value, to two decimal places, leading and ending zeroes removed for brevity. Black segments indicate coincident conditions not contained within this data subset.