Problems in separating aerosol and cloud in the Stratospheric Aerosol and Gas Experiment (SAGE) II data set under conditions of lofted dust: Application to the Asian deserts

Authors


Abstract

[1] Stratospheric Aerosol and Gas Experiment II aerosol data obtained at two wavelengths, 525 nm and 1020 nm, have been used for some time to identify the presence of cloud along the optical path from the Sun to the satellite instrument. Examination of data obtained over desert regions in the Northern Hemisphere, particularly the Taklimakan Desert, indicates that this separation method does not always operate correctly. In regions where there is expected to be a large amount of lofted dust, unexpectedly low values of mean extinction are found, combined with higher than expected amounts of cloud. These anomalous data have been analyzed in detail, and the discrepancy is plausibly shown to be due to faulty identification of lofted dust as cloud. Six Northern Hemisphere desert regions, together with three comparison regions, have been identified for study and the anomalies used to develop a description of the seasonal and altitudinal characteristics of the lofted aerosol over these regions.

1. Introduction

[2] The Stratospheric Aerosol and Gas Experiment (SAGE) II was designed for the measurement of aerosols and gases in the stratosphere using the solar occultation technique. In the absence of opaque cloud, measurements are possible down into the troposphere. Within the troposphere, the measured extinction may be due to aerosol or gas alone or may also include a contribution from thin cloud lying along the optical path from the Sun to the satellite instrument. During data processing, SAGE II solar irradiance measurements are binned in tangent altitude increments of 0.5 or 1.0 km. The data are then inverted under the assumption of horizontal stratification to give vertical profiles of gas concentrations and aerosol extinction coefficient [Chu et al., 1989]. The values of the extinction coefficient found in this way approximate to the average of all values within a particular atmospheric shell used in the inversion. Although the assumption of horizontal stratification is clearly a poor approximation when cloud is present, many useful data on cloud presence and behavior have been obtained. Kent and McCormick [1991] and Kent et al. [1993] have described a method of detecting the presence of cloud using the extinction data at wavelengths of 525 and 1020 nm. This method makes the assumption that there is little or no wavelength dependence of the extinction due to cloud, while the extinction due to aerosol at 525 nm is greater than that at 1020 nm. The method is applicable to altitudes above 6 km where data are available at both wavelengths. Below 6 km aerosol extinction data are only available at 1020 nm and alternative techniques, such as that developed by Wang et al. [1999], must be used.

[3] The assumption that the aerosol extinction at 525 nm is greater than that at 1020 nm, is essentially an assumption that the aerosol effective particle radius is less than about 0.25 μm [Kent et al., 1995] (the exact radius also depends upon the aerosol composition). Although aerosol particles at altitudes above 6 km are usually smaller than this size [Kent et al., 1995; Hofmann, 1993; Pueschel et al., 1994], it is not clear that this is true under conditions of lofted dust such as might be found over the world's deserts. This discrepancy may be particularly significant over the Taklimakan Desert, where dust is reported to be usually entrained to altitudes above 5 km [Sun et al., 2001] and thence transported for long distances.

[4] The main objective of the present study is to examine how departures from the commonly accepted behavior of SAGE II aerosol extinction data may be used to recognize and describe unusual aerosol behavior over selected regions of the globe. For this purpose, SAGE II tropospheric data have been examined over six desert areas in the Northern Hemisphere, together with three non-desert comparison areas. These areas are shown in Figure 1. The three non-desert areas were selected as follows: (1) Pacific Ocean: an area with no local surface dust, although the upper troposphere may contain dust transported from the Asian continent [Uematsu et al., 1983; Gao et al., 1992; Sun et al., 2001], (2) Western European Block: a region where the major aerosol source may be expected to be industrial rather than mineralogical [Bridgman, 1990], and (3) North Atlantic Dust Corridor: an area where dust is seasonally advected from north Africa at altitudes below 5 km [Prospero and Carlson, 1972; Guelle et al., 2000]. These nine regions vary considerably in area and in the case of the Taklimakan Desert the area is only of the order of 500,000 km2. Even when accumulated over several years, the number of SAGE II profiles found within this small area is unfortunately very low.

Figure 1.

Global map showing the Northern Hemisphere deserts and other areas used in the present study.

[5] The analysis has been carried out using a developmental version (6.0) of the SAGE II data. This has a more accurate determination of the ray path geometry than previous versions and a vertical resolution of 0.5 km as opposed to 1.0 km. One disadvantage of this version is an uncertainty in the low signal level used to define the occurrence of opaque cloud. Although not a major part of the present analysis, we have arbitrarily chosen to define opaque cloud as occurring at the highest altitude on the SAGE II 1020-nm vertical extinction profile where the extinction coefficient is greater than 10−5 m−1. SAGE II data are available from October 1984 to the present. The eruptions of El Chichon in 1982 and Pinatubo in 1991 injected considerable quantities of volcanic aerosol into the stratosphere, which in turn modified upper tropospheric aerosol in subsequent years. In order to avoid parts of the data set when this contamination was a major factor we have omitted data prior to and including spring 1986, and from and including, summer 1991 to spring 1995.

2. Aerosol Cloud Anomalies

[6] The extinction due to aerosol and cloud in the troposphere is highly variable because of inhomogeneities in both constituents. The method of Kent et al. [1993] for classifying individual data points into cloud-free and cloud-contaminated fractions is based on a study of the statistical properties of the inverted extinction coefficient. If the 525- and 1020-nm extinction coefficient data for the troposphere are subsetted by time, altitude and latitude, and plotted as shown schematically in Figure 2a, and as actual data in Figure 3a, it is found that the data points lie in two distinct data groups. One group lies close to the origin, has lower values for the extinction coefficients and a ratio of the two coefficients between 2 and 5. The other group shows larger extinction ratio values that are approximately equal at the two wavelengths. The first group is attributed to cloud-free aerosol and the second to aerosol plus cloud. The statistical properties of these subsets are examined and used to derive the parameters (slope and intercept) of a line that separates the two data groups when plotted as in Figures 2a or 3a. The parameters may then be used to decide the nature of any point within the data subset. The advantage of the method over attempting to categorize a single data point without external reference is that we are examining measurements made within a given region of space and time as an ensemble, and are able to determine the range of variation of the measured extinction coefficients for both the cloud-free aerosol and the aerosol/cloud mixture. The method, which then has a measurable degree of statistical significance, is used routinely to derive aerosol-cloud separation parameters for specified subsets of the SAGE II data and made available to coworkers. Since the derivation of these parameters requires the examination of the statistical properties of the subset for which it is derived, the latter cannot be too small. The standard subsets consist of one 3-month season of data (from a single year), at a specific altitude, and within a specified latitude band. For convenience, the data are divided into eight, 20°-wide latitude bands between 80°S and 80°N. No longitude subdivision is made, on the assumption that the major variation of aerosol properties in the upper troposphere is by altitude and latitude, rather than by longitude. It is clear that this assumption is subject to error and it would be possible to derive separation parameters for smaller subsets than those just described. It would not however be possible to derive meaningful values, for each season, year and altitude, for most, if not all, of the areas shown in Figure 1. In the data description that follows, the aerosol (cloud-free) and cloud (cloud-contaminated) fractions have been derived for each area using the separation parameters appropriate to the year and season, altitude and latitude of each profile within the area subsets. It has become apparent that the assumption of homogeneity within these time/latitude/altitude subsets, particularly that of longitudinal homogeneity, may under certain conditions be invalid. This paper describes when, and where, one form of this invalidity occurs and how it appears in the data set.

Figure 2.

(a) Model used by Kent et al. [1993] to describe the relationship between the aerosol and cloud extinction coefficients at 525 and 1020 nm (data ensemble). (b) Plot showing the relationship between the aerosol and cloud extinction coefficients at 525 and 1020 nm for the original model and its proposed modification (single data points).

Figure 3.

Plots of the 525-nm versus 1020-nm aerosol extinction coefficients for (a) the North Atlantic Dust Corridor and (b) the Taklimakan Desert. Both plots are for the Northern Hemisphere spring season, the altitude range 6–10 km, and the time period 1984–2000, omitting the periods of major volcanic contamination. The data points have been color coded as cloud-free or cloud-contaminated as determined by the model of Kent et al. [1993].

[7] An example of such an invalidity occurs when we consider the long-range transport of dust over the northern Pacific Ocean, which is observed at altitudes up to 8 km [Gao et al., 1992]. These observations are believed to be related to the seasonal pattern of sandstorms over China, which reach their maximum frequency in spring [Arakawa, 1969]. The analysis of Sun et al. [2001] indicates that the most likely source is the Taklimakan Desert, where the dust can be entrained to elevations above 5 km, travel north to 50°N and thence east over the northern Pacific Ocean. These results would lead us to expect that SAGE II aerosol extinction for middle and high and altitudes above 6 km would show a spring maximum and this is indeed found [Kent et al., 1995]. They would also lead us to expect that the highest extinction values might be observed over the Asian deserts and particularly the Taklimakan. In this case, the evidence is far less clear.

[8] Table 1 shows the mean 1020-nm spring aerosol extinction coefficient, for altitudes between 6 and 10 km, for the areas shown in Figure 1, using the method outlined above to distinguish cloud-free aerosol (data have been averaged for all years excluding those affected by volcanic contamination). The observed mean extinction coefficient over the Taklimakan Desert is lower than those over the central Asian and Gobi deserts and over Europe. Most anomalously, it is lower than the extinction coefficient over the Pacific Ocean. This is very surprising, as the enhanced springtime aerosol over the latter is generally believed to come from Asia. Almost certainly, this result is incorrect and the explanation may lie in the values, also shown in Table 1, for the total cloud amount seen by SAGE II. This latter number includes both nonopaque cloud (as determined by the method of Kent et al. [1993]) and opaque cloud as defined in the introduction. Here, the situation as regards the Taklimakan Desert is reversed, in that it shows the highest amount of apparent cloud (68%) occurring over any of the areas. Examination of the data for the other seasons shows a similar behavior, in that the apparent cloud over the Taklimakan is always the highest of the nine areas surveyed. There is no supporting evidence for this result. While the Taklimakan is a relatively small area surrounded on three sides by mountains, which could show its own local characteristics, deserts are not normally cloudy places. The Tarim Basin, containing the Taklimakan Desert is one of the driest areas in the world [Arakawa, 1969] and has little precipitation and very low cloud cover [Makra et al., 2003]. Satellite-derived cloud fields agree with surface observations in showing this to be a region with no obvious local cloud enhancements [Rossow and Schiffer, 1991; Ridout and Rosmond, 1996]. The implication of these results is that the division of data shown in Table 1 into aerosol and cloud is potentially incorrect, particularly in the case of the upper troposphere over the Taklimakan Desert.

Table 1. SAGE II Mean 1020-nm Springtime Aerosol Extinction Coefficient and Cloud Amount Over Selected Areas for Altitude Range 6–10 km
Area (Figure 1)Mean Extinction Coefficient, 10−7 m−1Total Cloud,%
Pacific Ocean5.7 ± 0.321 ± 2
Western European Block6.9 ± 0.238 ± 1
Sahara Desert4.2 ± 0.224 ± 1
Saudi-Arabian Peninsula3.8 ± 0.229 ± 2
Iranian and Indian Deserts4.3 ± 0.537 ± 4
Central Asian Deserts7.3 ± 0.344 ± 2
Taklimakan Desert5.4 ± 0.868 ± 5
Gobi Desert6.0 ± 0.763 ± 4
North Atlantic Dust Corridor3.0 ± 0.123 ± 2

[9] A further anomaly becomes apparent if we examine scatterplots of the 525-nm versus the 1020-nm aerosol extinction coefficient for the Taklimakan and other desert areas. Figure 3 shows scatterplots for the North Atlantic Dust Corridor and the Taklimakan Desert, the former being considered to be dust-free at altitudes above about 5 km. Both plots are for the Northern Hemisphere spring season, the altitude range 6–10 km, and the time period 1984–2000, omitting the periods of major volcanic contamination. The data have been color coded as aerosol and cloud, using the method of Kent et al. [1993]. The two plots are quite different in character, even allowing for the statistical spread in the data points. In Figure 3a the cloud data appear to stream away from the center of the aerosol distribution as shown schematically in Figure 2a. In Figure 3b in contrast, all the data points seem to lie about a single, somewhat curved, line, and the distinction between cloud and aerosol (which was made on the basis of the average behavior over all longitudes between 20°N and 40°N) appears rather arbitrary. Similar behavior is seen over some of the other Northern Hemisphere deserts. In the sections that follow, we show that this behavior of the relationship between the 525-nm and the 1020-nm extinction coefficients is typical of that expected if lofted desert aerosol is present.

[10] When placing the data presented in the next sections in the context of other aerosol and cloud measurements, two points should be borne in mind. The first is the extreme sensitivity of the solar occultation technique used by SAGE II to study aerosols as compared to most other atmospheric measurements. The instrument is capable of making accurate measurements of aerosol extinction coefficients less than 10−6 m−1 (equivalent visual range at 550 nm of the order of 1000 km), and the signal level falls below the measurement threshold if the extinction coefficient is greater than about 3 × 10−5 m−1. The second point is that, although the nominal vertical resolution of the measurement is 0.5 km, this resolution may not be reached when studying tropospheric aerosols and cloud. SAGE II data are inverted on the assumption of horizontal stratification over distances of the order of 100 km [Chu et al., 1989]. This assumption is not likely to be true for most tropospheric measurements, even those that are cloud-free. Bearing this in mind, some of the data averages presented have been downgraded to a vertical resolution of 1.0 km in order to reduce the statistical error.

3. Cloud Models

3.1. Basic Model

[11] The model of Kent et al. [1993] for distinguishing cloud-free and cloud-contaminated aerosol assumes that, if a given atmospheric shell contains both aerosol and cloud, the relationship between the extinction coefficient at 525 nm and that at 1020 nm may be expressed by the equations

equation image
equation image

where E0.5 is the total extinction coefficient at 525 nm; E1.0 is the total extinction coefficient at 1020 nm; E1.0,c is the extinction coefficient due to cloud at 1020 nm; E1.0,a is the extinction coefficient due to aerosol at 1020 nm; and k is the ratio of the aerosol extinction coefficient at 525 nm to that at 1020 nm (as noted earlier, we are assuming that the aerosols have an effective radius of less than about 0.25 μm). Assuming a certain amount of variation in aerosol properties and concentration and a much greater variation in cloud concentration, these relationships are shown graphically in Figure 2a. The dark gray region of the data distribution shows the location of cloud-free aerosol (E1.0,c = 0) and the light gray region that of cloud plus aerosol. The mean values of the data in each region are indicated by black dots, marked Basic Model in the case of cloud plus aerosol, and are connected by a dashed line at 45° to the axes (a second cloud plus aerosol point, marked Modified Model will be explained later). The relationships just described assume that the aerosol and cloud have optical characteristics that exist independently of each other and occupy space independently of each other. There is little that we can do about the first assumption but we can improve upon the latter assumption. The data inversion [Chu et al., 1989] is carried out assuming the atmospheric constituents to lie homogeneously within a series of concentric shells, which are then peeled apart. This is not the way in which cloud and aerosol are commonly distributed. Typically the ray path from the Sun will pass through regions of cloud-free aerosol as well as thin cloud. In the next section we describe a modification to this model that allows for the partial occupancy of the ray path by aerosol and cloud. This modification predicts a relationship between the 525-nm and the 1020-nm extinction coefficients that corresponds better with the measured values than does the original model.

3.2. Modification to the Basic Model

[12] Contributions to the total optical depth and mean extinction coefficient along the ray path will depend upon the relative amounts and volume of the shell occupied by each of the two components. If there is no cloud within the shell equations (1) and (2) become

equation image
equation image

If the shell is completely filled by thin cloud and aerosol is absent we have instead

equation image
equation image

In the general case, there is partial occupancy of the shell by both aerosol and cloud with aerosol occupying a volume fraction f, and cloud a volume fraction (1 − f), of the shell. Since the result of the inversion process is an average extinction coefficient for the shell, relationships (1) and (2) are then replaced by

equation image
equation image

In the case of cloud-free aerosol, f has a value of one and the relationships become those in equations (3) and (4). If aerosol volume fraction f is zero then the relationships become as in equations (5) and (6). The general case is illustrated in Figure 2b where, to avoid the confusion of a data ensemble, the locations of a single aerosol data point with and without a cloud fraction are shown for the original model and its modification. The coordinates and interpretations of the five labeled points in Figure 2b are as follows:

[13] A, (E1.0,a,k.E1.0,a): Aerosol without cloud, this point is independent of whether we are discussing the basic model or the proposed modification (equations (3) and (4))

[14] B, (E1.0,a + E1.0,c,k.E1.0,a + E1.0,c): Aerosol plus cloud, basic model (equations (1) and (2))

[15] C, (f.E1.0,a,f.k.E1.0,a): Aerosol component of the mean extinction in the modified model (derived from equations (7) and (8) by putting E1.0,c = 0)

[16] D, (f.E1.0,a + (1 − f).E1.0,c,f.k.E1.0,a + (1 − f).E1.0,c): Aerosol plus cloud, modified model (equations (7) and (8))

[17] E, (f.E1.0,a + (1 − f).E1.0,c, f.E1.0,a + (1 − f).E1.0,c + (k − 1)E1.0,a): A point on the 45° line from A immediately above point D. The abscissa is the same as for D, the ordinate is found by placing the point on the line through A and B.

[18] The main point of the above analysis has been to show that whereas, in the basic aerosol-cloud model the aerosol plus cloud point B lies on a 45° line from the aerosol point A, in the modified model the corresponding point D lies below this line. The amount of the displacement (ED) is easily shown to be equal to (1 − f)(k − 1).E1.0,a or (1 − f)((k − 1)/k).E0.5,a where E0.5,a is the aerosol extinction coefficient at 525 nm. Since f has a value between 0 and 1 and k is a positive number greater than 1, the displacement is expected to lie between 0 and E0.5,a. If we measure distances to be positive in the direction of the ordinate in Figure 2b then the displacement has negative sign. The analysis just presented has also been in terms of a single value for the aerosol extinction coefficient but may be extended to an ensemble of background aerosol extinction coefficients such as shown in Figure 2a. This leads to the third black dot in that figure marked Center of Aerosol/Cloud Mixture (Modified Model) which, as an average over a number of negative displacements, must lie below the 45° line from the aerosol center. This prediction of the modified model may also be tested against actual SAGE II data. In the case of the data shown in Figure 3a, the center of the Aerosol/Cloud Mixture is found to be displaced negatively with respect to the 45° line from the aerosol center by an amount approximately equal to 0.6 × E0.5,a, compatible with the predictions of the model.

[19] It might be supposed that we could apply the above analysis to determine values for the cloud fraction, f. This is not the purpose of the present paper; moreover, it should be recognized that the analysis presented is an approximation only, as it assumes that the background aerosol is the same whether cloud is present or not. The rather simplistic interpretation of f and (1 − f) as the cloud and aerosol volume fractions is also an approximation, as even statistical homogeneity is not likely within a real atmosphere and the ray path situation is extremely complicated. The purpose of this analysis is to show that, under the assumptions of this aerosol/cloud model, with a clear optical distinction between the aerosol and the cloud, the displacement as defined is negative. In the next section, where we discuss the problems that arise when larger aerosol particles are present we find, in contrast, that positive displacements may occur.

4. Optical Effects of Larger Aerosols

[20] In section 2 we discussed the possibility that many of the higher values of the extinction coefficient had been incorrectly identified as cloud. Figure 4a shows the relationship between E0.5 and E1.0 for an aerosol model with a particle density of 10 cm−3, a lognormal size distribution and an effective radius up to about 0.5 μm (see Kent et al. [1995] for definitions). The curve shown here has been calculated using Mie theory, for a dust aerosol, but the basic shape is fairly independent of the aerosol composition and refractive index chosen (see Kent et al. [1983] for a list of common materials and refractive indices). The use of Mie theory to calculate the optical properties of nonspherical particles is also an approximation but we are not concerned here with exact optical properties. It can be seen that, for the larger particle sizes, the relationship is somewhat similar to that of cloud in the aerosol-cloud models just described, but that the curve describing it rises above the straight line relationship of that model. In Figure 4b, the data in Figure 4a have been redrawn as they would be interpreted using the aerosol-cloud separation techniques described in the previous section. The portion of the aerosol extinction coefficient distribution curve near to the origin, now shown as a heavy line, is interpreted as aerosol, and the remainder of the dust aerosol curve is interpreted as cloud. The aerosol and cloud centers have been identified as was done previously and a 45° offset line from the aerosol center has been drawn. Because of the curvature in the extinction distribution, the cloud center is above the 45° offset line. We are also able to calculate a ratio that is exactly equivalent to the quantity (1 − f)((k − 1))/k, which was derived in the previous section. Two distances, a and d, are marked in Figure 4b; a is the mean 525-nm extinction coefficient for the aerosol subset (the equivalent of E0.5,a), and d is the departure of the dust cloud center from the 45° line from the aerosol center (the equivalent of the distance ED in Figure 2b: points corresponding to A, D and E in Figure 2b have been labeled as “A”, “D”, and “E” in Figure 4b). In direct contrast to the situation with small aerosol particles and cloud, where ED (or d) was negative, d is now positive. We are now in a position to distinguish between the situations shown in Figures 2 and 4 by looking at the magnitude and, more importantly, the sign of the ratio (d/a). If no large aerosol particles are present, (d/a) has the value −(1 − f)((k − 1))/k, which will lie between 0 and −1. On the other hand, if large aerosol particles are present without cloud, (d/a) will be positive, although given the lack of exact knowledge of the aerosol size distribution, it is not possible to quantify this ratio. In general, we might expect the data distribution over a desert region containing lofted dust to contain data points characteristic of the behavior shown in both Figures 2 and 4, as it is unlikely that all the lofted particles are large, or that cloud is entirely absent. Nevertheless, some positive bias should be evident in the value of (d/a) if a considerable amount of lofted dust is present.

Figure 4.

(a) The relationship between E0.5 and E1.0 for a dust aerosol model with a particle density of 10 cm−3, a lognormal size distribution, and a range of effective radii as shown in the figure. (b) The interpretation of the relationship shown in Figure 4a using the aerosol-cloud separation technique shown in Figure 2a. The portion of the aerosol extinction coefficient distribution curve near to the origin, now shown as a heavy line, is interpreted as aerosol. The remainder of the curve is interpreted as cloud. The aerosol and cloud centers are as identified, and a 45° offset line is drawn from the aerosol center. The distances a and d are used to further characterize the data distribution and points “A”, “D”, and “E” shown as corresponding to points A, D, and E in Figure 2b.

5. Comparison of Data and Models

[21] The dimensionless parameter (d/a) has been calculated for all the areas shown in Figure 1. As in Figure 3, data have been accumulated for altitudes from 6–10 km, and the mean values for (d/a) are shown in Figure 5 for each season and the entire SAGE II observational period, omitting times of volcanic contamination. The following features of the plots in Figure 5 may be noted:

Figure 5.

Plots showing the dimensionless parameter (d/a) for the four seasons and the areas shown in Figure 1.

[22] (1) The error bars are significant and indicative of the data spread.

[23] (2) With the exception of one data point (North Atlantic Dust Corridor in fall), all values of (d/a) are greater than −1.0, in agreement with predictions.

[24] (3) In winter, all values of (d/a) are between −1.0 and 0.0. This is in agreement with the predictions of the modified aerosol and cloud model, and anticipated to be the case in a season with relatively low amounts of lofted dust.

[25] (4) Values of (d/a) greater than zero are found as follows: Taklimakan Desert in spring, summer, and fall, Gobi Desert in spring and fall, and Iranian and Indian deserts in spring.

[26] The highest values of (d/a) occur over the Taklimakan Desert in spring, summer, and fall. This result is consistent, not only with the earlier data presented in this paper, but also with published data showing the Taklimakan Desert to be the major source of dust lofted into the upper troposphere of the Northern Hemisphere [Sun et al., 2001]. High values over the Gobi Desert are less expected, as dust is less often entrained to high altitudes from the Gobi than from the Taklimakan, but there may also be additional effects from dust raised over the Taklimakan and thence advected at high altitudes over the Gobi. The relatively high value of (d/a) over the Gobi Desert in winter (close to 0.0) is perhaps surprising as spring is normally noted as the peak of the dust storm activity. It is nevertheless consistent with the fact that large amounts of dust are observed in the atmosphere over eastern China in late winter [Arakawa, 1969]. The high value of (d/a) over the Iranian and Indian deserts in spring (and to a lesser extent in summer) is also interesting. This is consistent with observations presented for these areas by Ackerman and Cox [1989] and Middleton [1986], who show that there is widespread dust in suspension over these regions from spring through summer and, to a lesser extent, into fall. Ackerman and Cox estimate that these dust layers extend upward to the 600–400 mb level (4–7 km).

6. Additional Supporting Evidence From Within Sage II Data Set

[27] In this section, we describe an alternative form of analysis of the SAGE II aerosol and cloud data which supports the interpretation given above. Figure 6 shows the behavior of the 1020-nm aerosol/cloud extinction coefficient in Northern Hemisphere spring as a function of altitude between 0 and 10 km, over the nine selected regions. The extinction coefficient values have been divided into three groups depending upon their magnitude: low extinction (extinction coefficient < 10−6 m−1), high extinction (10−6 m−1 < extinction coefficient < 10−5 m−1), and cutoff values (>10−5 m−1). In terms of the discussion of the lofting of dust from the earth's surface, the most interesting parameter is the lowest altitude at which low extinction values begin to appear. Low values for the extinction coefficient are indicative of background aerosol conditions; cloud or lofted dust will appear as high extinction coefficient or cutoff values. Table 2a shows the lowest altitudes at which the low extinction coefficient values appear for all nine regions; Table 2b shows the seasonal behavior for the three regions that appear to have the greatest amounts of lofted dust.

Figure 6.

Individual plots showing the fractional distribution of 1020-nm aerosol/cloud extinction coefficients in Northern Hemisphere spring as a function of altitude between 0 and 10 km over the nine selected regions in Figure 1. The values of the extinction coefficients have been divided into three groups depending upon their magnitude: low extinction (extinction coefficient < 10−6 m−1), high extinction (10−6 m−1 < extinction coefficient < 10−5 m−1), and cutoff values (>10−5 m−1).

Table 2a. SAGE II Aerosol Extinction Behavior: Lowest Altitudes at Which Low Values of the 1020-nm Extinction Coefficient are Observed—Spring Behavior in All Areas
Area (Figure 1)Altitude, km
Pacific Ocean2.0
Western European Block2.5
Sahara Desert2.5
Saudi-Arabian Peninsula3.0
Iranian and Indian Deserts5.0
Central Asian Deserts3.0
Taklimakan Desert5.5
Gobi Desert3.0 and 5.5
North Atlantic Dust Corridor2.0
Table 2b. SAGE II Aerosol Extinction Behavior: Lowest Altitudes at Which Low Values of the 1020-nm Extinction Coefficient are Observed—Seasonal Behavior in Selected Areas
Area (Figure 1)Altitude, km
WinterSpringSummerFall
Iranian and Indian Deserts2.55.05.03.5
Taklimakan Desert3.55.56.04.0
Gobi Desert2.53.0 and 5.55.03.0

[28] The most surprising characteristic of the data shown in Figure 6 and Table 2a is that, over the Taklimakan and Indian and Iranian deserts, no low extinction coefficient values appear in spring for altitudes below 5 km. The implication of this result is that, at this time of year, the atmosphere is never clear at low altitudes over these deserts. This need not mean that dust storms must occur every day, but it does imply that they are of such intensity and frequency that suspended dust remains in the atmosphere between storms (we may also note that blowing dust, as opposed to actual dust storms, has an even greater frequency in this region [Gillette et al., 1993]). The profiles for the Gobi Desert are also unusual in that there are two altitude regions in which there are no low extinction coefficient values: 0.0 to 2.5 km and 4.5 to 5.0 km. Sun et al. [2001] note that dust from the Gobi Desert can only be entrained to an altitude of about 3.0 km, whereas it can reach to above 5.0 km over the Taklimakan Desert. The decrease in the number of low extinction coefficient values between 4.5 and 5.0 km suggests that, at these altitudes over the Gobi, dust is being advected from the neighboring Taklimakan. The lowest altitude at which low extinction coefficient values appear in any of these nine regions is 2.0 km over the Pacific Ocean and over the North Atlantic Dust Corridor. These values are compatible with observations using lidar that show the aerosol backscatter coefficient over the ocean to decrease rapidly above about 2 km [Alejandro et al., 1995]. The low altitude at which low extinction coefficient values appear over the Atlantic Dust Corridor is somewhat unexpected, as Saharan dust is advected over this region and would be expected to cause higher values for the extinction coefficient. The anomaly is more apparent than real, as the tail of low extinction values extending down to 2.0 km is quite small and the relative number of low extinction coefficient values is not significant below 3.0 km. A similar anomaly is observed over the Sahara Desert but once again there is a fairly small number of low extinction values below 3.5 or 4.0 km. Nevertheless, it is apparent that lofted dust over the Sahara does not reach altitudes of 5.0 km as frequently as it does over the Taklimakan and Iranian and Indian deserts (we may note that Middleton [1986] states that dust has been identified as rising to 9.0 km over the Rajasthan desert). The Saudi Arabian and Central Asian deserts show a behavior intermediate between those of the Sahara and the Taklimakan and Iranian and Indian deserts. Western Europe, where there is no reason to expect a dominant aerosol contribution from surface dust, shows a gradual decrease in the fraction of low extinction values down to 2.0 km.

[29] The seasonal behavior of the aerosol over the Iranian and Indian, Taklimakan, and Gobi deserts is shown in Figure 7, and the lowest altitudes at which low values of the aerosol extinction coefficient are observed are listed in Table 2b. Three features stand out: (1) the Taklimakan Desert shows the greatest degree of lofting in all four seasons; (2) the least amount of lofting occurs in winter, as would be expected, and (3) rather unexpectedly, the amount of lofted aerosol is approximately the same in spring and summer, at least in terms of the minimum altitude at which low extinction values are observed. This last finding appears somewhat at variance with other observations. It is fairly generally accepted that the peak dust storm season in the Taklimakan and Gobi deserts is spring. Arakawa [1969] lists observations made near the Chinese deserts and shows approximately three times as many storms between March and May as between June and August. The solution perhaps lies in the fact that, even in summer, storms are sufficiently frequent to keep dust opaque to SAGE II lofted up to an altitude of 5 km. Arakawa [1969] shows an average frequency of dust storms at Hami on the edge of the Taklimakan Desert of 10 for the months of June through August. Although this is not as great as the number shown for spring (17), it may still be sufficient to create a continuous haze that is opaque to SAGE II. The seasonal behavior of the minimum altitude at which low extinction values are observed over the Gobi Desert resembles that over the Taklimakan. This can perhaps be accounted for, as in the previous paragraph, in terms of advection from the neighboring Taklimakan. We may also note that over the Taklimakan, at altitudes between 5 and 10 km, there is a distinct difference between the extinction coefficient frequency distributions for spring and summer (see Figure 7). In spring, the fraction of the data showing low extinction values is considerably less than that showing high extinction values; in summer, the reverse is true. This is consistent with other observations and implies that a small part of the dust in spring is reaching altitudes of at least 10 km. The aerosol behavior over the Indian and Iranian deserts resembles that over the Taklimakan in that the minimum altitude is similar in spring and summer (5.0 km). Here, the agreement with other observations is better. Middleton [1986] and Ackerman and Cox [1989] present data showing that spring and summer are the peak dust seasons. The months having the greatest frequency of storms vary with location from March through to October. Only the westernmost stations show a summer decrease, which is associated with monsoon precipitation.

Figure 7.

Seasonal behavior of the aerosol/cloud extinction over the Iranian and Indian, Taklimakan, and Gobi deserts. Values of the extinction coefficient have been divided and color coded as in Figure 6.

7. Summary and Remarks

[30] In the previous sections, we have analyzed the behavior of the SAGE II aerosol data over several desert regions in the Northern Hemisphere, examining the division of data into cloud-free and cloud-contaminated subsets on the basis of a standard method developed by Kent et al. [1993]. The results of this division have been found to be anomalous in that, over regions where little cloud and high mean aerosol extinction coefficients are expected, the reverse is often the case. It has been suggested that large aerosol particles, with radii greater than about 0.25 μm are being erroneously identified as cloud. This interpretation is consistent with other observations showing large amounts of blowing and lofted dust in these regions, particularly over the Taklimakan Desert. Support for this explanation is also found in an examination of the statistical relationship between the SAGE II 525-nm and 1020-nm aerosol extinction coefficients. This analysis has involved a reexamination of the aerosol/cloud separation method of Kent et al. [1993]. Several points have emerged as a result of this study. The first is that, even within the range of applicability of this method, a slightly better approximation to actual atmospheric behavior is possible. The equations governing this improved approximation, which takes into consideration the relative volumes of the ray path occupied by aerosol and cloud, have been presented. We have also emphasized the fact that the SAGE II data inversion is made under the assumption of homogeneity within each atmospheric shell used in the inversion. While this may be a reasonable assumption within the stratosphere, it is clearly not so when tropospheric cloud is present. Because of the inhomogeneity, individual inverted data points are subject to significant error and the meaning of some of the parameters used in the models or derived from the data ensembles, such as the fraction of the volume of an atmospheric shell occupied by cloud, cannot be regarded as exact. It has also become apparent that the routine production and use of parameters to categorize SAGE II tropospheric data points as cloud-free aerosol and cloud-contaminated must be done with caution. This is particularly so where the same parameters are produced to be used over a range of longitudes and over varying land and sea surfaces. Departures from the model of Kent et al. [1993] and the modification presented here, due to the presence of large aerosols, have been quantitatively identified. These departures are presently regarded as anomalies, as the analysis is not at the point where an alternative improved model for the routine separation of aerosol and cloud in the data set, that incorporates the detection of large particles, can be proposed. We can simultaneously note however that, even if regarded as anomalous, such behavior may still be used to obtain new and useful information on the behavior of atmospheric aerosols.

[31] The greatest anomalies in the data set are found over the Taklimakan Desert; this is consistent with the published view that this desert lofts dust to greater altitudes than other Northern Hemisphere deserts and is responsible for most of the outflow of dust over the Pacific Ocean. In spring and summer, the atmosphere over the Taklimakan Desert is opaque to SAGE II measurements to an altitude of 5 km and there is evidence of lofted dust to altitudes as high as 10 km. The Gobi Desert also shows evidence of lofted dust but in this case much of it at the greater altitudes may be advected from the neighboring Taklimakan. A third region showing lofted dust is that of the Iranian and Indian deserts, again a region where dust storms are known to be of frequent occurrence in spring and summer. The fact that, throughout the spring and summer, the atmosphere over some of these deserts remains opaque to SAGE II measurements has to be interpreted in terms, not only of the prevalence of suspended dust, but also in terms of the sensitivity of the SAGE II measurement. SAGE II measurements are not possible if the extinction coefficient at 1020 nm becomes greater than about 3 × 10−5 km−1. This figure corresponds to a visual range at 550 nm of about 30 km. Values for visual range at altitudes of 5–10 km over these deserts do not appear in the literature and it would be interesting to have an independent confirmation or refutation of this value.

Acknowledgments

[32] We would like to thank the people at NASA Langley Research Center (LaRC) and Science Applications International Corporation (SAIC) who have been involved in the processing of the data used. We would in particular like to thank L. W. Thomason and J. W. Zawodny for discussions regarding use of the data set. Three of the authors (G. S. K. P.-H. W., and P. L. L.) have been supported under NASA contracts NAS1-18941 and NAS1-99129.

Ancillary