Volume 124, Issue 13 p. 7286-7307
Research Article
Free Access

A Multiwavelength Retrieval Approach for Improved OSIRIS Aerosol Extinction Retrievals

L. A. Rieger

Corresponding Author

L. A. Rieger

Institute of Space and Atmospheric Studies, University of Saskatchewan, Saskatoon, Saskatchewan, Canada

Correspondence to: L. Rieger,

[email protected]

Search for more papers by this author
D. J. Zawada

D. J. Zawada

Institute of Space and Atmospheric Studies, University of Saskatchewan, Saskatoon, Saskatchewan, Canada

Search for more papers by this author
A. E. Bourassa

A. E. Bourassa

Institute of Space and Atmospheric Studies, University of Saskatchewan, Saskatoon, Saskatchewan, Canada

Search for more papers by this author
D. A. Degenstein

D. A. Degenstein

Institute of Space and Atmospheric Studies, University of Saskatchewan, Saskatoon, Saskatchewan, Canada

Search for more papers by this author
First published: 13 June 2019
Citations: 21


The Optical Spectrograph and InfraRed Imaging System (OSIRIS) on board the Odin satellite has been used to provide vertically resolved aerosol extinction since 2001. The OSIRIS version 5.07 aerosol product has been used in numerous studies and now provides a 17-year record of global stratospheric aerosol. This work presents the new version 7 OSIRIS aerosol extinction retrieval. A multiwavelength aerosol extinction algorithm has been developed to reduce measurement geometry biases and improve extinction retrieval in the upper troposphere and lower stratosphere. The Chen et al. (2016, https://doi.org/10.5194/amt-9-1239-2016) cloud detection algorithm has been adapted for the OSIRIS wavelength range for improved cloud screening and polar stratospheric cloud detection, and comparisons after volcanic eruptions and with the CALIPSO-GOCCP product show promising results. The version 7 product shows comparable agreement with version 5.07 when compared to coincident SAGE II and SAGE III measurements, and improved agreement with CALIPSO time series. The algorithm has been applied to the complete set of OSIRIS measurements, and the new data set is now publicly available.

Key Points

  • A multiwavelength limb scatter aerosol retrieval has been developed and applied to OSIRIS data
  • The new algorithm reduces the dependence of the retrieved extinction on viewing geometry
  • Use of both shorter and longer wavelengths results in improved sensitivity in the upper troposphere and lower stratosphere

1 Introduction

In the 1960s, in situ measurements began of a layer of particles that extends from the tropopause to approximately 25 km in altitude (Junge et al., 1961). These particles are formed from trace gases including SO2 and carbonyl sulfide (OCS) that are transported from the troposphere into the stratosphere where they are converted to H2SO4 and combine with water to form liquid droplets (Brock et al., 1995; Hamill et al., 1997). Although OCS is the primary driver of the stratospheric aerosol during background periods (Sheng et al., 2015), even moderate volcanic eruptions, typically with a volcanic explosive index of three or four, can have a profound impact on aerosol levels (Vernier et al., 2011). The last two decades have been punctuated by several of these small-to-moderate eruptions and they have an important impact on climate, both through ozone depletion (Stone et al., 2017) and contributions to the radiative forcing (Solomon et al., 2011; Fyfe et al., 2013). The importance, as well as variability of the aerosol layer has meant that continuing measurements are of high importance for understanding the larger climate system.

Since Junge's measurements, a host of techniques have been used to study the stratospheric aerosol layer. Balloon measurements continue to be performed and provide an invaluable in situ record of aerosol size and concentration (Deshler et al., 2003). Additionally, numerous space-based remote sensing techniques have been developed to provide global coverage. These began with the occultation technique that was employed on a series of Stratospheric Aerosol and Gas Experiments (SAGE) in 1975. This set of instruments provided the first global measurements of aerosol extinction, during both the highly perturbed conditions after the Mount Pinatubo eruption, and during the quiescent period that followed. However, these occultation measurements ceased for over a decade in mid-2006 with the end of SAGE III/M3M, until they were resumed in 2017 with the launch of SAGE III/ISS. Between 2002 and 2012, the Global Ozone Monitoring by Occultation of Stars (GOMOS) experiment used stellar occultation measurements to retrieve vertical profiles of aerosols (Kyrölä et al., 2004). To help continue the record of stratospheric aerosols, satellite instruments employing the limb scattering technique have been used, including the Optical Spectrograph and InfraRed Imaging System (OSIRIS; Llewellyn et al., 2004), the SpectroMeter for Atmospheric CHartographY (SCIAMACHY; Bovensmann et al., 1999), and the Ozone Mapping and Profiler Suite Limb Profiler (OMPS-LP; Flynn et al., 2006). These instruments have added to both the long-term global record (Thomason et al., 2018) and studies of shorter-lived phenomenon such as volcanic eruptions (Bourassa, Robock, et al., Bourassa et al., 2012), meteoric events (Gorkavyi et al., 2013), and forest fires and continue to monitor stratospheric aerosol levels.

While limb scatter instruments provide high vertical resolution and good sensitivity to background levels of aerosol due to the long path lengths, the measurements are inherently complex. At every point along the instrument line of sight, light both directly from the Sun and from a diffuse component scattered by the ground and atmosphere can be scattered into the instrument. Each measurement depends not only on the extinction, as is the case for occultation measurements, but also on the scattering properties of the atmospheric constituents. This leads to three main challenges when retrieving aerosol from limb-scattered signals that can result in time-dependent biases in the retrieved products. These are the sensitivity to the assumed phase function, poor sensitivity and cloud contamination in the upper troposphere and lower stratosphere (UTLS), and sensitivity to assumptions about the aerosol profile at high altitudes. These three effects are discussed in more detail in the following sections 1.1, 1.2, and 1.3, but generally impart spatial and temporal biases into limb-scattering aerosol records. This has potential implications for climate modeling and long-term trends as the OSIRIS record is incorporated into the Global Space-based Stratospheric Aerosol Climatology being used in the Coupled Model Intercomparison Project 6 (Thomason et al., 2018). To reduce retrieval biases this work focuses on improving the OSIRIS retrieval algorithm through incorporation of multiple wavelengths.

1.1 Phase Function Sensitivity

The phase function, p, indicates the angular distribution of scattered light from a particle, and in this work normalized such that it has a value of 4π when integrated over all solid angles. The nature of limb scatter couples the aerosol extinction and phase function information together. Very approximately, the aerosol signal is a product of the extinction and phase function, although this is complicated by extinction along the path and multiple scattering, particularly in backscatter conditions where the single scattered signal is low. Typically, the phase function is determined by assuming an aerosol composition of spherical sulfate droplets consisting of ≈75% H2SO4 and 25% H2O with a unimodal lognormal particle size distribution, as given by the equation
where rg is the median radius, σg the distribution width, and N the aerosol number density concentration. However, bimodal and gamma distributions have also been used (Chen et al., 2018; Loughman et al., 2018). Errors in these assumptions translate to errors in the phase function used in the radiative transfer and carry through to errors in the retrieved extinction. As the value of the phase function depends on the single scattering angle (SSA), the error in the retrieved extinction will also be a function of the scattering angle. To minimize the error and obtain information on particle size, some retrievals attempt to determine one or more lognormal parameters (Malinina et al., 2018; Rieger et al., 2014); however, the information content in the visible and near-infrared wavelengths is limited and the coupled extinction-particle size retrievals have important limitations. The version 6 OSIRIS retrieval assumes a distribution width that is constant in space and time, and is constrained by the Infrared channel on OSIRIS which saturates under moderate aerosol loading. Due to this, version 6 is not produced on an operational basis, and the version 5 OSIRIS retrieval remains the standard product, and what is compared against in this work. The SCIAMACHY retrievals assume a fixed number density profile and have not been attempted outside of a limited range of scattering angles (Malinina et al., 2018). While both products provide important information on particle size, the limitations mean they cannot be used to construct complete global records of aerosol extinction. While the standard extinction-only products are even simpler, assuming a fixed size distribution at all locations and times, they provide coverage of lower altitudes for OSIRIS and all latitudes for SCIAMACHY, allowing for more complete records of aerosol extinction and easier use in global climatologies. The following work develops an extinction retrieval that can be applied at all altitudes and latitudes, while reducing some of the limitations of earlier extinction retrievals.

1.2 UTLS Sensitivity

The Rayleigh scattered signal is approximately proportional to atmospheric density and so increases exponentially with decreasing altitude. This can result in a relatively small fraction of the total radiance signal being attributable to aerosol in UTLS. Additionally, the sensitivity to aerosol decreases as the total atmospheric extinction increases due to a larger fraction of the light being scattered out of the instrument line of sight. Together, these effects result in rapidly decreasing sensitivity at lower altitudes. Longer wavelengths are generally used to reduce the Rayleigh signal and remain in an optically thin regime, providing better theoretical sensitivity at lower altitude. However, longer wavelengths also often have poorer signal to noise and increased stray light, limiting the extent to which they can be used, particularly in the infrared region. These artifacts depend strongly on the particular instrument but for limb-scattering instruments such as OSIRIS, SCIAMACHY, and OMPS-LP are often worse at increasing wavelengths, due in part to the diminishing scattered signal (Gottwald et al., 2011; Jaross et al., 2014). Depending on the retrieval and instrument the wavelength is usually chosen between 675 and 850 nm and is generally on the long side of the visible spectrum. This spectral region provides good sensitivity to aerosols while maintaining a relatively high signal-to-noise ratio over a wide range of stratospheric altitudes. While occultation and Lidar measurements often retrieve extinction in the 530-nm range, these wavelengths can have poor sensitivity at lower altitudes in limb scattering measurements, making longer wavelengths preferable in most cases. However, even at the longer wavelengths, lower altitudes can remain problematic and section 12 discusses improvements in sensitivity in the UTLS region.

Further complicating the UTLS is the possibility of clouds in the field of view. Sulfate aerosols in the UTLS can be an important component of the total aerosol optical depth, particularly in middle-to-high latitudes after moderate volcanic eruptions (Ridley et al., 2014), making this an important region for accurate measurements. However, the radiance signal from clouds and cloud/aerosol mixtures can appear similar to volcanically enhanced aerosols, so distinguishing them has proved challenging and many methods have been developed to screen clouds from limb aerosol records (Thomason & Vernier, 2013; Normand et al., 2013; Eichmann et al., 2016; Liebing, 2016; Chen et al., 2016). Section 10 implements an updated cloud detection algorithm applied to the OSIRIS data set.

1.3 High-Altitude Sensitivity

Limb scatter aerosol retrievals often use an altitude normalized radiance profile as the measurement vector (e.g., Bourassa, Rieger, et al., 2012; Loughman et al., 2018). The altitude normalization decreases sensitivity to upwelling radiation from surface and multiple scattering as well as biases in absolute calibration and radiative transfer modeling. However, it comes at the cost of increased sensitivity to stray light and any aerosol in the normalization range, strongly coupling errors at these higher altitudes to lower altitudes. See Rieger et al. (2018) for a more detailed description of this effect. Stray light tends to be larger at longer wavelengths due to the decreasing signal levels, favoring use of shorter wavelengths, although the magnitude of this effect is difficult to quantify as the precise magnitude of the stray light is generally unknown. Although not strictly addressed in this paper, the reliance on high-altitude (∼30–40 km) measurements is an important consideration in aerosol retrievals.

2 Algorithm Development

2.1 Overview

A multiwavelength aerosol retrieval is developed here to help address the issues of phase function dependence and lack of UTLS sensitivity when using limb-scattered aerosol measurements. The retrieval assumes a spherically homogeneous atmosphere for the retrieval of vertical profiles of aerosol extinction. Also assumed is a fixed aerosol particle size distribution. The state vector, x, used in the retrieval is the aerosol number density, which is converted to extinction using the assumed particle size. The inverse problem is solved using the Levenberg-Marquardt procedure (Levenberg, 1944; Marquardt, 1963) to update the state vector, x, on iteration n as
where y is the measurement vector, and F is the forward model used to simulate the measurement vector. The weight given to each measurement, W, is often the inverse of the measurement error covariance matrix, but this is not required. The damping factor, γ, is set to 0.1 for the duration of the retrieval. The retrieval is initialized with an a priori guess but is generally insensitive to this parameter within the retrieval domain, as no penalty is associated with divergence from the a priori. The exception to this is the effect from the a priori assumption in the normalization range, as discussed in section 3

2.2 The Measurement Vector

The measurement vector is constructed from a combination of measurements to decrease sensitivity to confounding variables including instrument calibration, upwelling radiation, and other atmospheric parameters. One of the simplest measurement vectors is a radiance, I, at a single wavelength, λ, normalized by a higher altitude. In this case the vector at altitude j is
where the range of altitudes from i to i+N−1 are used as the normalization. The logarithm of the radiance is often used to avoid the orders of magnitude difference in signal between the ground and upper range of the retrieval, although this is not generally required if the measurement covariance is also used. It should be noted that this logarithm creates a nonlinearity in the signal, such that even for an optically thin atmosphere the aerosol and Rayleigh contributions cannot be separated, potentially complicating single scatter analysis. A color ratio approach where the long-wavelength radiance is normalized by a shorter wavelength (usually near 470 nm) as well as a high altitude has also been used (Bourassa, Rieger, et al., 2012; Ernst et al., 2012). In this case,
The short wavelength can be either positively or negatively sensitive to changes in aerosol at the tangent point, making this normalization beneficial at times and detrimental at others. A series of measurement vector Jacobians is shown in Figure 1 for single wavelength measurement vectors at 470 and 750 nm as well as a short-wavelength normalized vector. The Jacobians are computed using an analytic approximation that includes both single scatter and an approximate multiple scatter component. Details of the algorithm are described by Zawada et al. (2017), and the approximate Jacobians agree very well with numerically computed terms, although do neglect the effect of aerosol below the tangent point. For these calculations an equatorial extinction profile from January 2007 from the GloSSAC climatology (Thomason et al., 2018) was used to simulate conditions with moderate volcanic enhancement. Below 17 km the aerosol was assumed to exponentially decay to a value of 4.6×10−4 km−1 at the ground. In forward scattering conditions, as shown in the top row, the 470-nm vector is positively sensitive down to ≈15 km, and nearly zero below this point, while the 750-nm vector is positive essentially down to the ground. In this case, the wavelength-normalized vector sees no change in aerosol extinction sensitivity at low altitudes and decreased sensitivity above 15 km. In backscattering conditions, as shown in the second row, both measurements are much less sensitive to aerosol due to the much smaller phase function. However, the 470-nm vector becomes negative below ≈20 km. Therefore, when a short wavelength is used as a normalization, the sensitivity of the vector is increased over the 750-nm measurement at lower altitudes. Above ≈20 km, for both measurement geometries, normalization always causes a decrease in sensitivity to aerosol. The magnitude of this effect depends slightly on the solar zenith angle, with smaller angles typically reducing the magnitude of the negative sensitivity, although the effect generally remains. Although this is a simple example, and a full analysis needs to incorporate measurement noise as is presented later, this indicates that a wavelength normalization can be useful in backscattering cases to improve the aerosol sensitivity in the UTLS region.
Details are in the caption following the image
The sensitivity of three different measurement vectors to a perturbation in aerosol extinction. The top row shows a measurement geometry with solar zenith angle of 85° and single-scattering angle of 60°. The bottom row shows the same solar zenith angle, but with a scattering angle of 120°. The left and center columns show a measurement vector from single wavelengths at 470 and 750 nm, respectively. The right column shows a short wavelength normalized vector, as defined in equation 4. The color of the line indicates the tangent point altitude of the measurement, as indicated by the colorbar. For reference, the gray lines indicate the sensitivity at the tangent altitude to changes at the tangent altitude.
The sensitivity also depends on particle size, and this can affect both the magnitude and sign of the measurement Jacobian. Although the number of particles with radii less than ≈50 nm in the stratosphere can be relatively large, their optical cross sections are generally too small to contribute meaningfully to aerosol extinction levels, and the effect on limb-scattering measurements is further dependent on the phase function and other atmospheric parameters. To examine this in more detail it is illustrative to briefly explore some optical properties of a lognormal distribution of particles. The top panel of Figure 2 shows the distribution of particle number density for a typical stratospheric lognormal distribution. The majority of particles are smaller than 100 nm, with the peak concentration around 65 nm for this particular distribution. However, this is not necessarily representative of where the extinction signal is large. Extinction can be determined directly as an integration over the size distribution,
where σ(r,λ) is the optical cross section computed from Mie theory, and also depends on the index of refraction. The center panel in Figure 2 shows ∂k/∂r at 750 nm for the same distribution. Over 95% of the extinction at 750 nm is generated by particles larger than 100 nm. This indicates for occultation measurements that these smaller particles play little to no role. However, for a limb-scattering measurement, the aerosol signal from the tangent point is approximately proportional to the extinction multiplied by the phase function. So, ignoring multiple scattering, the tangent point contribution to the aerosol measurement vector can be roughly estimated as
The proportionality comes from the fact that integration along the line of sight and multiplication by the incoming solar radiation is neglected. The bottom panel in Figure 2 shows ∂y/∂r for scattering angles of 30°, 60°, and 120°, again for 750 nm. In strongly forward scattering cases the signal is dominated by particles larger than ≈150 nm. However, for backscattering conditions, the signal is much weaker overall, with smaller particles playing a much more prominent role due to the more Rayleigh-like phase functions.
Details are in the caption following the image
The top panel shows the number density of a typical lognormal distribution with median radius of 100 nm and width of 1.5. The center panel shows the extinction at 750 nm for this distribution. The bottom panel shows the extinction distribution multiplied by the phase function for a range of scattering angles as an approximation to the limb scattering aerosol signal, also at 750 nm.
From equation 6 we have the sensitivity of our measurement vector y to a change in the number of particles at radius r as
Although, highly approximative, this gives a general indication of what particle sizes are important for limb scattering signals. Figure 3 shows this approximate ∂y(r)/∂n for three scattering angles for a vector at 750 nm. The dominant feature is the rapid increase in sensitivity to particles as radius increases due to the increase in the optical cross section. As the volume of aerosol increases with r3, this unsurprisingly shows that y is generally more sensitive to larger increases in aerosol loading than smaller increases. Equation 7 can be rewritten in terms of the change in aerosol volume, V, as
This has the physical interpretation of how a given mass of aerosol will change the measurement, y, if it is added to the atmosphere at different particle radii and has been used previously when looking at occultation measurements (e.g., Thomason & Poole, 1993). The quantity ∂y/∂V is shown in the center panel of Figure 3. Even with equal volume, particles with radii less than 50 nm have very little contribution to the signal at any scattering angle. While this interpretation is physically meaningful, for this work the retrieved quantity remains extinction, and so it is also useful to examine urn:x-wiley:jgrd:media:jgrd55565:jgrd55565-math-0009. This is shown in the bottom panel of Figure 3 for λret=750 nm and is calculated as
This indicates how sensitive the measurement vector is to changes in the extinction due to particles of size r and would ideally be a flat line; that is, the measurement sensitivity would be independent of the particle size causing the extinction. This is the case for occultation measurements and the reason extinction is such a robust quantity from the occultation measurement technique. For limb scattering, however, the variation can be quite large with measurements more than an order of magnitude more sensitive at some radii than others and also depends strongly on the viewing geometry.
Details are in the caption following the image
Analytic approximations to the 750-nm measurement vector sensitivity as a function of particle radius. The top panel shows the sensitivity to changes in number density, the middle panel to changes in aerosol volume, and the bottom panel to changes in extinction at 750 nm. As a reference, the black line in each panel shows the quantity k750/x. This is proportional to the change in optical depth, as an occultation instrument would measure.

Although this provides a rough estimate of what particle sizes are important, a more accurate way to compute the limb-scattering sensitivity to particle size variations is to model the signal using a reasonable background state, then add particles with a monodisperse size at a specific altitude, j, and model the signal once more. The difference between the two results at altitude i is then the numerical derivative ∂y(r,λ)i/∂nj for particles of radius, r. The number of particles in the perturbation can be easily adjusted as to keep the extinction perturbation constant across r, yielding ∂yi/∂kj,λ. Repeating this at various values of r can then be used to determine the sensitivity to particles of different sizes.

Figure 4 shows the sensitivity of a measurement vector, ∂yj/∂kj,λ, at four different wavelengths to a perturbation in extinction due to monodisperse particles. Each row indicates the sensitivity of a measurement with tangent altitude j to an extinction perturbation at that same altitude. The analytical sensitivities from equation 9, normalized to the peak numeric value, are plotted as dashed lines in the 28.5-km panels for reference. They agree very well with the numeric derivatives at high altitudes, although are substantially less accurate below ≈15 km and in backscattering conditions where large amounts of scattering out of the line of sight relative to the signal can cause negative values not captured in the simpler approximations. The 750-nm wavelength is more sensitive in general than the 470-nm vector, but particularly to larger particles and at lower altitudes, with a peak sensitivity to particles with a radius in the 100- to 200-nm range, depending on the geometry. The 470-nm measurements have a peak sensitivity in the 50- to 100-nm range, and at altitudes below 20 km the Jacobian is often negative for particles larger than 100 to 200 nm. This causes the wavelength-normalized vector to have less sensitivity to particles smaller than 100 nm.

Details are in the caption following the image
The sensitivity of three different measurement vectors to a perturbation in the aerosol size distribution. Left column shows the sensitivity for a forward scattering geometry and the right for a backscattering case. Each row indicates the sensitivity at a particular altitude to perturbations at that same altitude. The dashed lines show the analytic sensitivity calculated from equation 9.

Figure 4 also illustrates the sensitivity of the measurements to changes in particle size distributions. If these curves were flat, then any particle size distribution would produce the same measurement vector and the same retrieved extinction; it is the variability of the sensitivity with particle radius that causes the dependence on the assumed size distribution. Backscattering geometries in particular have a very sharp cutoff at particle sizes between 150 and 200 nm where sensitivity drops to nearly zero, making measurements highly sensitive to changes in the size distribution near this point. At lower altitudes the 470-nm measurement vector has a negative, but much broader response, reducing the dependence on the particle size distribution. Additionally, while the sensitivity of the wavelengths in the visible and near-infrared range are not unique enough to retrieve a particle size distribution, the response from each wavelength is slightly different, yielding the possibility that a measurement vector that uses a combination of wavelengths is less sensitive to particle size. Unfortunately, due to the limited wavelength range available from limb-scattering instruments, and the nonorthogonality of the measurements a flat response cannot generally be achieved, and indeed even if it could, it may decrease the sensitivity to extinction sufficiently that measurement noise swamps the signal. While a flat response cannot generally be attained, this does not preclude improving the measurement vector to reduce biases. A measurement vector that is consistently sensitive over a range of particle sizes across measurement geometries has the potential to reduce scattering angle dependencies and resultant seasonal and latitudinal biases. Instead of attempting to force a flat response at the expense of extinction sensitivity, the measurements at different wavelengths can be combined to maximize sensitivity to a reasonable size distribution. This will produce a measurement vector that has good sensitivity to extinction for realistic cases and will tend to produce measurement vectors that are similarly sensitive to particle size, so far as is possible with the available wavelengths. In the following section a retrieval is developed that uses a variable wavelength normalization to increase the sensitivity to extinction and help decrease the dependence on measurement geometry.

2.3 Implementation

The measurement vector, y, can be written as a linear combination of measurement vectors at individual wavelengths as
Although past limb scatter retrievals have used a(λ)=±1, for example, equation 4, this is not a strict constraint and aj(λ) can assume any value, including one that changes with altitude. Intuitively, the weight, aj(λ), should be related to the sensitivity of the measurement, as we would like to select measurements sensitive to aerosol without incorporating insensitive measurements that will increase noise or bias without adding information. The following section outlines a method of selecting weights to accomplish this. Stacking the measurement vectors at each individual wavelength, y(λ), into a single vector y, the combined vector can be written in matrix form as
where A is a matrix that produces a linear combination of measurement vectors at different wavelengths. In equation 2 this can be implemented by constructing the weighting matrix W such that
where Sϵ is the measurement error covariance matrix. Setting A=I reduces to the conventional urn:x-wiley:jgrd:media:jgrd55565:jgrd55565-math-0014 and serves to minimize the weighted square error of the residuals. However, as described above, to improve aerosol sensitivity and reduce the dependence on the measurement geometry, we would like to combine wavelengths that have high sensitivity to the extinction in addition to low noise on the measurement. Combining the measurements in such a way, instead of using y directly, allows for a measurement vector with similar sensitivities across measurement geometries. The matrix A is therefore constructed to produce a linear combination of wavelengths that minimize the diagonal elements of the measurement error covariance, Sx, which from Rodgers (2000) can be computed as
and G is the gain matrix, or in the case of equation 2,
If the Jacobian, K, is invertible, then this reduces to
The algorithm then proceeds as follows:
  1. Search for the weights A that minimize the diagonal elements of Sx. For this search each altitude is treated independently to help reduce the computation burden. As it is the relative weight of the vectors that matters, the norm of each column of A is set to one.
  2. Iteratively retrieve the aerosol extinction using equation 2 for five iterations. This is sufficient for convergence within measurement error under the vast majority of conditions, and measurements that do not converge are later discarded. Aerosol is retrieved from below the normalization altitude to above the altitude where the Jacobian falls below 2×10−4 km. Smaller than this and it is difficult to ensure that even the sign of the Jacobian is correct as it depends strongly on the particular size distribution chosen. The a priori values at and above the normalization region are scaled to match the value at the highest retrieved altitude. Below the retrieval values are tapered back to the a priori value.
  3. Repeat steps 1 and 2. After the first iteration the extinction is approximately correct, and so the second iteration couples the wavelength weights to the retrieved extinction.

The following section applies this algorithm to the OSIRIS measurements.

3 OSIRIS Version 7

OSIRIS (Llewellyn et al., 2004) was launched in 2001 on board the Odin satellite into a near-terminator orbit and continues operation. The spectrograph has a wavelength range of 280 to 810 nm with a spectral resolution of ≈1 nm. OSIRIS scans the limb from 7 to 75 km with measurements every 2 km and vertical resolution of 1 km at the tangent point. OSIRIS views in the tangent plane of the orbit which produces single scattering angles between 60° to 120° depending on the latitude and time of year.

Currently, OSIRIS aerosol retrievals are processed using the version 5.07 algorithm and includes retrieval of NO2, aerosol extinction, and ozone (Degenstein et al., 2009). Since the release of version 5.07 a pointing adjustment to the instrument line of sight to correct a drift in the satellite attitude has been calculated, and this correction is also included in version 7 (Bourassa et al., 2018). This correction has important ramifications for long-term ozone trends, but a smaller affect on aerosol profiles. Full descriptions of the algorithm are described by Bourassa et al. (2007) and Bourassa, Rieger, et al. (2012). For reference, the version 5.07 aerosol retrieval uses the MART relaxation technique that simplifies to the Chahine method (Chahine, 1972) for aerosol extinction as a single-measurement vector is used. Extinction is retrieved at 750 nm using the short wavelength-normalized vector shown in equation 4.

Version 7 retrieves only aerosol extinction, and so relies on additional external atmospheric constraints. Pressure and temperature profiles from the ECMWF ERA-Interim reanalysis (Dee et al., 2011) are interpolated to the OSIRIS scan location. Ozone values retrieved using the version 5.07 algorithm are used, as well as a NO2 climatology from the PRATMO photochemical box model (McLinden et al., 2000). Effective Lambertian reflectivity is retrieved at 675 nm before each aerosol iteration and assumed to be the same for all wavelengths used. The aerosol phase function used is computed using Mie scattering theory assuming a lognormal distribution with a width of 1.6 and median radius of 80 nm, with refractive indices from Palmer and Williams (1975) assuming a 75/25 mix of H2SO4/H2O, the same as previous versions. The inversion uses the SASKTRAN radiative transfer model to simulate the OSIRIS measurements (Bourassa et al., 2008; Zawada et al., 2015). For the OSIRIS version 7 retrieval wavelengths of 470, 675, 750, and 805 nm are used. Due to instrumental considerations wavelengths between 475 and 530 nm are not available, and wavelengths near 600 nm are not used to avoid strong ozone absorption. Additional wavelengths between 675 and 805 nm do not provide substantially different sensitivity to aerosols, so the choice of four wavelengths is a compromise between measurement noise and computational time.

Figure 5 shows an example of the the altitude-dependent weights for scans between 10°S and 10°N. Only measurements where Odin is traveling southward are used to isolate measurements with similar scattering angles. At high altitudes the aerosol loading is low, and the dominant source of error is the measurement noise, leading to the 470-nm measurement being weighted the most heavily. At lower altitudes the shorter wavelengths have poor sensitivity and the retrieval shifts to longer wavelengths. This is true until approximately 15 km. At this point, even the 805-nm measurements have minimal sensitivity to aerosol under backscattering conditions. However, here the 450- and 675-nm measurements become negatively sensitive; that is, increasing aerosol levels reduces the signal, and the shorter wavelengths can again be used. The altitude at which the retrieval shifts to longer wavelengths, and then back to shorter depends primarily on the scattering angle, with backscattering geometries favoring shorter wavelengths sooner. However, it also depends on aerosol loading, with enhanced aerosol conditions also favoring shorter wavelengths in the UTLS region.

Details are in the caption following the image
The top four panels show the weights used at each wavelength as a function of altitude and time for the OSIRIS descending node binned into 14-day averages. The last panel shows the solar scattering and solar zenith angles for the same node. OSIRIS = Optical Spectrograph and InfraRed Imaging System.

3.1 Scattering Angle Biases

One of the major goals of this work is to improve the sensitivity to the assumed particle size, and Odin's orbit provides a convenient test of this due to its terminator orbit. OSIRIS samples the same point on the globe approximately 12 hr apart, once on the ascending node and once on the descending. These two nodes can have drastically different scattering angles, and therefore, despite sampling the same location, have different retrieved aerosol values due to errors in the assumed aerosol phase function, dictated by the assumed particle size distribution.

Figure 6 shows the monthly averaged aerosol extinction retrieved at three altitudes and latitudes using the OSIRIS data, split into the ascending and descending nodes. The top row shows OSIRIS scattering angle for three latitude bands. The higher latitudes constantly measure both nodes throughout the mission; however, the orbital drift pushes the ascending node past the terminator for more tropical latitudes after about 2004. The middle row shows the extinction from version 5.07 retrieval, also grouped into ascending/descending conditions. There is a clear separation of the retrieved extinction between the two nodes that is dependent on the scattering angle. This dependence is stronger at middle and tropical latitudes and larger after 2006, when the stratospheric aerosol has been increased by a series of small volcanic eruptions (Vernier et al., 2011). In the tropics the two nodes cannot be compared beyond the first few years due to an orbital precession which pushes the ascending node past the terminator. However, there is still a clear seasonal cycle that correlates very well with the scattering angle. The results from version 7 are shown in the bottom row. Mean extinction values remain very similar to version 5.07; however, the separation between the nodes has been greatly reduced for all time periods. Although some differences remain in the tropics, the seasonal cycle has also been reduced substantially.

Details are in the caption following the image
Comparison of the monthly averaged aerosol extinction retrieved on the ascending (orange) and descending (blue) nodes in three latitude bands.
Reduction of the seasonal cycle is evident in most, but all locations. Figure 7 compares the differences between the ascending and descending nodes as a function of latitude and altitude. Here, the extinction difference has been computed as
where the extinction, k, and scattering angle, SSA, are the monthly mean values. The subscripts asc and desc denote the ascending and descending nodes, respectively. In version 5.07, for most regions in the stratosphere, a 1° change in scattering angle would cause approximately a 1% change in the retrieved extinction. In version 7, shown in the right panel, this dependence is generally reduced to ≈0.5% per degree. Hatched regions indicate where version 7 shows less dependence on the scattering angle than version 5.07. Improvements are seen everywhere except at ≈3–5 km above the thermal tropopause in the Northern Hemisphere, where version 7 shows a stronger dependence on scattering angle. This region generally has high values of aerosol which, when coupled with the low altitude, leads to poor sensitivity.
Details are in the caption following the image
Comparison of the zonally averaged extinction from the ascending and descending nodes, normalized by the difference in scattering angle, or urn:x-wiley:jgrd:media:jgrd55565:jgrd55565-math-0019%. The hatched region shows the areas where version 7 has less dependence on scattering angle. The gray line shows the mean thermal tropopause altitude.

3.2 Cloud Detection

Before improvements in the UTLS can be fully explored, clouds must first be filtered from the retrieved extinction. Chen et al. (2016) and Eichmann et al. (2016) developed similar techniques to filter clouds from limb-scattered data based on the assumption that clouds cause a steep vertical gradient in the radiance profile with longer wavelengths having a stronger response due to the larger size of cloud particles. As a filter criteria Chen et al. (2016) define the value as
where z is the altitude. For their application to the OMPS-LP instrument they used wavelengths of 674 and 868 nm and recommended a threshold value of 0.15 km−1 at these wavelengths. While this generally provides good results for OMPS-LP they note that during periods of increased volcanic loading this technique can flag what is likely to be aerosol as clouds; an effect that will be exacerbated when applying this technique to OSIRIS due to the more limited wavelength range available. Figure 8 shows the distribution of OSIRIS measurements in urn:x-wiley:jgrd:media:jgrd55565:jgrd55565-math-0021 and retrieved extinction space. The urn:x-wiley:jgrd:media:jgrd55565:jgrd55565-math-0022 is computed at 675 and 805 nm, and to make the values of urn:x-wiley:jgrd:media:jgrd55565:jgrd55565-math-0023 consistent with those discussed in Chen et al. (2016) a correction of 1.48 (the ratio of OMPS-LP to OSIRIS wavelength ranges) is applied to account for the smaller wavelength range of OSIRIS. The OSIRIS measurements are also interpolated to a 1-km grid to approximate the vertical sampling of the OMPS-LP instrument. The dashed line indicates the 0.15-km−1 cutoff used for the OMPS-LP instrument. Unfortunately, there is no clear separation between cloud and aerosol measurements, either according to extinction or urn:x-wiley:jgrd:media:jgrd55565:jgrd55565-math-0024. However, there is a weak correlation between extinction and urn:x-wiley:jgrd:media:jgrd55565:jgrd55565-math-0025, so values with high urn:x-wiley:jgrd:media:jgrd55565:jgrd55565-math-0026 also have a tendency to have large extinction, extending the tail of distribution when viewed in this two-dimensional space.
Details are in the caption following the image
Distribution of OSIRIS measurements for the duration of the mission in urn:x-wiley:jgrd:media:jgrd55565:jgrd55565-math-0027 and extinction space for a selection of altitude and latitude ranges. The color indicates the number of measurements with these values. The dashed line shows the threshold used in Chen et al. (2016), and the solid line shows urn:x-wiley:jgrd:media:jgrd55565:jgrd55565-math-0028 km−2. OSIRIS = Optical Spectrograph and InfraRed Imaging System.

To help improve the discrimination between volcanic aerosol and clouds a small modification is made to the Chen et al. (2016) cloud algorithm. The value urn:x-wiley:jgrd:media:jgrd55565:jgrd55565-math-0029 is multiplied by the extinction, k, before setting a threshold value. This incorporates the additional assumption that clouds will tend to have larger extinction values than aerosols, as has been used previously in SAGE II algorithms (Kent et al., 1993; Thomason & Vernier, 2013). The solid line in Figure 8 shows the line of constant urn:x-wiley:jgrd:media:jgrd55565:jgrd55565-math-0030 km−2 that is used in this work to discriminate between aerosol and cloud. In general, this leads to fewer measurements being flagged as clouds than the Chen et al. (2016) method, tending to increase the extinction after volcanic eruptions, even though some additional measurements with large extinction are now removed.

As a case study, Figure 9 compares the Chen et al. (2016) method with that employed in this work for the period around the 2009 Sarychev eruption. The left column shows the Chen et al. (2016) method applied to the OSIRIS data. The first panel is the urn:x-wiley:jgrd:media:jgrd55565:jgrd55565-math-0031 quantity computed at 675 and 805 nm for the period from March to December 2009 in 1-week averages between 50°N and 70°N. Note that while weekly averages are shown here, the flagging itself is done on a scan-by-scan basis. Clouds are clearly highlighted below ≈15 km until the Sarychev eruption on 15 June 2009, at which point the urn:x-wiley:jgrd:media:jgrd55565:jgrd55565-math-0032 value in the aerosol layer is often as large as where clouds are expected. The effect this has on the cloud flagging can be seen in the panel below. After the eruption, the early volcanic plume is often flagged as containing clouds, reducing the average extinction measured after an eruption. Although the 0.15-km−1 threshold could be tuned for the OSIRIS wavelengths to reduce aerosol-as-cloud misclassification, the similarity in values between clouds and volcanic aerosols makes improvements difficult without including many more clouds. The center column shows the updated cloud algorithm, using urn:x-wiley:jgrd:media:jgrd55565:jgrd55565-math-0033. Clouds are clearly visible below ≈15 km, and while the value of urn:x-wiley:jgrd:media:jgrd55565:jgrd55565-math-0034 does increase after the volcanic eruption, it is less likely to rise to the level typically seen in cloudy conditions, making the threshold easier to set. For this work a value of 7×10−4 km−2 was empirically determined to provide good discrimination.

Details are in the caption following the image
Comparison of the Chen et al. (2016) cloud detection algorithm applied to OSIRIS data, and the updated method used in this work. The left column shows results from the urn:x-wiley:jgrd:media:jgrd55565:jgrd55565-math-0035 quantity used in the Chen algorithm, and the center column from the updated method. The right column shows the difference in the effects on the final products. The top row shows the weekly averaged value of the cloud flag. The middle row shows the weekly averaged cloud fraction, and the bottom row the cloud-free extinction product. All plots are zonally averaged between 50°N and 70°N. The difference in cloud fraction is computed as urn:x-wiley:jgrd:media:jgrd55565:jgrd55565-math-0036. The difference in extinction is computed as urn:x-wiley:jgrd:media:jgrd55565:jgrd55565-math-0037%. The vertical gray line marks the eruption of Sarychev Peak on 15 June 2009, and the light gray line denotes the tropopause altitude. OSIRIS = Optical Spectrograph and InfraRed Imaging System.

The right column of Figure 9 shows the differences in the cloud fraction (row 2) between the two algorithms, and the difference in retrieved extinction (row 3). Cloud fraction is computed as the ratio of measurement that contains cloud to the total number. The updated algorithm generally flags clouds at slightly lower altitudes, placing more below the tropopause, denoted by the gray line. It also removes less aerosol after the Sarychev eruption, causing the zonally averaged extinction to increase by approximately 10% for the months following the eruption and up to 20% in isolated periods. It should be noted, however, that while this technique flags less aerosol as cloud, it is not a perfect classification. The inclusion of extinction in the threshold also means it will be more likely to miss flagging thin or spatially inhomogeneous clouds, as the smaller extinction values mean these cases are less likely to meet the threshold value. Any threshold technique is likely to misclassify some cases, especially in the UTLS where there may be a mix of both clouds and aerosol. Additionally, while fewer cases of aerosol are flagged as cloud, it is not a perfect discrimination, and high levels of aerosol such as immediately following an eruption are still occasionally flagged as clouds. Therefore, for studies involving the immediate evolution of volcanic plumes, where aerosol levels are high and still spatially inhomogeneous, it is recommended that the sampling of the OSIRIS instrument is taken into account; in the case of model comparisons by sampling at comparable locations and altitudes, or when comparing with other instruments through coincident comparisons, rather than the use of zonally averaged quantities when possible. The retrieved extinction values without cloud clearing applied are also provided in the final version 7 product for cases when distinction between clouds and aerosol may be ambiguous. It should also be noted that in the “cloud-free” extinction product, profiles are terminated above the cloud top, as interpretation of measurements that are below, but looking through the cloud are highly uncertain in terms of the retrieved aerosol.

It is also useful to briefly explore the retrieved cloud distributions as a check on the technique and chosen threshold. However, comparison of cloud measurements is complicated by the different sampling and sensitivities of various instruments. The long path lengths of limb instruments make them sensitive to thin cirrus clouds, but essentially blind to anything below a layer of cloud or thick aerosol. Conversely, nadir viewing instruments have less sensitivity to thin clouds but can penetrate to lower layers. Different satellite orbits also affect the local time of the measurements making climatologies harder to compare. Chepfer et al. (2010) used the Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observations (CALIPSO) data to develop the GCM (General Circulation Model) Oriented Cloud Calipso Product (GOCCP). While designed for comparison with GCMs it provides a convenient first test of the OSIRIS cloud flag as it has essentially global coverage of cloud fraction and cloud cover over the majority of the OSIRIS mission. This allows for sampling of the GOCCP at OSIRIS scan locations to avoid sampling biases. Comparisons with the CALIPSO-GOCCP data set shows good agreement, both in terms of altitude and spatial distribution of cloud fraction, as shown in Figure 10. The top row shows the zonally averaged cloud fraction between 2006 and 2018, with OSIRIS data on the left and CALIPSO-GOCCP on the right. Both data sets show a clear maximum in cloud fraction in the tropics near 15 km, with lows near ±25°, followed by a cloud layer that follows the tropopause. Generally, OSIRIS measures somewhat higher fractions of clouds and places them at slightly higher altitudes. This is at least partially due to the threshold set in the CALIPSO-GOCCP, which will only detect clouds with an optical depth greater than ≈0.03, missing many of the subvisual cirrus clouds. Horizontal inhomogeneity may also be playing a role. The long path lengths of a limb sounder means any cloud in the ≈200-km path length may result in a measurement being flagged as cloudy, while in truth only a portion of path contains clouds, biasing the cloud fraction high. Additionally, horizontal inhomogeneities may cause clouds to be assigned to lower altitudes if the line of sight passes through a cloud before or after the tangent point. Chepfer et al. (2013) investigated the differences between different CALIPSO cloud products and found that cloud fraction differences of 10–20% were not uncommon, with larger cloud fractions at higher altitudes in the tropics when different thresholds and averaging were used. Clouds in the tropics below ≈12 km are also underestimated by OSIRIS, likely due to high aerosol and clouds above these altitudes, which masks the cloud signature. The bottom panel shows the spatial distribution of cloud cover above 6.5 km from the two data sets over the same period. Cloud distribution is very similar in both data sets, although OSIRIS again measures somewhat higher cloud fraction, particularly in the tropics where more subvisual cirrus are expected.

Details are in the caption following the image
The top panels show the cloud fraction as a function of latitude from the OSIRIS version 7 (left) and CALIPSO-GOCCP data sets (right). The gray line indicates the mean tropopause altitude. The fractional cloud cover is shown in the bottom panels. All figures are computed from 2006 to 2017, the duration of the CALIPSO-GOCCP data set. OSIRIS = Optical Spectrograph and InfraRed Imaging System; CALIPSO = Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observations; GOCCP = GCM Oriented Cloud Calipso Product; GCM = general circulation model.

3.3 Polar Stratospheric Clouds

Although the OSIRIS orbit precludes measurements from being taken at high latitudes for most of the winter, measurements in the spring may contain polar stratospheric clouds (PSCs). As sulfate aerosols are often a primary component of these clouds, and they play an important role in ozone destruction, it is useful to flag PSCs separately from lower altitude ice and water clouds. To do this, PSCs are screened using the Chen et al. (2016) method with a threshold value of 0.12 with the additional constraint that the temperature at the tangent point must also be below 200 K. This method is used rather than the cloud detection presented above, as the inclusion of temperature allows for a reduced flagging of aerosol without the need to incorporate the retrieved extinction. Although 200 K is slightly above the typical formation temperature, it allows for some error in the ECMWF ERA-Interim climatology and variation along the line of sight. Figure 11 shows the weekly averaged extinction between 60°S and 90°S before and after removal of PSCs. Note that in the “cloud-free” panels PSCs as well as all aerosol below any PSC has been removed. The final panel shows the fraction of measurements that have been flagged as containing PSCs. Just after OSIRIS regains coverage in the austral spring, up to half of the measurements contain PSCs with essentially none remaining by the end of October. The exception to this is in 2015, when aerosol from the Calbuco Eruption in April 2015 produces PSCs well into November.

Details are in the caption following the image
Top panel shows the weekly averaged cloud-free extinction between 60°S and 90°S. The middle panel shows the extinction before cloud screening. The bottom panel shows the fraction of measurements that have been flagged as containing polar stratospheric clouds (PSCs).

3.4 UTLS Improvements

Combined with the cloud removal, the improved sensitivity of incorporating additional wavelengths allows for the version 7 algorithm to retrieve at lower altitudes than previously. Version 5.07 did not attempt retrievals below 10 km and has relied on various thresholds to determine lower limits depending on the use case. Version 5 of the OSIRIS data took a conservative approach and masked off any data below where the extinction exceeded ≈2.5 × 10−3 km−1. As noted in Fromm et al. (2014), this can cause apparent biases when comparing to records that include higher aerosol values. Other work such as Rieger et al. (2015) applied a much less conservative limit of 3×10−2 km−1 to help reduce these effects at the cost of including some clouds in the analysis. Version 7 does not limit the retrieval to above 10-km altitudes and consistently extends below the tropopause. Figure 12 shows a comparison between version 5.07 and 7 products in the UTLS after the Sarychev eruption. Each panel shows the change in average extinction relative to May 2009, the month preceding the eruption. The v5.07 product is shown in the left column, and immediately after the Sarychev eruption a clear increase aerosol extinction is present above 50°N. However, data in the tropics below the tropopause are heavily contaminated with clouds, and the 10-km cutoff results in data missing even above the tropopause at higher latitudes, with virtually no indication of what is happening below the aerosol plume. The version 7 data are shown in the center column and consistently extends down to 6.5-km altitudes. It shows a very similar evolution of the Sarychev plume above the tropopause; however, it also shows a clear decrease in aerosol levels in the troposphere outside of the tropics. CALIPSO data from the global space-based stratospheric aerosol climatology (GloSSAC) (Thomason et al., 2018) is shown in the far right column. CALIPSO backscatter at 532 nm has been converted to 750-nm extinction using a conversion factor of 30, consistent with the particle size used in the OSIRIS retrieval. The version 7 algorithm retrieves somewhat more variable extinction values in the tropical troposphere than CALIPSO. OSIRIS version 7 also retrieves somewhat lower aerosol values in the thickest part of the aerosol plume. This is likely an indication of some clouds remaining in the extinction product in the tropics, while some aerosol has been flagged as cloud at high extinction levels. Additionally, few extinction measurements exceed 0.01 km−1, due largely to limitations in the retrieval at large optical depths. Very large extinctions can be difficult to model, as assumptions about horizontal homogeneity and aerosol composition are likely less robust, and local minimums in the retrievals can lead to nonconvergence. As only converged profiles are reported in the final product, events with extinctions near or exceeding 0.01 km−1 will be underestimated when looking at averaged data. Despite the difficulty of the measurements, both instruments present a consistent picture of the UTLS region after a moderate volcanic eruption, with aerosol levels increasing rapidly above the tropopause, while altitudes below remain considerably cleaner than preeruption.

Details are in the caption following the image
Comparisons of upper troposphere and lower stratosphere measurements after the Sarychev eruption in June 2009. Monthly averaged extinction is shown as the difference from May 2009 values. OSIRIS v5.07 measurements are shown in the left column, version 7 in the center column, and CALIPSO-GloSSAC values in the right column. The gray line indicates the monthly mean tropopause altitude. OSIRIS = Optical Spectrograph and InfraRed Imaging System; CALIPSO = Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observations; GloSSAC = global space-based stratospheric aerosol climatology.

4 Validation

4.1 Sage Comparisons

SAGE II was launched in 1984 on board the Earth Radiation Budget Satellite and continued operation until 2005. It provided near-global sampling of stratospheric aerosol extinction approximately every month at wavelengths of 386, 452, 525, and 1,020 nm. For this study the SAGE II version 7 aerosol extinction (Damadeo et al., 2013) is used and converted to 750 nm by interpolating the 525- and 1,020-nm channels in logarithmic space. SAGE II data have been cloud cleared using the provided cloud flag.

SAGE III was launched on the Meteor-3 M (M3M) platform in 2001 into a polar orbit that provided coverage of the middle-to-high latitudes. SAGE III measured extinction at a range of wavelengths between 384 and 1,543 nm until 2006. A second SAGE III instrument was launched in February 2017 and placed on the International Space Station (ISS). SAGE III/ISS began operations in June 2017 with aerosol extinction processed using essentially the same algorithm as SAGE III/M3M. For this work the SAGE III/M3M comparisons use the version 4 extinction product at 755 nm (Thomason et al., 2010) and SAGE III/ISS uses the version 5.0 product at 756 nm. Note that the SAGE III data have not been cloud cleared. Using the 520- and 1,020-nm channels from SAGE III to interpolate to 750 nm, as was done with the SAGE II data, generally decreases the SAGE III extinction values by a few percent but does not meaningfully impact the comparisons, either in magnitude or standard deviation.

Scans are considered to be coincident if they are within ±2° latitude, ±10° longitude, and ±24 hr. Each line of sight through the limb spans ≈2° latitude, and an additional 1–4° latitude is spanned as the satellite orbits during the acquisition of a vertical profile, making tighter latitudinal criteria of little benefit. The median percent differences between OSIRIS version 7 and each of the SAGE instruments are shown in Figure 13 at three latitude bands. The solid line indicates the median, while the shaded regions indicate different percentiles of the data, as shown in the figure legend. The small numbers at the left of each panel indicate the number of coincident measurements at that altitude. For reference, the median version 5.07 differences are also shown in red.

Details are in the caption following the image
Coincident comparison of OSIRIS version 7 with SAGE II, SAGE III, and SAGE III-ISS. Differences are computed as (OSIRIS - SAGE)/SAGE × 100%. Solid lines show the median difference and shaded regions show various percentiles as indicated by the color bar. Version 7 results are shown in blue with version 5.07 comparisons shown in red as a reference. The numbers in the left of the panels indicate the number of coincident measurements at each altitude. OSIRIS = Optical Spectrograph and InfraRed Imaging System; SAGE = Stratospheric Aerosol and Gas Experiments.

Agreement between OSIRIS and the various SAGE instruments is generally very good, with biases of less than 10% for most regions above the tropopause. The exception to this is at high altitudes where OSIRIS has low bias with respect to SAGE and is thought to be due to sensitivity to stray light and nonzero aerosol in the OSIRIS normalization altitudes (Rieger et al., 2018). For the SAGE II and SAGE III/M3M comparisons there is little difference between the OSIRIS version 5.07 and 7 products, as the mean extinction does not change substantially between products when averaged over multiple years. However, the SAGE III/ISS comparison covers only a 6-month time span, limiting the range of OSIRIS scattering angles included in the comparison, and here improvements over version 5.07 are seen in the middle and high latitudes. Differences between SAGE III/ISS and OSIRIS in the tropics are currently under investigation.

In the future, moving to comparisons on equivalent latitude may be beneficial, as sampling in and outside of the polar vortex is evident in the SAGE III/ISS comparisons in the 30° to 90°S bin. The 90th percentile shows very large differences, a likely indicator that SAGE III/ISS is sampling inside of the relatively clean vortex, and OSIRIS outside. However, for 80% of the data, differences are within ±25%, indicating that when both instruments are sampling similar air masses agreement is generally very good.

4.2 Level 3 Comparisons

Although clouds and reduced sensitivity in the UTLS can make direct comparisons of averaged data difficult, the creation and validation of such climatologies is still an important task as they are often used both in model comparisons and as model inputs. Figure 14 shows OSIRIS data in monthly averages from 10°S to 10°N at three altitude levels. The shaded region indicates 1 standard deviation of the monthly mean values. Also plotted are the SAGE II data, averaged in the same way, as well as the CALIPSO extinction from GloSSAC, again converted using a backscatter to extinction factor of 30. This factor provides good agreement for most altitudes in the tropics, however, at higher latitudes, and particularly at altitudes above 25 km, causes CALIPSO to overestimate OSIRIS by up to a factor of 2. This may be an indicator of smaller particles at these higher extratropical altitudes. A lognormal distribution with a width of 1.6 and median radius of 50 nm would provide a Lidar conversion factor of 15, instead of the 30 used here, and greatly improve agreement in these regions (not shown). CALIPSO and SAGE II are displayed here as they, along with OSIRIS, provide the three main data sets used to construct GloSSAC. Additionally, neither SAGE II nor CALIPSO are limb scatter data sets, minimizing seasonal biases due to particle size assumptions, providing more independent data for time series comparisons. For reference, the version 5.07 data are also shown. Altitudes near 20 km remain largely unchanged between versions and both are in excellent agreement with both SAGE II and CALIPSO. The seasonal cycle in near-25 km has been greatly reduced in version 7, and OSIRIS now agrees much better with the CALIPSO extinction. Altitudes near 30 km agree very well for both versions of OSIRIS; however, large spikes in the version 5.07 data at the beginning of 2005, 2009, and 2014 have been reduced and now agree much better with CALIPSO. The cause of this is not currently known, but happens primarily when the quasi-biennial oscillation is transitioning from Easterly to Westerly, so it may be an indication of changing particle size coupled with the seasonal cycle in the OSIRIS scattering angle.

Details are in the caption following the image
Comparison of the monthly averaged version 7 OSIRIS aerosol record in the tropics compared to SAGE II, CALIPSO, and the OSIRIS version 5.07 record. The shaded region indicates 1 standard deviation of the monthly mean values. Gray triangles indicate the time of the largest volcanic eruptions during this time period. OSIRIS = Optical Spectrograph and InfraRed Imaging System; SAGE = Stratospheric Aerosol and Gas Experiments; CALIPSO = Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observations.

5 Conclusions

A new version of the OSIRIS aerosol extinction product has been processed using a multiwavelength retrieval. The data, as well as additional information on file formats, usage, and processing updates can be found at https://arg.usask.ca/docs/osiris_v7/. The new algorithm shows improved results when compared to previous versions as well as CALIPSO, SAGE II and SAGE III instruments. The dependence on the measurement geometry has been reduced when comparing measurements at different scattering angles, although some discrepancies remain just above the tropopause in the Northern Hemisphere, and the cause of this remains under investigation. Use of multiple wavelengths allows for retrievals at lower altitudes than previously, and agreement with CALIPSO measurements is promising even during enhanced aerosol conditions such as after the Sarychev eruption. However, with a limited wavelength range cloud clearing remains an imperfect exercise, and saturation of the retrieval immediately after eruptions may still result in low biases if zonal and temporal averages are used for comparisons.

Future work would benefit from improvements in a priori assumptions. Although the dependence on scattering angle has been reduced it has not been completely eliminated and improved particle size estimates would likely help this further. Additionally, this work did not address systematic biases due to extinction at higher altitudes where normalization precludes retrieval. These issues may be at least partially addressed through incorporation of measurements from different platforms. Investigation of using shorter wavelengths at higher altitudes may also be beneficial, in particular for instruments like OSIRIS where stray light at shorter wavelengths is minimal. However, while these improvements would help reduce remaining uncertainties, their impact on the larger climate was not addressed in this study, and future work will examine the possible significance of these errors.


This work was supported by the Natural Sciences and Engineering Research Council (Canada), the Canadian Space Agency. Odin is a Swedish-led satellite project funded jointly by Sweden (SNSB), Canada (CSA), France (CNES), and Finland (Tekes). The complete OSIRIS data set can be downloaded from https://arg.usask.ca/docs/osiris_v7/. SAGE and CALIPSO data were obtained from the NASA Langley Research Center EOSDIS Distributed Active Archive Center.