Detection of optical path in spectroscopic space-based observations of greenhouse gases: Application to GOSAT data processing



[1] We present a method to detect optical path modification due to atmospheric light scattering in space-based greenhouse gas spectroscopic sounding. This method, which was applied to the analysis of radiance spectra measured by the Greenhouse Gases Observing Satellite (GOSAT), is based on the path length probability density function (PPDF) and on retrieval of PPDF parameters from radiance spectra in the oxygen A-band of absorption at 0.76 μm. We show that these parameters can be effectively used to characterize the impact of atmospheric light scattering on carbon dioxide retrieval in the atmospheric carbon dioxide (CO2) absorption bands at 1.6 μm and 2.0 μm. The threshold for PPDF parameters is set so that the optical-path modification is negligible, and these settings are recommended as a basic guideline for selecting the clearest atmospheric scenarios. An example of data processing for six global GOSAT repeat cycles in April and July 2009 shows that PPDF-based selection efficiently removes CO2 retrieval biases associated with subvisible cirrus and sandstorm activities.

1. Introduction

[2] Atmospheric carbon dioxide (CO2) is generally accepted to be the dominant anthropogenic greenhouse gas. However, large uncertainties remain regarding CO2 natural global sources and sinks [Stephens et al., 2007]. Ground-based CO2 observations are fairly accurate but too sparse to sufficiently reduce uncertainties in source and sink characterization. Global monitoring of atmospheric carbon dioxide from space is expected to reduce uncertainties in estimating CO2 seasonal sources and sinks. Several studies have concluded that global, dense, and unbiased space-based remote-sensing measurements of CO2 column abundance at precisions of 1–10 ppmv (0.3–3%) will reduce the current uncertainties associated with estimates of CO2 sources and sinks [Rayner and O'Brien, 2001; Chevallier et al., 2007; Feng et al., 2009; Kadygrov et al., 2009; Baker et al., 2010]. The Greenhouse Gases Observing Satellite (GOSAT), which has been in orbit since 23 January 2009, has shown promise for improved surface flux inverse modeling [Chevallier et al., 2009].

[3] Atmospheric light-scattering by aerosols and clouds remains a major source of gas retrieval errors that could exceed the XCO2 precision limits [O'Brien and Rayner, 2002; Dufour and Bréon, 2003; Mao and Kawa, 2004; Aben et al., 2007; Reuter et al., 2010; Oshchepkov et al., 2008, 2009]. These errors arise from uncertainties in modifications of the optical path due to the high variability of aerosol, cloud, and ground surface optical properties. Thus, operational algorithms for data processing must accurately detect and account for the light-path modification through the atmosphere. It is therefore important to quantify, set up, and control the necessary threshold so that the path length modification is negligible or acceptable if aerosol and cloud characteristics are assumed or retrieved simultaneously with the gas level.

[4] There are several versions of algorithms in which the theoretical radiance is produced from the solution of the radiative transfer equation under modeled profiles of aerosol and cloud optical characteristics [Connor et al., 2008; Butz et al., 2009; Reuter et al., 2010]. The total aerosol and cloud optical thicknesses (AOT) that can be reliably derived from these algorithms or estimated independently from other satellites [e.g., Bréon et al., 2005] or from aerosol transport models [e.g., Takemura et al., 2000] are not always appropriate for characterizing aerosol and cloud impacts on gas retrievals because this impact also depends on surface albedo, vertical distribution of the optical characteristics, solar zenith angle, and other parameters. For example, near-ground aerosols with high AOT values might not modify the optical path significantly; therefore their spectra could agree with a polynomial fitting when a differential optical absorption spectroscopy (DOAS) technique is applied. Conversely, elevated aerosols or thin clouds with low AOT values might modify the optical path significantly, and thus the bias in gas spectroscopic sounding from space could be significant, depending on the surface albedo. Direct retrieval of the light path is possible using the recently developed photon path length probability density function (PPDF) method [Bril et al., 2007; Oshchepkov et al., 2008, 2009]. This method is based on a physically reasonable solution for the PPDF with a limited number of parameters that directly characterize the optical-path modification and which can be retrieved simultaneously with the gas amount.

[5] This study developed a method to detect the optical path from satellite measurements, such as those provided by GOSAT, and to quantify an admissible level of atmospheric light scattering when the path length modifications were negligible.

2. Methodology

2.1. Background

[6] The threshold level of the light-path modification beyond which estimation of gas levels is impractical depends on the ability of the retrieval method to correct aerosol and cloud effects. To define minimal threshold levels that could be applied to any retrieval algorithm, we focus on the simplest case of the clearest atmosphere, when the light-path modification is negligible. Under these conditions, DOAS could be applied to retrieve carbon dioxide levels from radiance spectra in the CO2 absorption bands. To control the optical-path modification, we used PPDF-based retrievals from radiance spectra in the oxygen A-band of absorption at 0.76 μm.

2.2. PPDF Radiative Transfer Model

[7] We considered the simplest PPDF model, in which a single boundary plane located at altitude h divides the atmosphere into two plane-parallel layers. To account for the optical-path modification, we analytically expressed the effective transmittance, equation image(C, h, α, ρ), through a set of PPDF parameters, h, α, ρ, and a state vector of the gas vertical profile, C [Bril et al., 2007; Oshchepkov et al., 2008] as follows:

equation image

where T1 = expequation image and T2 = expequation image; τ1 = equation imagek(h)dh and τ2 = equation imagek(h)dh are gaseous optical thickness in the lower layer and upper layer, respectively; δ = ρ · exp{−2(τ1 + τ2)}; Θ and Θ0 are the solar zenith and satellite angles, respectively; k(h) is the vertical profile of the gas absorption coefficient; and hc and ha are altitudes of the cloud or top of the aerosol layer and the top of the absorbing atmosphere, respectively.

[8] The light path can be modified by two mechanisms in this model. One is direct scattering of sunlight from the boundary plane to the satellite, and the other is due to multiple light scattering and reflections between the boundary plane and the ground surface. These effects are governed by the PPDF parameters α and ρ, respectively, and can result in either lengthening or shortening of the light path [Oshchepkov et al., 2008] compared with the direct path from the Sun to the surface to the satellite. PPDF parameters can be interpreted as follows [Oshchepkov et al., 2009]: α is the relative cloud/aerosol layer reflectivity, i.e., the ratio of photons scattered by the cloud/aerosol to the total number of photons within the view of the detector; and ρ is the scaled first moment of the PPDF under the cloud or within the aerosol layer, i.e., a parameter that describes path length modification due to multiple scattering/reflection between the ground surface and cloud/aerosol particles.

2.3. Retrieval Scheme and Quality Assessment

[9] We used the maximum a posteriori method with the Gauss–Newton iteration technique [Rodgers, 2000] to perform nonlinear fitting of the radiance spectra in both bands. PPDF parameters α, ρ, and h were retrieved in the oxygen A-band at a known temperature and surface pressure. The state retrieval vector in the CO2 absorption bands at 1.6 μm and 2.0 μm included a vertical profile of the carbon dioxide mixing ratio equidistantly distributed on a pressure scale and shifting factors of prior water vapor profiles. The retrieved CO2 profile was converted to column-averaged dry air mole fractions, XCO2, at a given pressure increment [Connor et al., 2008; Oshchepkov et al., 2008]. In each band, the low-frequency part of the inverted spectra was removed according to DOAS by adjusting the logarithm of radiance using a quadratic polynomial on the wavelength scale. The stretch factor was included to adjust the position of the wavenumber bins. The retrieval scheme was unified when processing the GOSAT measurements over land and for both regular and Sun Glint observation modes over the ocean.

[10] Finally, we performed a postprocessing quality assessment and dismissed those retrieved results for which the discrepancy between the forward model and the observed spectrum was significant (χ2 > 5 [Oshchepkov et al., 2008], signal-to-noise ratio (SNR) <100 for each spectral band, and degree of freedom for the signal (DFS) ≤1 [Rodgers, 2000]).

3. Data

[11] We used two sets of radiance spectra that were converted from raw interferograms measured by the GOSAT thermal and near-infrared sensor for carbon observation–Fourier transform spectrometer (TANSO-FTS) [Kuze et al., 2009] in the short wavelength infrared (SWIR) region: within 12,950–13,190 cm−1 in the oxygen A-band, within 6,185–6,272 cm−1 (1.61 μm band), and within 4,815–4,885 cm−1 (2.0 μm band), with ∼0.2 cm−1 spectral resolution. These spectra correspond to the P polarization component [Kuze et al., 2009]. A set of ∼5200 observation points from 20–28 April 2009 and a set of ∼4800 observation points from 20–28 July 2009 were chosen for data processing (three GOSAT global repeat cycles in each month). The locations of the selected single scans after the retrieval quality assessment (discussed in section 2.2) are plotted in Figure 1 (top) in terms of the total AOT that was calculated for each observation day by an offline three-dimensional aerosol transport model, the Spectral Radiation-Transport Model for Aerosol Species (SPRINTARS) [Takemura et al., 2000; Yoshida et al., 2010]. These observations fell into the category of cloud-free scenarios detected by a cloud and aerosol imager (CAI) [Kuze et al., 2009] when applying the TANSO-CAI cloud flag test developed by Ishida and Nakajima [2009]. However, Nakajima et al. [2008] noted that TANSO-CAI often fails to detect optically thin cirrus clouds because it does not have any thermal infrared channels sensitive to clouds in the upper troposphere. From the actual data processing, Yoshida et al. [2010] found that the TANSO-CAI cloud flag test tended to categorize sub-pixel-sized clouds as clear pixels over the ocean. Thus, the selected sets of GOSAT single scans could be contaminated by light scattering from thin cirrus clouds. To mitigate the aforementioned TANSO-CAI limitations, additional cloud detection methods (a TANSO-CAI spatial coherence test and a TANSO-FTS 2 μm band test) were applied to select cloud-free scenes in the operational GOSAT data processing [Yoshida et al., 2010; Oshchepkov et al., 2008]. We briefly describe these tests in section 4.3 when comparing them with PPDF-based thin cloud screening over the ocean.

Figure 1.

Global maps of the total aerosol optical thickness derived from SPRINTARS model (top) and surface albedo estimated at the gas window channels in the 1.6 μm band (bottom) for three global GOSAT repeat cycles from 20–28 April 2009 (left) and for three global GOSAT repeat cycles from 20–28 July 2009 (right). Values correspond to the color scale and the data set obtained after a postprocessing quality assessment.

[12] The prior gas profile for all observation points was uniform at 385 ppmv with a prior constraint of 30 ppmv standard deviation for all pressure levels, ensuring a DFS of ≥1. Surface pressure, temperature, and water vapor profiles were provided by the Japan Meteorological Agency (JMA) grid point value (GPV) data set. The gas cross-sections for theoretical transmittance were calculated using the HITRAN 2008 database [Rothman et al., 2009].

[13] For the algorithm evaluation we compared the retrievals with transport model simulations that are available globally and can be validated by comparison with the observed CO2 seasonal cycles. Simulated XCO2 values for each observation point were prepared with the National Institute for Environmental Studies global atmospheric tracer transport model [Maksyutov et al., 2008] (hereafter referred to as the NIES ATM). This model was run using the surface fluxes given by Law et al. [2008], 0.5 degree resolution wind from the JMA GPV operational analysis data set, and planetary boundary layer (PBL) height data from the European Centre for Medium-Range Weather Forecasts (ECMWF) operational analyses. The model performance at lower spatial resolutions was evaluated by comparison with observations of the mean seasonal cycle [Belikov et al., 2010] and synoptic scale variability [Patra et al., 2008]. To improve simulations of the seasonal cycle and trend, we also applied a surface flux correction on a monthly mean basis using inverse model-derived flux climatology following the procedure of Gurney et al. [2004], which matched the monthly mean observations with 1–2 ppmv. The stratospheric age of air bias in the model was corrected by replacing the stratospheric part with an observation-based climatological distribution following the profile observations of Aoki et al. [2003] and Andrews et al. [2001].

4. Results

4.1. Retrieval Bias and PPDF Parameters

[14] Initially, we verified the satellite-based XCO′2 retrievals by reference to modeled global XCO*2 data provided by NIES ATM [Maksyutov et al., 2008]. The bias ΔXCO2 = XCO′2 − XCO*2 for DOAS-based CO2 retrievals in the 1.6 μm and 2.0 μm bands at α = ρ = 0 is a suitable measure of the light-path modification.

[15] Figure 2 shows the bias ΔXCO2 as a function of α (top) and ρ (bottom) retrieved in the oxygen A-band. The color scale in Figure 2 shows the ground surface albedo retrieved over land at the gas window channels in the 1.6 μm band. Retrievals over sea are shown separately by open blue circles. Positive bias, or overestimation, of XCO2 (Figure 2, bottom) was pronounced for a bright surface (solid red circles) and increased as ρ increased. This bias resulted from multiple reflections of light between a bright surface and an aerosol or cloud layer, which lengthened the mean light path. Thus, DOAS overestimated the gas level. In contrast, aerosol and cloud light scattering over a dark surface resulted in underestimated gas retrievals (Figure 2, top). For a dark land or sea surface in Sun Glint observation mode, the optical path is modified by the photons that do not reach the surface because they are scattered by aerosols and clouds and detected by the satellite. Indeed, the fraction of photons that reach the detector after being scattered by aerosol/cloud and then reflected by the dark surface is small. In Sun Glint mode, the fraction of light that reaches the detector after a series of mirror reflections and aerosol/cloud scattering is also small. Thus, for both cases, light scattering shortens the path length, and the DOAS-based retrievals underestimate XCO2. The effects detected in the present study of space-based measurements were previously reported using synthetic radiance spectra numerically simulated for light scattering over land and sea surfaces [Aben et al., 2007; Oshchepkov et al., 2008].

Figure 2.

Scatterplot of XCO2 retrieval bias as a function of α (top) and ρ (bottom) for three global GOSAT repeat cycles from 20–28 April 2009 (left) and for three global GOSAT repeat cycles from 20–28 July 2009 (right). The color scale range shows average ground surface albedo in the 1.6 μm band; the bias over sea is plotted separately as open blue circles.

[16] The PPDF parameters α and ρ predominantly define negative-bias and positive-bias ΔXCO2, respectively, providing a simple method for selecting observation scenarios in which the impact of light-path modifications on gas retrievals is limited. It should be noted that simultaneous retrieval of h is important for providing nonbiased estimates of α and ρ parameters.

4.2. PPDF Selection

[17] Figure 3 shows that PPDF-based selection improved the latitudinal distribution of XCO2 retrievals with reference to the modeled data. The data over land were scaled by color, depending on the surface albedo. Admissible levels of both α and ρ parameters were chosen empirically to be 0.04, which is appropriate for Rayleigh light scattering in the oxygen A-band of absorption at 0.76 μm. Restriction of α ≤ 0.04 (Figure 3, middle) effectively removed underestimated XCO2 DOAS-based retrievals over sea (open blue circles). Further limitation of ρ ≤ 0.04 excluded the overestimated retrievals over land at high values of surface albedo (Figure 3, right). Latitudinal distributions of the selected retrievals averaged over 10° latitude grids (black curves) were within the scatterplot of modeled XCO2 (magenta crosses) and repeated the tendency of heightened values of XCO2 in the Northern Hemisphere in April (Figure 3, top) and the tendency of heightened values of XCO2 in the Southern Hemisphere in July (Figure 3, bottom). PPDF-based selection also reduced the spread in XCO2. (The bars in Figure 3 represent standard deviations over 10° latitude grids.) After averaging over latitude, the standard deviation decreased from 4.7 ppmv (Figure 3, top left) to 2.8 ppmv (Figure 3, top right) in April and from 4.9 ppmv (Figure 3, bottom left) to 2.8 ppmv (Figure 3, bottom right) in July. The remaining scenarios after PPDF-based selection were ∼30% and ∼45% of the total scenarios in April and July, respectively (following the postprocessing quality assessment).

Figure 3.

Latitudinal distribution of XCO2 DOAS-based retrievals without PPDF selection (left), assuming only α ≤ 0.04 (middle), and assuming both α ≤ 0.04 and ρ ≤ 0.04 (right) for three global GOSAT repeat cycles from 20–28 April 2009 (top) and for three global GOSAT repeat cycles from 20–28 July 2009 (bottom). The color scale range shows average ground surface albedo in the 1.6 μm band; retrievals over sea are plotted separately as open blue circles. Black curves correspond to XCO2 averaged over 10° latitude grids, and black bars show standard deviation.

4.3. Comparison With Other Techniques

[18] For reference and validation, we compare PPDF-based retrievals with other criteria that are currently used for thin cloud detection after cloud prescreening by the TANSO-CAI cloud flag test (discussed in section 3). The TANSO-CAI spatial coherence test deals with evaluation of the standard deviation of TANSO-CAI radiance within the TANSO-FTS instantaneous field of view (IFOV). If the standard deviation exceeds some threshold value, it is expected that the measurement scene contains sub-pixel-sized clouds. The threshold value has been empirically determined to separate eye-checked clear/cloudy scenes from TANSO-CAI images [Yoshida et al., 2010]. This test is applied only to observations over the ocean because the spatial variability of the ground surface albedo within the TANSO-FTS IFOV is relatively high for land scenarios [Yoshida et al., 2010]. The TANSO-FTS 2 μm band test looks for the existence of elevated scattering particles (mainly cirrus clouds) using the measurement radiance of the H2O-saturated absorption area of the 2.0 μm band (5150 to 5200 cm−1) [Oshchepkov et al., 2008; Yoshida et al., 2010]. The ten most H2O-absorptive channels are selected in this test to avoid contamination of the surface-reflected light [Yoshida et al., 2010]. If there is no light scattering in the upper layers, almost no radiance in this region is observed by TANSO-FTS. Because the H2O mixing ratio decreases exponentially with height under standard atmospheric conditions, the radiation scattered by elevated particles should not be completely absorbed by H2O and reaches the TANSO-FTS, depending on the partial column abundance of H2O above the scattering. Therefore, if the average radiance of the selected strong H2O-absorption channels exceeds the noise level radiance, light-scattering effects are expected to exist in the upper layers of the atmosphere.

[19] In both tests, all TANSO-CAI pixels or the signal value in the H2O-saturated absorption area of the 2.0 μm band are categorized as clear or cloudy cases by a 0 or 1 flag, respectively. These categories do not provide the level of atmospheric light scattering for each single scan. To quantify this level for comparison of these tests with PPDF-based selections, we categorized these flags as a function of the α parameter. This parameter is mainly responsible for optical path modification over a dark surface (discussed in sections 4.1 and 4.2) and thus can be compared with the TANSO-CAI spatial coherence test that works only over the ocean.

[20] Figure 4 shows α distributions of the GOSAT observation frequency for each test. The histograms in Figure 4 represent the number of GOSAT single scans that fall into each α bin (ln α, ln α + Δ ln α) when applying PPDF (α) retrievals (blue), the TANSO-CAI spatial coherence test (red), and the TANSO-FTS 2 μm band test (green). The α distributions of the TANSO-CAI spatial coherence test agree well with those obtained from PPDF-based retrievals (red and blue histograms). For the TANSO-FTS 2 μm band test, the agreement is limited by high values of α: α > 0.4 in April (Figure 4, left) and α > 0.15 in July (Figure 4, right). This is likely a result of too few elevated light-scattering particles to provide a sufficient signal in the H2O-saturated absorption area of the 2.0 μm band. The TANSO-CAI spatial coherence test reduces the XCO2 standard deviation from 5.3 to 3.62 ppmv for April and from 7.3 to 5.6 ppmv in July. PPDF-based screening (α ≤ 0.04) reduces the standard deviation more efficiently, up to 1.6 and 2.2 ppmv, respectively (open blue circles in Figure 3, left to right), and is applicable for observations besides those over the ocean. Other advantages of the PPDF-based selection are that it uses the same instrument (TANSO-FTS) by analyzing radiance spectra in the oxygen A-band at 0.76 μm and detection of optical path modification between a surface and cloud or aerosol layer, ρ, which is not available from the TANSO-CAI spatial coherence test or the TANSO-FTS 2 μm band test.

Figure 4.

Alpha distribution of GOSAT over-ocean single-scan frequency when applying PPDF (α) retrievals (blue), the TANSO-CAI spatial coherence test (red), and the TANSO-FTS 2 μm band test (green).

4.4. Global Distributions

[21] Global maps of the modeled XCO2 as well as DOAS-based retrievals before and after PPDF-based selection are presented in Figure 5 (top to bottom). A comparison of the bottom and middle plots shows that PPDF-based selection α < 0.04 and ρ < 0.04 removed overestimated XCO2 mostly over the Sahara and Arabian Peninsula deserts, whereas underestimated gas levels were removed over oceans in tropical regions. These biases represent lengthening of the optical path due to light scattering caused by sandstorm activity over a bright surface [Houweling et al., 2005] and shortening of the optical path due to light scattering caused by high-altitude subvisible cirrus in tropical regions over a dark surface [Aben et al., 2007]. These statements are supported by high values of surface albedo (Figure 1, bottom), the total AOT (Figure 1, top), and the ρ parameter (Figure 6, bottom), over the Sahara and Arabian Peninsula deserts. The shortening of the optical path due to light scattering caused by high-altitude subvisible cirrus over the ocean (dark surface) is supported by the TANSO-CAI spatial coherence test (discussed in section 4.3) and high values of the α parameter in tropical regions (Figure 6, top).

Figure 5.

Global maps of XCO2 (values correspond to the color scale) according to NIES ATM (top), DOAS-based retrievals (middle), and DOAS-based retrievals with PPDF selection α < 0.04 and ρ < 0.04 (bottom) for three global GOSAT repeat cycles from 20–28 April 2009 (left) and for three global GOSAT repeat cycles from 20–28 July 2009 (right).

Figure 6.

Global maps of α (top) and ρ (bottom) for three global GOSAT repeat cycles from 20–28 April 2009 (left) and for three global GOSAT repeat cycles from 20–28 July 2009 (right). Values correspond to the color scale and the data set obtained after a postprocessing quality assessment.

5. Summary and Concluding Remarks

[22] We present an original method to detect optical path modification due to atmospheric light scattering from satellite measurements such as GOSAT. The PPDF-based method is very rapid, as analytical representation of the modeled spectra through PPDF parameters excludes time-consuming radiative transfer computations due to light scattering from the retrieval process. The retrieval methodology is unified for all bands because the DOAS-based technique for the CO2 bands immediately follows from the PPDF-based scheme, assuming α = ρ = 0.

[23] To quantify admissible levels of atmospheric light scattering, we chose the simplest case when optical path modification is negligible, allowing a simple DOAS-based technique to be applied for CO2 retrievals. Note in this connection that ignoring the optical path does not imply that there is no aerosol in the atmosphere. Indeed, certain aerosol components, such as those near the ground, can be accounted for by polynomial spectral fitting when applying the DOAS technique. The threshold for the allowed optical path modification should obviously be increased if aerosol/cloud effects are included in the retrieval scheme, such as in the weighting function-modified DOAS (WFM-DOAS) [Buchwitz et al., 2000] or if aerosol or cloud is retrieved simultaneously with gas levels. We recommend more detailed studies of threshold values for these cases. However, we suggest that selection of observation conditions under negligible optical path modification is useful for testing any retrieval method for spectroscopy, meteorological data, instrumental line shape, and any other factors that are not associated with atmospheric light scattering.

[24] Initial application of optical-path detection to GOSAT observations for six global repeat cycles in April and July 2009 showed that both gas and PPDF retrievals had a clear physical interpretation. PPDF-based selection provided acceptable retrievals of column-integrated levels of carbon dioxide in comparison with those predicted by the atmospheric tracer transport model.


[25] GOSAT is a joint effort promoted by the Japan Aerospace Exploration Agency (JAXA), the National Institute for Environmental Studies (NIES), and the Ministry of the Environment (MOE), Japan. The authors thank the members of the NIES GOSAT Project and ACOS team for their useful comments.