Embed Size (px)
Citation preview
Seediscussions,stats,andauthorprofilesforthispublicationat:https://www.researchgate.net/publication/228875405
OptimalestimationappliedtotheretrievalofaerosolloadusingMSG/SEVIRIobservations
ARTICLEinPROCEEDINGSOFSPIE-THEINTERNATIONALSOCIETYFOROPTICALENGINEERING·OCTOBER2007
ImpactFactor:0.2·DOI:10.1117/12.737729
CITATIONS
6
READS
29
4AUTHORS,INCLUDING:
SebastienC.Wagner
EuropeanOrganisationfortheExploitation…
18PUBLICATIONS82CITATIONS
SEEPROFILE
Y.Govaerts
Rayference
110PUBLICATIONS1,596CITATIONS
SEEPROFILE
AlessioLattanzio
EuropeanOrganisationfortheExploitation…
43PUBLICATIONS338CITATIONS
SEEPROFILE
Availablefrom:AlessioLattanzio
Retrievedon:05February2016
Optimal estimation applied to the retrieval of aerosol loadusing MSG/SEVIRI observations
S. Wagnera, Y.M. Govaertsb, A. Lattanzioc and Ph. Wattsb
aWagner Consulting, Darmstadt, Germany;bEUMETSAT, Darmstadt, Germany;cMakalumedia, Darmstadt, Germany
ABSTRACT
Using the principle of reciprocity, observations acquired by the SEVIRI radiometer on-board the Meteosat SecondGeneration satellites provide multi-angular and multi-spectral measurements that can be used for retrievinginformation on both the atmospheric aerosol load, and the Earth surface. The purpose of the presented newLand Daily Aerosol algorithm developed at EUMETSAT is to derive simultaneously the mean daily troposphericaerosol load and the land surface properties from the SEVIRI observations. The algorithm is based on theOptimal Estimation theory. The aerosol load is calculated through the optical depth parameter, for variousclasses of aerosols over land surfaces, and is inferred from the inversion of a forward radiative transfer modelagainst daily-accumulated observations in the 0.6, 0.8 and 1.6 SEVIRI bands. These daily time series providethe angular sampling used to discriminate the radiative effects that result from the surface anisotropy, fromthose caused by the aerosol scattering. Results of comparisons with AERONET data are presented to validatethe modelling approach and the algorithm that resolves the inversion problem. The retrieval error is analysed,together with the effects on the retrieval quality of updating in time the prior information.
Keywords: optimal estimation, aerosols, land, SEVIRI, diagnostic tools
1. INTRODUCTION
One of the major problems related to the retrieval of tropospheric aerosol load over land surfaces from space-borne imager observations consists in the discrimination of the contribution of the observed signal reflected bythe surface from the one scattered by aerosols. It is thus equivalent to solving a radiative system composed ofminimum two layers, where the upper layer(s) include(s) aerosols and the bottom one(s) represent(s) the surface.This problem is further complicated by the intrinsic anisotropic radiative behaviour of natural surfaces and theatmosphere/surface radiative coupling. The accuracy with which it is possible to retrieve aerosol properties,i.e.,to determine the radiative properties of the upper layer(s), is therefore intimately related to the determinationof the underling surface characteristics. As the number of independent observations is not large enough to fullycharacterise the radiative properties of the observed medium, i.e., the problem is ill-posed, it is necessary toconstrain the problem by providing some assumptions or a priori knowledge on this medium. Typically, thisadditional knowledge concerns the lower layer(s), i.e., the surface properties, as the primary objective is todetermine the characteristics of the upper aerosol layer.
Different approaches have been proposed so far to define a priori information on surface properties accordingto the type of radiometers or the nature of the retrieval algorithms. To retrieve aerosol load from Geostation-ary Operational Environmental Satellite (GOES) data, Knapp et al.1 determines the surface contribution from
Further author information: (Send correspondence to S. Wagner)S. Wagner: E-mail: [email protected], Telephone: +49 6151 8077344Y.M. Govaerts: E-mail: [email protected]. Lattanzio: E-mail: [email protected]. Watts: E-mail: [email protected]
Proceedings of the SPIE Europe Remote Sensing, 17 – 20 September 2007, Florence, Italy
temporal compositing of visible imagery, where darker pixels correspond to less atmospheric attenuation andsurface reflectance is deduced from the composite using radiative transfer. Remer et al.2 assumes a surfacealbedo ratio between two difference spectral bands deriving aerosol load from the Moderate Resolution ImagingSpectroradiometer (MODIS) observations. This method has been subsequently improved to relate this ratio tothe amount of vegetation.3 The algorithm proposed by Diner et al.4 based on the Multiangle Imaging Spec-troRadiometer (MISR) data assumes a spectral invariance of the normalised shape of the surface BidirectionalReflectance Factor (BRF).
An original method is proposed here where surface anisotropy and atmospheric scattering properties arecharacterised simultaneously, explicitly accounting for the radiative coupling between these two systems. Thealgorithm exploits the frequent Spinning Enhanced Visible and Infrared Imager (SEVIRI) observations to char-acterise the surface BRF. The use of a priori information, assuming that the surface albedo temporal variationsare much slower than aerosol ones, is mathematically rigorously combined with the information derived fromSEVIRI observations in the framework of the Optimal Estimation (OE) method.5,6 The retrieval approach isbriefly described in Section (2). The algorithm evaluation experiments are described in Section (3). At last, theresults of the validation test are presented in Section (4). This paper analyses in detail the impact of improvingthe surface properties prior knowledge on the accuracy of the retrieved aerosol load.
Figure 1. LDA forward model and measurement vector schematic representation.
2. RETRIEVAL METHOD
2.1 Forward model
The main objective of the Land Daily Aerosol (LDA) algorithm is to derive mean daily tropospheric aerosoloptical depth (AOD) for various aerosol classes over land surfaces. The aerosol properties are inferred from the
Figure 2. Left panel: Location of the AERONET stations used in this study. The ∗ symbol indicates the restricted list ofstations for which the time series analysis was performed shown on Fig. (6). Right panel: Aerosol assymetry factor as afunction of the single scattering albedo for the various classes used in LDA.
inversion of a forward radiative transfer model against daily accumulated observations in the 0.6, 0.8 and 1.6SEVIRI bands. This model expresses the Top-of-Atmosphere (TOA) BRF ym in a given spectral band as a sumof the atmospheric contribution and the contribution specifically due to surface scattering effects. In addition,the gaseous absorption are treated separately from the molecular and aerosol scattering-absorbing effects, sothat the system can be represented at the end by a 3-layer model as in Fig. (1). It explicity accounts for thesurface anisotropy and its coupling with atmospheric scattering.7 The properties of this model, defined by thestate vector x, include the surface reflectance ρs(λ) in the three SEVIRI bands and the aerosol optical thicknessxa = τ normalised at 0.55µm . The surface reflectance is represented by the RPV model8 as a function of:ρ, the amplitude of the surface contribution, k the modified Minnaert contribution, Θ the Heynyey-Greensteinparameter that characterise the surface anisotropic properties, and h the hot spot parameter. In our case h isconstant. These parameters are all wavelength dependent. Aerosol optical thickness is retrieved at 0.55µm forthree standard pre-defined aerosol classes (e.g., large particles,9 small non-absorbing10 and continental11) shownin Fig. (2, right panel). All these aerosol models assume spherical particles. The costs associated with thecomputation of the forward model is reduced by pre-computing the TOA BRF for a limited number of opticalthicknesses, i.e., 0.05, 0.1, 0.2, 0.3, 0.4, 0.6, 0.8, 1.0, 1.5. No interpolation is performed between these values.
2.2 Measurement vector
The proposed algorithm capitalises on the capability of SEVIRI to acquire data every 15 minutes to performan angular sampling of a same pixel under various solar geometries (Fig. 1). Using the principle of reciprocityapplied to the three SEVIRI solar channels,12 the temporal accumulation is thus used to form a virtual multi-angular and multi-spectral measurements y that can be used for retrieving simultaneously information on theatmospheric aerosol load, and the Earth surface reflectance. This approach assumes that the aerosol load doesnot vary during the course of the day. The error induced by this assumption is converted into a equivalent aerosolautocorrelation noise which is added to the estimated measurement uncertainty.13
2.3 Optimal Estimation
The basic principle of Optimal Estimation (OE) is to maximise the probability of the retrieved atmospheric stateconditional on the value of the measurements and any a priori knowledge on the observed medium. Maximising
probability is equivalent to minimising a cost function which combines these two pieces of information. Thisfunction writes
J(x) =(ym − y
)S−1
y
(ym − y
)T +(x− xb
)S−1
x
(x− xb
)T (1)
where x is the state vector (retrieved model parameters), ym the modelled SEVIRI observations, y the measure-ment vector, Sy is the measurement error covariance matrix, xb the a priori knowledge, and Sx the a priori errorcovariance matrix. A Marquardt-Levenberg descent algorithm is used for the minimisation of J(x) which adoptsthe steepest descent away from the solution and makes use of the Newtonian descent near the solution. The re-trieval is performed using several randomly-chosen independent first guess values in order to reduce the possibilityof finding a solution x corresponding to a local minimum.
The quality of the retrieval is expressed by the probability Px to find a solution x with a cost J(x) greater thanthe cost that has been calculated. A high Px value means that there is a high probability not to find a solutionat lower cost. The retrieval error is based on the OE theory, assuming a linear behaviour of ym in the vicinityof the solution x. Under this condition, the Hessian ∇2J(x) at the minimum of J approximates the inverse ofthe a posteriori error covariance matrix Sε. This matrix provides information on the correlation between theretrieved model state variables of the solution state vector x extracted from a posteriori error covariance matrix.Comparison of the values in the a priori and a posteriori covariance matrices thus expresses the knowledge gainfrom the measurement system and its associated uncertainties.
2.4 a priori information definitionAs already mentioned, one of the major issues in retrieving the aerosol load over land surfaces is to separate theaerosol contribution from the surface one. In order to improve the retrieval of the aerosol load, it is thereforeessential to characterise the surface as well as possible. To do so, LDA uses a mechanism to update the priorknowledge xb on the surface. This prior update procedure relies on the assumption that the surface albedotemporal stability is higher than the aerosol load one. Hence, a timeseries analysis can be performed on theprevious day retrievals, seven in the present case, to provide a priori information on the expected amplitude andshape of the surface BRF. LDA is able to detect a sudden change in the surface albedo (i.e., because of snow, orrain). However, if the lifetime of this change is not long enough with respect to the period set for the analysis,the corresponding values of the surface variables will be discarded.
3. EVALUATION EXPERIMENTS
In order to evaluate the efficiency of the retrieval procedure and prior update mechanism, comparisons withAERONET14 data were made over the time period extending from 15 February 2005 to 15 April 2005. 56AERONET stations with data available during that period over the Meteosat Second Generation (MSG) discare shown in Fig. (2).
A series of experiments are defined to evaluate the algorithm performance, comparing the retrieved aerosoloptical thicknesses with those from the AERONET dataset. The reference experiment is made without updatingthe prior information on the surface. The default prior information is constant from a day to another and isgiven with a very large error covariance Sx so that xb has no significant impact on J(x). Sensitivity analysis isperformed using, from a day to another, updated prior information on the surface, as described in Section (2).The results of these retrievals are compared with the cases without prior update of the surface reflectance.
The effect of xb and Sx on the retrieved state vector x is first illustrated with a visual inspection of atime series retrieved over the Dakar AERONET station (Section 4.1). To further analyse the performanceof the algorithm, and in particular the impact of the assumption on the daily stability of the aerosol load,a posteriori autocorrelation matrices associated to the solution x are compared with matrices derived from thereference experiment (4.2). Finally, a quantitative analysis is performed on the temporal evolution of the rootmean square error (RMSE) between LDA retrieval and AERONET data. This last analysis is limited to stationsthat have a complete record during the two-month time interval, and that are indicated with the ∗ symbol onFig. (2, left panel). The overall impact is also analysed considering all AERONET observations available duringthe investigation period (Section 4.3). In all these experiments, only retrievals with a probability Px larger than0.9 are considered for analysis.
Figure 3. Comparison with AERONET data for Dakar. Effects on the AOD and the BHR (presented: channel VIS08)of updating the prior information. Dashed line with × symbols: no update of the prior information. Plain line with 2
symbols: update of the prior information. Vertical bars in the lower panel indicate the daily variability of AERONETretrieved AOD.
4. RESULTS
4.1 Effects in time of updating the prior information
Time series analysis were performed on the available AERONET data in order to illustrate the impact of theprior update on the surface BRF and aerosol load retrieval. To update the prior knowledge on the surface clearlyled to a stabilisation of the surface parameters, and therefore a stabilisation of the surface albedo. As a result,the contribution of the surface is better separated from the contribution of the aerosols. Fig. 3 illustrates inthe case of Dakar the benefits on both the surface BRF and the AOD of updating the prior information onthe surface parameters. For clearness, only the BiHemispherical Reflectance (BHR) in the VIS08 channel ispresented. However, a similar stabilisation of the surface contribution occurs in the VIS06 and NIR16 channels.As the surface parameters are stabilised, the retrieved AOD fits better the AERONET retrieved AOD than inthe case when no update of the a priori information is made.
4.2 A close look at the autocorrelation matrices
The a posteriori error covariance matrix, noted Sε, provides useful information on the correlation between thestate variables. In order to interpret the non-diagonal terms of Sε a rescaling is performed, leading to theautocorrelation matrix Σ whose components are:
σij =Sε,ij√
Sε,ii · Sε,jj
(2)
Figure 4. Example of averaged autocorrelation matrices. Left panel: no update of the prior information. Right panel:with updated prior information. Grey levels range from 0 (in white), up to 1 (in black). Positive correlations are indicatedwith the symbol +. Negative correlations are indicated with the symbol −. ρ is the amplitude of the surface contribution.k is the modified Minnaert contribution. θ is the Heynyey-Greenstein contribution (surface anisotropy). The indices 1,2,and 3 stand respectively for the channels VIS06, VIS08, and NIR1.6.
Figure 5. Example of averaged relative error matrices. Left panel: no update of the prior information. Right panel: withupdated prior information. Values ranges from 0 (in white) up to a predefined threshold (here 100) (in black). ρ is theamplitude of the surface contribution. k is the modified Minnaert contribution. θ is the Heynyey-Greenstein contribution(surface anisotropy). The indices 1,2, and 3 stand respectively for the channels VIS06, VIS08, and NIR1.6.
Such that −1 ≤ σij ≤ 1. When σij → 0 uncertainties on the state variables i and j are not correlated. If σij → 1then uncertainties on the state variables i and j are correlated. And, if σij → −1 then uncertainties on the statevariables i and j are anticorrelated . The analysis if the autocorrelation matrices Σ derived for each station withand without update of the prior information essentially reveals four different types of impact:
1. Only the correlations between the amplitude (ρ) of the surface contribution and the AOD τ550 are reduced,
2. Only the correlations between the anisotropy surface (θ) and the AOD τ550 are reduced,
Figure 6. RMSE time series analysis between retrieved aerosol optical thickness from SEVIRI observations and AERONETdata shown with a ∗ symbol in Fig. (2). The solid line represents the reference experiment, i.e., without update of thea priori information on surface reflectance. The dashed lines represents the RMSE with update of the a priori information.
3. Both correlations between τ550 and the amplitude (ρ), and τ550 and the surface anisotropy (θ) are reduced
4. The update of the prior has little impact on the retrieval.
Fig.(4) illustrates the benefits of updating the a priori knowledge in the case where both the crossed termsτ / amplitude (ρ), and τ / surface anisotropy (θ) are reduced. The presented matrices are an averaged value ofthe results obtained for all the stations where such a behaviour was observed. In the case of Fig.(4), it is clearthat the overall correlations are decreased when improving the a priori information. The correlation between theerrors on τ , and the surface variables is reduced by at least 37% (with a maximum of 94% for the crossed termsτ / k) in the VIS08 and NIR16 channels. In the VIS06 channel, the decrease ranges between 4% and 31%. Ingeneral, the lowest gain is observed for the crossed terms τ / ρ, or τ / θ, showing that the coupling between theaerosol contribution and the surface contribution to the TOA signal is made through both the amplitude of thesurface contribution and the anisotropic properties of the surface. The anticorrelation between τ and the surfaceparameters ρ and θ is an evidence of the compensation mechanism that couples the surface contribution to theaerosol one (if one contribution is too small, the other one will compensate). It demonstrates the importanceof characterising the surface contribution as accurately as possible to retrieve the best possible aerosol signal.For the diagonal block illustrating the correlation between the surface parameters of a same channel, the gainobtained by improving the a priori knowledge ranges between 10% and 46%, with the lowest values for the crossedterms ρ / θ. For the extra-diagonal blocks, some negative gains have been locally observed (i.e., an increasingcorrelation despite an improved a priori information) but they correspond to negligible absolute values in anycase when compared with the order of most of the matrix values.Fig.(5) presents the corresponding relative errors for the simulations without update of the a priori information,and the simulations with improved a priori knowledge. The significant decrease of the error terms when thea priori information on the surface is improved is clear.
4.3 Quantitative effects of updating the priorFig. (6) represents the RMSE time series analysis between retrieved aerosol optical thickness from SEVIRIobservations and AERONET data shown with a ∗ symbol in Fig. (2). When no update of the a priori takes
Figure 7. AERONET versus LDA AOD for 0.1 bins using all station shown in Fig. (2) from 15 February 2005 to 15 April2005. The symbol 3 represents the reference experiment. The symbol 2 represents the experiment with a priori update.Left panel: all LDA aerosol classes. Centre panel: only AOD for which LDA retrieved the large particle class. Rightpanel : idem but for the small non-absorbing class.
place (solid line), the overall trend of the RMSE is to remain constant during the period of interest, despite sharpvariations from a day to another. Conversely, when a priori information on the surface is provided, the RMSEdecreases in time (dashed line). This figure shows the positive impact of this mechanism, where the “memory”on the state of the surface is provided to the algorithm allowing a 35% reduction of the RMSE.
This effect is further analysed in Fig. (7) considering now all AERONET stations in the MSG disc. On theaverage, the a priori mechanism tends to reduce the RMSE and the slope of the regression line (left panel). Lesslarge particles are found when a priori information on the surface is provided. This mechanism has howeverlimited impact in this case (central panel). Conversely, retrievals are much better when LDA identify the smallnon-absorbing aerosol as the best class (right panel). These results demonstrate the importance of the aerosolclass definition in the accuracy of the retrieval. In particular, it is necessary to consider correctly the shape oflarge aerosol particles when estimating the phase function15.16
5. CONCLUSIONS
An original method has been proposed to retrieve simultaneously aerosol load and surface reflectance, usingthe SEVIRI observations. The algorithm has proved to be capable of separating the aerosol contribution fromthe surface one in the overall TOA observed signal. The method that has been put in place, based on OptimalEstimation, insures a rigorous control of the system errors, and allows a quality check on the retrieved information.Comparisons with AERONET time series over a period of two months, and at 56 locations around the Earthdisk covered by the SEVIRI instrument have demonstrated the robustness of the algorithm. They have alsoshown the stabilisation effect on the surface properties of improving the a priori knowledge on the surfacestate variables, and the resulting improvement of the AOD retrieved values when compared to AERONET. Theanalysis of the autocorrelation matrices has shown the strong coupling between surface and aerosol contributions,and the compensation mechanism that occurs between the two. This demonstrates the necessity of retrievingsimultaneously the surface reflectance and the aerosol load. The retrieval error has shown to decrease significantlywhen the a priori information on the surface variables is improved and better constrained. The benefits in timeof updating the prior can be summarised by the decreasing trend of the RMSE. These comparisons have alsoshown the limitations of our aerosol class definitions in the case of large particles. The use of specific phasefunctions for non-spherical aerosols will be part of future work on the LDA algorithm.
REFERENCES1. K. R. Knapp, R. Frouin, S. Kondragunta, and A. Prados, “Toward aerosol optical depth retrievals over
land from goes visible radiances: determining surface reflectance,” International Journal of Remote Sens-ing 26(18), pp. 4097–4116, 2005.
2. L. A. Remer, Y. Kaufman, D. Tanr’e, S. Mattoo, D. Chu, J. Martins, R.-R. Li, C. Ichoku, R. Levy,R. Kleidman, T. Eck, E. Vermote, and B. Holben, “The MODIS aerosol algorithm, products and validation,”Journal of the Atmospheric Sciences 62, pp. 947–973, 2005.
3. R. Levy, L. Remer, S. Mattoo, E. Vermote, and Y. Kaufman, “Second-generation operational algorithm:Retrieval of aerosol properties over land from inversion of moderate resolution imaging spectroradiometerspectral reflectance,” Journal of Geophysical Research 112(D13211), p. doi:10.1029/2006JD007811, 2007.
4. D. J. Diner, J. V. Martonchik, R. A. Kahn, B. Pinty, N. Gobron, D. L. Nelson, and B. N. Holben, “Usingangular and spectral shape similarity constraints to improve misr aerosol and surface retrievals over land,”Remote Sensing of Environment 94(2), pp. 155–171, 2005.
5. Rodgers C.D., Inverse Methods for Atmospheric Sounding. Theory and Practice., vol. Vol 2 of Series onAtmospheric, Oceanic, and Planetary Physics, World Scientific, 2000.
6. Tarantola A., Inverse Problem Theory. Methods for data fitting and model parameter estimation, Elsevier,3rd edition ed., 1998.
7. B. Pinty, F. Roveda, M. M. Verstraete, N. Gobron, Y. Govaerts, J. V. Martonchik, D. J. Diner, and R. A.Kahn, “Surface albedo retrieval from meteosat - 1. theory,” Journal of Geophysical Research 105(D14),pp. 18099–18112, 2000.
8. H. Rahman, B. Pinty, and M. M. Verstraete, “Coupled surface-atmosphere reflectance (CSAR) model. 2:Semiempirical surface model usable with NOAA advanced very high resolution radiometer data,” Journalof Geophysical Research 98, pp. 20791–20801, Nov. 1993.
9. E. Shettle, “Optical and radiative properties of a desert aerosol model,” in Radiation in the Atmosphere,A. Deepak, ed., pp. 74–77, (Perugia, Italy), 1984.
10. L. A. Remer, Y. Kaufman, and B. Holben, “The size distribution of ambient aerosol particles: Smoke versusurban/industrial aerosol,” in Global Biomass Burning, pp. 519–530, MIT Press, Cambridge, 1996.
11. E. Shettle and R. Fenn, “Models for the aerosols of the lower atmosphere and the effects of humidityvariations on their optical properties,” Tech. Rep. AFGL-TR-79-0214, U.S. Air Force Geophysics Laboratory,U.S. Air Force Geophysics Laboratory, 1979 1979.
12. A. Lattanzio, Y. Govaerts, and B. Pinty, “Consistency of surface anisotropy characterization with Meteosatobservations,” Advanced Space Research , p. doi:10.1016/j.asr.2006.02.049, 2006.
13. Y. Govaerts and A. Lattanzio, “Retrieval error estimation of surface albedo derived from geostationarylarge band satellite observations: Application to Meteosat-2 and -7 data,” Journal of Geophysical Re-search 112(D05102), p. doi:10.1029/2006JD007313, 2007.
14. B. N. Holben, T. F. Eck, I. Slutsker, D. Tanre, J. P. Buis, A. Setzer, E. Vermote, J. A. Reagan, Y. J.Kaufman, T. Nakajima, F. Lavenu, I. Jankowiak, and A. Smirnov, “Aeronet - a federated instrumentnetwork and data archive for aerosol characterization,” Remote Sensing of Environment 66(1), 1998.
15. J. Wang, X. Liu, S. A. Christopher, J. S. Reid, E. Reid, and H. Maring, “The effects of non-sphericityon geostationary satellite retrievals of dust aerosols,” Geophysical Research Letters 30(24), pp. 2293–2293,2003.
16. O. Dubovik, A. Sinyuk, T. Lapyonok, B. N. Holben, M. Mishchenko, P. Yang, T. F. Eck, H. Volten,O. Munoz, B. Veihelmann, W. J. van der Zande, J. F. Leon, M. Sorokin, and I. Slutsker, “Application ofspheroid models to account for aerosol particle nonsphericity in remote sensing of desert dust - art. no.d11208,” Journal of Geophysical Research-Atmospheres 111(D11), pp. 11208–11208, 2006.
