Operational Retrieval of Atmospheric Temperature, Moisture ...€¦ · radiosondes of atmospheric temperature, moisture, and ozone proﬁles. Evaluation of retrieved total precipitable

1072 VOLUME 42J O U R N A L O F A P P L I E D M E T E O R O L O G Y

q 2003 American Meteorological Society

Operational Retrieval of Atmospheric Temperature, Moisture, and Ozone from MODISInfrared Radiances

SUZANNE W. SEEMANN AND JUN LI

Cooperative Institute for Meteorological Satellite Studies, University of Wisconsin—Madison, Madison, Wisconsin

W. PAUL MENZEL

Office of Research and Applications, NOAA/NESDIS, Madison, Wisconsin

LIAM E. GUMLEY

Cooperative Institute for Meteorological Satellite Studies, University of Wisconsin—Madison, Madison, Wisconsin

(Manuscript received 14 June 2002, in final form 31 January 2003)

ABSTRACT

The algorithm for operational retrieval of atmospheric temperature and moisture distribution, total columnozone, and surface skin temperature from the Moderate Resolution Imaging Spectroradiometer (MODIS)longwave infrared radiances is presented. The retrieval algorithm uses clear-sky radiances measured by MODISover land and ocean for both day and night. The algorithm employs a statistical retrieval with an option fora subsequent nonlinear physical retrieval. The synthetic regression coefficients for the statistical retrieval arederived using a fast radiative transfer model with atmospheric characteristics taken from a dataset of globalradiosondes of atmospheric temperature, moisture, and ozone profiles. Evaluation of retrieved total precipitablewater vapor (TPW) is performed by a comparison with retrievals from the Geostationary Operational En-vironmental Satellite (GOES) sounder, radiosonde observations, and data from ground-based instrumentationat the Atmospheric Radiation Measurement (ARM) Program Cloud and Radiation Test Bed (CART) inOklahoma. Comparisons over one and one-half years show that the operational regression-based MODISTPW agrees with the microwave radiometer (MWR) TPW at the ARM CART site in Oklahoma with an rmseof 4.1 mm. For moist cases, the physical retrieval improves the retrieval performance. For dry atmospheres(TPW less than 17 mm), both physical and regression-based retrievals from MODIS radiances tend to over-estimate the moisture by 3.7 mm on average. Global maps of MODIS atmospheric-retrieved products arecompared with the Special Sensor Microwave Imager (SSM/I) moisture and Total Ozone Mapping Spectrom-eter (TOMS) ozone products. MODIS retrievals of temperature, moisture, and ozone are in general agreementwith the gradients and distributions from the other satellites, and MODIS depicts more detailed structure withits improved spatial resolution.

1. Introduction

The development of global climate and weather mod-els requires accurate monitoring of atmospheric tem-perature and moisture, as well as trace gases and aero-sols. Until recently, continuous monitoring of changesin these parameters on a global scale has been difficult.The Moderate Resolution Imaging Spectroradiometer(MODIS) offers a new opportunity to improve globalmonitoring of temperature, moisture, and ozone distri-butions and changes therein. MODIS was launched on-board the National Aeronautics and Space Administra-tion (NASA) Earth Observing System (EOS) Terra and

Corresponding author address: Suzanne Wetzel Seemann, CIMSS/SSEC, University of Wisconsin—Madison, Madison, WI 53706.E-mail: [email protected]

Aqua platforms on 18 December 1999 and 4 May 2002,respectively. The instrument is a scanning spectrora-diometer with 36 visible (VIS), near-infrared (NIR), andinfrared (IR) spectral bands between 0.645 and 14.235mm (King et al. 1992). For reference, Table 1 lists theMODIS infrared spectral bands. Figure 1 shows thespectral responses of the MODIS infrared bands relativeto an atmospheric emission spectrum computed by aline-by-line (LBL) radiative transfer model (RTM) forthe U.S. Standard Atmosphere, 1976.

The wide spectral range, high spatial resolution, andnear-daily global coverage of MODIS enable it to ob-serve the earth’s atmosphere and continuously monitorchanges. MODIS retrievals of atmospheric water vaporand temperature distributions are intended to advanceunderstanding of the role played by energy and watercycle processes in determining the earth’s weather and

AUGUST 2003 1073S E E M A N N E T A L .

TABLE 1. MODIS infrared band specifications: bandwidth (mm), spectral radiance values (W m22 sr21 mm21), reference temperature (K), andthe noise-equivalent temperature difference NEDT (K). Information was obtained online at http://modis.gsfc.nasa.gov/about/specs.html.

Primary use BandBandwidth

(mm)

Spectral radiance(and reference

temperature) (K)Required

NEDT (K)

Surface temperature 20212223

3.660–3.8403.929–3.9893.929–3.9894.020–4.080

0.45 (300)2.38 (335)0.67 (300)0.79 (300)

0.052.000.070.07

Temperature profile 2425

4.433–4.4984.482–4.549

0.17 (250)0.59 (275)

0.250.25

Cirrus clouds/water vapor 272829

6.535–6.8957.175–7.4758.400–8.700

1.16 (240)2.18 (250)9.58 (300)

0.250.250.05

Ozone 30 9.580–9.880 3.69 (250) 0.25Surface temperature 31

3210.780–11.28011.770–12.270

9.55 (300)8.94 (300)

0.050.05

Temperature profile 33343536

13.185–13.48513.485–13.78513.785–14.08514.085–14.385

4.52 (260)3.76 (250)3.11 (240)2.08 (220)

0.250.250.250.35

FIG. 1. MODIS infrared spectral response functions (numbered by MODIS band) and nadir-viewing brightnesstemperature spectrum of the U.S. Standard Atmosphere, 1976 computed by LBL RTM.

climate. MODIS temperature and moisture products canbe used together with other satellite measurements innumerical weather prediction models in the regionswhere conventional meteorological observations aresparse. Determination and validation of various land,

ocean, and atmospheric products, such as sea surfacetemperature, land surface temperature, and ocean aero-sol properties, require temperature and moisture pro-files, as well as total ozone at MODIS spatial resolutionas ancillary input.


The advantage of MODIS for retrieving the distri-bution of atmospheric temperature and moisture is itscombination of shortwave and longwave infrared spec-tral bands (3–14.5 mm) that are useful for soundingand its high spatial resolution that is suitable for im-aging (1 km at nadir). The increased spatial resolutionof MODIS measurements delineates horizontal gradi-ents of moisture, temperature, and atmospheric totalozone better than companion instruments, such as theGeostationary Operational Environmental Satellite(GOES) sounder [10-km resolution for single field-of-view (FOV) retrievals], Advanced Microwave Sound-ing Unit (AMSU; 45-km resolution)-A, High-Resolu-tion Infrared Radiation Sounder (HIRS; 19-km reso-lution), and Atmospheric Infrared Sounder (AIRS; 15-km resolution). However, the MODIS broadbandspectral resolution provides only modest informationcontent regarding vertical profiles. True sounder ra-diances with higher spectral resolution, such as AIRSand the Cross-Track Infrared Sounder (CrIS) of theNational Polar-Orbiting Environmental Satellite Sys-tem (NPOESS), contain more information about theatmospheric vertical distribution of temperature andmoisture. Because of the limited spectral resolution ofMODIS, the strength of its retrieved products lies inthe resolution of horizontal gradients and the distri-bution of retrieved quantities in integrated vertical lay-ers (as opposed to vertical profiles).

This paper details the operational MODIS algorithmfor retrieving atmospheric temperature and moisture dis-tributions, total column ozone burden, and integratedtotal precipitable water vapor. This algorithm is alsoreferred to as the MOD07 algorithm. The retrievals areperformed using clear-sky radiances measured by MOD-IS within a 5 3 5 field of view (approximately 5-kmresolution) over land and ocean for both day and night.The MOD07 algorithm is operational at the GoddardDistributed Active Archive Center (GDAAC) process-ing system. A version of the algorithm is also availablefor local processing of MODIS data, such as that re-ceived from a direct broadcast antenna, with the Inter-national MODIS/AIRS Processing Package (IMAPP)developed at the Space Science and Engineering Center(SSEC) at the University of Wisconsin—Madison (in-formation available online at http://cimss.ssec.wisc.edu/;gumley/IMAPP/IMAPP.html).

Section 2 describes the retrieval algorithm. Section 3outlines the implementation of the retrieval, includingcloud detection, radiance bias adjustments, the trainingdataset, and a technique for eliminating the IR short-wave surface emissivity uncertainty in the retrievals.Section 4 presents some evaluation of MODIS atmo-spheric products. A discussion of retrieval errors andother issues affecting the atmospheric retrievals is insection 5. Section 6 offers some conclusions and plansfor future work.

2. Algorithm development

The MODIS atmospheric temperature, moisture, andozone retrieval algorithm is a statistical synthetic re-gression with the option for a subsequent nonlinearphysical retrieval. The retrieval procedure involves lin-earization of the radiative transfer model and inversionof radiance measurements. To derive the statistical syn-thetic regression coefficients, MODIS infrared band ra-diances are calculated from radiosonde observations ofthe atmospheric state, generating an ensemble of com-puted radiances with associated observed atmosphericprofiles. The radiative transfer calculation of the MODISspectral band radiances is performed using a transmit-tance model called Pressure-Layer Fast Algorithm forAtmospheric Transmittances (PFAAST) (Hannon et al.1996; Eyre and Woolf 1988); this model has 101 pres-sure-level vertical coordinates from 0.05 to 1100 hPa.The fast transmittance model uses LBL RTM calcula-tions obtained with the LBL RTM, version 6.01, programand the high-resolution transmission molecular absorp-tion spectroscopic database HITRAN 2000 (Rothman etal. 1998). The calculations take into account the satellitezenith angle, absorption by well-mixed gases [includingnitrogen, oxygen, and carbon dioxide (CO2)], water va-por (including the water vapor continuum), and ozone.The retrieval algorithms are developed in this section.

a. Radiative transfer model and its linearization

If scattering by the atmosphere is neglected, the trueclear radiance exiting the earth–atmosphere system fora given MODIS IR band with center wavenumber y isapproximated by

ps

R 5 «B t 2 B dt(0, p)y s s E0

ps

1 (1 2 «) B dt* 1 R9, (1)E0

where Ry is the spectral radiance with wavenumber y,t(0, p) is the total transmittance from the top of theatmosphere to the atmospheric pressure p, « is the sur-face emissivity, B is the Planck radiance, which is afunction of pressure p, subscript s denotes surface, t*5 /t is the downwelling transmittance, and R9 rep-2ts

resents the contribution of reflected solar radiation inthe infrared region. The reflected infrared solar radiationis considered negligible for bands with wavelengths lon-ger than 4.0 mm during the day.

If the MODIS-observed radiance (m denotesmRy

MODIS) of each band is known, then can be con-mRy

sidered a nonlinear function of the atmospheric prop-erties, including the temperature profile, water vapormixing ratio profile, ozone mixing ratio profile, surfaceskin temperature, and surface emissivity. That is, 5mRy

Ry (T, q, O3, Ts, «, . . . . . .) 1 sy , where sy is the in-strument noise and other sources of error, or, in general,


mR 5 R(X) 1 s, (2)

where the vector X contains L (levels of atmosphere)atmospheric temperatures (T), L atmospheric water vapormixing ratios (q), L atmospheric ozone mixing ratios (O3)(the water vapor or ozone is expressed as the logarithmof the mixing ratio in practical applications), one surfaceskin temperature, two infrared surface emissivities at 909and 2500 wavenumbers, and Rm contains N (number ofMODIS spectral bands used) observed radiances.

To linearize Eq. (1), the first-order variations of thePlanck and spectral radiances dB 5 (]B/]T)dT and dR5 (]R/]Tb)dTb are used, where Tb is the brightnesstemperature. The linearized form of Eq. (1) is

p ps s

dTb 5 W dT 1 W dT dp 1 W d lnq dpy T s E T E qs

0 0

ps

1 W d lnO dp, (3)E o 33

0

where Tby is the brightness temperature for the MODISIR spectral band with center wavenumber y; WT, Wq, andW are the weighting functions (sensitivity functions) ofo3

the atmospheric temperature profile, water vapor mixingratio profile, and ozone mixing ratio profile, respectively.The weighting functions can be calculated efficiently froma given atmospheric state (Li 1994; Li et al. 2000):

W 5 b t «, (4a)T s ss

]t ]t*W (p) 5 2b 1 b(1 2 «) , (4b)T ]p ]p

p ps s] lnt ] lnt]T ]Tq qW (p) 5 (T 2 T )«t b 2 2(1 2 «) bt* dp 1 b[t 1 (1 2 «)t*] dp , (4c)q s a s s E E5 6[ ]]p ]p ]p ]p0 p

p ps s] lnt ] lnt]T ]To o3 3W (p) 5 (T 2 T )«t b 2 2(1 2 «) bt* dp 1 b[t 1 (1 2 «)t*] dp , (4d)o s a s s E E3 5 6[ ]]p ]p ]p ]p0 p

where b(p) 5 ]B/]T/]R/]Tb and tq and t are the watero3

vapor and ozone component transmittance functions, re-spectively. Figure 2a shows the temperature-weightingfunctions for the MODIS IR bands 20–25 and bands27–36 calculated for a U.S. Standard Atmosphere, 1976.Bands 27 and 28 are water vapor absorption bands.However, these bands also provide information aboutthe atmospheric temperature if there is enough moisturein the atmosphere, and provided that the water vaporprofile is known or estimated accurately. Bands 33–36are CO2 absorption bands that also provide atmospherictemperature information; they can be used to infer thecloud properties with a known atmospheric temperatureprofile (Frey et al. 1999; Li et al. 2001). The moisture-weighting functions in Fig. 2b show that bands 27 and28 provide information about the distribution of mois-ture in the troposphere, and that the window bands 29,31, and 32 also provide some moisture information dueto weak water vapor absorption. The weighting func-tions in Figures 2a and 2b were computed without anycontrast between surface air temperature and surfaceskin temperature. Under these conditions, MODIS haslimited skill in retrieving boundary layer temperatureand moisture information. However, if contrast existsbetween the surface air temperature and surface skintemperature, information about the boundary layermoisture structure can be retrieved with more success.A skin temperature of 5 K warmer than the surface air

temperature was used in computing the moisture-weighting functions in Fig. 2c, where increased sensi-tivity to moisture can be seen near the surface, partic-ularly in the window channels (bands 29, 31, and 32).

b. Statistical synthetic regression retrieval processing

A computationally efficient method for determiningthe distribution of atmospheric temperature, moisture,and ozone from satellite sounding measurements usespreviously determined statistical relationships betweenobserved or modeled radiances with corresponding at-mospheric profiles. In the regression procedure, tem-perature and moisture are regressed together against ra-diances from CO2, water vapor, and window bands. Thismethod is often used to generate a first guess for aphysical retrieval algorithm, as is done in the Interna-tional Television and Infrared Observation Satellite(TIROS) Operational Vertical Sounder (TOVS) Pro-cessing Package (ITPP; Smith et al. 1993). The statis-tical regression algorithm for atmospheric temperatureis described in detail in Smith et al. (1970), and is sum-marized below for cloud-free skies.

The general inverse solution of Eq. (2) for the at-mospheric profile can be written as (Smith et al. 1970)

X(i, k) 5 A(i, n)Y(n, k), (5)

where i is the number of parameters to be retrieved, kis the number of profiles and surface data in the training


FIG. 2. MODIS-weighting functions based on a U.S. Standard At-mosphere, 1976 for the MODIS infrared bands 20–25 and 27–36. (a)Temperature-weighting functions are expressed in []Tb/]T( p)] andscaled by d(lnp). (b) Water vapor mixing ratio weighting functionsare in {]Tb/][lnq( p)]} and scaled by d(ln p). (c) The water vapormixing ratio weighting functions as in (b), except with skin temper-ature 5 K greater than the surface air temperature.

TABLE 2. Predictors and their uncertainty used in the MOD07 re-gression procedure. BT is brightness temperature and units are spec-ified in the table.

Predictor

Noise used inMOD07 algorithmfor Terra MODIS

Postlaunch NEDTaveraged overdetectors (K)

Band-25–24 BT

Band-27 BTBand-28 BTBand-29 BTBand-30 BT

0.75 K

0.75 K0.75 K0.189 K0.75 K

0.163 (band 24)0.086 (band 25)

0.3760.1930.1890.241

Band-31 BTBand-32 BTBand-33 BTBand-34 BTBand-35 BT

0.167 K0.192 K0.75 K0.75 K0.75 K

0.1670.1920.3080.3790.366

Band-36 BTSurface pressurePercentage of landMonthLatitude

1.05 K5 hPa0.01%0.10.18

0.586————

sample, and n is the number of channels and other pre-dictors used in the regression procedure. The statisticalregression algorithm seeks a ‘‘best fit’’ operator matrixA that is computed using least squares methods with alarge sample of atmospheric temperature and moisturesoundings and collocated radiance observations. Mini-mizing the difference between synthetic observationsand the regression model,

]2|AY 2 X| 5 0, (6)

]A

yields

21T TA(i, n) 5 X(i, k)Y (k, n)[Y(n, k)Y (k, n)] , (7)

where (YT Y) is the covariance of the radiance obser-vations and (YT X) is the covariance of the radianceobservations with the atmospheric profile.

In the MODIS regression procedure, the primary pre-dictors [Y in Eq. (5)] are infrared spectral band bright-ness temperatures. The algorithm uses 12 infrared bandswith wavelengths between 4.465 and 14.235 mm (seeTable 2). Quadratic terms of all brightness temperaturepredictors are also used as separate predictors to accountfor the nonlinear relationship of moisture to the MODISradiances. Surface emissivity effects in the shortwavewindow bands are mitigated by regressing against banddifferences [e.g., instead of BT(4.5 mm) and BT(4.4mm), the difference, BT(4.5 mm) 2 BT(4.4 mm) is usedin the regression, where BT represents brightness tem-perature]. See section 3d for more discussion on surfaceemissivity uncertainty. In addition to the primary pre-dictors, estimates of surface pressure, latitude, month,and percentage of land are also used as predictors toimprove the retrieval.

Ideally, the radiance predictors Y would be taken fromactual MODIS measurements and used with time- and


space-collocated radiosonde profiles X to directly derivethe regression coefficients A. In such an approach, theregression relationship would not involve any radiativetransfer calculations. However, radiosondes are onlyroutinely launched 2 times per day at 0000 and 1200UTC simultaneously around the earth; Terra passes oc-cur at roughly 1100 and 2300 LST each day. It is, there-fore, not possible to obtain many time- and space-col-located radiosondes and MODIS radiances that are glob-ally distributed at a wide range of locations. Alterna-tively, synthetic regression coefficients can be generatedfrom MODIS radiances calculated using a transmittancemodel with profile input from a global temperature andmoisture radiosonde database. The advantage of thisapproach is that it does not need MODIS radiances col-located in time and space with atmospheric profile data;it requires only historical profile observations. However,it involves the radiative transfer calculations and an ac-curate forward model in order to obtain a reliable re-gression relationship. Any uncertainties (e.g., a bias ofthe forward model) in the radiative calculations willinfluence the retrieval. To address model uncertainties,radiance bias adjustments have been implemented in theretrieval algorithm, as discussed in section 3b.

In the MODIS retrieval algorithm, the National Oce-anic and Atmospheric Administration (NOAA)-88bglobal dataset of radiosonde observations is used fortraining the regression. The NOAA-88b dataset waschosen for this purpose to build on previous experiencewith the sounding algorithm from the Advanced TOVS(ATOVS; Li et al. 2000). See section 3c for more in-formation about the NOAA-88b dataset. Other radio-sonde datasets such as the Thermodynamic Initial GuessRetrieval (TIGR)-3 are also being tested for possibleoperational implementation. The original NOAA-88bdata contains 7547 globally distributed clear-sky radio-sonde profiles of temperature, moisture, and ozone,along with observations of surface temperature andpressure. Additional radiosonde observations have beenadded to this dataset for the MODIS retrieval algorithm,as described in section 3c. The radiative transfer cal-culation of the MODIS spectral band radiances is per-formed with the PFAAST model for each profile fromthe training dataset to produce a temperature–moisture–ozone profile–MODIS radiance pair. The synthetic re-gression coefficients [see Eq. (7)] are generated usingthe calculated radiances and the matching atmosphericprofile. For each training profile, 680 coefficients aregenerated, corresponding to different local zenith anglesfrom nadir to 658, with an increment of approximately0.18.

The estimated MODIS instrument noise was addedto the calculated spectral band radiances before creatingthe coefficients. The noise was randomly generated witha Gaussian distribution, a standard deviation equal tothe NEDT in Table 2, and an average of zero. The noiseused in the algorithm is larger than estimates of post-launch detector noise to account for variability between

the 10 detectors (striping). The correlation in the noisebetween the spectral bands was not considered in theregression, and it is assumed that any impact of spectralnoise correlation on the retrievals should be small. Thisimpact will be further studied in future work.

The predictands, or the parameters to be retrieved byregression, include the temperature profile, logarithm ofthe water vapor mixing ratio profile, logarithm of theozone mixing ratio, surface skin temperature, and long-wave and shortwave infrared surface emissivities. Toperform the retrieval, Eq. (5) is applied to the actualMODIS measurements, where Y is now the observedMODIS radiances. Integration of the retrieved profilesyields the total precipitable water (TPW) or total columnozone. Other atmospheric parameters, such as layer tem-perature and moisture and stability indices, can also becalculated from the predictands. The retrieved water va-por mixing ratio at each pressure level is checked forsaturation and the mixing ratio at any level with relativehumidity greater than 100% is set equal to the saturationmixing ratio at that level. In addition to integrating re-trieved mixing ratio profiles to obtain TPW, TPW is alsoretrieved directly by integrating the training data profilesand including TPW as a predictand.

To analyze the information contribution from variouspredictors to the retrieval, a simulation study was per-formed. In the simulation, 90% of the NOAA-88b pro-file data samples were used for training, while radiancescalculated from the remaining 10% of the profiles wereused for testing the retrieval algorithm. The noise addedin the simulation is the same as the noise used in theoperational algorithm for Terra (see Table 2). Figure 3shows the root-mean-square error (rmse) of the retrievedvertical profiles of temperature and moisture comparedwith the actual profiles for the 844 independent regres-sion retrievals. The most accurate MODIS temperatureretrievals are between 800–400 hPa, where the rmse isapproximately 1 K. Near the surface, rmse increases to2 K. Moisture retrieval accuracy decreases with heightfrom an rmse maximum of 1.5 g kg21 at the lowestlevels. The rmse of the ozone profiles reaches a maxi-mum of 0.75 ppmv at the highest levels.

The impact of various predictors on the total precip-itable water retrieval was also investigated using retriev-als from the simulated data. Figure 4 shows comparisonsbetween actual TPW (integrated from the 844 indepen-dent test profiles) and TPW retrieved from radiancescalculated from the same profiles. For the simulationresults in Fig. 4a, the same NEDT and predictors wereused as in the operational Terra MODIS algorithm. Cal-culating the coefficients without adding artificial noiseto the calculated radiances (Fig. 4b) leads to improve-ment of the rmse from 4.15 mm with added noise to2.9 mm without. However, for retrievals with true MOD-IS radiances, it is necessary to add artificial noise to theradiances calculated in the synthetic coefficients to ac-count for striping and other instrument noise. Removingthe quadratic terms (individual brightness temperatures


FIG. 3. Rmse of actual compared with retrieved profiles of (a) temperature (K), (b) mixing ratio(g kg21), and (c) ozone (ppmv). Profiles were taken from an independent sample of 10% of theNOAA-88b data (844 profiles) and retrievals were performed on radiances computed from eachprofile.

squared) from the predictor list shows an increase in thermse to 3.2 mm (not shown) from 2.9 mm, computedwith no added noise.

c. Nonlinear physical retrieval processing

The statistical regression algorithm has the advantageof computational efficiency, numerical stability, andsimplicity. However, it does not account for the physicalproperties of the radiative transfer equation (RTE). Aftercomputing atmospheric profiles from the regressiontechnique, a nonlinear iterative physical algorithm (Liet al. 2000) applied to the RTE often improves the so-lution. The physical retrieval approach is described inthis section.

The physical procedure is based on the regularizationmethod (Li et al. 2000), wherein a penalty function de-fined by

p m p 2 p p 2R(X ) 5 \R 2 R(X )\ 1 g \X 2 X \0 (8)

is minimized to improve the fit of the MODIS spectralband measurements to the regression first guess. In Eq.(8), Xp is the atmospheric profile to be retrieved, ispX0

the initial state of the atmospheric profile or the firstguess from regression, Rm is the vector of the observedMODIS brightness temperatures used in the retrievalprocess, R(Xp) is the vector of calculated MODISbrightness temperatures from an atmospheric state (Xp),and g is the regularization parameter that can be de-

termined by the discrepancy principle (Li and Huang1999; Li et al. 2000). The solution provides a balancebetween MODIS spectral band radiances and the firstguess.

A comparison between the regression-based first-guess brightness temperatures and the physical retrievalbrightness temperatures is presented in Fig. 5. The re-sults are based on simulated data similar to that detailedin section 2b for Figs. 3 and 4, except that retrievalswere performed on only the 394 moist profiles withTPW greater than 17 mm. Only moist cases were usedin this comparison because the current physical retrievalalgorithm has very little effect on dry cases; the mois-ture signal is weak in the MODIS IR radiance mea-surements for a dry atmosphere. The rmse of actualminus retrieved brightness temperature decreased fromthe regression-based guess to the physical retrieval. Themost significant improvement is apparent in the watervapor bands 27 and 28, with the rmse decreasing from0.8 to 0.2 K for band 27 (6.7 mm), and from 0.6 to 0.2K for band 28 (7.3 mm). The improvements evident inthe brightness temperature calculations using the phys-ical algorithm are also apparent in the retrieved prod-ucts. For the same 844 simulated radiances as in Fig.4, the rmse of the actual TPW minus that retrieved usingthe physical retrieval was 3.6 mm with added noise and2.5 mm without, compared with 4.15 and 2.9 mm forthe regression retrievals with and without noise, re-spectively.


FIG. 4. Comparison of retrieved TPW (mm) with actual TPW froman independent sample of 10% of the NOAA-88b profile data (844profiles) for the MODIS algorithm (a) with noise added to the ra-diances calculated for the coefficients and (b) without noise. Rmsefor the retrievals with noise is 4.3 mm and it is 2.9 mm without noise.The solid line in each figure shows a 1-to-1 correspondence betweenretrieved and actual TPW.

FIG. 5. Rmse (K) for MODIS bands 24, 25, 27–29, and 31–36 ofobserved minus synthetic regression retrieval brightness temperatures(solid lines) and observed minus physical retrieval brightness tem-peratures (dashed lines). Brightness temperatures were computedfrom profile retrievals of the same simulated radiances as in Figs. 3and 4, except using only moist cases with TPW greater than 17 mm.

In summary, a nonlinear physical retrieval showssome improvement over a synthetic regression retrievalfor MODIS retrievals. However, the physical retrievalrequires more computation time and, like the syntheticregression used in the MODIS processing, is dependenton the accuracy of the radiative transfer model.

3. Operational implementation

The operational MODIS retrieval algorithm consistsof several procedures that include cloud detection, av-eraging clear radiances from 5 3 5 FOV areas, biasadjustment of MODIS brightness temperatures to ac-count for forward model and instrument errors, synthetic

regression retrieval, and an option to perform a physicalretrieval. Because of computer limitations, the MODISMOD07 retrieval algorithm that is operational at theGDAAC processing system includes only the syntheticregression retrieval. A version of the algorithm with thephysical retrieval will be available for MODIS directbroadcast processing through IMAPP.

a. Cloud detection algorithm

MODIS atmospheric and surface parameter retrievalsrequire clear-sky measurements. The operational MOD-IS cloud-mask algorithm (Ackerman et al. 1998) is usedto identify pixels that are cloud free. The MODIS cloud-mask algorithm determines if a given pixel is clear bycombining the results of several spectral threshold tests.A confidence level of clear sky for each pixel is esti-mated based on a comparison between observed radi-ances and specified thresholds. The MODIS atmosphericretrieval algorithm requires that at least 5 of the 25pixels in a 5 3 5 field-of-view area have a 95% orgreater confidence of clear by the cloud mask. The re-trieval for each 5 3 5 field-of-view area is performedusing the average radiance of only those pixels that wereconsidered clear. Experimentation with MODIS retriev-als for different requirements for the minimum numberof clear pixels showed that requiring that 5 of 25 pixelsbe clear maximized retrieval coverage, while minimiz-ing undetected cloud contamination. A more strict re-quirement would further reduce cloud contamination ef-fects but also lead to fewer retrievals. Since the decisionto perform a retrieval depends on the validity of thecloud-mask algorithm, cloud contamination may occurif the cloud mask fails to detect a cloud, or the retrievalmay not be processed if the cloud-mask falsely identifiesa cloud.

b. Radiance bias adjustment in the retrievalprocessing

The forward model–calculated radiances have biaseswith respect to the MODIS-measured radiances. Thereare several possible causes, including calibration errors,


FIG. 6. Average observed minus calculated brightness temperaturebiases for Terra MODIS IR bands 24, 25, 27–29, and 31–36 from72 clear-sky cases at the SGP ARM CART site from Apr 2001 toAug 2002. Biases for radiance calculations using skin temperatureobserved by the IRT (black bars). Synthetic regression-derived skintemperature was used for the calculated radiances in the biases shownas gray bars. No bias is shown for band 30 because of insufficientozone observations for input to the forward model.

FIG. 7. Comparison of skin temperature (K) computed by MODISsynthetic regression (y axis) with that observed by the SGP CARTIRT (x axis) for the same cases used in Fig. 6. The dotted line showsa 1-to-1 correspondence.

spectral response uncertainty, temperature and moistureprofile inaccuracies, and forward model error. The syn-thetic regression and physical retrieval methods use bothmeasured and calculated radiances and thus require thatthis bias be minimized. Some of the techniques devel-oped for computing GOES sounder radiance biases withrespect to the forward model (Hayden 1988) were adapt-ed to the MODIS atmospheric temperature, moisture,and ozone retrieval algorithm. Bias adjustment for ra-diative transfer calculation of MODIS spectral band ra-diances is demonstrated to have a positive impact onthe atmospheric product retrievals.

MODIS radiance bias calculations are routinely com-puted for the Southern Great Plains (SGP) AtmosphericRadiation Measurement (ARM) Cloud and RadiationTest Bed (CART) for clear scenes with MODIS sensorzenith angle less than 358. Observed MODIS radiances,averaged from a 5 3 5 field-of-view area, are comparedwith those computed using the 101-level PFAAST mod-el, with temperature and moisture profile input fromNational Center for Environmental Prediction (NCEP)’sGlobal Data Analysis System (GDAS) global analysisdata. Skin temperature and emissivity are estimated byregression with MODIS radiances. To establish credi-bility for the synthetic regression-derived skin temper-ature input, actual observed skin temperature from theSGP ARM CART ground-based downward-looking in-frared thermometer (IRT) measuring the radiating tem-perature of the ground surface was also used. Figure 6shows that, on average over 72 clear-sky day and nightcases from April 2001 to August 2002, the observedminus calculated radiance biases computed using theregression-based skin temperature differ very little fromthose computed using the IRT skin temperature. A com-parison of the two skin temperatures for the same casesused in computing the biases shows good agreement(Fig. 7); the rmse of the IRT skin temperature comparedwith MODIS regression-based skin temperature is 1.75K, and the slope of a linear best fit is 1. Most CARTsite MODIS radiance biases (expressed as observed mi-nus calculated BT) shown in Fig. 6 are positive, indi-cating that, on average, the observed MODIS brightness

temperatures are warmer than those predicted by themodel. Case-to-case variability for each spectral bandcan be seen in Fig. 8. The high variability in the radiancebias for band 27 is partially due to the significant de-tector-to-detector differences, or striping, that exists inthe radiance data for this band. Some of the variabilityin this band is also related to differences in the atmo-spheric moisture; band-27 radiance bias calculationsvary as a function of TPW. For this reason, a differentband-27 radiance bias correction is applied to moistscenes (TPW . 17 mm) than to dry scenes (TPW .17 mm).

A comparison of MODIS products at the SGP ARMCART site with and without the radiance bias correc-tions confirms a significant improvement with the biascorrections. The rmse for TPW sensed by the CARTsite microwave radiometer (MWR) compared withMODIS TPW for 80 clear-sky cases between April 2001and October 2002 decreased from 5.9 to 4.1 mm whenthe bias corrections were applied (not shown). The im-provements were primarily apparent for moist cases; theaverage MWR minus MODIS TPW difference for the43 cases with TPW greater than 17 mm was 5.6 mmwithout bias corrections, as compared with 1.24 mmafter correction. Because the MODIS retrieval algorithmis applied globally, the biases computed at the SGPARM CART site are not appropriate for application atother latitudes and for other ecosystem types. Thus, ra-diance bias corrections have been computed for otherregions of the globe; however, they are less well vali-dated.

To compute the global radiance biases, observedMODIS brightness temperatures were compared withcalculated brightness temperatures for clear-sky scenesdistributed globally with a MODIS-viewing zenith angle.308. Calculations of brightness temperatures were per-


FIG. 8. Calculated vs observed brightness temperatures (K) for each MODIS IR band used in the retrieval algorithm. Each dot represents 1 ofthe 72 CART cases used in Figs. 6 and 7. The dotted lines show a 1-to-1 correspondence between calculated and observed BT.

formed as outlined above for the SGP CART site withskin temperature and emissivity estimated from regres-sion of MODIS radiances, and temperature and moistureprofiles taken from NCEP GDAS analysis. Becausethere are known difficulties in retrieving skin temper-ature and emissivity over the desert, these cases wereexcluded from the global averages. The global radiancebiases are separated into six latitudinal zones each forland and ocean. The average global radiance bias cal-culations for northern midlatitude land agree fairly wellwith those calculated at the SGP CART site; the rmseof MODIS minus MWR TPW for 80 cases increasedonly 0.3 mm when using the global radiance bias insteadof the SGP CART biases (not shown).

The radiance bias corrections applied in the opera-tional MODIS atmospheric retrieval algorithm will beupdated regularly to account for adjustments in the in-strument calibration and improvements in the forwardmodel. Future versions of the algorithm will include amore advanced global bias scheme that uses a regressionbased on predictors that characterize the air mass (e.g.,

brightness temperatures, surface skin temperature, TPW,and atmospheric layer thickness) similar to that em-ployed on the TOVS (Eyre 1992; Harris and Kelly2001).

c. Adjustments to the NOAA-88b training dataset

The NOAA-88b dataset contains 7547 globally dis-tributed clear-sky radiosonde observations. Profiles oftemperature, moisture, and ozone and surface data fromthis dataset were used to compute the synthetic regres-sion coefficients for the MODIS statistical retrieval. Thetemperature and mixing ratio profiles are from radio-sondes launched in 1998 at a wide range of locationsdistributed globally; ozone was derived from solar back-scatter ultraviolet (SBUV) collocated profiles. All pro-files were extended to 0.005 hPa and interpolated to 101levels by researchers at NOAA–National EnvironmentalSatellite, Data, and Information Service (NESDIS) whocreated the dataset.

A surface emissivity and skin temperature were as-


FIG. 9. Histogram of (a) observed MODIS 11-mm brightness tem-perature from four daytime granules on 2 Jun 2001 (0830, 0835,1010, and 1150 UTC) over the North African deserts, and (b) com-puted from a forward model calculation using the original NOAA-88b training profiles and surface data as input. Histogram is expressedin number of profiles for each BT11 (K).

FIG. 10. Frequency of occurrence of 11-mm brightness temperaturein the original NOAA-88b training dataset (dashed) and the extendedtraining dataset, including the additional desert radiosondes (solid).

signed to each training profile to be used as input to theforward model because observed surface information isnot included in the NOAA-88b profile dataset. The char-acterization of the surface used in the calculation of thesynthetic coefficients is described below. Infrared emis-sivity spectra were assigned randomly to the trainingprofiles with a Gaussian distribution, standard deviationof 0.05, mean of 0.95 for longwave IR bands (8–13mm), mean of 0.88 for the 6.7-mm band, and mean of0.84 for shortwave IR bands (4–5 mm). The differencebetween the lowest-level radiosonde temperature andthe skin temperature to be assigned to the profile wasalso generated randomly with a Gaussian distribution,but with a mean of 0 K and standard deviation of 10K. To better represent the range of conditions sensedby MODIS, a second skin temperature was assigned toeach training profile in a similar manner. This createda second set of radiance–profile pairs with different skintemperatures, thus, doubling the amount of training dataused in the regression. Future work will improve uponthis choice of skin temperature and emissivity by em-ploying techniques with more physical bases, includingthe use of ecosystem-dependent emissivities (Wilber etal. 1999) and deriving physical relationships betweensurface air and skin temperature.

To limit the retrievals to training data with physicalrelevance to the observed conditions, the NOAA-88bdata were partitioned into seven zones based on the 11-mm brightness temperatures (BT11) calculated from theprofiles. The seven zones are BT11 .245, 245–269,269–285, 285–294, 294–300, 300–310, and .310 K.Any profiles corresponding to BT11 within 2 K of a

zone end point were included in both closest zones.When each retrieval is performed, it uses only that sub-set of the training data in the same zone as the observedMODIS BT11.

After partitioning, there were insufficient trainingdata in the NOAA-88b dataset for the very warm sur-faces (the last two zones, BT11 .300 K). Figure 9ashows that many measurements over the North Africandeserts with BT11 .300 K had little to no training datacorresponding to a similar BT11 (Fig. 9b). To addressthis problem, new radiosonde data from the North Af-rican desert regions for January–December 2001 wereadded to the training dataset. Spread equally throughthe 12 months, 900 radiosonde observations met thecriteria of relative humidity ,90% at each level andphysically reasonable behavior; profiles of temperatureand moisture from these radiosondes were added to theNOAA-88b dataset. A skin temperature was assignedto the new profiles by adding a positive increment tothe temperature of the lowest level in the profile. Theincrement was derived by taking the absolute value ofa random number from a Gaussian distribution withmean of zero and standard deviation of 15 K. Thesewarm skin temperatures were necessary to generatebrightness temperatures in the desert region, represen-tative of conditions observed by MODIS, and to addmore training in the highest BT11 zones. Figure 10compares the frequency of occurrence of the 11-mmbrightness temperatures for both the original and en-hanced NOAA-88b dataset with the extra desert profiles.

Results from a study of retrievals using simulated data(similar to that described in section 2b for Figs. 3 and4) illustrate the benefit of separating the training datainto brightness temperature zones and adding the warmdesert profiles. For the 844 independent cases with nonoise added to the radiances used in the coefficients,the rmse of retrieved TPW compared with actual TPWdecreased to 2.9 mm with BT zones and new profiles(see Fig. 4b) from 4.2 mm without (not shown). Themost improvement was evident for moist cases withTPW .30 mm.

One example of the improvement in retrievals fromMODIS data after partitioning the BT11 into sevenzones and adding desert training data is presented in


FIG. 11. (top) Terra MODIS images from 1735 UTC 20 Aug 2001. (a) True color using MODIS reflectance frombands 1, 4, 3 as red, green, and blue, respectively; (b) MODIS band-31 radiance. Images obtained from Universityof Wisconsin—Madison’s direct broadcast (http://eosdb.ssec.wisc.edu/modisdirect/). (bottom) Comparison of TPW(mm) from the same day retrieved from (c) GOES-8, (d) MODIS with the 11-mm BT zones, and (e) MODIS withoutthe 11-mm BT zones. The MODIS granule began at 1735 UTC and the GOES at 1800 UTC.

FIG. 12. An 11-mm brightness temperature divided into the sameseven zones used in the retrieval algorithm for the same case as shownin Fig. 11. The areas that show the most improvement in the com-parison in Fig. 11 are in the warmest two brightness temperaturezones.

Fig. 11. True color reflectance and 11-mm radiance im-ages from 20 August 2001 are shown in Figs. 11a and11b, respectively. A large area of clear sky exists fromTexas through Oklahoma and southeastern Kansas. Fig-ures 11c–e compare GOES-8 TPW with MODIS TPWretrieved with and without the BT zones and new pro-files. The MODIS results with the zones and new pro-files compare considerably better with the GOES. Thearea in Kansas and Oklahoma that shows the most im-provement has a warm BT11 that falls within the highesttwo classes (see Fig. 12). This is consistent with resultsof other cases; the most significant improvements oc-curred for scenes with BT11 in the two highest classes.

d. Surface emissivity for IR 4.5-mm spectral bands

In most situations the infrared emissivity assigned tothe NOAA-88b training dataset accounts for the globalvariations in emissivity. However, for some regions,such as the deserts of North Africa, the surface emis-


FIG. 13. Measurements of emissivity spectra (%) for two silicatescommonly found in desert regions: cyclosilicates (solid line) andtectosilicates (dashed line). Data used in this figure were obtainedthrough the ASTER Spectral Library at the Jet Propulsion Laboratory,California Institute of Technology, Pasadena, CA (available online athttp://speclib.jpl.nasa.gov).

FIG. 14. Difference between the BT calculated at a given emissivityand that calculated with emissivity equal to 0.4 (BT increment; K)for bands 24 and 25 individually (dash–dot and dash, respectively),and for the difference between bands 25 and 24 (solid line). Cal-culations used a standard U.S. midlatitude summer atmosphere.

sivity has the potential to be significantly lower at 4than at 11 mm (Salisbury and d’Aria 1992). Figure 13shows measurements of emissivity spectra from dataobtained through the NASA Jet Propulsion Laboratory’sspectral library (information available online at http://speclib.jpl.nasa.gov) for two silicates commonly foundin desert regions: cyclosilicates and tectosilicates. Theemissivity at the 4.5-mm spectral region is extremelylow for both minerals, indicating that emissivities in thedesert may be as low as 30%–40%. Because emissivitiesthis low are not included in the training dataset, thesurface characteristics in the 4.4- and 4.5-mm bands(MODIS bands 24 and 25) were not accurately repre-sented and the MODIS retrievals were excessively moistin the desert regions. To remedy this problem, the dif-ference between these two bands is used as a singlepredictor instead of using bands 24 and 25 indepen-dently; this subtraction removes most of the surfaceemissivity signal in the regression equation. The BTdifference between band 25 and band 24 is found to bemuch less sensitive to the surface emissivity change thanthe BT of either of the bands independently. Figure 14illustrates the advantages of this approach by showingthe BT increment, defined as the difference between theBT calculated at a given emissivity and that calculatedwith emissivity equal to 0.4. The BT increment for theindividual bands 24 and 25 is greater than 2.0 K for asurface emissivity of 0.85, while the BT increment forthe difference between them is less than 0.8 K. Takingthe BT difference between band 25 and band 24 insteadof using the bands individually reduces the error in theretrieval due to surface emissivity uncertainties over thedesert region. In addition to the surface emissivity ef-fects on 4.5-mm region, there are also effects of nonlocalthermodynamical equilibrium on that region for the ra-

diative transfer modeling. Taking the difference betweenbands 24 and 25 also minimizes this effect in the re-trieval.

4. Evaluation of MOD07 products

Atmospheric retrievals from MODIS have been com-pared with products from other observing systems atthree spatial scales: (a) a fixed point with ground-basedmeasurements at the SGP ARM CART, (b) the conti-nental scale with GOES sounder products, and (c) theglobal scale with retrievals from the Special Sensor Mi-crowave Imager (SSM/I) and Total Ozone MappingSpectrometer (TOMS).

a. Comparison of MODIS temperature and moisturewith ARM CART observations

Specialized instrumentation at the SGP ARM CARTsite in Oklahoma facilitates comparisons of MODIS at-mospheric products with other observations collocatedin time and space. Terra passes over the SGP CARTdaily between 0415 and 0515 UTC and between 1700and 1800 UTC. Radiosondes are launched 3 times perday at approximately 0530, 1730, and 2330 UTC. Ob-servations of total column moisture are made by theMWR every 40–60 s. An additional comparison is pos-sible with the GOES-8 sounder (Menzel and Purdom1994; Menzel et al. 1998) that retrieves TPW hourly.

MODIS retrieved products were compared for 80clear-sky cases from April 2001 to October 2002. Man-ual inspection of visible and infrared images excludedany scenes with the possibility of cloud contamination.MODIS sensor zenith angle was less than 508 to theCART site for all cases. To reduce noise due to striping,all retrievals in a 3 3 3 retrieval area were averaged inthis comparison. TPW from MODIS synthetic regres-sion retrievals, the GOES-8 sounder, radiosondes, andthe MWR are compared in Fig. 15. MODIS shows gen-eral agreement with the MWR for these cases; however,GOES-8 sounder and radiosonde observations show bet-ter agreement with the MWR. The rmse between MOD-


FIG. 15. Comparison of TPW from MODIS synthetic regression (shaded diamond), GOES-8(circle), and radiosonde (x) with the SGP ARM CART MWR (mm); 80 cases from Apr 2001 toOct 2002 are shown in the comparison. The dotted line shows a 1-to-1 correspondence.

FIG. 16. Comparison of (a) temperature (K) and (b) mixing ratio(g kg21) on 1 Aug 2001 from the average of nine MODIS profilesin a 3 3 3 retrieval area surrounding the SGP ARM CART site at1705 UTC (solid), and a radiosonde launched at 1728 UTC (dashed).In this situation in which the temperature and moisture profiles aresmooth, MODIS captures the vertical structure fairly well.

IS and MWR TPW collocated in time and space is 4.1mm for regression retrievals, as compared with 1.9 and1.2 mm for GOES-8 and radiosondes, respectively. Forthe 43 cases with moist atmospheres (TPW . 17 mm),MODIS compares better to the MWR; the rmse betweenMODIS and MWR TPW is 3.9 mm, as compared with2.5 mm for GOES and MWR TPW (not shown). On

average, for these moist cases, MODIS is 1.2-mm drierthan the MWR. For dry atmospheres, MODIS consis-tently overestimates the total column moisture; MODISis on average 3.7 mm more moist than the MWR forthe 37 dry cases with MWR TPW less than 17 mm.Stephens et al. (1994) observed similar behavior inTOVS total column water vapor and attributed it to alimitation of low spectral resolution infrared soundersin subsidence regions. Use of different band-27 biascorrections for moist and dry cases (see section 3b)improved the retrievals for moist cases but had littleeffect on the dry cases.

An example comparison of temperature and moistureprofiles from a radiosonde at the SGP CART with MOD-IS regression retrievals of temperature and moisture isshown in Fig. 16. For atmospheres with a fairly mono-tonic, smooth temperature and moisture distribution,such as that shown here, MODIS retrievals compare wellto radiosonde observations. However, in situations withisolated layers of sharply changing temperature or mois-ture, MODIS is not able to capture the finer-scale struc-ture. Improved sounding capability is expected from thehigh spectral resolution AIRS on Aqua (Susskind et al.1998).

b. Continental-scale comparisons between MODISand GOES TPW

On the continental scale, MODIS TPW was comparedwith GOES-8 and GOES-10 sounder retrievals of TPW


FIG. 17. Total precipitable water (mm) for 2 Jun 2001 over North America retrieved by (a), (d) MODIS syntheticregression (b), (e) MODIS physical, and (c), (f ) GOES-8 and GOES-10 combined. (top) (a)–(c) Daytime retrievals(four MODIS granules from 1640–1830 UTC; GOES at 1800 UTC) and (bottom) (d)–(f ) nighttime retrievals (fiveMODIS granules from 0435–0625 UTC; GOES at 0600 UTC). The slight discontinuity visible in Oklahoma in theMODIS daytime retrievals occurs where granules from the two subsequent overpasses, separated by 1 h and 40 min,intersect.

over the continental United States and Mexico. GOESTPW has been well validated (Schmit et al. 2002) andhas a resolution at the subsatellite point of 10 km. GOESuses radiances measured from a 3 3 3 field-of-viewarea (approximately 30-km resolution) to retrieve oneatmospheric profile, while MODIS has nadir resolutionof 1 km and uses a 5 3 5 field-of-view area (5-kmresolution). Unlike the MODIS retrieval, GOES hourlyradiance measurements are supplemented with hourlysurface observations of temperature and moisture as ad-ditional information in the GOES retrieval. MODIS andGOES retrieval procedures also use different first-guessprofiles; GOES uses a numerical model forecast, whileMODIS uses the previously described synthetic regres-sion retrieval.

Figure 17 compares MODIS TPW with TPW re-trieved by the GOES-8 and GOES-10 sounders overNorth America for 2 June 2001 during the day and atnight. MODIS and GOES show fairly good agreementexcept the daytime MODIS TPW retrieved by regressionis drier than GOES over Oklahoma, Arkansas, and theGulf of Mexico. TPW retrieved by physical retrievalshows better agreement with GOES in these areas.

c. Global comparisons of MODIS products withSSMI and TOMS

Global TPW from MODIS atmospheric retrievals iscompared with TPW from the Defense MeteorologicalSatellite Program (DMSP) SSM/I (Alishouse 1990; Fer-raro et al. 1996; Wentz 1997) for 22 May 2002 in Fig.18. The SSM/I (resolution 12.5 km) TPW retrievals forclear or cloudy skies over ocean use the 22- and 37-GHz microwave channels. MODIS and SSM/I showsimilar patterns of TPW distribution and similar mag-nitudes, however, MODIS retrievals are somewhat lessmoist over tropical oceans. Some of the differences canbe attributed to MODIS TPW retrievals occurring onlyin clear pixels and, thus, not capturing the moist en-vironment associated with clouds. Other differencesmay be attributed to the time differences between thetwo satellite overpasses.

In order to predict the evolution of ozone on time-scales of a few days to 1 week, reliable measurementsof ozone distribution are needed. The NASA GoddardSpace Flight Center (GSFC) TOMS instrument (Bow-man and Krueger 1985; McPeters et al. 1998) onboardthe Earth Probe (EP) satellite measures backscattered


FIG. 18. (top) MODIS TPW (mm) and (bottom) SSM/I TPW (mm) from the F-14 satellite on 22 May 2002. Retrievalsfrom ascending and descending passes were averaged to obtain these values. The color scale is the same for bothMODIS and SSM/I and is shown below the two images. SSM/I data were obtained through the Remote Sensing SystemsSSM/I Web page at http://www.ssmi.com. MODIS data were degraded to 25-km resolution from the original 5-kmresolution in this figure. White areas in the MODIS image are areas where no retrievals were performed because ofthe presence of clouds. SSM/I retrievals are performed for both clear and cloudy skies, but they are not performedover land.

ultraviolet solar radiation and cannot provide measure-ments at night. High-spatial-resolution IR radiance mea-surements at 9.6 mm (band 30) from MODIS allowozone estimates during both day and night.

MODIS total column ozone retrievals are comparedwith TOMS ozone from 22 May 2002 in Fig. 19. Thegeneral distribution of ozone from TOMS is similar tothat from MODIS. A quantitative comparison for thesame day between TOMS and MODIS ozone (Fig. 20)shows percent errors of less than 10% for latitudes be-

tween 6208. Errors increase toward the Poles, with thehighest errors at the South Pole, where very few trainingdata exist.

5. Discussion

Retrieval accuracy, computation efficiency, and re-trieval validation are important considerations when ap-plying the operational algorithm to real-time MODISdata processing. The accuracy of the retrievals depends


FIG. 19. Total column ozone (Dobson units) for 22 May 2002 for (top) MODIS and (bottom) TOMS. TOMS datawere obtained from the NASA GSFC TOMS Web page at http://toms.gsfc.nasa.gov/ozone/ozone.html. The color scaleis the same for both MODIS and TOMS and is shown below the two images.

FIG. 20. Percent error for MODIS ozone retrievals in comparisonwith TOMS ozone as a function of latitude for the same day as shownin Fig. 19.

on the instrument calibration, navigation, and coregis-tration in the infrared bands; the MODIS is assumed tobe calibrated within the instrument noise, navigatedwithin one FOV, and coregistered within two-tenths ofa FOV. Several sources of errors must be addressed.First, forward model errors can influence the retrievals.These may be due to atmospheric transmittance calcu-lation error because the transmittance is computed by afast regression procedure rather than by an accurate line-by-line model; they may be due to inaccurate or insuf-ficient representation of the atmospheric temperatureand moisture profiles in the training dataset; or they maybe due to the surface uncertainties, such as surface el-


FIG. 21. Aqua MODIS TPW (mm) from 20 Jul 2002 with (a) 1-km retrievals and (b) averaged up to 10 km; 10-km retrievals wereperformed when at least 80% of the pixels are clear.

evation, emissivity, and skin temperature. Improvementof the forward model is important for deriving theMODIS products with a high level of accuracy.

In addition, the MODIS instrument detector noise andcalibration error (observation error) can have an impacton the retrieval accuracy. Large observation errors resultin poor MODIS atmospheric products. Detector-to-de-tector differences in the spectral response functionswithin a band have been shown to produce 0.5%–1.0%differences in radiance measurements in the IR thermalbands. This is seen as detector striping (called bandingby some investigators) within a scene. Well-calibratedradiance measurements without (or with reduced) strip-ing noise are expected to improve the quality of theMODIS atmospheric products presented in this paper.Cloud detection errors may also have a negative impacton the retrieval products. Regions with cloud contam-ination in the MODIS retrievals show inaccurate mois-ture and temperature.

Computation efficiency realized by regression meth-ods enables fast retrieval of atmospheric products. Thephysical retrieval takes more computation time, there-fore, it is not possible to apply both the regression andphysical retrieval procedures to process global MODISdata operationally at the Distributed Active ArchiveCenter (DAAC). Regional MODIS data received innear–real time by direct broadcast MODIS stations canbe processed using both algorithms.

An advantage of MODIS over lower spatial resolutioninfrared sounders is that it can provide good retrievalcoverage under broken cloud conditions and, thus, de-lineate the horizontal gradients of moisture distribution.To illustrate the value of higher resolution, Fig. 21 com-pares MODIS TPW retrievals from the Aqua satellitewith full 1-km resolution to retrievals averaged up to10-km spatial resolution. The reduced-spatial-resolutionretrievals degrade the gradients resolved at 1 km andreduce some of the product coverage. Full-resolutionMODIS 1-km retrievals are important for studying thefuture operational sounder and image systems, such asthe planned geostationary Hyperspectral EnvironmentalSuite (HES) and Advanced Baseline Imager (ABI).

A challenge is to maintain adequate global trainingdata so that the synthetic regression algorithm producesaccurate atmospheric retrievals. A further challenge isto accommodate special regions, such as the Sahara,where the surface characteristics are unique (very warmand dry, and the surface IR emissivity is lower than inother regions). The operational MODIS algorithm is stillevolving in order to meet these challenges.

6. Conclusions and future work

This paper describes the operational atmospheric re-trieval algorithms for global DAAC and direct broadcastMODIS data processing. Issues regarding training pro-files, noise performance, forward model bias, and lowsurface emissivity over desert regions are addressed.

Comparison of MODIS TPW with ground-based in-strumentation at the SGP ARM CART site reveals thatMODIS agrees with the MWR, with an rmse of 4.1 mm.For moist scenes with TPW greater than 17 mm, MODISis on average 1.2 mm drier than the MWR, with an rmseof 3.9 mm; for dry scenes, MODIS is on average 3.7mm more moist than the MWR. The physical retrievalsshow improvement of 0.8 mm in rmse over the syntheticregression retrievals for moist cases. More comparisonswith ground-based instrumentation are planned; thesewill include other ARM CART sites in the tropical west-ern Pacific and in Barrow, Alaska. On large scales,


MODIS products show good agreement in spatial dis-tribution when compared with GOES, SSM/I andTOMS.

The atmospheric MOD07 algorithm performs retriev-als of atmospheric temperature and moisture layers, totalprecipitable water, and total column ozone. WhileMODIS is not per se a sounding instrument, the infor-mation content of the high-spatial-resolution infraredmultispectral radiance observations can improve upona priori definitions of atmospheric state by providingbetter delineation of horizontal gradients.

Future work to improve the algorithm will includeenhancing the training database with more radiosondeobservations, particularly in under-represented polar ar-eas. In addition, use of new training datasets, such asTIGR-3, in the synthetic regression will be tested. Sur-face emissivity based on ecosystem type will be inves-tigated to improve the characterization of the surface inthe training data. A more physical basis for assigninga skin temperature to each training profile will also beinvestigated. Improvements to the radiance bias correc-tions are planned, including adding a seasonal variationto the global radiance bias values, accounting for small-scale variations in bias, and developing a regressionrelationship between the radiance bias and measuredradiances. Mitigation of the relatively high level of noisedue to nonuniform detector-to-detector response (strip-ing) will be investigated.

Terra MODIS algorithms have been adapted to thesecond MODIS instrument onboard the Aqua platform.The new platform will double the frequency of globalcoverage and allow for more consistent monitoring oftemperature, moisture, and ozone. Early results fromAqua MODIS show considerable reduction in striping.The smoother fields retrieved from Aqua MODIS pro-vide better depiction of gradients and allow full use ofthe high spatial resolution measurements. Future workwill focus on validation and applications of the AquaMODIS temperature, moisture, surface temperature, andozone retrievals. In addition, retrievals based on a com-bination of MODIS and AIRS radiances from Aqua willbe investigated to take advantage of the high spectralresolution (hence, high vertical spatial resolution) ofAIRS and the high horizontal spatial resolution ofMODIS.

Acknowledgments. The authors thank the MODISgroup at the University of Wisconsin—Madison CIMSSfor their assistance in this work, including Chris Moellerfor help with instrument calibration and noise estimates;Kathy Strabala for applying the MOD07 algorithm intothe IMAPP software; Richard Frey, Steve Ackerman,and Hong Zhang for assistance with the cloud mask;and Bryan Baum (NASA/LaRC) for useful discussions.Tim Schmit of NOAA/NESDIS helped with the com-parisons with GOES sounder retrievals. A note of thanksgoes to Hal Woolf of CIMSS who provided the MODISfast radiative transfer model. Rich Hucek of the MODIS

Science Data Support Team assisted with implementingthe operational MOD07 algorithm at the GSFC DAAC.We gratefully acknowledge the helpful comments bythree anonymous reviewers. ARM CART site data usedin this work were obtained from the Atmospheric Ra-diation Measurement Program sponsored by the U.S.Department of Energy, Office of Science, Office of Bi-ological and Environmental Research, EnvironmentalSciences Division. Measurements of emissivity spectrawere obtained through the ASTER Spectral Library atthe Jet Propulsion Laboratory, California Institute ofTechnology, Pasadena, California. SSM/I TPW datawere obtained from Remote Sensing Systems. TOMSdata were obtained through NASA GSFC. This workwas funded by NASA through NAS5-31367 and NAG5-9389.

REFERENCES

Ackerman, S. A., K. I. Strabala, W. P. Menzel, R. A. Frey, C. C.Moeller, and L. E. Gumley, 1998: Discriminating clear sky fromclouds with MODIS. J. Geophys. Res., 103, 32 141–32 157.

Alishouse, J. C., S. Snyder, J. Vongsathorn, and R. R. Ferraro, 1990:Determination of oceanic total precipitable water from the SSM/I. IEEE Trans. Geosci. Remote Sens., 28, 811–816.

Bowman, K. P., and A. J. Krueger, 1985: A global climatology oftotal ozone from the Nimbus-7 Total Ozone Mapping Spectrom-eter. J. Geophys. Res., 90, 7967–7976.

Eyre, J. R., 1992: A bias correction scheme for simulated TOVSbrightness temperatures. ECMWF Tech. Memo. 186, 28 pp.

——, and H. M. Woolf, 1988: Transmittance of atmospheric gasesin the microwave region: A fast model. Appl. Opt., 25, 3244–3249.

Ferraro, R. R., N. C. Grody, F. Weng, and A. Basist, 1996: An eight-year (1987–1994) time series of rainfall, clouds, water vapor,snow cover, and sea ice derived from SSM/I measurements. Bull.Amer. Meteor. Soc., 77, 891–906.

Frey, R. A., B. A. Baum, W. P. Menzel, S. A. Ackerman, C. C. Moeller,and J. D. Spinhirne, 1999: A comparison of cloud top heightscomputed from airborne lidar and MAS radiance data using CO2

slicing. J. Geophys. Res., 104, 24 547–24 555.Hannon, S., L. L. Strow, and W. W. McMillan, 1996: Atmospheric

infrared fast transmittance models: A comparison of two ap-proaches. Proc. Conf. on Optical Spectroscopic Techniques andInstrumentation for Atmospheric and Space Research II, Denver,CO, SPIE, 94–105.

Harris, B. A., and G. Kelly, 2001: A satellite radiance bias correctionscheme for radiance assimilation. Quart. J. Roy. Meteor. Soc.,127, 1453–1468.

Hayden, C. M., 1988: GOES-VAS simultaneous temperature-mois-ture retrieval algorithm. J. Appl. Meteor., 27, 705–733.

King, M. D., Y. J. Kaufman, W. P. Menzel, and D. Tanre, 1992:Remote sensing of cloud, aerosol, and water vapor propertiesfrom the Moderate Resolution Imaging Spectrometer (MODIS).IEEE Trans. Geosci. Remote Sens., 30, 2–27.

Li, J., 1994: Temperature and water vapor weighting functions fromradiative transfer equation with surface emissivity and solar re-flectivity. Adv. Atmos. Sci., 11, 421–426.

——, and H.-L. Huang, 1999: Retrieval of atmospheric profiles fromsatellite sounder measurements by use of the discrepancy prin-ciple. Appl. Optics, 38, 916–923.

——, W. Wolf, W. P. Menzel, W. Zhang, H.-L. Huang, and T. H.Achtor, 2000: Global soundings of the atmosphere from ATOVSmeasurements: The algorithm and validation. J. Appl. Meteor.,39, 1248–1268.

——, C. C. Schmidt, J. P. Nelson, T. J. Schmit, and W. P. Menzel,


2001: Estimation of total ozone from GOES sounder radianceswith high temporal resolution. J. Atmos. Oceanic Technol., 18,157–168.

McPeters, R. D., and Coauthors, 1998: Earth Probe Total Ozone Map-ping Spectrometer (TOMS) data products user’s guide. NASAReference Publ. 1998-206895, 64 pp. [Available from NASACenter for AeroSpace Information, 800 Elkridge Landing Rd.,Linthicum Heights, MD 21090.]

Menzel, W. P., and J. F. W. Purdom, 1994: Introducing GOES-I: Thefirst of a new generation of geostationary operational environ-mental satellites. Bull. Amer. Meteor. Soc., 75, 757–781.

——, F. C. Holt, T. J. Schmit, R. M. Aune, A. J. Schreiner, G. S.Wade, and D. G. Gray, 1998: Application of GOES-8/9 sound-ings to weather forecasting and nowcasting. Bull. Amer. Meteor.Soc., 79, 2059–2078.

Rothman, L. S., and Coauthors, 1998: The HITRAN molecular spec-troscopic database and HAWKS (HITRAN Atmospheric Work-station). J. Quant. Spectrosc. Radiat. Transfer, 60, 665–710.

Salisbury, J. W., and D. M. d’Aria, 1992: Emissivity of terrestrialmaterials in the 8–14 mm atmospheric window. Remote SensEnviron., 42, 83–106.

Schmit, T. J., W. F. Feltz, W. P. Menzel, J. Jung, A. P. Noel, J. N.Heil, J. P. Nelson, and G. S. Wade, 2002: Validation and use of

GOES sounder moisture information. Wea. Forecasting, 17,139–154.

Smith, W. L., H. M. Woolf, and W. J. Jacob, 1970: A regressionmethod for obtaining real-time temperature and geopotentialheight profiles from satellite spectrometer measurements and itsapplication to Nimbus 3 ‘‘SIRS’’ observations. Mon. Wea. Rev.,98, 582–603.

——, H. M. Woolf, S. J. Nieman, and T. H. Achtor, 1993: ITPP-5—The use of AVHRR and TIGR in TOVS data processing. Proc.Seventh Int. TOVS Study Conf., Igls, Austria, International Ra-diation Commission, 443–453.

Stephens, G. L., D. L. Jackson, and J. J. Bates, 1994: A comparisonof SSM/I and TOVS column water vapor data over the globaloceans. Meteor. Atmos. Phys., 54, 183–201.

Susskind, J., C. Barnet, and J. Blaisdell, 1998: Determination of at-mospheric and surface parameters from simulated AIRS/AMSU/HSB sounding data: Retrieval and cloud clearing methodology.Adv. Space Res., 21, 369–384.

Wentz, F. J., 1997: A well-calibrated ocean algorithm for SSM/I. J.Geophys. Res., 102, 8703–8718.

Wilber, A. C., D. P. Kratz, and S. K. Gupta, 1999: Surface emissivitymaps for use in satellite retrievals of longwave radiation. NASATech. Publ. NASA/TP-1999-209362, 35 pp. [Available online athttp://techreports.larc.nasa.gov/ltrs/.]

Documents

Operational Retrieval of Atmospheric Temperature, Moisture ...€¦ · radiosondes of atmospheric temperature, moisture, and ozone proﬁles. Evaluation of retrieved total precipitable