Cloud detection and clearing for the Earth Observing System ......Cloud detection and clearing for...

s

Cloud detection and clearing for the Earth ObservingSystem Terra satellite Measurements of Pollutionin the Troposphere ~MOPITT! experiment

Juying X. Warner, John C. Gille, David P. Edwards, Dan C. Ziskin, Mark W. Smith,Paul L. Bailey, and Laurie Rokke

The Measurements of Pollution in the Troposphere ~MOPITT! instrument, which was launched aboardthe Earth Observing System ~EOS! Terra spacecraft on 18 December 1999, is designed to measuretropospheric CO and CH4 by use of a nadir-viewing geometry. The measurements are taken at 4.7 mmin the thermal emission and absorption for the CO mixing ratio profile retrieval and at 2.3 and 2.2 mmin the reflected solar region for the total CO column amount and CH4 column amount retrieval, respec-tively. To achieve the required measurement accuracy, it is critical to identify and remove cloudcontamination in the radiometric signals. We describe an algorithm to detect cloudy pixels, to recon-struct clear column radiance for pixels with partial cloud covers, and to estimate equivalent cloud topheight for overcast conditions to allow CO profile retrievals above clouds. The MOPITT channel radi-ances, as well as the first-guess calculations, are simulated with a fast forward model with inputatmospheric profiles from ancillary data sets. The precision of the retrieved CO profiles and totalcolumn amounts in cloudy atmospheres is within the expected 610% range. Validations of the cloud-detecting thresholds with the moderate-resolution imaging spectroradiometer airborne simulator dataand MOPITT airborne test radiometer measurements were performed. The validation results showedthat the MOPITT cloud detection thresholds work well for scenes covered with more than 5–10% cloudcover if the uncertainties in the model input profiles are less than 2 K for temperature, 10% for watervapor, and 5% for CO and CH4. © 2001 Optical Society of America

OCIS codes: 280.0280, 280.1120, 010.7030, 010.0010.

1. Introduction

The Measurements of Pollution in the Troposphere~MOPITT! instrument1 that is aboard the Earth Ob-erving System ~EOS! Terra spacecraft is designed to

measure tropospheric CO and CH4. This instru-ment scans the Earth at nadir with a spatial resolu-tion of 22 km 3 22 km and achieves a global coveragewithin approximately three days. The MOPITT in-strument is a gas correlation radiometer that mea-sures the CO profile by use of atmospheric and

J. X. Warner, J. C. Gille, D. P. Edwards, D. C. Ziskin, M. W.Smith, and P. L. Bailey are with the Atmospheric Chemistry Di-vision, National Center for Atmospheric Research, P.O. Box 3000,Boulder, Colorado 80307-3000. The e-mail address for J. X.Warner is [email protected]. L. Rokke is with the Office ofData Assimilation, NASA Goddard Space Flight Center, Green-belt, Maryland 20771.

Received 28 July 2000; revised manuscript received 20 Novem-ber 2000.

0003-6935y01y081269-16$15.00y0© 2001 Optical Society of America

terrestrial thermal radiation in the spectral regionnear 4.7 mm and the CO and CH4 total columnamounts by use of reflected solar radiation in thespectral region near 2.3 and 2.2 mm, respectively.The thermal channel measurements, as well as thereflected solar radiation during daytime, will be usedto retrieve CO profiles in the troposphere at sevennominal pressure levels, and the reflected solar radi-ation will be used to retrieve CO and CH4 total col-umn amounts.2 The anticipated accuracy is 10% forthe CO measurement and 1% for the CH4 columnamount.

Tropospheric CO is among the most importanttrace species in the atmosphere, in part because of itschemically active nature. The CO and CH4 concen-trations directly affect the concentration of the OHradical through the atmospheric oxidation of CO andthe reaction of CH4 with OH, and hence affect therate at which many natural occurring and anthropo-genic trace species are removed from the atmosphere.In addition, oxidation of CO in the presence of NOx isa major contribution to tropospheric ozone produc-

10 March 2001 y Vol. 40, No. 8 y APPLIED OPTICS 1269

tS

msn

oesrrtaecap

otb

1

tion. Because CH4 and ozone are greenhouse gasesaffecting the Earth’s energy balance, the concentra-tions of CO and CH4 indirectly affect the global cli-mate and air quality. CO is produced near theEarth’s surface mainly from naturally occurringchemistry processes and human activities, and it hasa lifetime of approximately two months. This makesCO a good tracer for studying dynamic circulationand convection in the atmosphere. CH4, in contrast,with its lifetime of approximately ten years, can be-come more or less uniformly mixed over the entireEarth. However, it has been reported that CH4 con-centration has increased at a rate of 1.3% per yearuntil recently.3 In recent years the rate of the CH4increase has appeared to slow to approximately 0.6%per year4 or less. Measurements of CO and CH4concentrations from MOPITT will provide continuouscoverage in both space and time, which is currentlylacking.

One of the major challenges related to spaceborneoptical sensors that measure tropospheric propertiesis the treatment of cloud contamination in the instru-ment field of view ~FOV!. MOPITT measurementsat the spectral regions of 2.2, 2.3, and 4.7 mm aresensitive to attenuation by clouds. Cloud-relatedproblems in satellite remote sensing have been stud-ied and addressed for many years, and a few exam-ples are given below. Wylie et al.5 studied globalcloud coverage using four years of High-ResolutionInfrared Radiation Sounder ~HIRS! data with a spa-tial resolution of 20 km 3 20 km at nadir and foundthat only approximately 23% of the pixels were cloudfree. Because MOPITT has a FOV of 22 km 3 22km, a large percentage of its global data set will becovered, or partially covered, with clouds.

The procedures of working with the cloud-contaminated data include two steps: to identify thetropospheric scenes covered, or partially covered, byclouds and to estimate clear-sky retrievals fromcloud-contaminated pixels. The former is generallycalled cloud detection, which involves defining theobservable quantity that discriminates betweencloudy and clear scenes and determining a value thatrepresents the contrasts. The latter is called thecloud clearing or removing process. The most com-mon of the cloud detection techniques are the thresh-old methods that make use of radiance variations inwavelength, space, or time.6,7 Much research oncloud detection was initiated from processing the ad-vanced very high resolution radiometer ~AVHRR!data and the Geostationary Operational Environ-mental Satellite ~GOES! images, and multispectralechniques were often used.8,9 The Internationalatellite Cloud Climatology Project ~ISCCP!10–13

has developed a cloud detection scheme that usesvisible and infrared window radiances and their spec-tral, space, and time variations. The moderate-resolution imaging spectroradiometer ~MODIS! cloud

ask14 uses a number of thresholds for 14 differentpectral channels and the combination of these chan-els to identify clouds in the FOV.To retrieve atmospheric properties in the presence

270 APPLIED OPTICS y Vol. 40, No. 8 y 10 March 2001

f clouds in an instrument FOV involves either thestimation of clear column radiances from cloudy ob-ervations or the iteration of the cloud clearing andetrieval steps. Smith15 developed the N* method toemove the cloud effect in the process of retrievingemperature profiles. Some other algorithms werelso based on the N* method, such as that by Smitht al.16 that used the collocated AVHRR and HIRSy2hannels to provide estimated clear column radiancend calculate cloud cover contrast between adjacentixels. Chahine17 developed a method in which the

temperature profiles, cloud heights, and cloudamounts are derived simultaneously in an iterativerelaxation retrieval scheme. It uses knowledge ofadjacent pixels without a priori temperature infor-mation or cloud-free observations; however, it alsorequires a careful choice of instrument spectral cov-erage. Many other investigators have developed al-gorithms to handle operational satellite observationsthat are contaminated with clouds, as summarized byRizzi et al.18

In this paper we describe the MOPITT cloud algo-rithm, which includes the detection of cloudy pixels,clearing of clouds for pixels with nonuniform cloudcover, and the determination of cloud top heights forovercast opaque conditions to allow retrievals aboveclouds. In Section 2 we briefly summarize the in-strument characteristics, the forward radiance mod-els, and the ancillary data sets that are used in thestudy. In Section 3 we describe the cloud detectionalgorithm, and in Section 4 we discuss the MOPITTcloud clearing processes. In Section 5 we presentsome validation results of the cloud detection thresh-olds using the MODIS airborne simulator ~MAS!19

and MOPITT airborne test radiometer ~MATR!20

measurements. Finally, in Section 6 we summarizethe current status of the MOPITT cloud algorithmand discuss the planned future improvements.

2. Review of Instrumentation and Forward Models

A. MOPITT Instrument Characteristics and ChannelSignal Sensitivities

MOPITT takes measurements at 2.2-, 2.3-, and4.7-mm spectral regions by use of gas correlation ra-diometers. A gas correlation radiometer filters theoutgoing radiation at the top of the atmosphere~TOA! through a cell containing the target gas that isbeing measured in the atmosphere. This radiome-ter modulates the cell pressure, in the case of thepressure-modulated cell ~PMC!,21 or the cell length,for a length-modulated cell ~LMC!.22 In each mod-ulating cycle the detector outputs two signals corre-sponding to the two states of the modulating cell.An average ~A! signal is obtained when these twosignals are averaged, and a difference ~D! signal isbtained when the difference of the two signals areaken. The D signals generally include the contri-utions from the atmospheric CO or CH4, and the A

signals represent mainly the background radiance~including the contribution from the target gases!.The MOPITT instrument includes two PMC’s and

ditt

Mp

i

l

ta

n

Table 1. MOPITT Correlation Radiometer Channel Characteristics

two LMC’s for measurement of the CO profile and twoLMC’s each for the CO and CH4 total columnamounts. There are a total of eight channels used inMOPITT measurements, corresponding to the eightmodulating cells, and in each channel there is an Asignal and a D signal. The MOPITT channel infor-mation is listed in Table 1, in which the filter band isgiven in full width half-maximum ~FWHM!.

In the CO thermal channels, the A signals are dom-inated by the surface emission and can therefore beused to determine the surface characteristics.23

These characteristics can then be applied to the Dsignals to isolate the atmospheric information. Inthe solar CO and CH4 channels, the A signals are

etermined effectively by the product of the solarntensity and the surface reflectivity transmittedhrough the atmosphere. The D signals are sensi-ive to the absorption by CO and CH4 as well as to the

solar intensity and surface reflectivity. The DyAsignal ratios are used to obtain the total columnamount of CO and CH4. To obtain this ratio, thesurface reflectance and the solar radiance terms arecanceled to first order, and the ratio mainly dependson the atmospheric parameters. Sensitivity studiesalso show that the effect of interfering constituents isgreatly removed when one performs the DyA signalratio.

B. Forward Models and Ancillary Data Sets

The MOPITT operational fast radiative transfer~MOPFAS! model, which is used to simulate thechannel radiances, is described in detail by Ed-wards et al.24 Only a brief review is given here.

OPFAS uses an approach similar to the opticalath transmittance regression scheme25 from pa-

rameters calculated by a line-by-line model~GENLN2!.26 This fast model has a computed signalaccuracy within approximately 0.5% of a full line-by-line calculation.

Spectral response functions are used for each A andD channel to simulate the MOPITT signals. Eachsignal can be expressed as

SiA,D 5 *

Dn

I~n!fiA,D~n!dn, (1)

ChannelNumber Target Gas Cell Type

C

1 CO LMC2 CO LMC3 CO PMC4 CH4 LMC5 CO LMC6 CO LMC7 CO PMC8 CH4 LMC

where fiA,D~n! is the A or D signal response function,

which includes the channel blocking filter and gascorrelation response. In Eq. ~1!, I~n! is the mono-chromatic radiance reaching the TOA and is de-scribed below.

Assuming the atmosphere to be plane parallel, wecan write the monochromatic radiance for clear con-ditions at the TOA, Iclear~n!, in the absence of scatter-ng, as

Iclear~n! 5 *ps

0

B@n, T~p!#dt~n, p, usat!

dpdp

1 ε~n!B~n, Ts!t~n, ps, usat!

1Fd~n!

p@1 2 ε~n!#t~n, ps, usat!

1Fsol~n!

p sec usolt~n, ps, usol!@1 2 ε~n!#

3 t~n, ps, usat!. (2)

The first term on the right-hand side of Eq. ~2!represents the atmospheric thermal emission be-tween the lower boundary at pressure ps and theocation of the satellite. B~n, T! is the Planck func-

tion at wave number n and temperature T, and t~n,p, usat! is the transmittance between pressure p andhe satellite as a function of the satellite zenithngle usat. The second term refers to the surface

emission that reaches the satellite, where ε~n! is thesurface emissivity neglecting the viewing-angle de-pendence and Ts is the surface temperature. Thethird term describes the downwelling thermal flux,Fd~n!, that is reflected by the surface to the satellite.The fourth term is the reflected solar radiation,where Fsol~n! is the solar irradiance incident at theTOA atmosphere, and t~n, p, usol! is the transmit-tance from the TOA to the surface at solar zenithangle usol.

Assuming that a single-layer opaque cloud with aonunity emissivity is present in the FOV, the radi-

r Wavelength~mm!

Filter BandFWHM

~mm!Nominal Cell

Pressure ~kPa!

4.617 0.111 202.334 0.022 204.617 0.111 7.52.258 0.071 804.617 0.111 802.334 0.022 804.617 0.111 3.82.258 0.071 80

ente


cTdber

wccd

rt

~artidc

1

ative transfer equation for TOA cloudy radianceIcloud~n! becomes

Icloud~n! 5 *pc

0

B@n, T~p!#dt~n, p, usat!

dpdp

1 εc~n!B~n, Tc!t~n, pc, usat!

1Fdc~n!

p@1 2 εc~n!#t~n, pc, usat!

1Fsol~n!

p sec usolt~n, pc, usol!@1 2 εc~n!#

3 t~n, pc, usat!. (3)

In Eq. ~3! the parameters with subscript c refer tothose at the cloud level. Similar to Eq. ~2!, the firstterm on the right-hand side in Eq. ~3! represents theatmospheric emission from the cloud top to the TOAwith a cloud emissivity εc~n!. The second term is theloud thermal emission that reaches the satellite.he third term shows the thermal radiation emittingownward from the atmosphere above the clouds,eing reflected by clouds and transmitted to the sat-llite. The final term represents the solar radiationeflected by clouds that reaches the satellite.

Often an instrument FOV is only partially coveredith clouds, and, in addition, a large percentage of

louds are transmissive clouds.5 When a FOV isovered partially with thick clouds, the observed ra-iance can be expressed as

I~n! 5 ~1 2 n!Iclear~n! 1 nIcloud~n!, (4)

where n is the cloud fraction in the FOV. A trans-missive cloud can be simulated with a similar equa-tion when the scattering effect of the clouds can beneglected, and in this case n stands for the product ofcloud fraction and cloud emissivity.

Although Eqs. ~1!–~4! can generally be used for allMOPITT channels, there is virtually no thermal ef-fect in the solar regions at 2.2 and 2.3 mm, and onlythe solar term is included in the calculation. How-ever, if solar effects are neglected in the thermal cal-culations at the 4.7-mm region, it causesapproximately a 2% error in the average channelsand a 1% error in the difference channels.24

For simulation of instrument signals, another MO-PITT forward model, namely, the MOPITT absorp-tion ~MOPABS! model, is used in this study.24

MOPABS uses an absorption look-up table to obtainCO and CH4 instrument signals for any observationaltitude. This code has the capability to simulateinstrument signals from aircraft measurements, aswell as the satellite measurements, and it includesfilter functions for MOPITT, MAS, and MATR chan-nels. MOPABS has accuracies comparable to a line-by-line model, but it is considerably faster and isoften used as a research tool to analyze the radiativetransfer process for MOPITT-related problems.

Both MOPFAS and MOPABS take into account thecontributions from H2O, CO2, O3, N2O, CO, and CH4,


and they require as input the vertical distributions ofatmospheric temperature, surface temperature,emissivity, and reflectivity. Currently, CO2, O3, andN2O profiles from climatological data sets are used.The real-time meteorological data, i.e., temperatureand water-vapor distributions, are provided by theNASA Data Assimilation Office ~DAO!. The surfaceeflectance in the 2.2–2.3-mm spectral region was re-rieved from the Landsat thematic mapper.27 The

4.7-mm thermal emissivity distribution over the globeis determined from the Fu and Liou28 radiative trans-fer model parameters that are assigned to each U.S.Geological Survey scene type.29 This is described byWilber et al.30 There is no seasonal variation in thecurrent long-wave emissivity data set, and additionalresearch is necessary. A range of cloud emissivity~or reflectivity! from 0.2 to 0.8 is assigned on the basisof the type of cloud under study.

3. Cloud Detection

Many existing cloud detection techniques cannot beapplied directly to MOPITT cloud problems becauseof the limitation of the number of MOPITT spectralchannels and the spatial resolution. For example,the multispectral technique has been used widely inthe community to identify certain types of clouds6,24;however, it requires measurements at certain spec-tral bands. Also, statistical techniques based onpatterns of the neighboring pixels are often used todetect heterogeneous clouds over the ocean, whenthere is adequate spatial resolution.31,11 Neverthe-less, MOPITT signals provide information that can beused to detect and clear clouds by measuring theconcentration of trace species above the clouds. Useof this information forms the core of the MOPITTcloud algorithm and is discussed in Subsection 3.B.A threshold method is also used to compare the ob-served radiance against the model-calculated clearcolumn radiance, which is discussed in Subsection3.A.

A. Threshold Method

In general, clouds are characterized as colder andhaving higher solar reflectance than the Earth’s sur-face. The temperature differences between cloudyand clear scenes appear in the 4.7-mm channel radi-ances, and the differences in the boundary reflectanceare revealed mainly from the solar channels. TheMOPITT cloud detection threshold method uses bothsolar and thermal channels for daytime passes andthermal information only for nighttime passes. TheA signals from the 2.3-mm band LMC at a cell pres-sure of 800 mb ~denoted as ch6A! and the A signalsfrom the 4.7-mm LMC at a cell pressure of 800 mbch5A! are used for threshold tests. The A signals ofll four thermal channels, or all four solar channels,espond to the surface or cloud similarly. Thereforehe A signal from only one thermal, or solar, channels necessary to detect clouds if the FOV’s from alletectors are collocated. Because there is no exactollocation in the MOPITT instrument, however, it

fect

sT

,

may be decided at a later time that all channels areneeded in cloud detection.

By studying the radiance sensitivities to cloudcover, and including the anticipated uncertainties inthe clear radiance estimate, we determined a set ofthresholds to distinguish the clear from the cloudyscenes. The cloud cover is simulated by use of frac-tions from 0 to 100% and cloud top pressures fromapproximately 900 to 200 hPa. Single-layer, opaqueclouds are assumed with a thermal emissivity of 0.98and a solar reflectivity of 50%. MOPITT cloudthresholds, based on observed channel radiance andmodel-calculated clear column radiance, are then se-lected for daytime as

for nighttime,

In inequalities ~5! and ~6!, Rch5A and Rch6A are aver-age radiances from ch5A and ch6A, respectively.Mid-latitude and high latitude are defined as thosegreater than 35° in the summer hemisphere and 30°in the winter hemisphere for this study. If a tem-perature inversion occurs in the atmosphere, theclouds will be warmer than the underlying surface,hence the Rch5A ratio between cloudy and clear sceneswill be greater than 1. A threshold of 1.1 is used todetect clouds under this condition. During daytimethe solar channel information is used, and the clouddetection is not completely dependent on the temper-ature contrasts; therefore the additional threshold forthe thermal channel to detect clouds under temper-ature inversion conditions is not necessary. Boththe differences and the ratios of the observed and themodel-calculated radiances are used in the cloud de-tection to reflect different aspects of the contribu-tions. For example, when the ratio is taken of thesolar channel average signals, certain solar-flux-related terms cancel, and more emphasis is placed onthe surface characteristics and the atmospherictransmittances. As shown in Eqs. ~2! and ~3!, theourth terms on the right-hand side of these twoquations represent the solar radiation terms forlear and cloudy atmospheres, respectively. Whenhe ratio of these terms for clear and cloudy atmo-

5Rch6Aobserved

Rch6Acalculated. 1.5, or

Rch5Acalculated 2 Rch5Aobserved $ 0Rch5Aobserved

Rch5Acalculated# 0.93 ~or . 1.1

5Rch5Acalculated 2 Rch5Aobserved $ 0Rch5Aobserved

Rch5Acalculated# 0.97 ~or . 1.1

pheres is taken, the solar irradiance incident at theOA Fsol~n! and sec~usol! are canceled.The one-day coverage of MOPITT channel radi-

ances are simulated with the DAO 1 August 1998meteorological data set. Cloud information, cloudfractions, cloud top pressures, and cloud tempera-tures provided by the DAO are used in the simula-tion. Note that the quality of this cloud informationis not important in the simulation; however, it pre-vents us from using it in the retrieval of real obser-vations. Although only single-layer clouds areincluded in the simulations, the existence of multi-layer clouds will not change the cloud detection con-clusions. It will, however, affect the cloud clearing

processes. Polar regions ~latitude . 65° N or260° S! are excluded in this simulation because of

the frequent temperature inversions in the atmo-sphere and to avoid the effect on the daytime signalsof possible snow and ice coverage. Real-time snowand ice covers are not currently available to us, andthe cloud detection over snow or ice surfaces will bediscussed in a separate study.

Cloud detection by use of the threshold method wasperformed on the 1 August 1998 simulated data, andthe results are summarized in Table 2. The total inthe second column refers to the cloud cover in thesimulation, and the third ~detected as clear! andfourth ~detected as cloudy! columns are the detectionresults. The cloud cover is categorized as clear

Table 2. Cloud Detection Results of the Threshold Method withoutNoise

Cloud Cover Total ~%!Detected asClear ~%!

Detected asCloudy ~%!

Clear 50.46 50.46 0.0,5% cloud cover 2.19 1.58 0.62.5% cloud cover 47.35 0.42 46.93.5% cloudy, excluding

nocturnal low cloudsa46.98 0.055 46.93

aLow clouds are below the 75-kPa level.

.005, or

at mid-latitude and high latitude!;

(5)

.005, or

at mid-latitude and high latitude!. (6)


la1eTA

1

scenes ~50.46%!, scenes covered with less than 5%clouds ~2.19%!, more than 5% clouds ~47.35%!, andscenes with more than 5% cloud covers, but nocturnallow clouds ~46.98%! are excluded. Almost all pixels~more than 99%! containing more than 5% cloudswere detected by use of the MOPITT thresholds.There are, however, a very small number of pixels~0.8%! that are covered with more than 5% cloudsthat could not be detected. The majority of the un-detected pixels come from low clouds at night, whichis shown by the differences between the third andfourth rows in Table 2. Under these conditions,there is not enough thermal contrast between thesimulated cloudy radiances and the calculated clearcolumn radiances.

We evaluated the accuracy of the cloud detectionusing the threshold method by comparing the re-trieved CO amounts from a clear simulation and acloudy simulation. The left upper panel of Fig. 1 isa box chart showing the percentage differences be-tween the clear and the cloudy runs for the retrievedtotal column CO and the CO mixing ratio at sevenlevels for pixels identified as clear by MOPITT cloudthresholds. Therefore the data shown on the graphare those pixels with cloud contamination, whichwere erroneously classified as clear. Note that there

Fig. 1. Effect of the threshold uncertainties on the retrieved totalof undetected cloudy pixels for the total column CO amount and tsame as the upper panel except that input random error to the moas the left panels, except that the results are from the N* cleareretrieval and Bottom and L-1 to L-6 show the levels of retrieved C210 to 220%; green, 20 to 30%, 220 to 230%; blue, 30 to 40%, 2


are approximately 760,000 simulated pixels per day.Nevertheless, failure to detect these types of scenesas described above does not impact the retrieved re-sults significantly. As shown in Fig. 1 in the leftupper panel, almost all data are within a 610% dif-ference for the total column CO amount ~labeled astotal! and for the level retrievals at 700 hPa ~L-2!, 500hPa ~L-3!, 350 hPa ~L-4!, and 250 hPa ~L-5!, whichmeets the accuracy requirement for MOPITT. Atthe lowest two levels ~Bottom and L-1! and the topevel ~150 hPa!, however, the relative errors frompproximately 6% of cloudy pixels are beyond the0% accuracy requirement, and a few individual pix-ls at the bottom level reach the 30–40% range.his is related to MOPITT retrieval sensitivities.s discussed by Pan et al.,2 the retrieval profile will

have smaller uncertainty in the free tropospheric lay-ers typically within 10% of a rms error, but will havelarger uncertainties for the boundary layer and thetop tropospheric layer, which may reach 10–20%, de-pending on the noise-equivalent radiance. Wang etal.,32 in their MOPITT instrument sensitivity stud-ies, pointed out that the retrieval errors are mostpronounced in the first layer ~0–2 km! and the lastlayer ~12–14 km! because of reduced instrument sen-

n CO amount. The left upper panel shows the percentage errorsixing ratios at seven retrieval levels. The left lower panel is thes included in the cloud detection. The right panels are the samediances. For the x-axis labels, Total indicates total column COxing ratios. Orange, 210 to 0%; red, 0 to 10%; yellow, 10 to 20%,to 240%; purple, 40 to 50%, 240% to 250%.

columhe mdel id ra

O mi30%

c

C

w

Table 3. Cloud Detection Results by the Threshold Method with

sitivity to CO and that any radiance errors get mag-nified in these two layers.

There are two major sources of uncertainties in theestimation of the reference clear column radiance:model uncertainties and the random error introducedby the input data. Although the model systematicbias can be validated and compensated for, the inputdata uncertainties could lead to larger errors thanpresented here; therefore the error of the input datahas to be considered in the determination of the clouddetection thresholds. The A signals of the MOPITTinstrument are most sensitive to the surface temper-ature and the emissivity or reflectivity, and the sur-face temperature provided by the DAO reanalysis issubject to random error. Input data random error isadded to the clear radiance estimation when the in-put profiles and the surface quantities are perturbedrandomly. The error is assigned on the basis of aGaussian distribution with a zero mean and a stan-dard deviation at the expected error levels. The an-ticipated error level for the temperature profiles andthe surface temperature over land is 2 K, and for thesurface temperature over ocean it is 0.5 K; a 10%uncertainty is included in the water-vapor profiles,and 5% is included in the profiles of CO and CH4.

The performance of the cloud detection thresholds,with consideration of the random error in the inputmeteorological data, is listed in Table 3. The majordifference between the tests with and without ran-dom error occurs for clear pixels and pixels with lessthan 5% cloud cover. Many pixels ~approximately13%! that are clear are classified as cloudy, and thisreduces the retrieval coverage. The pixels that arecovered by more than 5% clouds are nearly unaffectedby the input noise, provided that the noise levels donot exceed those described above. The cloud detec-tion accuracy including the input data random erroris also evaluated with the CO retrieval results and isshown in the left lower panel of Fig. 1. Comparedwith the left upper panel, the range of uncertaintiesof CO retrieval caused by the cloud detection errorstays approximately the same with and without in-clusion of the input data random error in the clouddetection. When the random error is included in thereference radiance, however, the number of undetec-ted cloudy pixels decreases. The reason is thatmany pixels with less than 5% cloud cover are de-

Random Input Error



Clear 50.46 43.72 6.74,5% cloud cover 2.19 1.18 1.01.5% cloud cover 47.35 0.46 46.89.5% cloudy, exclud-

ing nocturnal lowcloudsa

46.98 0.09 46.89


tected as cloudy; hence they are not included in theretrieval.

The uncertainties of the surface emissivity and re-flectivity are tested similarly to that of the surfacetemperature. Approximately 1% of the emissivityand reflectivity variation is equivalent to an 0.5 Ksurface temperature change. When both sources ofuncertainties are present, the model calculation errorincreases slightly; however, this will not change theconclusions of the cloud detection accuracy describedabove. In summary, on the basis of the simulationresults the MOPITT cloud detection thresholds arecapable of detecting scenes with more than 5% cloudcover. The input data uncertainties remain a majorconcern with this technique, especially if the inputprofile errors exceed the anticipated values.

B. Equivalent Cloud Top Pressure Estimation from CH4

Total Column Amount

The MOPITT instrument measures total column CH4at 2.2 mm through reflected solar radiation, and thehannel DyA signals are used to retrieve the CH4

total column amount.23 CH4 can be approximatedas a uniformly distributed gas in the troposphere to agood accuracy ~approximately 2–3%!. Therefore themeasured total column CH4 amount is related di-rectly to the altitude of the surface. When an opti-cally thick overcast cloud is presented in the FOV, themeasured column of CH4 represents the portionabove the cloud.

A simple exponential curve of CH4 DyA signalsversus cloud top pressures is obtained when we fitsimulated data in each narrow range of solar zenithangles and satellite zenith angles. The cloud toppressure for a pixel with complete cloud cover is writ-ten as

Pcloud 5 a 1 exp~bRch4DyA 1 c!, (7)

where the parameters a, b, and c are obtained from anonlinear least-squares fitting; and Rch4DyA is the

H4 DyA channel radiance. A set of parameters isdetermined for every 5-deg solar zenith angle and10-deg satellite zenith angle. An equivalent cloudtop is determined by Eq. ~7! when a FOV is partiallycovered with clouds or when the clouds are opticallythin. The estimated equivalent cloud top pressurefor each pixel location is then compared with thesurface pressure to detect clouds. Thresholds of 50mb is used for solar zenith angles less than 35° and of100 mb for solar zenith angles greater than 35°.Larger values of thresholds are used for larger solarzenith angles where Rch4DyA signals spread in greaterranges. An example of a fitting is given in Fig. 2,where the solar zenith angles span the range of 25°–30° and the satellite zenith angles are 0°–10°. Thesolid curve shows the fitting function, which repre-sents the data with good accuracy. The parametersa, b, and c for the range of solar zenith angles 0°–80°and satellite zenith angles 0°–30° are listed in Tables4~a!, 4~b!, and 4~c!, respectively. These parameters

ere calculated with the MOPABS model to ensure


Table 4~a!. Parameter a Used to Estimate Equivalent Cloud Top

Table 4~b!. Parameter b Used to Estimate Equivalent Cloud Top

Table 4~c!. Parameter c Used to Estimate Equivalent Cloud Top


high accuracy. In the range of solar zenith anglesbetween 0° and 20°, the function varies slightly andwas fitted with one set of parameters.

Equation ~7! is predetermined from model calcula-tions by use of a one-day global simulation. Thusthe parameters include a large range of atmosphericconditions. Table 5, which is similar to Tables 2 and3, summarizes the detection results by use of the CH4channel technique. The performance of this tech-nique is slightly worse than that of the thresholdmethod without the input data random error. Ap-proximately 6% of the cloudy pixels were undetected,some of which have more than 50% cloud cover. Al-most all these pixels are covered with low-level cloudsin which the differences of the cloud top pressure andthe surface pressure are within the range of thethreshold ~note the cases excluding low clouds!. Theadvantage of this technique is that it does not rely onthe input meteorological data to predict clear columnradiance, and therefore it eliminates most of the un-certainties associated with the reference radiancecalculation. The disadvantage is that this techniquecan be used only in the daytime.

MOPITT cloud detection assigns confidence levelsto the output products, so data users can estimate theuncertainties involved in data processing that aredue to the cloud algorithm. In general, daytimedata are assigned with higher confidence becauseboth solar and thermal channels are used. Data

Pressure

Solar ZenithAngles ~deg!

Satellite Zenith Angles

0°–10° 10°–20° 20°–30°

75–80 9.37509537 9.35650063 9.3456096670–75 9.65037441 9.55518627 9.5390291265–70 9.65175819 9.69774342 9.6685466860–65 9.77284908 9.78327179 9.7463731855–60 9.82933807 9.82208157 10.031369250–55 10.4652643 10.3090706 10.211434445–50 10.5770931 10.0749979 10.131051140–45 10.7087765 10.6183786 10.765327535–40 10.6508646 11.0678282 10.604973830–35 10.3976059 10.541935 10.353645325–30 11.1586599 11.0067167 10.765626920–25 11.1539001 11.4247379 11.10930920–20 11.3551416 11.2557335 11.313098

Table 5. Cloud Detection Results by Use of Equivalent Cloud TopPressures



Clear 51.67 49.11 2.56,5% cloud cover 1.81 0.33 1.47.5% cloud cover 46.52 2.54 43.99.5% cloudy, excluding

low cloudsa44.02 0.035 43.99


Fig. 2. Example of a fitting of the cloud top pressures as a func-tion of CH4 DyA signals in the range of 25°–30° solar zenith anglesand 0°–10° satellite zenith angles. The solid curve represents thefitting, and the fitting data are shown as dots.

Pressure



0°–10° 10°–20° 20°–30°

75–80 2298.001007 2306.795105 2310.64764470–75 2558.743530 2440.747070 2441.99829165–70 2560.893005 2546.792847 2537.73358260–65 2668.972534 2647.250977 2639.43872155–60 2636.217041 2692.088318 2536.14227350–55 2471.006226 2501.777008 2531.19329845–50 2484.940002 2662.728516 2621.70965640–45 2489.939148 2479.331512 2429.58511435–40 2561.581909 2411.882294 2548.83270330–35 2727.869629 2629.792969 2688.85235625–30 2475.089325 2506.804779 2571.02081320–25 2513.109375 2425.240967 2496.771820–20 2459.871826 2483.929382 2442.511932

Pressure



0°–10° 10°–20° 20°–30°

75–80 277.0613632 276.3436890 276.112899870–75 274.5113373 274.5829849 274.299385165–70 271.9215240 273.7236633 273.372329760–65 271.3478317 272.2258453 271.633186355–60 272.4910965 271.2615738 280.665222250–55 292.4623108 287.8229752 284.755546645–50 294.3021393 277.353096 279.914169340–45 297.0175247 295.2946167 2100.62959335–40 293.2850189 2108.035858 292.846908630–35 282.9045334 288.8233032 283.035469125–30 2107.804504 2103.384636 296.025802620–25 2106.610794 2116.15789 2106.4425280–20 2113.224983 2110.118034 2113.047668

So4

fct

crstefac

rafs1citdCdNordaFtnbf

smulwaaawc

ttpqicccccrMsat

over the ocean have a higher confidence level whenthe threshold method is used because the input data,particularly the surface quantities, are less uncer-tain. Tests with values close to the thresholds willbe assigned lower confidence levels. Both the clouddetection threshold and the equivalent cloud toppressure methods are used by the MOPITT level-2processor. The pixels that are classified as clear byboth methods are assigned higher confidence levelsand those classified by only one method are assignedlower confidence levels.

The MOPITT cloud detection algorithm has somelimitations. It may be difficult to detect thin cirrusclouds if the effect on the radiance is not greater than3–7%. Detection of clouds at night depends solely onthe temperature contrasts between the clouds andthe surface. If the clouds are low, have cold sur-faces, or a temperature inversion is present withinthe FOV, the threshold method may not work well.Different approaches are proposed and are discussedin Section 6.

4. Cloud Clearing

The retrieval of tropospheric CO and CH4 in the pres-ence of clouds is made possible by two techniques:one to estimate a clear column radiance by use ofradiances from neighboring pixels and the other toidentify overcast opaque cloud tops and retrieveabove clouds. The first technique makes use of theN* method introduced by Smith15 and is discussed in

ubsection 4.A. The procedures to determine anvercast opaque cloud are described in Subsection.B.

A. N* Technique

The N* method assumes that two adjacent cloudypixels possess the same radiative and cloud physicalproperties, and they differ only by the amount ofcloud cover. N* is defined as the ratio of the cloudcover in two adjacent pixels; hence it is independentof spectral frequency and channel. N* can be calcu-lated from a reference channel as

N* 5Robs1~ref! 2 Rclear~ref!Robs2~ref! 2 Rclear~ref!

, (8)

where Rclear~ref ! is the clear column radiance andRobs1~ref ! and Robs2~ref ! are the observed radiancesrom pixel 1 and 2, respectively, for the referencehannel. N* is then applied to all the other channelso reconstruct the clear column radiance with

Rclear~i! 5Robs1~i! 2 N*Robs2~i!

1 2 N*, (9)

where i indicates the ith channel.In the MOPITT cloud clearing algorithm, the clear

column radiance for the reference channel, which isused to calculate N*, is obtained from forward modelcalculations. The average signals, ch5A for thermaland ch6A for solar, are used as reference channels forcloud clearing applications. Cloud signatures can

appear differently in the solar and the thermal bandsdepending on the temperature and reflectance of thecloud. In a comparison of the N*’s calculated fromthe reference channel with N*’s from the ratio of theloud covers, the solar CO channels show better cor-elation than the thermal channels. Therefore theolar channel was chosen to calculate N*’s for day-ime cases, whereas at nighttime the thermal refer-nce signals are used. Because the MOPITTootprints are relatively large ~22 km 3 22 km!, twodjacent pixels are used to resolve heterogeneouslouds.

Note that in the calculation of the clear columnadiances, the uncertainties in the observed radi-nces are magnified by a factor 1y~1 2 N*!. There-ore for accuracy considerations N* should be muchmaller than 1 ~assuming that N* is always less thanby taking the ratio that uses the radiance with less

loud cover over that with more cloud cover!. Hencet requires that the two pixels possess adequate con-rasting cloud covers. N* limits of less than 0.6 foraytime and 0.5 for nighttime are used to ensure aO measurement accuracy of 10%. To control ran-om errors in the N* calculations, for daytime the*’s from ch5A and ch6A are compared with each

ther, and those that differ by more than 0.1 areejected. This constraint ensures the quality of theata passing through the N* calculations, although itlso limits the area where the technique can be used.or example, in a daytime low-cloud case when theemperatures of the cloud cover and the surface doot differ significantly, the solar reflectance can be aetter indicator of clouds, and the N*’s calculatedrom the solar and thermal channels can be different.

The right panels of Fig. 1 show graphs that areimilar to the left panels with pixels cleared by the N*ethod and without input data random error ~rightpper panel! and with input data random error ~right

ower panel!. More than 95% of the data points fallithin the 10% accuracy range for the CO columnmounts and more than 90% for the CO mixing ratiot L-3 and L-4. CO profile retrievals at the bottomnd at levels 1, 2, and 6 have relatively larger errors,here almost 30% of pixels fail outside the 10% ac-

uracy range and some reach as high as 40–50%.When the input data random error is introduced in

he N* calculations, the magnitude of the uncertain-ies is approximately the same, with 5% more dataoints falling outside the range of the accuracy re-uirement. The anticipated error levels are addedn the same way as described in Subsection 3.A. Itan be concluded that the accuracy of the model cal-ulations, as well as that of the input parameters, isrucial to the N* process. In fact, for the referencehannels the N* cleared radiances are exactly thosealculated by the model. Nevertheless, because theeference channels are two of the A channels and theOPITT retrieval uses the thermal D channels or the

olar DyA ratio channels, the retrieved information ispproximately independent of the model calcula-ions. However, the errors introduced in the N* val-


rcktwc

nhtswIafaitsschepcesmeiPerscaatissPm

srstd1itTalteTfcctfiaflu

1

ues from the reference channel will propagate to allother channels in the form of random errors.

The clear pixels that are erroneously categorized ascloudy because of random error in the input data ~seeSubsection 3.A! will not pass the N* limit becausethere is not enough contrast in the observed radi-ances, and they will be discarded from the CO re-trieval. In addition, because the N* methodequires high contrast of cloud cover between adja-ent pixels and because of the large pixel sizes ~22m 3 22 km!, the N* technique is used mainly nearhe edges of more uniform clouds and areas coveredith broken clouds. The global coverage of N*

leared pixels is thus limited.A major disadvantage of this cloud clearing tech-

ique by use of MOPITT data alone is the relativelyigh uncertainties in the cleared radiances. Ideallyhe reference channel for the N* method can be cho-en from the channels that possess relatively narroweighting functions and are away from the surface.

n this ideal case the radiance difference between twodjacent pixels is dominated by the cloud cover dif-erences, and the cloud heights for the two pixels arepproximately the same. Unfortunately the weight-ng functions for the MOPITT radiances are rela-ively broad, and the signals are most sensitive to theurface properties. With these limitations, the as-umption of two adjacent pixels having the samelear and cloud radiative properties will not alwaysold. Because the surface characteristics of two pix-ls can be different, the radiance difference of the twoixels partially covered with clouds will include thisontribution. When the signals from these two pix-ls are used to correct for the cloud contributions, theurface characteristics of the two pixels are beingodified. Although several constraints are added to

nsure an acceptable quality, the range of applications limited. Another limitation of the current MO-ITT cloud clearing technique is the inability to prop-rly identify and treat multilayer clouds. Theesolution of multilayer clouds requires a number ofpectral bands whose weighting functions are suffi-iently narrow and peak at different altitudes in thetmosphere. Because MOPITT weighting functionsre relatively broad and the signals are most sensi-ive to the surface properties, there is not enoughndependent information from the MOPITT channelignals to remove multilayer clouds. Future re-earch is proposed in Section 6 to improve the MO-ITT cloud clearing accuracy related to the N*ethod by the incorporation of other data sources.

B. Determination of Overcast Opaque Cloud Tops

A large majority of the clouds over the globe arerelatively uniform and optically thick in the MOPITTspectral bands. It is possible to use the cloudy ob-servations to retrieve CO concentrations above cloudswhen an optically thick and uniformly distributedcloud top can be determined. After individual pixelsare identified as cloudy, three steps are taken to de-termine an optically thick overcast cloud top.

First, an equivalent cloud top pressure is obtained


for each pixel under consideration from the CH4channel DyA signals by use of Eq. ~7!. Second, apatial area of 1° 3 1° in latitude and longitude sur-ounding a pixel is selected, and a mean CH4 DyAignal and a standard deviation of all pixels withinhe area are calculated. When the ratio of the stan-ard deviation to the mean for this pixel is less than% and when there are more than 30% of the pixelsn the surrounding area having the same properties,his pixel is assigned as having uniform cloud cover.his area of 1° 3 1° in latitude and longitude containspproximately 20–25 pixels at the mid-latitudes andow latitudes and fewer at higher latitudes. Thehird step is to determine the opacity of the cloud toxclude cases with optically thin uniform cloud cover.he radiance contributed from the clouds to the TOA

or the thermal channel ch5A is modeled and thenompared with the observed radiance of the samehannel. The cloud top temperature for the simula-ion is interpolated from the DAO temperature pro-le on the basis of the estimated cloud top pressure,nd the cloud top emissivity is set to 1. If the dif-erence of the thermal ch5A signals for this pixel isess by a small amount ~0.005 in channel radiancenit!, the pixel is considered optically thick.Figure 3~a! shows the histogram of the number of

Fig. 3. ~a! Cloud cover of pixels determined as having overcastopaque tops ~shaded area! compared with all pixels in the simu-lation. ~b! Accuracy of the CO profile retrieval above opaque cloudevaluated by a comparison with the retrieval under clear skies.Orange, 210 to 0%; red, 0 to 10%; yellow, 10 to 20%, 210 to 220%;green, 20 to 30%, 220 to 230%; blue, 30 to 40%, 230% to 240%;purple, 40 to 50%, 240% to 250%.

Table 6. MAS Bands Used to Simulate MOPITT Bands ~in micrometers!

pixels detected as having overcast cloud tops for asmall area under testing in the range of 30 °S to70 °N latitude and 45 °W to 90 °W longitude duringdaytime. The shaded area indicates those pixelsconsidered as overcast cloud tops, and the area underthe dashed curve indicates the cloud fractions of allpixels included in the simulation. The pixels de-tected as having overcast cloud tops generally includemore than 90% cloud cover in each pixel. About halfof the simulated overcast cloudy pixels are detected.The MOPITT operational algorithm retrieves COprofiles at the bottom and at six vertical levels in thetroposphere. Only the pixels with cloud top pres-sures below the 400-mb level are considered, so thatthe retrievals for at least two levels, excluding the toplevel, can be obtained above the clouds. Under thiscondition, most clouds are optically thick, which elim-inates the complication raised from thin cirrusclouds. Multilayer clouds, including the cases withcirrus clouds above thick opaque clouds, were notstudied with this technique.

We evaluated the accuracy of the CO profile re-trieval above opaque clouds by comparing it with theretrieval under clear skies. Figure 3~b! shows thepercentage differences between the two runs at the350-hPa ~L-4! and 250-hPa ~L-5! levels. Approxi-mately 80% of the pixels at L-5 meet the MOPITTmeasurement requirement. At the levels too closeto the cloud tops or the TOA, the accuracy is too lowfor MOPITT products. Although the CO mixing ra-tio is available at only one level in the tropospherewhen an overcast opaque cloud is present, this willadd valuable information in the assimilation of theglobal CO concentrations.

5. Validation of MOPITT Cloud Detection Thresholdsagainst Aircraft Measurements

A. Validation against MAS Data

MAS is an imaging spectrometer flown on a NASAER-2 aircraft.19 It has 50 spectral bands rangingfrom the visible through the IR. The pixel resolu-tion is 50 m 3 50 m in spatial extent. An array of440 3 440 MAS pixels was averaged to simulate asingle 22 km 3 22 km MOPITT pixel. MAS datawere selected for this experiment from the WinterCloud Experiment33 performed over the Wisconsinregion in January and February 1997. Three MASspectral bands were used, two to match MOPITTbands and a third for visual cloud detection. TheseMAS bands are summarized in Table 6, taken from the

BandMASBand

MASPeak

MAS Width atHalf-Maximum

MOPITTCenter

MOPITTWidth

Solar CO 24 2.353 0.048 2.334 0.022Thermal CO 35 4.542 0.152 4.617 0.111IR 45 11.009 0.472 NyAa NyA

aNyA, not applicable.

Winter Cloud Experiment campaign web site ~http:yyltpwww.gsfc.nasa.govyMASywince_bands50.html!.

The flight occurred on 12 February 1997. Fromthis flight we selected track 6, which began at18:24:28 and ended at 18:53:58 Greenwich MeanTime ~GMT!. This track began over eastern Minne-sota ~clear sky!, then continued above Lake Superior.Initially the lake was ice covered with a small patchof clouds. As the flight proceeded east, the ice brokeup into mostly open water with some ice. Cloudsappeared in appreciable amounts over the lake wa-ter, and the flight was divided into 25 MOPITT pix-els; analysis was performed on the cloudy pixels overwater. An image of this section of the flight is pre-sented in Fig. 4. The pixels were numbered from 0to 24, but only the last seven were analyzed. Theseseven pixels are over the lake water, so the surfacecharacteristics are better known.

Once these pixels were identified, each was ana-lyzed independently. We estimated the cloudamount from MAS measurements using a combina-tion of methods. These methods ranged from ourcreating histograms of the data ~number of pixelsversus radiance in various bands! to subjective guess-ing. The success of various methods depended onthe contrast between cloudy pixels and the surface, aswell as factors such as the homogeneity of the under-lying surface. The average radiance of all the pixelswas calculated to simulate what the MOPITT instru-ment would have seen, and the clear-sky radiance

Fig. 4. MAS data divided into MOPITT pixels over Lake Supe-rior, 12 February 1997 at 18:45 GMT. Solco is the solar CO band,Thermal is the thermal CO band, and IR indicates a long-waveband; their properties are listed in Table 6.


a

Tl2slsTtttMbttt5

ls~bdfcsBasdbappc

papaps

MAS

1

was also estimated. The latter was estimated to bethe peak radiance of the histogram of solar CO ~i.e.,the most likely reflected radiance! and the pixel withthe highest radiance in the thermal channel. Histo-grams of a representative MOPITT pixel ~pixel 20!re presented as Figs. 5~a! and 5~b!. The histogram

of the solar CO channel illustrates that the surfaceradiance is sharply peaked and easily identifiable.This peak radiance is selected to represent the clear-sky radiance in the solar CO channel. The broad-peaked histogram of the thermal channel indicatesthat the radiances from the surface and the cloudsblend together. Thus distinguishing clouds from the

Fig. 5. ~a! Histogram of solar CO MAS pixels within MOPITTixel 20. Note that the surface radiance peak is easily identifi-ble. ~b! A histogram of thermal MAS pixels within MOPITTixel 20. Note that the surface radiance peak is easily identifi-ble. ~b! A histogram of thermal MAS pixels within MOPITTixel 20. Note how clouds form a continuous distribution with theurface.

Table 7. Average and Clear-Sky Radiances from

MOPITT PixelNumber

Thermal CO Radiance

Average MAS Clear Mo

218 0.17759 0.22719 0.17917 0.23820 0.18264 0.23421 0.18267 0.23522 0.17589 0.23123 0.17233 0.24124 0.18443 0.234


cold lake surface solely on the basis of temperature isnearly impossible. The highest radiance from a MO-PITT pixel is selected to represent the clear-sky ra-diance in the thermal channel, and these radiancesare presented in Table 7.

Clear-sky radiances were also calculated with theMOPABS model. The profile of CO2, O3, N2O, CO,and CH4 was the mean reference atmosphere used byMOPABS.24 The atmospheric profiles of water va-por and temperature were obtained from a soundingreleased as part of the Winter Cloud Experiment 7campaign, approximately 10 min after the imageswere taken from a location approximately 400 kmsouth ~089° 24.419 W, 43° 04.179 N at 19:06:34 GMT!.

he model-derived clear-sky radiances were calcu-ated with the input of the surface temperature set to70.0 K, considering a cold lake water surface withome floating ice and a lower temperature at theowest temperature sounding level ~266.96 K!. Theurface emissivity was set to 0.98 for both IR bands.he differences between the MAS clear radiances andhe MOPABS-calculated clear radiances are rela-ively large for the thermal channel. A major con-ribution is the difference in spectral bands used forAS observation ~peak wavelength at 4.542 mm! and

y the MOPABS model ~at 4.451 and 4.602 mm!. Forhe solar CO channel, both the MAS observation andhe MOPABS model use the same spectral band, andhe difference in the clear radiance is approximately.7%.This validation experiment tests the currently se-

ected thresholds for MOPITT. The ratio of ob-erved radiance to clear-sky radiance is calculatedand referred to as the ratio test!. The differenceetween the clear-sky radiance and the observed ra-iance for the thermal channel is also calculated ~dif-erence test!. The ratio and difference tests are thenompared with the predetermined thresholds, whichhould indicate whether the pixel is cloudy or clear.ecause the thresholds are made for MOPITT radi-nces, and MAS and MOPITT filter band widths arelightly different ~see Table 2!, the thresholds for theifference test are adjusted. For ratio tests, theandwidth variation is canceled when radiance ratiosre taken. The results of each of these tests areresented in Table 8. Because each of these sevenixels exceeded all thresholds, these pixels would belassified as cloudy.

Data and Model Calculations @in W~m2ysr!ymm#

Solar CO Radiance

lear Average MAS Clear Model Clear

7 0.75562 0.14 0.1487 0.7611 0.14 0.1487 0.66466 0.14 0.1487 0.69193 0.14 0.1487 1.03229 0.14 0.1487 1.02086 0.14 0.1487 0.49867 0.14 0.148

del C

0.250.250.250.250.250.250.25

awv

vE1ctm1btuaaFaT

caap

Table 8. MAS and Model Radiance Ratios ~ObservedyCalculated! and

adiom

All seven pixels were covered with clouds and de-tected as cloudy. Pixel 24, with the least cloudcover, was estimated with a cloud fraction of 15%.Therefore these threshold techniques provide a goodfoundation for detecting clouds under operational cir-cumstances. Clearly, there will be some challengingscenes that any automated algorithm will fail.These tests also provided insight into the conserva-tive nature of the selected thresholds. Under theseconditions, it is more likely that clear pixels would bedesignated as cloudy than vice versa. A bias in thisdirection will produce higher-quality data in the ini-tial stages of MOPITT production.

B. Validation against MATR Data

MATR is an airborne gas filter correlation radiometerthat measures tropospheric CO profiles with three-channel or CH4 column amounts with one channel.This instrument is designed to collect data to testMOPITT retrieval algorithms and to validate MO-PITT observations. The channel information islisted in Table 9. These channels are similar to fourof the MOPITT channels ~see Table 1!. The maindifference is that the MATR LMC’s operate with 2-and 10-mm nominal path lengths, whereas the MO-PITT LMC’s operate with 4- and 20-mm nominalpath lengths. MATR can operate at a maximum ofthree channels at a time, and MATR channel 4 usesthe same optics as MATR channel 1, except the cellcontent is CH4 instead of CO and MATR channel 4uses a different bandpass. When MATR is onboardan aircraft at an altitude of 12 km, its ground instantFOV is approximately a 1.2-km-diameter circle. Fordetails on instrument design and construction, dataacquisition, and calibration information, see Smithand Shertz.20

The MATR observation under cloudy conditions

Differences ~Calculated 2 Observed! Compared with Threshold Valuesfor Several Cloudy Pixels

PixelNumber

ThermalRatio

Solar CORatio

ThermalDifference

18 0.6910 5.1055 0.0794119 0.6972 5.1426 0.0778320 0.7107 4.4909 0.0743621 0.7108 4.6752 0.0743322 0.6844 6.9749 0.0811123 0.6705 6.8977 0.0846724 0.7175 3.3694 0.07257Thresholds ,0.93 .1.5 .0.045

Table 9. MATR Correlation R

ChannelNumber Target Gas Cell Type

C

1 CO LMC2 CO LMC3 CO PMC4 CH4 LMC

was taken on 2 March 1998 over Oklahoma. ACessna Citation II operated by the Department ofEnergy Remote Sensing Laboratory in Las Vegas,Nevada, was used for the MATR flight. The dataanalyzed were taken from 16 to 18 UTC, which isequivalent to 10:00 a.m. to 12:00 p.m. Central Stan-dard Time ~CST!. The flight track covers the area of36° N to 38° N latitudes and 96° W to 98° W longi-tudes transecting the cloud and radiation test bed~CART! site, as shown in Fig. 6 where the startingnd ending times of the track are labeled. Dataere collected in several groups, leaving short inter-als for in-flight calibrations.The cloudy and clear scenes were recorded by a

ideo camera attached on the aircraft, facing thearth’s surface. Along the track it was clear from0:00 to 10:19 a.m. local time, with scattered thinlouds between 10:19 and 10:21:10 a.m. and withhicker clouds beginning at 10:21:20 a.m. It stayedostly clear, with a few thin patches of clouds, until

0:47:42 a.m. Then the FOV’s started to be coveredy some broken clouds and transformed into consis-ent cloudy scenes at 10:48:49 a.m. It stayed cloudyntil 11:18 a.m., when it became clear. At 11:23.m. small broken clouds appeared, and at 11:30:15.m. the scenes were covered with solid clouds.rom 11:45 a.m. clouds started to break up, but inlmost all FOV’s the cloud cover was more than 50%.he FOV’s were clear again at 11:59:20 a.m.We calculated the reference clear radiances for

loud detection using MOPABS. The temperaturend water-vapor profiles were taken at the CART sitet 10:04 a.m. CST. The MOPABS mean referencerofiles for CO2, O3, N2O, CO, and CH4 are used and

described in Subsection 5.A. The surface skin tem-perature was not measured over the CART site dur-

Fig. 6. MATR flight track of 2 March 1998.

eter Channel Characteristics

r Wavelength~mm!

Filter BandFWHM ~mm!

Cell Pressure~kPa!

2.334 0.022 804.617 0.110 804.617 0.110 7.52.268 0.077 80

ente


ottatTdogercIisostbct

othsfcapwm

1

ing this time but was estimated. Based on theMATR clear radiances at 10:05 a.m., a surface tem-perature of 282 K was estimated, which is approxi-mately 8 K higher than the lowest-layer temperaturefrom the sounding. This value is reasonable be-cause in the late morning the radiative heating of theEarth’s surface is high, but the heating of the atmo-sphere is not at its peak. Based on another sound-ing taken at 2:27 p.m., there was a 4 K temperatureincrease in the lowest layer of the atmosphere from10 a.m. to 12 p.m., so a surface temperature of 286.0K is used at 12 p.m. The surface temperatures be-tween 10 a.m. and 12 p.m. are interpolated linearly.Surface emissivity of 0.92 for the thermal channeland solar reflectance of 15.5% are used in the calcu-lation. These values are based on the surface scenetypes provided by the U.S. Geological Survey and thespectral emissivity from Fu and Liou,28 as docu-mented by Wilber et al.30

Figure 7 shows the cloud detection results; the ab-scissas in all three panels represent the time of theobservation, and the ordinates represent the quanti-ties used for cloud detection. The vertical solid linesindicate the times when transitions between clearand cloudy occurred as indicated by the videotapes.The dotted horizontal lines are the thresholds foreach test. In Fig. 7~a!, the ordinate represents the

Fig. 7. Cloud detection results validated against MATR data forthree different tests: ~a! the solar channel ratio test ~observedycalculated radiance!, ~b! the thermal channel ratio test ~observedycalculated radiance!, and ~c! the thermal channel difference test~calculated 2 observed radiance!. Dotted lines are the thresholdsfor each test, and the vertical solid lines show the transition timesbetween clear and cloudy. The open circles depict those pixelsdetected as clear, and the solid circles are detected as cloudy by theMOPITT cloud thresholds.


ratio of the observed radiances over that calculatedfor 2.3-mm signals. A threshold of 1.5 is used. Thepen circles depict those pixels detected as clear, andhe solid circles are detected as cloudy by any one ofhe thresholds. Some pixels observed between 11:30.m. and 12:00 p.m. were detected as cloudy by thehermal thresholds but not by the solar channels.hese pixels may have been under broken cloud con-itions ~as verified by the videotapes! where the shad-ws affected the average radiances. Shadows on theround appear to be darker and colder than the av-rage surface, and thus, when averaged with cloudadiances, they show stronger signals in the thermalhannels and weaker signals in the solar channels.n Fig. 7~b!, the thermal ~4.6-mm A signals! ratio tests shown, and a threshold of 0.93 is used. Figure 7~c!hows the thermal difference test where a thresholdf 0.005 is used. Most of the pixels detected by theolar channel threshold were also detected by thehermal ones. It is necessary to stress that a testased on information from one channel is not suffi-ient, and all thresholds are necessary for cloud de-ection.

This validation test proves that MOPITT thresh-lds are sensitive enough to detect cloud contamina-ion in a FOV if the cloud optical depth is relativelyigh and if the clear column radiance calculations areufficiently accurate. Validations with observationsrom more difficult cloud types, such as thin cirruslouds, and complicated surface characteristics, suchs snow or ice-covered surfaces, are not currentlyossible. Future research will be directed to copeith these problems. Testing the cloud detectionethod with the CH4 total column amount will be

carried out when the MATR CH4 instrument is readyto take measurements.

6. Discussion

In this paper we summarized the current status ofthe MOPITT cloud algorithm. Based on the simu-lations, both the cloud detection and the cloud clear-ing algorithms, combined with the retrievalalgorithm, provide CO measurements approximatelywithin the required accuracy. The sensitivity of theCH4 accuracy to the uncertainties in the cloud algo-rithm will be tested at a later time. The comparisonof our cloud detection technique by use of a thresholdmethod with MAS and MATR data confirms that theMOPITT thresholds are appropriately set to detectthe differences in radiance caused by the presence ofclouds.

Once MOPITT is operational, real-time informa-tion will be collected and analyzed, making it possibleto develop algorithms based on temporal and spatialconsistency tests with CO and CH4 distribution in-formation. Future research also includes the collec-tion of the retrieved long-wave surface emissivity andits seasonal variations under clear skies and the ad-dition of information to the input data sets. Thecloud detection and clearing algorithm will be ad-justed to handle the snow or ice-covered surfaces,once real-time snow and ice information is available

crease in tropospheric methane, 1978–1987,” Science 239,
from the DAO or other data sources. Specifically,when the clear column radiance reference is esti-mated, an empirical adjustment to the input data tothe radiance model can be used.
As stated by Rossow,6 “Cloud detection methodsthat use specified clear radiances or radiance varia-tions to determine clear conditions will all have dif-ficulty obtaining a definitive separation of the cloudyand clear conditions in these ‘low contrast’ cases.”To solve the problems associated with low contrastsin our algorithm, additional information is necessary.Thin cirrus clouds show slight signatures at both the2.2-mm ~or 2.3-mm! solar and 4.7-mm thermal chan-nels. These signals are too weak to be detected bythe MOPITT cloud detection routines. The detec-tion of thin cirrus clouds in MOPITT data processingwill rely on additional data sources, and an empiricalcorrection is necessary to remove its effects. A sep-arate study will be carried out and discussed at alater time.

The MODIS cloud mask provides clear and cloudyinformation by performing a series of cloud detectiontests. The algorithm uses information from 14channels ranging from visible to long-wave thermalchannels, and the pixel sizes at nadir range from250 m to 1 km. Using multispectral techniques andtemporal and spatial tests, the MODIS cloud maskcan detect clouds in difficult situations, includingthin cirrus clouds, high clouds, low clouds, cloudsoccurring at night, and clouds over complicated sur-faces. A more realistic set of N* calculations canalso be derived by use of MODIS cloud mask prod-ucts. In each MOPITT pixel, there are approxi-mately 484 MODIS pixels, and a statistical summaryof the clear and cloudy scenes can be obtained. A N*value can be directly calculated on the basis of itsdefinition. This will eliminate the N* uncertaintiesassociated with the MOPITT signal sensitivities dis-cussed in Subsection 4.A. The results of experi-ments that use MODIS data to improve the MOPITTcloud algorithm will be included in future publica-tions.

The National Aeronautics and Space Administra-tion ~NASA! EOS Program funded this research un-der contract NAS5-30888. Meteorological data wereprovided by NASA’s DAO. The authors thank Mer-rit Deeter for providing the retrieval codes and C.Craig, L. Mayer, and C. Cavanaugh for providingMOPITT level-1 data simulations. Thanks also goto Brian Johnson and Jinxue Wang for National Cen-ter for Atmospheric Research internal reviewing ofthis paper.

References1. J. R. Drummond, “Measurements of Pollution in the Tropo-

sphere ~MOPITT!,” in The Use of EOS for Studies of Atmo-spheric Physics, J. C. Gille and G. Visconti, eds. ~North-Holland, Amsterdam, 1992!, pp. 77–101.

2. L. Pan, J. C. Gille, D. P. Edwards, and P. Bailey, “Retrieval oftropospheric carbon monoxide for the MOPITT experiment,” J.Geophys. Res. 103, 32277–32290 ~1998!.

3. D. R. Blake and F. S. Rowland, “Continuing worldwide in-

1129–1131 ~1988!.4. L. P. Steele, E. J. Dlugokenchy, P. M. Lang, P. P. Tans, R. C.

Martin, and K. A. Masarie, “Slowing down of the global accu-mulation of atmospheric methane during the 1980s,” Nature~London! 358, 313–316 ~1992!.

5. D. P. Wylie, W. P. Menzel, H. M. Woolf, and K. I. Strabala,“Four years of global cirrus cloud statistics using HIRS,”J. Clim. 7, 1972–1986 ~1994!.

6. W. B. Rossow, “Measuring cloud properties from space: areview,” J. Clim. 2, 201–213 ~1989!.

7. A. H. Goodman and A. Henderson-Sellers, “Cloud detectionanalysis: a review of recent progress,” Atmos. Res. 21, 203–228 ~1988!.

8. E. Liljas, Processed Satellite Imageries for Operational Fore-casting ~Swedish Meteorological and Hydrological Institute,Norrkoping, Sweden, 1984!.

9. P. Minnis and E. F. Harrison, “Diurnal variability of regionalcloud and clear sky radiative parameters derived from GOESdata. I. Analysis method,” J. Clim. Appl. Meteorol. 23, 993–1011 ~1984!.

10. W. B. Rossow, A. W. Walker, and L. C. Garder, “Comparison ofISCCP and other cloud amounts,” J. Clim. 6, 2394–2418~1993!.

11. G. Seze and W. B. Rossow, “Time-cumulated visible and infra-red radiance histograms used as descriptors of surface andcloud variations,” Int. J. Remote Sens. 12, 877–920 ~1991!.

12. W. B. Rossow and L. C. Garder, “Cloud detection using satel-lite measurements of infrared and visible radiances forISCCP,” J. Clim. 6, 2341–2369 ~1993!.

13. W. B. Rossow and L. C. Garder, “Validation of ISCCP clouddetections,” J. Clim. 6, 2370–2393 ~1993!.

14. S. A. Ackerman, K. I. Strabala, W. P. Menzel, R. A. Frey, C. C.Moeller, and L. E. Gumley, “Discriminating clear sky fromclouds with MODIS,” J. Geophys. Res. 103, D24, 32141–32157~1998!.

15. W. L. Smith, “An improved method for calculating tropo-spheric temperature and moisture from satellite radiometermeasurements,” Mon. Weather Rev. 96, 387–396 ~1968!.

16. W. L. Smith, H. M. Woolf, S. J. Nieman, and T. H. Achtor,“ITPP-5—The use of AVHRR and TIGR in TOVS data process-ing,” presented at the Technical Proceedings of the SeventhInternational TOVS Study Conference, Igls, Austria, 10–16February 1993.

17. M. T. Chahine, “Remote sounding of cloudy atmospheres.Part I: The single cloud layer,” J. Atmos. Sci. 31, 233–243~1974!.

18. R. Rizzi, C. Serio, G. Kelly, V. Tramutoli, A. McNally, and V.Cuomo, “Cloud clearing of infrared sounder radiances,”J. Appl. Meteorol. 33, 179–194 ~1994!.

19. M. D. King, W. P. Menzel, P. S. Grant, J. S. Myers, G. T.Arnold, S. E. Platnick, L. E. Gumley, S.-C. Tsay, C. C. Moeller,M. Fitzgerald, K. S. Brown, and F. G. Osterwisch, “Airbornescanning spectrometer for remote sensing of cloud, aerosol,water vapor and surface properties,” J. Atmos. Oceanic Tech-nol. 13, 777–794 ~1996!.

20. M. W. Smith and S. R. Shertz, “Current plans and status ofMOPITT algorithm test radiometer ~MATR!,” in Earth Observ-ing System, W. L. Barnes, ed., Proc. SPIE, 2820, 78–86 ~1996!.

21. F. W. Taylor, “Pressure modulator radiometry,” in Spectro-scopic Techniques, G. A. Vanasse, ed. ~Academic, New York,1983!, Vol. 3, pp. 137–197.

22. J. R. Drummond, “Novel correlation radiometer: the length-modulated radiometer,” Appl. Opt. 28, 2451–2452 ~1989!.

23. L. Pan, D. P. Edwards, J. C. Gille, M. W. Smith, and J. R.Drummond, “Satellite remote sensing of tropospheric CO andCH4: forward model studies of the MOPITT instrument,”Appl. Opt. 34, 6976–6988 ~1995!.


24. D. P. Edwards, C. Halvorson, and J. C. Gille, “Radiative trans-

2

2

set,” IGBP ~International Geosphere–Biosphere Project!

1

fer modeling for the EOS Terra satellite Measurement of Pol-lution in the Troposphere ~MOPITT! instrument,” J. Geophys.Res. 104, 16755–16775 ~1999!.

5. L. M. McMillin, L. J. Crone, M. D. Goldberg, and T. J.Kleespies, “Atmospheric transmittance of an absorbing gas.4. OPTRAN: a computationally fast and accurate transmit-tance model for absorbing gases with fixed and with variablemixing ratios at variable viewing angles,” Appl. Opt. 34, 6269–6274 ~1995!.

6. D. P. Edwards, “GENLN2: a general line-by-line atmospherictransmittance and radiance model: version 3.0 descriptionand users guide,” Rep. NCARyTN-3671STR ~National Centerfor Atmospheric Research, Boulder, Colo., 1992!.

27. Z. Z. Yu and J. R. Drummond, “A global surface reflectivitydata set for the 2.2–2.35 mm region,” Int. J. Remote Sens. 19,331–344 ~1998!.

28. Q. Fu and K. N. Liou, “On the correlated k-distribution methodfor radiative transfer in nonhomogeneous atmospheres,” J.Atmos. Sci. 49, 2139–2156 ~1992!.

29. A. Belward and T. Loveland, “The DIS 1-km land cover data


NewsLetter 27, 7–9 ~1996!.30. A. C. Wilber, D. P. Kratz, and S. K. Gupta, “Surface emissivity

maps for use in satellite retrievals of longwave radiation,”NASA Tech. Paper 1999–209362 ~NASA Langley ResearchCenter, Hampton, Va., 1999!.

31. D. R. Brooks, E. F. Harrison, P. Minnis, J. T. Suttles, and R. S.Kandel, “Development of algorithms for understanding thetemporal and spatial variability of the Earth’s radiation bal-ance,” Rev. Geophys. 24, 422–438 ~1986!.

32. J. Wang, J. C. Gille, P. L. Bailey, J. R. Drummond, and L. Pan,“Instrument sensitivity and error analysis for the remote sens-ing of tropospheric carbon monoxide by MOPITT,” J. Atmos.Oceanic Technol. 16, 465–474 ~1999!.

33. S. A. Ackerman, C. C. Moeller, W. P. Menzel, J. D. Spinhirne,D. Hall, J. R. Wang, H. E. Revercomb, R. O. Knuteson, E. W.Eloranta, A. W. Nolin, and M. D. King, “WINCE, a wintercloud experiment,” in Proceedings of the Tenth Conference onAtmospheric Radiation ~American Meteorological Society, Dal-las, Tex., 1999!.

Cloud detection and clearing for the Earth Observing System ......Cloud detection and clearing for...

Documents

Transcript of Cloud detection and clearing for the Earth Observing System ......Cloud detection and clearing for...