DECEMBER 1993 HOU ET AL. 833Comparison of an Experimental NOAA AVHRR Cloud Dataset with Other Observed and Forecast Cloud Datasets YU-TAI HOUGeneral Science Corporation, Laurel, MarylandKENNETH h. CAMPANA AND KENNETH E. MITCHELL National Meteorological Center, Washington, D.C. SHI-KENG YANGResearch and Data Systems, Corporation, Greenbelt, Maryland LARRY L. STOWENalionat Environmental Satellite. Data, and Information Service, Washington, D.C.(Manuscript received 16 September 1992, in final form 27 April 1993) ABSTRACT CLAVR [cloud from AVHRR (Advanced Very High Resolution Radiometer)] is a global cloud datasetunder development at NOAA/NESDIS (National Environmental Satellite, Data, and Information Service).Total cloud amount from two experimental cases, 9 July 1986 and 9 February 1990, are intercompared withtwo independent products, the Air Force Real-Time Nephanalysis (RTNEPH), and the International SatelliteCloud Climatology Project (ISCCP). The ISCCP cloud database is a climate product processed retrospectivelysome years after the data are collected. Thus, only CLAVR and RTNEPH can satisfy the real-time requirementsfor numerical weather prediction (NWP) models. Compared with RTNEPH and ISCCP, which only use twochannels in daytime retrievals and one at night, CLAVR utilizes all five channels in daytime and three at nightfrom AVHRR data. That gives CLAVR a greater ability to detect certain cloud types, such as thin cirrus andlow stratus. Designed to be an operational product, CLAVR is also compared with total cloud forecasts fromthe National Meteorological Center (NMC) Medium Range Forecast (MRF) Model. The datasets are mappedto the orbits of NOAA polar satellites, such that errors from temporal sampling are minimized. A set of statisticalscores, histograms, and maps are used to display the characteristics of the datasets. The results show that theCLAVR data can realistically resolve global cloud distributions. The spatial variation is, however, less than thatof RTNEPH and ISCCP, due to current constraints in the CLAVR treatment of partial cloudiness. Resultssuggest that if the satellite cloud data is available in real time, it can be used to improve the cloud parameterizationin numerical forecast models and data assimilation systems.1. Introduction The important role of clouds in long-range climatestudy and in short-range and medium-range forecastinghas been recognized over the past several decades bymany atmospheric researchers (Manabe 1969; Hermanet al. 1980; Ramanathan et al, 1989; Wilson andMitchell 1986; Stephens and Greenwald 1991). Anumber of cloud parameterization schemes have sincebeen developed for numerical atmospheric models(e.g., Slingo 1980, 1987; Sundqvist et al. 1989). However, incomplete understanding of cloud physical processes and lack of an accurate three-dimensional de Corresponding author address: Dr. Yu-Tai Hou, GSC/NMC, National Meteorological Center, Washington, DC 20233.scription of cloud distributions are still believed to bethe major sources of uncertainty in radiative processesaffecting model performance. To reduce this uncertainty, the National Meteorological Center (NMC) andother forecast centers are increasing their use of traditional and newly available real-time cloud observations to validate and improve model cloud-radiationparameterizations, as well as improve model initialanalyses by the four-dimensional data assimilation ofatmospheric fields inferred from cloud observations(e.g., winds, humidity, and cloud water). An experimental satellite-derived cloud dataset isbeing developed at NOAA/NESDIS (National Environmental Satellite, Data, and Information Service) tohelp satisfy these cloud-modeling requirements in realtime. The cloud data retrieval system is referred to asthe "CLAVR" [ cloud from AVHRR (Advanced Veryc 1993 American Meteorological Society834 JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY VOLUME 10High Resolution Radiometer)] algorithm. It is a sequential, multispectral decision-tree threshold algorithm using information from NOAA satellites (Stowe1991; Stowe et al. 1991 ). The algorithm is based onthe following differences between the radiative andphysical properties of clouds and the underlying surface: magnitudes of reflected and emitted radiation(contrast), wavelength dependence, and spatial variability. Because CLAVR uses all five channels in daytime retrievals and three channels at night (Stowe etal. 1991 ), it is better able to detect certain cloud types,such as thin cirrus and low stratus, compared with otheravailable global cloud analyses that commonly use onlytwo channels in daytime and one at night (Schiffer andRossow 1983; Hamill 1992). A preliminary evaluation of the CLAVR cloudproduct has been made for two ( 1 day) datasets: a winter day, 9 February 1990; and a summer day, 9 July1986. These two cases were used as test days for initialCLAVR algorithm development and testing. Independent cloud analyses from either the U.S. AirForce Real-Time Nephanalys!s (RTNEPH) or theInternational Satellite Cloud Climatology Project(ISCCP) are used in the evaluation. However, no directcomparisons are made here between ISCCP andRTNEPH because the ISCCP analyses for the case of1990 were not available at the time of this study andthe RTNEPH data for the case of 1986 were not suitably at hand "in house," though the latter are obtainable for follow-up studies from the U.S. Air Force archive described by Zamiska (1986). Data fields fromthe NMC global analysis and forecast model are alsoavailable for comparison. While the initial evaluationis restricted to total cloud amount, additional productssuch as layered cloud amount, cloud-top temperatureand pressure, cloud radiative properties, and cloud typeare planned. The next section briefly describes the cloud datasetsused in this study and the choice of spatial and temporalscales for the comparison. Sections 3 and 4 show comparisons among the datasets either with histogramanalysis or statistical scores. Section 5 displays a number of two-dimensional cloud maps from the datasets.2. Datasets The phase I CLAVR "test" cloud datasets (CLAVRI) are derived from NOAA-9 and NOAA-11 polar-orbiting satellites. They provide global total cloud distributions twice daily, once for ascending segments ofthe orbits (mostly from daytime retrievals) and oncefor descending segments (mostly from nighttime retrievals). Day-night retrievals switch at a solar zenithangle of 84.3-. The local equator crossing times areapproximately 1400-1600 for the ascending segmentsof the orbits and 0200-0400 LST for the descendingsegments. The retrieved cloud information is mappedto a longitude-latitude equal-angle grid of 0.5- resolution. During daytime retrievals, information fromall five AVHRR channels (two visible, three infrared )is used, but only infrared channels are available fornighttime retrievals (Stowe et al. 1991 ). Included inthis study are cloud amount maps for both ascendingand descending segments on the two test days, or atotal of four CLAVR-I cloud datasets. Over the polar regions, cold snow-covered surfacespose more difficulties for satellite cloud retrievals(Stowe et al. 1991 ). In a study of the U.S. Air Force3D-Nephanalysis (3DNEPH, the predecessor of thecurrent RTNEPH), Henderson-Sellers (1985) notedthat the strong temperature inversion over polar regionsprobably contributes to an apparent underestimate ofpolar clouds by satellites, relative to estimates madefrom surface observations. Furthermore, a seriousoverestimate of antarctic cloud was also noted by Stoweet al. (1989) due probably to the poor specification ofsurface temperature in the 3DNEPH database. Therefore, for the cloud data intercomparison, all data arerestricted to the region between 60-N and 60-S. Wheremissing data exist in a cloud dataset, data from all ofthe other datasets are excluded from the comparisonsat those locations. The cloud distribution diagnosed from forecastsmade by a research version of NMC's Medium RangeForecast (MRF) Model (Kanamitsu 1989 ) is also usedin the comparison. The model has a horizontal triangular truncation at wavenumber 80 (horizontal resolution of approximately 2-), and has 18 unevenlyplaced vertical sigma layers. The model cloud prediction scheme is described in detail by Campana (1990),and is a version of that developed by Slingo ( 1987 ) forthe European Centre for Medium-Range WeatherForecasts (ECMWF). Clouds are parameterized frommodel variables either as stratiform (primarily modelrelative humidity, and adjusted in the low-cloud domain by model vertical velocity and lapse rate) or ascumuliform (using convective precipitation rate).Clouds may be one of the three types: low, middle, orhigh. Their shortwave radiative properties are fixed,different for each cloud type, while the longwave emissivity of high clouds is a function of latitude. To obtain the necessary MRF cloud data for mapping into satellite orbits, the model radiation (bothlongwave and shortwave) and cloud parameterizationsare computed once at each model hour during a 72-hforecast period. A characteristic of the model forecastis that moisture and relative humidity require a numberof forecast hours to reach model equilibrium. Thisspinup time directly affects the model cloud generationscheme in middle to high latitudes and takes about twomodel days to complete (Campana 1990). Thus, onlythe cloud information from the 48-72-h forecastingperiod is used for the comparisons. The model forecastbegins 48 h prior to the desired day, a period longenough to account for much of the model spinup, butshort enough so that the forecast is still a reasonablesimulation of the atmosphere. MRF cloud data areDECEMBER 1993 HOU ET AL. 835collected from the 24 forecast hours that cover either9 February 1990 or 9 July 1986. The RTNEPH cloud dataset (Kiess and Cox 1988;Hamill et al. t992) is an automated global cloud analysis produced routinely by the U.S. Air Force. It provides real-time analyses of cloud amount, cloud type,and cloud height at 48 km (25 n mi) horizontal resolution every 3 h on two hemispheric polar stereographicgrids (true at 60- latitude). It is used for the wintercase comparison. Much of the information for theRTNEPH cloud analysis comes from the Defense Meteorological Satellite Program (DMSP) measurements;however, conventional observations and limited datafrom partial orbits of NOAA polar-orbiting satellitesare included in certain regions, such as the tropics [lessthan 5% of the total; J. Pereira (1992), personal communication]. Though the cloud analysis is updatedevery 3 h, some data points may not changedue to lack of more recent observational informatiomFurthermore, local equatorial crossing times in 1990are approximately 0600 and 1800 for one of the DMSPsatellites, and 0930 and 2130 LST for the other (Hamillet al. t992). As a result, very little visible channel information can be used because, at these times, the solarelevation is too low to be suitable for reliable clouddetection. Cloud data from the ISCCP database (Schiffer andRossow 1983; Rossow et al. 1985) are used for thesummcr case of 9 July 1986. The ISCCP data containscloud distribution on a relatively coarse, 250-km equalarea global grid, which may bc transformed into anequal-angle grid of 2.5- resolution. Information in theISCCP cloud database is updated eight synoptic timesdaily, primarily with retrievals from geostationary satellites, although polar-orbiting satellite data are usedto fill gal~s predominantly in the polar latitudes. Each of the four cloud datasets used in this studyhas its own map projection and time sampling scheme.Unification of horizontal map projection and data timeis necessary for proper intercomparisons. An equalangle latitude-longitude grid with 1- resolution is selected for the comparison, giving a total of 43 200 gridpoints in the area~between 60-N and 60-S. The 0.5-resolution CLAVR-I data are bilineady interpolated(4 to 1 average) to the new grid. Originally, cloud datafrom the MRF model are located on a computationalgrid (unevenly distributed in latitude), having approximately a 2- horizontal resolution. In this case, alinear interpolation method is used to remap modelcloud data. Special treatment, however, is made bothto avoid spreading of cloud cover near cloud edges~ ~The special treatment used in the MRF model reduces thespreading of cloud data whenever only one or two of the four gridpoints surrounding the desired location have cloud fraction greaterthan zero. The requirement is that the interpolated location must bewithin a specified radius of the cloud-covered points to be influencedby these points.and to preserve its area-weighted global mean amount.The RTNEPH cloud data are binned into 1 o boxes(60-N-60-S). The average value of each of the boxesis used to represent the cloud fraction in that area. Theremapped data are checked against the original datato ensure that the process preserves the global meanamount. The ISCCP cloud data are first transformedfrom its 2.5- equal-angle projection to a 0.5- grid projection by the replication method described in theISCCP data documentation (i.e., use the cloud amountin an original 2.5- grid box to represent each of the 250.5- subgrid boxes). This preserves large-scale datastatistical properties; however, the original coarse gridded data will not contain the high-resolution cloud details of CLAVR-I (see Fig. 9). The finer grid data arethen mapped to 1 - resolution by bilinear interpolation. To make these different cloud datasets comparablewithin a reasonable time window, while avoiding possible errors introduced by time interpolation, we decided to choose a temporal grid defined by the NOAAsatellite orbit times carried in the CLAVR-I data. TheRTNEPH and ISCCP datasets with 3-h synoptic updatetime intervals are mapped to within 1.5 h of the NOAAsatellite observing time. This mapping method, however, still presents undesirable time differences betweenCLAVR-I and RTNEPH. The latter uses mainlyDMSP satellites with local observation times 4-5 haway from the NOAA satellite sampling times (Hamillet al. 1992). More discussion on this issue is givenlater. The MRF model cloud dataset, which has a 1-htime interval, is mapped to the NOAA satellite observing time within half an hour.3. Cloud distributions--Histogram analysis Figure 1 shows histograms of global total cloud distributions for ascending segments of the orbit for (a)9 February 1990 and (b) 9 July 1986, and descendingsegments of the orbit for (c) 9 February 1990 and (d)9 July 1986. The fraction of cloud cover is categorizedby a 0.1 (i.e., 10% ) interval, except that a 0.05 (or 5%)level is used at the clear-sky and overcast ends. In thefigure, cloud distributions from the CLAVR-I,RTNEPH, ISCCP, and MRF model are labeled asCLV, RTN, ISC, and MRF, respectively. The daytimesatellite cloud data are more reliable than nighttimebecause visible channel information is included in thecloud retrieval. Thus, most comparisons presented inthis study will emphasize the ascending cases. The appearance of a cloud distribution histogram isa ftmction of the resolution of observations. At twoextreme cases, for example, the cloud histogram willshow only two spikes at clear and complete cloudyconditions for point observations, or will show onespike peaked at about 0.5 value for global mean observations. In general, satellite-observed cloud histograms are found between the two extreme examples,and commonly show a U-shaped distribution. In the836 JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY VOLUME 10 50 50 J NORR-11 RSCENDING SEGMENTS FEB 9, 1990 60N - 605 GLOBRL ARER / ~,o MERN VRLUE t,~0~ MRFg O.BS9 VALID OATR POINTS 430'76~ 30~J -ill RTN ~ 0.808 - 05-.15 .25-.35 .~5-.55 .65-.75 .85-.95 0.0-,05 .15-.25 .35-.~5 .55-.65 .75-.85 .95-1.0 CLOUD CRTEOORY ~c] 50 150 NOAA-11 DESCENOIN~ SEDHENTS FEB 9, 1990 60N - 605 GLOBAL AREA J / 40 MEnN VRLUE tt0D MRFJ-1 O.SSO VnLm mT~ POINTS 43200 / ~~/ ~~ CLV~ 0,560 MISSING DRTR POINTS 0 / ~8 so ~ ~o ~: ~m ~ o.s.~ ~~ ~ zo ~o ~: . 0 ' ~ m , m , m ~ m , ].05-. 15 .25-.35 .'~5-.55 .65-.75 .85-.950.0-.05 . 15-.25 .35-.45 .55-.65 .75-.85 .95-1.0CLOUD CRTEGORYso[ 150J NORA-9 ASCENDING SEGMENTS JUL 9, 1986~OI MERN VRLUE J~ CLV~ 0.556 MISSING anTn POINTS Sn61 J3-11 ISC~ -'847 ~13- Ol, .05-. lS .25-.35 .ti5-.5S .85-.75 .85-.95 0.0-.05 .15-.25 .35-.~5 .55-.55 .75-.85 .95-1.0 CLOUD CATEGORY(0)50 NORR-9 DESCENDING SEGMENTS JUL 9, 1986 6ON - 605 GLOBRL RRER MERN VRLUE ,~0 MRF 0 0.338 VRLID DRTR POINTS 39841 ~ CLV~ 0.531 MISSING DRTR POINTS 9559 0.05-. 15 .25-.95 .45-.55 .65-.75 .85-.950.0-.05 - 15-,25 . 35-. '~5 - 55-. 65 . 75-. 85 .95-1.0CLOUO CATEGORY lC'lG. 1. Histograms of global (60-N-60-S) total cloud distributions. Data for MRF model, CLAVR, RTNEPH, and ISCCP are labeledas MRF, CLV, RTN, and ISC, respectively. Data are mapped on: (a) ascending segments of the orbits of 9 February 1990, (b) ascendingsegments of the orbits of 9 July 1986, (c) descending segments of the orbits of 9 February 1990, and (d) descending segments of the orbitsof 9 July 1986. Mean values give area-weighted mean cloud amounts for the corresponding datasets.figure, each cloud dataset displays a unique cloudamount distribution histogram. All the datasets showdepartures from the U-shaped histogram distribution.The MRF model cloud data are skewed toward thecloud-free condition (cloud fraction less than 0.05),indicating that the current model cloud parameterization scheme generates too little cloud cover. Almost35% of the grid points are covered by little or no clouds,and less than 7% of these are covered completely byclouds (cloud fraction greater than 0.95). The CLAVR-I cloud amount distribution, on theother hand, peaks in the center of the diagram, at approximately 0.5 cloud fraction. The CLAVR-I algorithm is primarily designed to separate cloud-freeAVHRR GAC (global area coverage) pixels from allothers. To estimate cloud amount, cloud fractions of0%, 50%, and 100% are assumed, respectively, for eachof the three cloud classifications (cloud-free, mixedcloudy, and cloudy) assigned to each of the 2 x 2 arraysofGAC pixels (Stowe et al. 1991 ). As seen in the figure,more than 20% of the grid points have cloud fractionin the range of 0.45-0.55, in contrast to less than about8% of total occurrences displayed in this range by theother cloud datasets. This occurs because "mixedcloudy" is the predominant classification from theCLAVR-I algorithm. Future phases of the CLAVR alDECEMBER 1993 HOU ET AL. 837gorithm will incorporate energy balance methods (e.g.,Coaldey and Baldwin 1984) between cloudy and clearpixcl radiances and time series analyses of previousday's clear-sky radiance statistics to provide a very precise separation of these mixed-cloudy arrays into aclear, partly cloudy, or cloudy classification for eachpixel. The RTNEPH total cloud amount distributionsshown in panels (a) and (c) of Fig. 1 display a biastoward the overcast condition, with more than 30% ofthe global total occurrences having cloud fractiongreater than 0.95. This may result from the U.S. AirForce's design philosophy, which attempts to maximizethe probability ofdetecting cloud (Hamill et al. 1992).The ISCCP cloud amount distributions shown in panels (b) and (d) of the figure display a pattern similarto the RTNEPH data, except for slightly lower frequencies of occurrence at both the high and low endsof the cloud fraction scale. Considering the differences in spatial resolution and season, RTNEPH andISCCP datasets are generally in good agreement. Areaweightcd mean values of global total cloud amount arealso shown in the figures for each of the comparisoncases. The M1LF model cloud data produces global meancloud values less than 0,36,2 while the CLAVR-Idata produces global mean cloud values in the rangeof 0.53-0.56. On the other hand, the values of globalmean cloud fraction are between 0.59 and 0.65 for theRTNEPH cloud data and the ISCCP cloud data. Each cloud dataset also has different cloud amountdistributions over land and ocean regions. For example,for ascending segments of the orbit, Figs. 2a and 2cshow cloud amount distribution histograms over landand ocean, respectively, for 9 February 1990, and Figs.2b and 2d show similar histograms for the case of 9July 1986. In general, all the cloud datasets show atendency for greater cloud cover over ocean than overland. The three satellite cloud distributions show moreovercast eondilions and less cloud-free conditions overocean than over land. For MRF model clouds, differcnccs in cloud coverage between ocean and land aremore evenly spread over all cloud fraction categories.l lowever, significantly less cloud-free conditions areobserved over the oceans. Regardless of differences incloudiness, satellite cloud-retrieval algorithms arethemselves different over land and ocean backgrounds.These differences suggest that improving a model cloudscheme (tuning), based on real-time satellite data,should be done separately for land and ocean.4. Cloud statistical scores To compare statistical characteristics of the clouddistribution for the four cloud datasets, we adapt andextend the computation of cloud comparison skill 2 The global-mean cloud value of the current operational T126model is approximately 0.42, a value still too low.scores applied at Air Force Global Weather Central(AFGWC) (Trapnell 1992). First, a series of contingency tables (or matrices) and chance tables are calculated for each pair of cloud datasets being compared.Each entry in a contingency table (or a chance table)represents the number of occurrences of cloud (or thechance of cloud occurrence in the chance table) in 1of 11 cloud fraction categories for each set. In additionto a global region (60-N-60-S), tables are presentedfor a set of somewhat arbitrarily selected regions inorder to assess statistical characteristics associated withchanges in latitude, in scene background (land, ocean),and in geographic regions favoring formation of certaincloud types. From the tables, comparison skill scoresare then calculated. We use seven scores to show statistical characteristics of cloud datasets. Six of them areselected from Trapnell (1992), and the seventh, theS-60 score, is added here as a natural counterpart toone of the six. Considering the lack of accessibility ofthe cited report, and the lack of an in-depth explanationof scores therein, we include a detailed description ofthe scoring system in the Appendix. Tables 1-4 show selected comparison scores for several specified regions for ascending segments for 9 Feb~ruary 1990 and 9 July 1986, and descending segmentsfor 9 February 1990 and 9 July 1986, respectively. TheS2o score gives an estimate of how well two cloud datasets agree. It is the percent of points where the cloudamount in the two cloud fields differ by less than approximately 20% (i.e., percent of agreement). The scorevalue ranges from 0 to 1. The larger the score, the betterthe agreement. The S-60 score, on the other hand, givesan estimate of how poorly the cloud distributions agree.It is the percent of points where the cloud amount inthe two cloud fields differs by more than 60% (i.e.,percent of disagreement). With the score value rangingfrom zero to one, a large S-6o value indicates largedisagreement. The bias score Sbias is used to compareoverall cloud amounts between two cloud datasets. Alarge value (either positive or negative) indicates thatthere is a large difference in overall cloud amount. Thesign of Sbias score shows the direction of the overallbias. Another very useful score is called the Heidikescore &,. The score value, ranging from 0 to 1, measureshow closely two compared cloud datasets are relatedto each other. The larger the value ofSh, the less likelyit is that the two cloud datasets are statistically independent of each other. A large Sh score usually impliesthat most of the other scores will indicate favorableagreement between the datasets. Three additionalscores are included in the table. They are the rootmean-square (rms) error Srms, the mean absolute errorSabs, and the correlation score Scott, respectively. Sincethese are commonly used in many other applicationswith the same statistical meanings, no further descriptions are given here. Scores shown in Tables 1-4 consist of results fromascending and descending segments of the orbit for838 JOUR,NAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY VOLUME 105OqO~ scoomklJ~2cn 1[5o90 NORA-11RSCENOING SEGMENTS PE8 9, 1990 I 60N - 6OSGLOORL LAND POINTS MEAN VALUE MRF~ 0.319 VALID DATA POINTS 11420 CLV8 0.488 MISSING DATA POINTS 29 RTN ~ O. 597 ,~l,lil,F~,F .05-.15 .25-.35 .45-.55 .65-.75 .85-.950.0-.05 .15-.25 .35-.45 .55-.65 .?5-.85 .95-1.0 CLOUD CATEGORY20lO 1NORA-9MRF 0CLV ~ISC ~ ASCENDING SEGMENTS JUL 9, 198660N - 605 GLOBAL LAND POINTSMEAN VALUE 0.253 VRLIO DATA POINTS 10786 0.427 MISSING DRTR POINTS 6630.5365O4O3O .05-.15 .25-.35 .45-.55 .85-.75 .85-.950.0-.05 .15-.25 .35-.45 .55-.65 .75-.85 .95-1.0 CLOUO CATEGORY5O 4O~ 2c 1c(c)NORA-11 ASCENDING SEGMENTSFEB 9, 1990 60N - 605 GLOBAL OCEAN POINTS MEAN VALUEMRF0 0.363 VALID DRTR POINTS 31656CLV~ 0.54? MISSING DRTR POINTS 95RTN~ 0.619 .05-.15 .25-.35 .45-.55 .65-.75 .85-.950.0-.05 .15-.25 .35-.45 .55-.65 .75-.85 .95-1.0 CLOUO CATEGORY5O3O20 20 o5-m~ NORA-9 ASCENDING SEGMENTS JUL 9, 1986 J 60N - 605 GLOBAL OCEAN POINTS40 MEAN VALUE MRF0 0.364 VALID DRTR POINTS 26953 CLV~ 0,606 MISSINO DRTR POINTS 4798SO ISC~ 0.6780.0-.055O lO ,FI,~ , o.05-.15 .25-.35 .45-.55 .65-.75 .85-.95 .15-.25 .95-.45 .55-,65 .75-.85 .95-1.0 CLOUD CRTEGORY FIG. 2. Histograms of global (60-N-60-S) total cloud distributions. Data for MRF model, CLAVR, RTNEPH, and ISCCP are labeledas MRF, CLV, RTN, and ISC, respectively. Data are mapped on: (a) land areas of ascending segments of 9 February 1990, (b) land areasof ascending segments of 9 July 1986, (c) ocean areas of ascending segments of 9 February 1990, and ( d ) ocean areas of ascending segmentsof 9 Jury 1986.both the winter case of 9 February 1990 (RTNEPH,CLAVR-I, and MRF datasets) and the summer caseof 9 July 1986 (ISCCP, CLAVR-I, and MRF datasets).In each of the cases, three cloud datasets are comparedto each other in pairs. For each pair, statistical scoresare calculated over the global region. Additionally,scores for two smaller regions are given in the table,which shows relatively good and bad examples ofagreement. Generally speaking, results from comparisons of theCLAVR-I cloud data to either the RTNEPH cloud dataor the ISCCP cloud data [section (a) in the tables] arebetter than results from comparisons of the MRF modelclouds to any of the observational cloud databases[sections (b) and (c) in the tables]. In the global region,for instance, values of S20 score range from 0.547 to0.677 for comparisons between the CLAVR-I clouddata and the RTNEPH or ISCCP data. Those valuesare higher than the 0.378-0.426 score values obtainedfrom model cloud comparisons with any of the observational cloud databases. One of the main reasons forthe low MRF scores is attributed to the large amountof clear sky in the forecast model. The S2o score valuesassociated with the CLAVR-I cloud data, however,DECEMBER t993 HOU ET AL. 839 TABLE 1. Selected statistical scores computed from pairwise comparisons among the CLAVR, RTNEPH, and MRF model total clouddatasets. Data are mapped on NOAA-tl ascending segments of theorbits of 9 February 1990. For each comparison pair, scores arc computed over global area (60-N-60-S). Scores for two selected smallerregions are also included to show examples o- relatively good andbad comparisons. (a) RTNEPH vs CLAVR Western NorthScore Global Europe AmericaS20 0,573 0.638 0.422S-~0 0.062 0.037 0.269$'~ 0.304 0.263 0.452Sb~ 0.062 0.067 -0.233Sa~ 0.234 0.225 0.361Scoff 0.547 0.591 0.242S~ 0.334 0.362 0.141 (b) MRFT80 vs CLAVRScore Global Australia East Asia520 0.402 0.495 0.271S-6o 0.188 0.118 0.454Srms 0.412 0.349 0.549Su~s -0.173 -0.114 -0.379Sabs 0.335 0.274 0.471S~o~, 0.255 0.430 O. 113Sh 0.! 13 0.221 0.019 (c) MRFT80 vs RTNEPH Western North TropicalScore Global America Oceansrelatively large for the MRF model cloud data comparisons, which, in general, supports the preceding discussion. When comparing data between RTNEPH andCLAVR-I, calculated values for the global region aregenerally 0.5 or better for $~orr, and 0.3 or better forSh. Even higher scores are yielded when comparing theISCCP cloud data to the CLAVR cloud data, resulting in corresponding score values of 0.7 and 0.46.Although these scores are significantly lower than theideal values of one, they have twice the magnitudesobtained in comparisons involving the MRF modelcloud data. Despite the fact that neither the RTNEPHnor the ISCCP database represents absolute "truth,"the comparison indicates that the CLAVR-I data is inreasonable agreement with these independent clouddatasets. Future phases of the CLAVR development,which will separate mixed-pixel classification into a TABLE 2. Selected statistical scores computed from pairwise comparisons among the CLAVR, ISCCP, and MRF model total clouddatasets. Data are mapped on NOAA-9 ascending segments of theorbits of 9 July 1986. For each comparison pair, scores are computedover global area (60-N-60-S). Scores for two selected smaller regionsare also included to show examples of relatively good and bad comparisons. (a) ISCCP vs CLAVR Eastern Western NorthScore Global Pacific America S20 0.677 0.813 0,570$20 0.416 0.560 0.377 5_60 0.021 0.005 0.027S~60 0.311 0.211 0.329 S~ 0.244 0.187 0.272Srms 0.488 0.404 0.509 Su~ 0.092 -0.016 0.183Su~ -0.235 -0.171 -0.238 S~ 0.185 0.141 0.226Sa~s 0.380 0.287 0.407 S~or* 0.700 0.788 0.568Scor~ 0.249 0.532 0.110 S~ 0.465 0.669 0.209$'~ 0.124 0.311 0.050show only moderately good agreement with the other Scoreobserved cloud databases, due to the Pact that cloudfractions from the CLAVR4 algorithm are heavily S2obiased toward 0.5 from frequently occurring mixed- s-60pixel array classifications. Nevertheless, the $-60 scores Srm~ Sbiassuggest that the geographical location of clouds detectedby the CLAVR-I algorithm is generally in excellentagreement with the other observations. In contrast, theS--6o scores are much higher in those comparisons withmodel-generated clouds, implying that model cloudsdo not match well with the satellite-derived values.spite the fact that CLAVR-I clouds have a different Scorecloud amount distribution than either the RTNEPHor ISCCP datasets (recall Figs. I and 2), values of the &0 S 60Su~s score are quite small for most of the cases. On theother hand, negative score values prevail when comparing model clouds with the other datasets, showingthat the current model cloud scheme severely underestimates cloud amount. Values of Srm~ and S~b~ are (b) MRFT80 vs CLAVR CentralGlobal AsiaIndian Ocean0.422 0.591 0.3270.185 0.082 0.2620.408 0.302 0.468-0.179 -0.101 -0.2700.326 0.224 0.3910.307 0.462 0.0790.132 0.298 0.044(c) MRFT80 vs ISCCP SouthernGlobal Europe Atlantic0.378 0.464 0.3260.322 0.219 0.4460.494 0.428 0.579-0.293 -0.299 -0.4090.396 0.326 0.4720.258 0.549 0.1140.092 0.222 0.018840 JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY VOLUME 10 TABLE 3. Selected statistical scores computed from pairwise comparisons among the CLAVR, RTNEPH, and MRF model total clouddatasets. Data are mapped on NOAA-11 descending segments of theorbits of 9 February 1990. For each comparison pair, scores are computed over global area (60-N-60-S). Scores for two selected smallerregions are also included to show examples of relatively good andbad comparisons. (a) RTNEPH vs CLAVR Eastern NorthScore Global Australia AmericaS20 0.547 0.604 0.453S-60 0. 100 0.066 0.250S~s 0.335 0.291 0.439Sbias 0.022 0.022 --0.127$~b~ 0.252 0.214 0.336S~o~ 0.492 0.666 0.202S~ 0.301 0.391 0.098 (b) MRFT80 vs CLAVRScore Global Europe East AsiaS20 0.407 0.498 0.287~60 0.211 0.180 0.490S~r~s 0.430 0.376 0.607S6~as -0.193 -0.036 -0.400Sabs 0.343 0.288 0.510S~o,~ 0.281 0.478 0.031Sh 0.125 0.248 0.006 (c) MRFTS0 vs RTNEPH Western North TropicalScore Global America Atlanticover the same region but for the summer case of 9 July1986. Comparisons among CLAVR-I, ISCCP, andMRF model cloud datasets display three different cloudamount distributions, which act to degrade the statistical scores. Differences generally are greater for comparisons involving the MRF model clouds. As shownin Tables 1-4, scores related to the model cloud datain some "bad" examples are so poor that the Sh valueindicates that the two referenced cloud distributionsare almost totally unrelated (e.g., Sh score lessthan 0.1 ). Further inspection of the tables shows that theCLAVR-I cloud distribution agrees better with theISCCP distribution than the RTNEPH distribution.As an example, correlations between ISCCP clouds andCLAVR clouds are fairly high, giving Scott valuesaround 0.7 or higher, in contrast to a value less than TABLE 4. Selected statistical scores computed from pairwise comparisons among the CLAVR, ISCCP, and MRF model total clouddatasets. Data are mapped on NOAA-9 descending segments of theorbits of 9 July 1986. For each comparison pair, scores are computedover global area (60-N-60-S). Scores for two selected smaller regionsare also included to show examples of relatively good and bad comparisons. (a) ISCCP vs CLAVR Northern midlatitude SouthernScore Global land Atlantic S2o 0.669 0.757 0.571S20 0.422 0.551 0.31 I 8-60 0.031 0.016 0.035S-6o 0.321 0.220 0.457 &ms 0.254 0.21 I 0.284Srm~ 0.492 0.410 0.579 Sb~s 0.060 0.027 0.118S~ -0.215 -0.104 -0.266 S~, 0.190 0.152 0.214S~ 0.381 0.289 0.482 S,o~ 0.697 0.795 0.282S~o~ 0.250 0.430 -0.152 S~ 0.472 0.595 0.146S~ 0.128 0.262 -0.067precise spectrum of cloud fraction, may help to reduce Scorethe difference.Over the two smaller regions, scores change greatly &ofrom "good" to "bad." For instance, in section (a) of 8-60Table 1, comparison of the RTNEPH to the CLAVR SbiasI cloud data over the European region shows that theS20, S-6o, Sbias, and S~orr scores are all in the good range &o=of 0.638, 0.037, 0.067, and 0.591, respectively. Thismay result from large areas of layered cloud. Over thewestern part of North America, however, most of thescores are poor, especially for the $-60 and Sbias scores.Figure 3a shows cloud distribution histograms over this Scoreregion for the case of 9 February 1990. As seen in thefigure, most of the CLAVR-I cloud data are in large S-~0cloud amount categories (cloud fraction greater than &m~0.5 ), whereas a considerable portion of the RTNEPHdata is defined as clear sky or near-clear sky. Note that Sa~better agreement exists between the RTNEPH data and &o~rthe MRF model data. Figure 3b shows another example (b) MRFT80 vs CLAVR Central IndianGlobal Asia Ocean0.406 0.594 0.3250.191 0.152 0.2190.414 0.354 0.456-0.198 -0.121 -0.2330.334 0.248 0.3810.299 0.477 0.0380.120 0.304 0.012(c) MRFT80 vs ISCCPGlobal IndianEurope Ocean0.426 0.589 0.3050.293 0.153 0.3880.472 0.349 0.536-0.241 -0.088 -0.3520.366 0.243 0.4470.291 0.451 0.1140.131 0.284 0.024DE('EMBER 1993 HOU ET AL. 841 (R) ~Ott'q' NORR-tl RSCENgINO SEGMENTS FEB 9, I990 II NEST NORTHERN RMERICR REGION ,t0 tl MEAN V~LUE~ ~ ,RF~ 0.,88 VRLID ORT~ POINTS 1~11~ II CLVB 0.626 MISDING DRTR POINTS O~0~_.,~ o~o H ~m ~ .n i[ ~ o ....(B)1239 72NORR-9 RSCENDING SEGMENTS JUL 9, 1986NEST NORTHERN RHERICR REGION MERN VRLUEMRF0 0,26? VRLID DRTR POINTSCLVB 0.520 MISSING DRTR POINTSISC~ 0.7q7.05-.15 .25-.35 .~5-.55 .15-.25 .75-.85- 65-. 75 .85-. 95 0 0 .05-.15 .25-.35 .q5-.55 .65-.75 .85-.950.0-.05 .15-.25 .35-,q5 .55-.65 .75-.85 .g5-1.o 0.0-.O5 .35-.~5 .55-.65 .gS-hO CLOUO CRTEGOR- CLOUD CRTEGOR-5O3O2O10 Fi(h 3. ttistograms of total cloud distributions over the western part of the North American region. Data for MRF model, CLAVR,RTNEPH, and ISCCP are labeled as MILV, CLV, RTN, and ISC. respectively. Data are mapped on: (a) ascending segments of 9 February1990, and (b) ascending segments of 9 July t986.0.6 for the correlation with RTNEPH clouds- A possine explanation for the differences in score values issatellite observation time. Except for relatively fewconventional observation points, most RTNEPH observations come from DMSP polar-orbiting satellitesthat typically have 0600-t800 or 0930-2130 localcrossing times and rarely coincide with the data's assigned synoptic times (Hamill et al. t 992). ISCCP datacontains cloud information mostly from 3-h geostationary satellites with some data from NOAApolar-orbiting satellites to fill gaps. As a result, theCLAVR-I cloud distribution, which comes solely fromNOAA afternoon polar satellites, agrees better with theISCCP cloud distributions than with RTNEPH data.5. Two-dimensional cloud distributions Figures 4 and 5 show maps of total cloud distributionfor ascending segments ofthc orbit on 9 February 1990and 9 July 1986, 3 respectively. Panel (a) displays cloudamount from the CLAVR cloud dataset, panel (b)from the RTNEPH (or ISCCP), and panel (c) fromthe MRF model. As expected from the statistical analyses, when CLAVR is compared to the RTNEPH (orISCCP) cloud distribution, it shows less spatial variability, especially over tropical latitudes. Nevertheless, 3 Two missing segments of satellite orbit are found in the 9 July1986 display (Fig. 5). The one crossing the Atlantic and northernAIkiea is from the CLAVR-I dataset, which also masks out the samearea in other datasets with mapping processes. The other missingsegment (crossing the Pacific Ocean) comes from 1SCCP data. Datapoints within these areas are eliminated from comparisons.locations of the most prominent cloud structures inthe two satellite-retrieved distributions coincide wellwith each other, resulting in very small S-60 scores inTables 1-4. Despite less cloudiness over most of theglobe by comparison, the MRF model cloud schemegenerates reasonably well defined cloud patterns insome regions, especially along the midlatitude stormtracks. Thus we believe that further improvement tothe model cloud parameterization can be made throughtuning it with a set of cloud observations. Disagreements are also obvious in other regions, especially inlow latitudes where model convective clouds becomedominant. There, the MRF model clouds do not showlarge-scale organized cloud patterns. In the currentMRF model cloud scheme, there are two types of generation mechanisms (Campana 1990). Stratiformclouds are diagnosed from the model moisture distribution, which generally follows the model's large-scaleairmass movement. Convective clouds, on the otherhand, are computed from the model convectionscheme (Kanamitsu et al. 1989), which dependsstrongly on the local surface conditions and the staticstability of the air mass within a vertical column. Themodel cloud scheme allows only one type of cloud permodel point, and the convective mechanism takes precedence over the stratiform one. Thus, the model convective cloud clusters prevail over low-latitude regions,where the localized convection process is usually veryactive. Figures 6-8 show some detailed examples of thethree cloud data distributions over different regions ofthe Pacific Ocean. Figure 6 displays clouds over thenorthern midlatitude Pacific Ocean. In the figure, all842 JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY VOLUME 10 (R) CLFIVR TOTRL60N~30N-'~~'~-! O ": "'"'":"~30S~605 ' ;':'~ l~O CLOUD NORR-1i RSCENDING SEGMENT FEB 9, 1990 60N F~/!BCE!~lr-" F'~I ~''' ' "~"""~ ''~' ' '--~' - '"' ' .,~~, .,.,.~ON ~f-' -~?~ - . n '~]~~ ~,,'~' ' ~~,'~,~ ~~/:. ;~.~,,,..::~ ............ ~~~"~.~~~ ....... ~o~ 60S150~ 120~ 90~ 60~ 30~ 0 30E 60E 90E 120E iSOE t80 :'=::~.:~i~iii ' . MISSING DATA 0 10 90 50 70 90 tO0 8) RTNEPH TOTAL CLOUD NORA-t1 ASCENDING SEGMENT FEB 9, 199060N 60N30N 30N0 0309 305605 60S 180 tSON 120N 90N 60N 30N 0 30E 60E 90E 120E 150E t80 i.:~ :.... ::...i:~i~il : MISSING ORTR 0 10 30 50 70 90 100(C) MRFTSO TOTAL CLOUD NORA-11 ASCENDING SEOHENT FEB 9, 1990 030530S60560N 60N3ON 30N 0605 180 1501,4 120i4 90N 60Fl 30N 0 30E 60E 90E 12OE 150E 180 "--:::.....:':~. '..::' .:i: MISSING DATA O 10 30 SO 70 90 tO0 FIG. 4. Comparisons of global (60-N-60-S) total cloud amount distribution. (a) CLAVR total clouddistribution, (b) RTNEPH total cloud distribution, and (c) MRF model total cloud distribution. Data aremapped on NOAA-11 satellite ascending segments of 9 February 1990.three show large frontal cloud systems in the easternpart of the ocean. The shape and location of the cloudsystems produced by the MRF model cloud generationscheme match quite well with the other two satelliteobservations. In the central and western parts of thenorthern Pacific, both the MRF and CLAVR-I cloudsdisplay less cloudiness than the RTNEPH. Nevertheless, the agreement between the MRF model and satellite observations is still encouraging. Figure 7 showsa similar comparison for the southern Pacific Ocean.Again, the MRF model cloud scheme compares favorably with observations in predicting large-scalecloud structures. In fact, clouds from the model presentan amazingly look-alike large-scale wave pattern acrossthe southern Pacific Ocean, although agreement withobservations is less over the western third of the area.DECEMBER 1993 HOU ET AL. 84360N30S~0S 1DO~//,-,~/~z/~ :' ~1~ -_,,~i--?~/'/) ,;.': .r. ,~.:: ::4. -. ~-,--/~!.- !: ':E ' ":.iiii': tSON 120W 90N 6ON 30N(A) CLRVR TOTAL CLOUD NORA-9 ASCENDING SEGMENT JUL 9, 1986 60N.. ....:i~-'~~'~'~~,~-:... ~.... .~~o~ " 0 0 30E 60E OE 120E iSOE 180 '-- '""-.---:'~- :.~i;i~ MISSING DRTR 0 t0 30 50 ?0 90 10060N305805 180B) ISCCP TOTAL CLOUD NORA-9 RSCENDING SEGMENT JUL 9, 1986 ... 150N 120N 90N 80N 30N 0 30E 60E 90E 120E ISOE 180 E::i:~i!:2i!i, .: MISSING OBTB 0 10 30 50 70 0 tOO(C) NRFT80 TOTAL CLOUD NORA-9 ASCENDING SEGMENT JUL 9, 1986....,,.,,. ~i~..~......~ ..... ~::~,.,.~:,~:..:~: .-~?.:-~ ....... ~o~':/..:~ '~:: ~~ :~ :~. ~s ,.:..~,~ ......... -~ '?" ' .... ' ..... :, ,.~,~.~ ..:~ (:;~,.':~',:.:,~ ..;. , ~ ;~ ,'.'*~..,,;. ~,~. - - ?, '.%..,~.,,.,.,,~,,,~,:.~, ~.,,, ~'"" ~::~g ;..:.~"'" ..... :~...~ :;:.C-:'~ ~:~:;,..;,~2~"&',?-,-', .:,,;~1,- '.,-, ... , ,~"-'" -'. ..... - ,,~/'~' '''-' ~--., ' ~ e',:' ~ 1 ~,:,~.. ~'~;;:'~ ;:~ s~-~ ~..::~ ~'~ ~ - ~ ~ :~:~,:~ ~8ON30N 0 30N 330S 30S605 605 180 lSON 120N 90N 80N 30N 0 30E 60E 90E 1ZOE 150E 180 '-':....::-::~!!~!!!. ~ - HISSING ORTR 0 10 30 50 70 90 tOO FiG. 5. Comparisons of global (60-N-60-S) total cloud amount distribution. (a) CLAVR total clouddistribution, (b) ISCCP total cloud distribution, and (c) MRF model total cloud distribution. Data aremapped on NOAA-9 satellite ascending segments of 9 July 1986. Even though model clouds agree well in cloud pattern, this does not necessarily result in good comparisonscores. For example, over the southern Pacific Ocean,where the model cloud scheme appears quite reasonable, comparison between the model data and theRTNEPH data yields S20, S-6o, &ms, and Scott scoresof 0.448, 0.272, 0.453, and 0.257, respectively. In contrast, comparison between the two satellite cloud databases yields score values of 0.579, 0.049, 0.285, and0.551, respectively. Over the tropical Pacific Ocean (Fig. 8), however,agreement between cloud datasets becomes dramatically worse. Though the CLAVR-I cloud map showsless contrast than the RTNEPH cloud map, the similarities in large-scale cloud distribution can still be seen.On the other hand, the MRF model cloud scheme is844JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY (R) CLRVR TOTAL CLOUD NOAR-i1 RSCENDING SEGMENT FEB 9, 1990 150E 185E 180 165N 150W 135H 120N 0 10 30 50 70 90 100VOLUME 1060NqSN30N(B) RTNEPH TOTRL CLOUDNOAR-11 ASCENDING SEGMENT FEB 9, 1990 50NqSN30N150E 165E 180 165N 150H 135N . 120H 0 10 30 50 70 90 100 (C) MRFT80 TOTRL CLOUD NORA-11 ASCENDING SEGMENT FEB 9, 199060N 60N qSN BON150E 165E 180 1BSN 150W 135N 120N 0 10 30 50 70 90 100F]G.6. Similarto Fig. 4 but ~rthe no,hem midlatitude Pacificregion.not able to produce a similar cloud distribution pattern.Clouds from the model become scattered clusters, diagnosed from model convective processes. Despite theoverall underestimate of cloud amount by the MRFmodel, these areas that strongly disagree will certainlybe a major reason for degraded comparison scores(Tables 1-4). Figure 9 shows an example for 9 July 1986 over theeastern Pacific region. Some missing data points arefound in the ISCCP cloud database (mainly west of120-W), but that does not affect the following discussion. Despite the fact that the ISCCP data has a coarserhorizontal resolution (2.5-) than the CLAVR-I data,comparison of the two satellite cloud datasets showsthat cloud patterns and locations are in good correspondence, except over the western United States. Although the CLAVR-I cloud map shows less spatialvariability, the comparison scores are impressive, givingDECEMBER 1903 ttOU ET AL. 845 IR) CLRVR TOTnL CLOUO NOnR-11RSCENDING SEGMENT FEB 9, 1990 ".'::La~. - ...... - -~, ? -:..~ '" :i: .::~//.xx' .: ...... ' '"~55GOS 150E 90W3O516bE 180 165N 150W 135W 120N 105N F----Y':' 0 10 30 50 70 ~0 100455(B) RTNEPH TOIRL CLOUD305'~qSS.~805::.~ 150E 60575W NOFtR-11 RSCENOING SEGMENT FEB 9, 1990 !!!1GSF 180 lSSW 150W 135N 120W lOSW 90W 75W30S45560~0 10 30 50 70 90 100(el MRFTBO TOTRL CLOUDNOQR 11RSCENOING SEGMENT FEB 9, 199030545S60S150E 0 t0 30 50 70 90 100FIG. 7. Similar to Fig. 4 but ~r the southern midlatitude Pacific region.S20, S 60, S~n~, S~i~, and S~o~ of 0.813, 0.005, 0.187,-0.016, and 0.788, respectively. In contrast, the MRF model clouds compare lessfavorably. Over the southern Pacific, where the model'sstratitbrm cloud mechanism is dominant, large-scalecloud structures are present. The patterns and locationsof those cloud organizations, however, are differentt?om the satellite observations, suggesting that furtherimprovement in the parameterization of model stratusclouds is needed. The deficiency in large-scale cloudpattern is considered to be caused mainly by imperfections in both the initial model conditions and thelarge-scale model airmass flows and distributions.Again, over the intertropical convergence zone (ITCZ)and Mexico, model convective clouds become scatteredclusters, not at all like those large cloud systems shownin the satellite retrievals.6. Concluding remarks Comparisons of the experimental NOAA/NESDISCLAVR-I cloud data with RTNEPH, ISCCP, andNMC MRF model cloud datasets have been made fortwo test cases. The preliminary results show that the846 JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY VOLUME 10CLAVR TOTAL CLOUD NOAA-ll ASCENDING SEGMENT FEB 9, 1990 ~-~~. ~/~~ ?,l ss--~------ ... .....~ ..~ .. .:;!! 135E I50E 185E 180 185bl 15014 135N 120N 105N 90N ?5W 5 :::.. :':...-..: L-':~ ~:: . . 0 10 30 50 70 90 100 (B) RTNEPH TOTAL CLOUD NORA-11 ASCENDING SEGMENT FEB 9, 199015N 15N151 120E 135E 150E 165E 180 165N 150N 135N 120N 105N 90N 75N 0 lO 30 50 70 90 100(C) MRFTSO TOTAL CLOUD NORA-11 ASCENDING SEGMENT FEB 9, 199015N 0155 120E15N 0 15S135E 150E 165E 180 165W 150N 135N 120N ]05N 90N 75N- ..--::'......:~x~i ..... 0 10 30 50 70 90 100 ~G.8. Similarto ~g. 4 but ~rthetropical Pacific region.CLAVR-I cloud data is capable of providing goodglobal cloud distributions at high horizontal resolution.The main difficulty with the CLAVR-I product is itstendency to favor a cloud amount of 50%. Future workwill improve the interpretation of mixed cloud conditions and provide a multilayer cloud analysis. The current NMC MRF model cloud predictionscheme is capable of producing generally reasonablelarge-scale cloud patterns in midlatitudes, especiallyalong frontal zones and storm tracks. However, evenin midlatitude, several significant weaknesses are apparent. The model exhibits an early period spinup behavior, wherein the area-mean cloud amounts increasewith time. Even following the spinup, the MRF areamean cloud amounts show a significant negative bias,both in midlatitudes and the tropics. In lower latitudes,the model's cloud patterns are forced by the model'sconvective parameterization and compare poorly withthe satellite retrievals. Here the model convective cloudclusters are too small and localized, reflecting characteristics of the model's tropical convective rainfall. Statistical comparisons of model and "observed"clouds, presented in section 4, indicate that work isneeded to improve the model's diagnosis of both cloudamount and location. The serious shortcomings resultfrom imperfections in the model cloud parameterization scheme and in other model physical parameterizations, especially convection. There is active researchat NMC to improve the model physical parameterizations and to develop a more physically based, explicitcloud scheme. In the context of diagnostic cloudschemes, Mitchell and Hahn (1989, 1990) present anautomated, objective adjustment procedure (tuning)that eliminates model cloud forecast bias and cloudamount spinup in extratropical latitudes. In theirscheme, cumulative frequency distributions of(a) observed layered cloud fraction (e.g., the RTNEPH cloudanalysis) and (b) model forecast layer relative humidity-)ECEMBER 1993 HOU ET AL. 847 NORR-9 RSCENDING SEGMENT JUL 9, 1990 MRFT80 TOTRL CLOUD ISCCP TOTRL CLOUD CLRVR TOTRL CLOUD30N ..... 3ON tZON 105N 90N 7SN 120N tOSN 90N 75N 120N 105N 90N 75N MISSING DRTR 0 I0 50 50 70 90 100FIG. 9. Comparisons of total cloud amount distribution over the eastern Pacific, South to Noah American reg~n. Data am mapped on NOAA-9 mteltite ascending segments of 9 July 1986.are mapped one-to-one to yield critical relative humidity values and cloud-relative humidity functions.Their method is currently being tested with RTNEPHdata at NMC. When CLAVR becomes operational, itwill be used continuously to improve the model diagnostic cloud parameterization and data assimilationsystems on a real-time basis. Despite the inevitable differences between the threecloud analyses due to data sources, timeliness, and retrieval methods, these differences are significantlysmaller than differences from MRF forecasts, showingthat state of the art cloud analyses are of sufficientquality to improve cloud forecasts. Acknowledgments. The authors wish to thank Captain John Percira, U.S. Air Force Liaison Officer toNMC and NESDIS, for providing satellite orbiting information of the RTNEPH data used in this study.The authors would also like to thank Dr. Paul A. Davisof NESDIS, Dr. Peter M. Caplan of NMC, Capt. Norman H. Mandy of USAF, and three anonymous reviewers for their valuable comments and suggestionsfor improving this manuscript. APPENDIX Statistical Comparison Scores The system of statistical comparison scores used inthis study is adapted and modified from the one appliedat AFGWC (Trapnell 1992). Score values are calculatcd from a set of cloud amount contingency tables.A brief description of the structure of a contingencytable and definitions of comparison scores are givenbelow.a. Contingency table Cloud amount has been divided into 11 categoriesover the entire cloud fraction range from 0 to 1. Theinterval value of cloud amount categories is 0.1 excepta value of 0.05 is used at the two ends of the spectrumdesignated for clear sky and overcast conditions. Fortwo compared cloud datasets, F1 versus F2, cloudgroupings are denoted as FIi and F2j with i andj running from I to 11. A contingency table (or mmrix)then is established for each of the specified geographicregions.Contingency table: Chance table: F2./ F2.~ nl,l ...... /'/1,11 Cl,1 ...... Cl,ll/'/11,1 ..... /'/11,11Cll,1 ..... ClI,11 An entry nid in the table represents the number ofoccurrences in the region that the first cloud datasetFI falls in the ith cloud amount category and the sec848 JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY VOLUME 10ond dataset F2 in the jth category. Then the numberof occurrences for each of the cloud categories for theF1 and F2 datasets is defined as follows: 11 11 ni = ~ hi,j, nj = ~ tli,j. (A1) j=l i=1The total number of occurrences for all cloud categoriesin the given region will be: 11 II II 11 N= Z ni = ~ n~= Z Z ni,~. (A2) /=1 j=l j=l i=1In practice, a large region, such as the global area, maybe divided into smaller regions focusing on particularstudy interests. Contingency tables are calculated foreach of those smaller regions. Thus, one can easily establish a new contingency table for an area containingtwo or more nonoverlapping smaller regions by simplysumming over the corresponding entries in the tablesof smaller regions. From an established contingency table, a corresponding chance table (matrix) may also be establishedin a very similar fashion by defining each of the entriesas: Cij - ninj _ Nni nj = NPi~ = NPij, (A3) ' N NN 'where P~j may be interpreted as the joint probabilityon two statistically assumed independent databases,that one has clouds in the FIi category while the otherone has clouds in the F2j category.b. Statistical comparison scores After establishing the contingency and corresponding chance tables, a number of useful statistical comparison scores can be derived. Listed below are somescore definitions derived from a contingency table. Aset of corresponding definitions of chance scores mayalso be obtained by simply replacing every nij in thedefinitions with the corresponding c~ quantity. 1 ) S20 SCORE The S2o score is j~ll 1 ~ j=i+l S2o = ~ Z ni,j. (A4) i=1 j=i-I j~lThe S2o score gives an estimate of how well two clouddatasets agree. Its value represents the fraction of pointswhere the cloud amount in the two cloud fields differby less than about 20%. The larger the score (maximumvalue is one), the better the agreement in cloud amountin a given area. Here a threshold value of 20% is setbecause, in most cases, exactly the same distributionfrom two different cloud data sources is not expected. 2) 5_60 SCORE The S60 score is 1 s ~t s ~ S-6o=~(Z~ Ill,j+ Z ~ tli,j). (A5) i=1 j=i+6 j=l i=j+6In contrast to the S2o score, the S-6o score gives anestimate of how poorly the two cloud datasets agree.It represents the percent of points where the cloudamount in the two cloud fields differ by more than60%. The smaller the score (minimum value is zero),the better the agreement between the two cloud datasets. Because this score tracks large cloud amount differences for each grid point, it is a good indicator ofagreement in cloud locations. 3) ~SCORE The Sh score is &o - C20 Sh = (A6) 1 -- C20The Sh score (called the Heidike score) is the relativedifference between S20 and C20 (chance of S2o score).The score measures how closely the two cloud datasetsare related to (or statistically dependent on) each other.In a perfect agreement between two cloud datasets, theS20 score is equal to 1, and in that limit, Sh approachesits maximum value of 1. The score is larger when theS20 score (i.e., actual frequency of occurrence of cloudswithin 20% difference for the two datasets) is far greaterthan the C20 score (i.e., chance of joint occurrence oftwo assumed statistically independent cloud datasets).The larger the Sh score, usually the better the agreementbetween two datasets. 4) ROOT-MEAN-SQUARE DIFFERENCE The root-mean-square (rms) difference is = [ 1" F2j)2]'/2. (A7) Srms [~/j--~li=l ~ ni-(FliIn its definition, FI~ and F2j are the mean values ofcloud amount from the two datasets in the ith andjthcloud amount categories, respectively. 5) BIAS SCORE The bias score is 1 ~l~ Sbias = ~/~ ni,j(Fli -- F2/). (A8) j=l- i=1The bias score is useful in checking the difference (orbias) of overall mean cloud amount given by the twocloud datasets. A large positive (negative) value indicates that the overall mean cloud amount given by1993 HOU ET AL. 849dataset F1 is much larger (smaller) than that givenby F2. 6) IVIEAN ABSOLUTE DIFFERENCE The mean absolute difference is defined as Sab~=~j~= i=~nijIFli--F2~" (A9)The S~b~ scorc shows the mean magnitude of absolutedifferences of cloud amount between two datasets in agi-~n area. 7) CORRELATION COEFFICIENT The correlation coefficient ~ defined asScott ~ GxY t 11 a.,~ ~ ~ Z [n~,~(Fl~ - FI)(F2~ - F2)I, a~, [~ i~ ni(Fli - ~)2 ,/2, = [ 1" _ ~)2],/2, (At0)whom Fl='~niFli, ~=--~n~F2; (All)are the average cloud amount values for the two clouddatasets. The higher the score value, thc better thea~eement in overall cloud patterns (amount and disthbufion) of the two cloud fields.REFERENCESCmnpana, K. A., 1990: Radiation and cloud parameterization at the National Meteorological Center. Proc. ECMWF/WCRP Ik?~rkshop on Clouds, Radiative Tranafi, r and the ttydrological C.,~x'le. Reading, England, ECMWF, 313-340.Coakley, J. A., Jr., and D. G. Baldwin, 1984: Towards the objective analysis of clouds from satellite imagery data. J. Climate Appl. Meteor.. 23, 1065-1099.Itamilk T. M., R. P. d'Entremont, and J. T. Bunting, 1992: A de scription of the air force real4ime nephanalysis model. Wea. Forecasting, 7, 288-306.ttenderson-Sellem A., 1985: Layer cloud amount for January and July 1979 from 3D-nephanalysis. J. CfimateAppl. Meteor., 25, 118-132.Herman, G., M.-L. Wu, and W. Johnson, 1980: The effect of clouds on the earth's solar and infrared radiation budget. J. Atmos. Sci., 37, 1251-1261.Kanamitsu, M., 1989: Description of the NMC global data assimi lation and forecast system. Wea. Forecasting, 4, 335-342.Kiess, R. B., and W. M. Cox, 1988: The AFGWC automated real time cloud analysis model. AFGWC/TN-88/001, Air Force Global Weather Central, Offutt AFB, NE, 82 pp.Manabe, S., 1969: Climate and the ocean circulation: I. The atmo spheric circulation and the hydrology of the earth's surface. Mort. Wea. Rev., 97, 739-774.Mitchell, K. E., and D. C. Hahn, 1989: Development of a cloud forecast scheme for the GL baseline global spectral model. AFGL TR-89-0343, AFGL, Air Force Systems Command, Hanscom AFB, MA, 147 pp. , and --, 1990: Objective development of diagnostic cloud forecast schemes in global and regional models. Seventh Conf on Atmospheric Radiation. San Francisco, CA, Amer. Meteor. Soc., J 138-J 145.Ramanathan, V., R. D. Cess, E. F. Harrison, P. Minnis, B. R. Bark storm, and D. Hartman, 1989: Cloud-radiative forcing and cli mate: Results from the earth radiation budget experiment. Sci ence, 243, 57-63.Rossow, W. B., E. Kinsella, A. Wolf, and L. Gardner, 1985: Inter national Satellite Cloud Climatology Project ( ISCCP): Descrip tion of reduced resolution radiance data. WMO/TD No. 58, World Meteorological Organization, Geneva, 132 pp.Schiffer, R. A., and W. B. Rossow, 1983: The International Satellite Cloud Climatology Project (ISCCP): The first project of the World Climate Research Program. Bull. Amer. Meteor. Soc., 64, 779-784.Slingo, J. M., 1980: A cloud parameterization scheme derived from GATE data for use with a numerical model. Quart. J. Roy. Meteor. Sot'., 106, 747-770.---, 1987: The development and verification of a cloud prediction scheme for the ECMWF model. Quart. J. Roy. Meteor. Soc., 113, 899-927.Stephens, G. L., and T. J. Greenwald, 1991: The earth's radiation budget and its relation to atmospheric hydrology: 2. Observations of cloud effects. J. Geophys. Res, 96, 15 325-15 340.Stowe, L. L., 1991: Cloud and aerosol products at NOAA / NESDIS. Palaeoclimatol., Palaeoecol. [Global and Planetary Change sec tion], 90, 25-32.--, H. Y. M. Yeh, T. F. Eck, C. G. Wellemeyer, H. L. Kyle, and the Nimbus-7 Cloud Data Processing Team, 1989:Nimbus-7 global cloud climatology, Part II: First year results. J. Climate, 2, 671-709.--, E. P. MaClain, R. Carey, P. Pellegrino, G. Gutman, P. Davis, C. Long, and S. Hart, 1991: Global distribution of cloud cover derived from NOAA/AVHRR operational satellite data. Adv. Space Res., 2~, 51-54.Sundqvist, H., E. Berge, and J. Kristjansson, 1989: Condensation and cloud parameterization studies with a mesoscale numerical weather prediction model. Mort. Wea. Rev., 117, 1641-1657.Trapnell, R. N., Jr., 1992: Cloud curve algorithm test program. Rep. PL-TR-92-2052, Phillips Lab., Hanscom AFB, MA, 170 pp.Wilson, C. A., and J. F. B. Mitchell, 1986: Diurnal variation and cloud in a general circulation model. Qaart. J. Roy. Meteor. Soc., 112, 347-369.Zamiska, A., 1986: RTNEPH USAFETAC Climatic Database Users ttandbook No. 1. USAFETAC/User Handbook-86/001, USAF Environmental Technical Applications Center, Operating Lo cation A, Asheville, N.C., 15 pp.