JANUARY 1988 RICHARD W. REYNOLDS 75A Real-Time Global Sea Surface Temperature Analysis RICHARD W. REYNOLDSClimate Analysis Center, NMC/NWS/NOAA. Washington, DC(Manuscript received 7 May 1987, in final form 25 August 1987) ABSTRACT A global monthly sea surface temperature analysis is described which uses real-time in situ (ship and buoy)and satellite data. The method combines the advantages of both types of data: the ground truth of in situ dataand the improved coverage of satellite data. The technique also effectively eliminates most of the bias differencesbetween the in situ and satellite data. Examples of the method are shown to illustrate these points. Sea surface temperature (SST) data from quality-controlled drifting buoys are used to develop error statisticsfor a 24-month period from January 1985 through December 1986. The average rms monthly error is 0.78-C;the modulus of the monthly biases (i.e., the average of the absolute value of the monthly biases) is 0.09-C.1. Introduction In February 1985 the World Meteorological Organization and the U.S. National Weather Service established a global sea surface temperature (SST) data center at the U.S. National Meteorological Center (NMC)in support of the World Climate Research Program'sTropical Oceans and Global Atmosphere (TOGA) effort. The SST center collects in situ (ship and buoy)and satellite SST measurements in real time and usesthese data to produce analyses of monthly mean globalSST on a 2-deg latitude-longitude grid for the 10-yearTOGA period (1985-94). The activities of NMC inthis task are divided between two groups: the ClimateAnalysis Center, responsible for technical guidance,and the Ocean Products Center, responsible for operations. The purpose of this paper is to describe the presentanalysis techniques and to give a preliminary evaluation of their accuracy by use of quality-controlled SSTdata from drifting buoys.2. SST analyses Three analyses are produced by NMC: an in situ, asatellite, and a "blended" analysis. All analyses arecomputed relative to the monthly SST climatology ofReynolds and Roberts (1987) which is discussed at theend of this section.a. The in situ analysis The SST data used in the in situ analysis are obtainedfrom the NMC archive of surfa~e marine observations.These data consist of all ship and buoy observations Corresponding author address: Dr. Richard W. Reynolds, ClimateAnalysis Center/NMC/NWS/NOAA, WWB, Washington, DC 20233.available to NMC on the Global TelecommunicationSystem (GTS) within 10 h of observation time. Thisarchive is accessed daily to extract and save all newSST observations. The monthly distribution of in situ observations (seeFig. 1) is adequate to describe the SST patterns between30-S and 60-N except in the central and eastern trop- -ical and South Pacific. However, the individual observations are subject to large errors in both temperatureand position and thus further analysis is needed. Theprocessing includes eliminating questionable values,averaging the monthly values on a 2-deg grid, converting the means to anomalies (by subtracting themonthly climatological mean), interpolating missingvalues, applying a spatial median filter, replacing median values by original gridded values in regions witha high density of observations and, finally, smoothinglinearly in space. (Complete details are given in appendix A.) The most important step of the in situ analysis procedure is the application of the nonlinear filter basedon medians which was developed by Tukey (see Rabiner et al., 1975) and which is applied spatially. Theuse of medians rather than weighted means results inthe objective elimination of extreme values instead ofsmoothing the effect of the extremes over a larger region. The application of the filter (see appendix B forthe algorithm) is made in several steps with differentlength scales of up to 8 deg and degrades the original2-deg resolution to roughly 6 deg. The gridded valuesare filtered without regard to the number of observations that were used to compute the average. Sincevalues obtained from a larger number of observationsshould be more accurate than those from a smallernumber, the median filtered value was replaced in moredense reporting areas with the original gridded value.This technique enhances the gradients in some of the76 JOURNAL OF CLIMATE VOLUME 1120- 150- 180- 150- 120- 90- 60- 30ow 0- 30-E 60- 90-60-30-N 0o30-S60-120- 150- 180- 150- 120- 90- 60- 30- W 0- 30- E 60- 90- FIG. 1. Distribution of surface marine in situ (ship and buoy) observations received over the GTS for October 1986. (Drifting buoys may be distinguished as nearly continuous wiggly lines.)60-30-N0o30-S60-better sampled coastal upwelling areas. An example ofthe in situ anomaly field after this processing is shownin Fig. 2.b. The satellite analysis The use of satellite data can significantly improvethe in situ analysis, especially in regions of sparse insitu data. At this time (see Njoku et al., 1985) the multichannel sea surface temperature (MCSST) techniqueof McClain et al. (1985), using the advanced very highresolution radiometer (AVHRR) on the NOAA polarorbiting satellites, is one of the more accurate SST retrieval methods. When comparing these measurementswith conventional observations, it is important to notethat the initial satellite measurement is a "skin" temperature (i.e., the temperature of a surface layer of lessthan a millimeter), while the in situ observations are"bulk" temperatures (i.e., the temperature of a surfacelayer on the order of meters). To correct for this difference, the satellite algorithms are "tuned" by regression against quality-controlled drifting buoy SST measurements. These regressions differ between day andnight because different AVHRR channels are used.However, the regressions are not a function of globallocation or season. The total number of MCSST satellite retrievals overthe globe from 1982 through 1986 has varied from alow of two hundred thousand to over three million permonth. This large variation is due to satellite hardwarefailures and to interpretation difficulties related to cloudcover and atmospheric aerosols. However, recent retrievals (see Fig. 3) have given excellent global monthlycoverage. The present NMC archive of MCSST retrievals isaccessed daily. (Beginning in February 1985 the datawere also separated into daytime and nighttime categories.), The analyzed field is computed using techniques similar to those used for the in situ analysis; thedetails are described in appendix C. However, becauseof the large number of observations, the median filteredvalue is replaced by the original gridded value in mostregions. Thus, the most important step in the satelliteprocedure is a linear smoothing using a two-dimensional ( 1-2-1 ) binomial filter. The linear smoothing isneeded for the later blending .of the two analyses because (as discussed in subsection 2c) the first and second derivatives of the satellite field must be computed.An example of the resulting satellite anomaly field isshown in Fig. 4. Because the global satellite coverage is superior tothe in situ coverage, the satellite field has potentiallybetter accuracy. This is demonstrated by the tendencyof the Southern Hemisphere satellite anomalies to bemore coherent in space and time than the in situanomalies. However, direct comparisons of Figs. 2 and4 show general similarities but with important differences of over I-C (see Fig. 5) which occur even innorthern midlatitude regions where the in situ coverageis also good. Figure 6 shows a monthly time series ofthe in situ and satellite analyses (1982-86) in two eastern Pacific regions: one in the tropics and one in northern midlatitudes. Generally, the satellite analysis is approximately 0.5-C colder than the in situ analysis.JANUARY 1988 RICHARD W. REYNOLDS 77120- 150- 180- 150- 120- 90- 60- 30- W 0- 30- E 60- 90-N ,, N 120- 150- 180- 150- 120- 90- 60- 30- W 0- 30- E 60- 90- FIG. 2. Analyzed in situ SST anomaly field for October 1986. The contour interval is I-C. Values less than -1 -C arc shaded; values greater than 1 -C are stippled.However, important reversals in these tendencies persist for several months. Both analyses clearly show thewarming ENSO (El Nifio/Southern Oscillation) signalof 1982-83 in the tropical Pacific. These comparisons show that although SST information is found in both types of data, there are differences between them. Biases in ship intake temperatureshave been well documented (e.g., see Barnett, 1984).(The biases are thought to be positive by several tenthsof a degree Celsius although there is disagreement about120- 150- 180- 150- 120- 90- 60-. 30oW 0o 30-E 60- 90-~ __ _,.,~,~.~ -"-' ~ ----~ .~-"'--~-~--m.-_~o~? ~.~o~__...~.._~,~_-. ~-~. ~ -~---7"~-~o2~~,~'~, ~ ~ ~'- -.-~ ~~?00 .~ ~%~ ~/ _ ~'" 100 100 ~f ~~ ~ ~ ~~ ~o ~ - ~ ~ u .... ~M ~ ~ 1000 % ~~ ~ '~ :~ .... - ~ ~ ~~~nnn .~ ~- ~.'~ooo~ ~o~=u ~ ~~~ ~?,~ ~ ~ ~ ~.~~% ~~ ~oo ~o~o ~~~~~~oo . ~ ~F~ oSr~ :~, o ~o to$o ' ~ ~ .~o ~( o ~~ ~ ~ ~ o.o 100 ~ ~r 'J~~ ~0o ~ " ~_~~ > ~o~ ~l~~~,oo~-~!20 e o ~_~ O ~100 '~ ~:,60- 60-30- 30-N N 0o 0o30- 30-S S60- 60- 120- 150- 180- 150- 120- 90- 60- 30- W 0- 30- E 60- 90- FIG. 3. Number of satellite (day plus night) observations available on a 2-deg grid for October 1986. The contours are 10 (heavy line), 100 and 1000. Values greater than 1000 are stippled.78 JOURNAL OF CLIMATE VOLUME I60-30-N0o30-S60-120- 150- 180- 150- 120- 90- 60- 30-W 0- 30- E 60- 90-120- 150- 180- 150- 120- 90- 60- 30-W 0- 30-E 60- 90-60-30-N0o30-S60-F~G. 4. As in Fig. 2 except for analyzed satellite SST anomaly field for October 1986.the exact value.) However, satellite SST retrievals alsohave biases as discussed above. To illustrate the spatialscales of the satellite bias, monthly satellite fields wereanalyzed using only daytime or only nighttime observations. A typical difference (see Fig. 7) shows largecoherent zonal regions where nighttime satellite temperatures were more than 0.5 -C greater than daytimetemperatures. Because this would be unlikely to occurin the open ocean for an entire month (especially inthe tropics), biases in the satellite field may be inferred. The causes of the satellite biases are only partiallyunderstood. Perhaps the best known example is the60-30-N0o30-S60-120- 150- 180- 150- 120- 90- 60- 30-W 0- 30- E 60- 90-120- 150- 180- 150- 120- 90- 60- 30- W 0- 30- E 60- 90- FIG. 5. As in Fig. 2 except for difference between in situ and satellite SST analyses (in situ - satellite) for October 1986.60-30-N0o30-S60-J^NUARY 1988 RICHARD W. REYNOLDS 7940-1 20--1t l I I I I I I I I I I I I I I I I I I I I- ~/ \'~,~' ~ NJfio-3area'.. x_/x',. ,., .,,-,"1. ,,,, /.i Blend - - - In situ Satellite ~ Namias area--2I I I I I I I I I I I I I I I I I I I II 3 6 9 12 3 6 9 12 3 6 9 12 3 6 9 12 3 6 9 1982 1983 1984 1985 1986 Month/yearFIG. 6. Time series of in situ (dashed line) and satellite (solid line) SST anomalies for January 1982 to September 1986. The blendedvalues (see text) are dotted. The Pacific regions are Ni~o 3 (5-S-5-N, 90-W-150-W) and Namias (30-N-50-S, 150-W-165-W).SST biases which followed the March-April 1982eruptions of E1 Chich6n. The 'aerosols from the eruptions resulted in negative biases in the SST retrievalsof over 2-C relative to the in situ reports (Strong, 1983).Within days after the eruptions, the aerosols (with theassociated negative SST biases) were spread by the atmosphere along the latitude of the volcano, approximately 10-N. In the following months, they were gradually spread to other latitudes. (This can explain muchof the satellite to in situ bias in 1982-83 in the tropicaltime series of Fig. 6 as well as the delayed bias in themidlatitude series.) Another source of negative bias occurs when cloud detection algorithms fail and the temperatures of cloud tops are mixed with SSTs. The causesof the biases are further complicated by the global"tuning" method which assumes that the relationshipbetween "skin" and "bulk" temperatures are time andspace independent. However, it is not the purpose ofthis work to explain satellite biases but only to identifythat they exist. The analysis method which follows isdesigned to minimize their effects.c. The blended analysis The method described in this section "blends" thetwo types of observations by using the in situ analysisto define "benchmark" temperature values in regionsof frequent in situ observations and the satellite analysisto define the shape of the field in regions with little orno in situ data. This is done by requiring that the SSTfield satisfy Poisson's equation (see Oort and Rasmusson, 1971) in spherical coordinates. The blended field,qb, is set equal to the in situ field at grid points wherethere are a sufficient number of in situ observations todefine the analysis adequately. ("Sufficient" has beenempirically defined--see below for more details--as80 JOURNAL OF CLIMATE VOLUME I120- 150- 180-. 150- 120- 90- 60- 30-W 0- 30- E 60- 90- ~ ~_ -,~~ x,'"', , _~'~~ -,~..-_~ -"--"'--"'---'-~:--,_ .....~-~-~:~q~l~ ~~',.~~..?~.,, _~60- -. ~ ~,~ - '~~._ , ~;:~'~~~0 .~ 60--- ~lr~-~~~~ ~~ '~~o~_ -~ -r~ -, ,~~~~~~~-fC --~ ~ o - _ ~ v ~ . 'o ~30-'- 'X~~c ~~~~~~~ ~~~;~r~ ~0 - 30- o.,.- '" '" ~"' , [-~u~ ~~~'~~~'~~o.~/ 2 ~ % ~.oo ~ ~:~~T~~~~ -~~o~-~ oo~"~ :,~~ ~~:~Z ~m ~ _ ~~ ~ - - ~'(~ ~ ~ -~ ~N'u~ '~_~o.s% ~~o.5 ~ _ ~ ~ .b30- ~5.~. 5 0~0 0 0~~o ?~';.5/ ~ C ~0~s ~ ~ %~ - _~ ~. ~- 4 ~ ~ > ~ ~ ~ ~ ~ 5L~~ ~-~ ~O~s ;~n,-~~ ' :~' ~ , ~-~,:-t--:-~ '~ ~~~~~~'_ ~~~~~~ ~ t~ '-' -~ t~ ~ ~ (_z ._.._~-., ~,*~-~~)~~~ ~'~~~-1 __-~~~(~1:..~60~- ~ -~.~1 ~F)~-h' ~"~~~-~(~~' 7~- - , u.u -0.5 - 60-120- 150- 180- 150- 120- 90- 60- 30-W 0- 30-E .60- 90- FiG. 7. Difference between satellite analyses using daytime or nighttime observations (day - night) for October 1986. Otherwise as in Fig.2, except that the 0-C contour is indicated by a heavy line and additional contours are added at -0.5-C (dashed line) and 0.5-C (solid line).five or more observations per grid box per month.) Atthe other grid points - is determined by solving theequation 72- = ,o, (l)subject to the internal boundary conditions imposedby the in situ benchmark values. The forcing term, p,is defined to be the Laplacian of the satellite analysis-7 2S) in regions of sufficient satellite observations (empirically defined--see below--to be ten or more observations per month) and 0 elsewhere. At grid pointswith less than ten satellite observations, therefore, thesatellite field cannot locally affect the final result. Atother grid points, the satellite field can only affect thesolution via the Laplacian. Thus large-scale biases inthe satellite field (i.e., biases with a Laplacian of zero)cannot affect the final solution ~ in this aspect of theanalysis. To obtain a well-posed solution to (1), conditionsmust also be specified at an external boundary whichcompletely encloses the interior region. This is doneby initially defining the poleward limits of all SSTmeasurements. At these limits - is set equal to the insitu analysis, in regions of sufficient (five or more) insitu observations, or otherwise, set to the satellite analysis. Biases in the satellite data can thus enter theblended analysis when the external boundary conditions are derived from satellite observations. In thiscase the effects of the satellite biases'spread towardlower latitudes where they are finally eliminated by insitu internal boundary points.. The technique involves two important choices whichcan only be defended empirically. The first choice requires that five or more in situ observations be availablelocally before a gridded value from the in situ analysiscould be used as an internal or external boundary value.This criterion was selected experimentally and is.acompromise between a blended field dominated by thesatellite analysis (if more in situ data were required) orthe in situ analysis (if less in situ data were required).The second choice was to set p equal to the local Laplacian of the satellite analysis only if at least ten satellite observations had been available there; otherwisep was set to 0. This criterion was selected after notingthat the difference between in situ and satellite fieldstended to be large when the number of monthly satelliteretrievals per grid point became very low. The restriction is equivalent to linearly interpolating the satellitefield across regions where the field is unreliable. It wastested during late 1982 when the effect of the E1 Chich6n aerosols on the satellite retrievals produced anerroneous negative trough in the satellite anomaly fieldalong 10-N. Because the trough was associated with avery low number of satellite retrievals, the ten-observation restriction allowed the trough to be automatically eliminated from the blend. However, during themore recent TOGA period, the restriction has minimaleffect because of the limited number of interior oceanregions with less than ten satellite observations (e.g.,see Fig. 3). For computational convenience, the blended analysis is also computed for grid points on land. SinceJANU^RY 1988 RICHARD W. REYNOLDS 81there are no land observations, the solution automatically reduces to X72qb = 0. However, along most of thecoastal regions, the grid values become in situ internalboundary points because of the high density of shiptraffic. This effectively isolates solutions on land fromthose on the sea. The Poisson method was chosen primarily becauseit objectively eliminates both the bias and the largescale gradient of the satellite field between the internalboundary points. This procedure allows the shape ofthe satellite field to be matched to the boundary pointsmore effectively than by correction of the satellite biasalone. The Poisson equation also has the importantbenefit of behaving well numerically. Thus, when theequations are expanded by finite differences into a setof linear algebraic equations, they can be solved iteratively to obtain a unique solution. This behavior isespecially convenient since the set of equations varyas the boundary points respond to changes in themonthly distribution of observations. The solution wasdone by sequential overrelaxation with a relaxationcoefficient of 1.6 (e.g., see Thompson, 1961). For eachcomplete iteration of all interior points, the solutionwas defined to have converged when the maximumabsolute value of the individual grid point residualswas less than 0.001 -C. The convergence took less than300 iterations and was only weakly affected by the relaxation coefficient or by the convergence criterion. Afinal smoothing was done by a linear binomial (1-21) filter in both the north/south and east/west directions. An example of the blended field is shown inFig. 8. The effect of the blended procedure can be seen byexamining the in situ, satellite and blended anomalies(Figs. 2, 4 and 8) as well as the difference between thein situ and the blended fields (Fig. 9) and the satelliteand the blended fields (Fig. 10). These figures showthat the satellite analysis has almost no effect on theblend from approximately 60-N to the equator in theAtlantic and Indian oceans and from approximately55- to 20-N in the Pacific Ocean. In these regions thein situ data are sufficiently dense so that almost all ofthe grid values become internal boundary values. Inthe remaining areas the blended field is a mix of thetwo input analyses so that the blend retains the averageanomaly value from the in situ field with the greatercoherence of the satellite field. (The change in the coherence is evident by the smoothness contrast betweenFigs. 9 and 10.) To illustrate the behavior of the blendwith time, the blended anomalies have been includedin the time series of Fig. 6. In the midlatitude seriesthe in situ and blend are almost identical due to thelarge number of internal in situ boundary points.However in the tropical time series, where the in situdata are more sparse, the blend uses the in situ field asa reference in spite of changing satellite biases.d. SST climatology The monthly SST climatology of Reynolds andRoberts (1987) has been used in the SST analyses. Although complete details of the processing can be foundthere, a summary is included here. An initial climatological monthly analysis was completed using the in situ data from the Comprehensive60-30-N0o30-S60-120- 150- 180- 150- 120- 90- 60- 30- W 0- 30- E 60- 90-120- 150- 180- 150- 120- 90- 60- 30- W 0- 30- E 60- 90- ~G. 8. As in Fig. 2 except for analyzed blended SST anomaly for October 1986.60-30-N0o30-S60':'82 JOURNAL OF CLIMATE VOLUME60-30-N0-30-S60-120- 150- 180- 150- 120- 90- 60- 30-W 0- 30-E 60- 90-120- 150- 180- 150- 120- 90- 60- 30-W 0- 30-E 60- 90-60-30-N030-S60-FIG. 9. As in Fig. 2 except for difference between in situ and blended SST analyses (in situ - blend) for October 1986.Ocean-Atmosphere Data Set (COADS) of Slutz et al.(1985) for the period 1950-79. This was done by firstcombining all data for the same month, without regardfor year, on a 2-deg grid. Then following a medianfilter procedure similar to that of the in situ analysisabove, monthly climatological fields were obtained.However, because of the lack of data in the SouthernHemisphere, the climatological fields were extendedusing ice and satellite data. The ice data (obtained froma 10-yr dataset from the Glaciological Data Center,Boulder, Colorado) were used to produce monthlyfields which indicate the percentage of time that each60-30-N0-30-S60-120- 150- 180- 150- 120- 90- 60-. 30-W 0- 30- E 60- 90-60-30-N0o30-S60-120- 150- 180- 150- 120- 90- 60- 30-w 0- 30- E 60- 90- FiG. 10. As in Fig, 2 except for difference between satellite and blended SST analyses (satellite - blend) for October 1986.JANUARY 1988 RICHARD W. REYNOLDS 83grid point was covered by ice during the 10 years. Ifsea ice was present at least 50% of the time, the monthlyclimatological SST at that point was set equal to-1.8-C, the freezing point of seawater at a salinity of35 ppt. (A similar procedure was used by Alexanderand Mobley, 1976.) At the non-ice grid points, themonthly SST climatology was set equal to the preliminary COADS in situ analysis if at least ten observationshad been available there. In the remaining interior regions, the grid points were found by solving (1) wherethe forcing term was determined by a monthly satelliteclimatology. The satellite climatology was computedby averaging 4 yr (1982-85) of the monthly satelliteanalyses described above. The effect of satellite SSTbiases was minimal in the climatology since the satelliteanalysis was only used to determine the shape of thesolution in internal regions of the field and was notused to define external boundary conditions as in theoperational procedure described above.3. Verification In this section, quality-controlled drifting buoy dataare used to provide objective error statistics for all threeanalyses during the first 2 years of the TOGA period.The quality-control procedure (see appendix A forcomplete details) only accepts buoy data which passcertain tests on the SST measurements themselves andon the buoy speed and position. The error statistics arecalculated by first computing a monthly averaged temperature (and position) for each quality-controlled buoyand comparing it to a value at the same location whichwas obtained by interpolation from all three of themonthly analyzed SST fields. (Although some buoysmove distances of 1000 km or more in a month, theSST changes are relatively small because the buoys tendto drift with water of the same physical characteristics.)To ensure that the verifying analyses are as independentof the buoys as possible, special versions of the in situand the blended analyses were computed by withholding the drifting buoy data. The satellite analysis, andtherefore the verifying version of the blend, are notcompletely independent of the buoy observations sincethe satellite algorithms are "tuned" at 6- to 12-monthintervals by regression against the drifting buoy data.(This "tuning" is done apefiodica!ly when new satellitesare made operational or when errors between the satellite retrievals and the buoy temperatures suggest thatthe satellite calibrations may have changed.) Monthly biases and rms errors between the qualitycontrolled buoy data and the satellite and the specialin situ and blended analyses have been computed forall 24 months from January 1985 to December 1986.The results (abbreviated in Table 1) show that themodulus of the monthly buoy-to-analysis biases (i.e.,the average of the absolute value of the monthly biases)varies from 0.09- to 0.15-C. The table indicates thatthe modulus of the buoy to blend bias is slightly betterthan the others. The average rms buoy to analysis errorvaries from 0.78- to 1.09-C. In contrast to the modulus, the rms buoy to analysis error is the smallest forthe satellite analysis although the blended analysis isvery similar. In both cases the buoy to in situ analysishas the worst modulus of the bias and the worst rmserror, primarily because of the lack of high latitudeSouthern Hemisphere ship data which, in turn, resultsin an ill-defined SST field there. The buoy distribution(see Fig. 1) is not uniform over the globe since it wasdesigned to complement other in situ data. (Duringthis 2-yr period more than two-thirds of the buoys werelocated in the Southern Hemisphere.) The tropical Pacific, especially its western portion,is a region where the accuracies of the SSTs are of particular concern for diagnosing and predicting ENSOphenomena. Therefore, statistics comparing the analyses to drifting buoys were examined for the Pacificbetween 20-N and 20-S. In this case the average rmserror for all analyses reduced to less than 0.5-C. Thiserror is less than the globally averaged error becauseof the contribution to the latter from large differencesbetween the buoys and the analyses in higher latitudesnear strong oceanographic fronts. However, since thenumber of buoys in the tropical Pacific was as low assix per month, the rms statistics there should be usedwith caution.4. Concluding remarks Details have been presented of an SST analysis whichblends both in situ and satellite data. The method usespreliminary in situ and satellite analyses as input fields.The in situ analysis is used as ground truth to provide"benchmark" temperatures in regions of frequent insitu observations; the satellite analysis is used to definethe shape of the final field between the benchmarks.Examples have been presented which suggest that theblended technique is an effective way to utilize the improved satellite coverage while eliminating much ofthe bias between in situ and satellite data. Comparisons using drifting buoy data showed thatthe modulus of the buoy to blend monthly biases wasless than 0.1 -C while the average rms buoy to blenderror was less than 0.8-C. Although these results indicated that the blend was an improvement over thein situ analysis, they did not clearly indicate that theblend was superior to the satellite analysis. This resultmay seem surprising when contrasted with the satelliteto in situ biases (e.g. see Fig. 6) which have been discussed earlier. Further comparisons using additionalSST analyses and additional SSTs from bathythermographs are now being completed in a cooperativeeffort with the United Kingdom Meteorological Office.and will soon be ready for publication. The statisticalresults show that the satellite biases are nonzero at the5% significance level. The reason that the drifting buoydata cannot strongly distinguish whether the satellite84 JOURNAL OF CLIMATE VOLUME 1TABLE 1. Global bias and rms statistics between monthly SST analyses and monthly averaged SST from quality-controlled driftingbuoys for January 1985 through December 1986. The bias is defined as (buoy - analysis); only every third month is shown. Number of Type 'of Bias rmsMonth Year drifting buoys comparison (-C) (-C)January 1985 59 in situ .04 0.97 satellite .01 0.74 blend .01 0.87April 1985 84 in situ -. 17 1.01 satellite -. ! 5 0.58 blend -.02 0.72July 1985 77 in situ -.23 1.07 satellite -. 13 0.63 blend .00 0.81October 1985 79 in situ -.30 1.34 satellite -.26 0.97 blend -. 14 1.00January 1986 101 in situ -. 15 0.85 satellite -.29 0.63 blend -. 16 0.61April 1986 135 in situ -.02 0.83 satellite -.05 0.72 blend .02 0.71July 1986 123 in situ -.17 1.13 satellite .04 0.81 blend .01 0.81October 1986 110 in situ .03 1.05 satellite ' -. 18 0.99 blend -.04 1.01Average (24 months) 99 in situ -. 13 1.09 satellite -.09 0.74 blend -.01 0.78'Modulus (24 months) 99 in situ .15 - satellite .12 - blend .09 -or the blended analysis is superior is almost certainlydue to their use in the aperiodic "tuning" of the satellitealgorithms. The blended analysis is a continuing effort to obtainthe best real-time SST fields for the TOGA period(1985-94). Each blended field is carefully monitoredeach month using the drifting buoy data and the insitu and satellite analyses as diagnostic tools. The analysis may be modified in the future when the accuracyof the technique can be improved or when required bychanges in the available observations. Acknowledgments. I am grateful to E. Rasmussonfor the encouragement to begin this work. I would.alsolike to thank W. Gemmill, M. Halpert, C. Nelson, andL. Roberts for their computer expertise. M. Bottomley,C. Folland, W. Gemmill, R. Legeckis, and D. Parkerprovided helpful criticism in early drafts of this paper.I also wish to thank the editor, R. Rosen, and his reviewers for their help in improving the final version. APPENDIX A In Situ Analysis Procedure The in situ analyzed field is computed as follows: 1) All drifting buoy observations are separated fromthe other in situ data and sorted by buoy identification.(This is done because the drifting buoy data have ahigh frequency of real time observations in regionswhich may have little other in situ observations.)Monthly time series of buoy position, speed, and SSTare produced and smoothed with a five-point temporalmedian filter to eliminate some spikes in the data. Ifone of the time series for a buoy fails any of the following gross error tests, all data for the buoy are rejectedfor the month.' (a) Tests on each monthly time series of buoy position are: (i) Any absolute change'in position between adjacent points is greater than 3- lat or 3- long. (ii) The buoy position does not change during themonth.. (iii) The monthly standard deviation of position isgreater than 10- lat or 10- long. (iv) The buoy is located over land. (b) Tests on each monthly time series of buoy speedare: (i) Any individual buoy speed is greater than 5 m/s. (ii) The monthly standard deviation of speed isgreater than 3 m/s.JANUARY 1988 RICHARD W. REYNOLDS 85 (c) Tests on each monthly time series of buoy SSTare: (i) Any absolute change in SST between adjacentpoints is greater than 5-C. (ii) The buoy SST does not change. (iii) The monthly standard deviation of SST isgreater than 4-C. (iv) The monthly mean SST is not within four standard deviations of the monthly climatological SST. 2) The individual observations from fixed buoys orships are discarded if they are located over land or ifthey differ from the local climatological mean by morethan four standard deviations. 3) All the remaining in situ observations for themonth (approximately 100 000) are used to computean initial gridded field by arithmetically averaging theobservations within a 2-deg lat and long (centered oneven values). The gridded values are then converted toanomalies by subtracting the climatological monthlymean. 4) Monthly grid point SST values are then discardedif they fail any of the following screening tests:(a) The absolute value of the anomaly is greater than8-C. (b) The absolute value of the anomaly is greater than6-C, the number of observations is 2, and the locationis either north of 60-N or south of 30-S. (c) The absolute value of the anomaly is greater than3-C, the number of observations is 1, and the locationis either north of 60-N or south of 30-S. (d) The number of observations is I and there areno observations in any of the four neighboring boxesto the north, south, east, or west. (e) The magnitude of the difference between thegridded 2-deg anomaly value and the nearest grid pointin a separate analysis of the anomaly field on a coarsergrid was more than 4-C. (This test was designed toeliminate gridded values which disagreed strongly withtheir neighbors. The separate analysis was completedon a 4-deg lat and long grid using all the steps in thisprocedure except this one. The new analysis, althoughsmoothed, should be accurate enough to allow anyoriginal 2-deg gridded value to be eliminated if it differed greatly from the smoothed field. Since anomalieswere used throughout, the effect on strong gradient regions, e.g., western boundary currents, was minimal.) 5) All grid points without an assigned value (i.e.,either no observations were available or the grid valuewas discarded) are filled by interpolation or extrapolation using an objective analysis based on the iterativedifference-succestive correction method of Cressman(1959). 6) The spatial median filter (described in the maintext and appendix B) is applied. 7) The median filtered value is replaced by the original arithmetically averaged value for all grid pointshaving at least 30 observations. For grid points with15 to 30 observations, a linear combination of the filtered and original values is used so that there is noeffect on the median filtered value for 15 observations,but there is total replacement by the original value for30 observations. 8) A final linear smoothing using a binomial ( 1-21) filter in both the north/south and east/west directionis applied. As described in section 2, the median filter (step 6)is the most important step in the in situ analysis. Thedata screening (step 4) is designed to eliminate questionable observations in regions of sparse data. If theseobservations are not eliminated, their effects are spreadover larger regions by the filling of misting data (step5). If such regions exceed the maximum spatial widthof the median filter (eight degrees), they are unaffectedby the filtering and remain in the final product. Thereplacement procedure (step 7) permits grid values witha large number of observations to have a greater influence on the analyzed field. This enhances the gradientsin some of the well-sampled coastal upwelling regions.The linear filter (step 8) smoothes the field to producethe final in situ field. APPENDIX B Median Filter Algorithm The median filter was obtained from unpublishednotes by John Tukey (see Rabiner et al., 1975, for ageneral discussion). Given a time series of n data pointszi where i -- I, 2, - -., n, the filter function f(zi) isdefined as follows:Yi+3/2 = median(z, zi+~, 2i+2, 2i+3) for i= 1,2, ...,n-3,xi+3/2 = median(yi+l/2, Yi+3/2, Yi+5/2) for i=2,3, ...,n-4,vi+~ = median(x/+l/2, xi+3/2) for i=3,4, -..,n-4,andf(2i+l) ~' median(v/, vi+~, 1)i+2) for i=4,5, ...,n-5. To filter the data, the function is applied twice. Thefirst use defines a change in zi, Az~, as Azi=zi-f(zi) for i=5,6, ...,n-4,andAzi=0 for i= 1,2,3,4 and i=n-3, n-2, n- l,n.The second gives the final filtered series ui as ui=f(zi)+f(Azi) for i=5,6, --.,n-4.86 'JOURNAL OF CLIMATE VOLUME 1 In this paper the filter was applied spatially, first inthe east/west and then in the north/south direction.Since the filter was undefined at the end of each spatialseries, the length of each series was increased by addingten points to both ends. In the east/west direction thespatial series then overlapped along latitudinal circles.The extra values needed in this direction were obtainedby repeating values in a cyclic manner. The extra valuesin the north/south direction were obtained by repeatingthe original first and last values at the beginning andend of the series, respectively. APPENDIX C Satellite Analysis Procedure The satellite SST analysis is computed as follows. 1) All satellite SST observations for the month (bothday and night) are arithmetically averaged on a 2-deglat and long grid and converted to anomalies by subtracting the climatological monthly mean. 2) Grid point SST values are then discarded if theyfail any of the following screening tests: (a) The absolute value of the anomaly is greater than8-C. (b) The absolute value of the anomaly is greater than5-C and the number of observations is less than 30. (c) The absolute value of the anomaly is greater than2-C and the number of observations is less than 10. (d) The number of observations is 3 or less. (e) The magnitude of the difference between thegridded 2-deg anomaly, value and the nearest grid pointin a separate analysis of the anomaly field on a coarser4-deg grid was more than 4 oc. (As in the in situ analysis,this test was designed to eliminate gridded values whichdisagreed strongly with their neighbors.) 3) All grid points without an assigned value (i.e.,either no observations were available or the grid valuewas discarded) are filled by interpolation or extrapolation using an objective analysis scheme based on theiterative difference-successive correction method ofCressman (1959). 4) The spatial median filter is applied. 5) The median filtered value is replaced by the original arithmetically averaged for all grid points havingat least 100 observations. For grid points boxes with30 to 100 observations, a linear combination of thefiltered and original values is used so that there is noeffect on' the median filtered value for 30 observationsbut there is.total replacement by the value for 100 observations. (The number of observations for replacement are higher for the satellite 'analysis than for thein situ analysis because the satellite observations weremade from one instrument while the in situ observations were made from different instruments with assumed independence of observational errors.) 6) A final linear smoothing using a binomial ( 1-21) filter in both the north/south and east/west directionsis applied. In spite of the similarity with the in situ analysis,the effect of the median filter is small in the satelliteanalysis because the number of satellite observationsresults in the replacement (step 5) of approximately70% of the median values by the original arithmeticallyaveraged values. Because of the replacement, the finallinear smoothing (step. 6) has the most influence onthe final satellite field. REFERENCESAlexander, R. C., and R. L. MoNey, 1976: Monthly average sea surface temperature and ice-pack limits on a 1 o global grid. Mon. Wea. Rev., 104, 143-148.Barnett, T. P., 1984: Long-term trends in sea surface temperatures over the ocean. Mort. Wea. Rev., 112, 303-313.Cressman, G. P., 1959: An operational objective analysis system.Mon. Wea. Rev., 87, 367-374. McClain, E. P., W. G. Pichel and C. C. Walton, 1985: Comparative performance of AVHRR-based multichannel sea surface tem perature. J. Geophys. Res., 90, 11 587-11 601.Njoku, E. G., T. P. Barnett, R. M. Laurs and A. C. Vastano, 1985: Advances in satellite sea surface temperature measurements and oceanographic applications. J. Geophys. Res., 90, 1 t 573-11 586.Oorl, A. H., and E. M. Rasmusson, 1971: Atmospheric Circulation Statistics. NOAA Professional Paper No. 5, 323 pp. [Available from US Government Printing Office, Washington, DC, 20402, Stock No. 0317-0045.]Rabiner, L. R., M. R. Sambar and C. E. Schmidt, 1975: AppliCations of nonlinear smoothing algorithm to speech processing. IEEE Trans. on Acoust. Speech Signal Process, ASSP-23, 552-557. [Available from IEEE, 345 East 47 Street, New York, NY 10017.]Reynolds, R. W., and L. Roberts, 1987: A global sea surface tem perature climatology from in situ, satellite and ice data. Trop. Ocean-Atmos. Newslett., 37, 15-17. [Available from Rosentiel School of Marine and Atmospheric Science, 4600 Rickenbacker Causeway, Miami, FL 33149.]Slutz, R. J., S. J. Lubker, J. D. Hiscox, S. D. Woodruff, R. L. Jenne, D. H. Joseph, P. M. Steurer and J. D. Elms, 1985: COADS, Comprehensive Ocean-Atmosphere Data Set. Release 1,262 pp. [Available from Climate Research Program, Environmental Research Laboratories, 325 Broadway, Boulder, CO 80303.]Strong, A. E., 1983: Satellite-derived sea surface temperature errors due to El Chichtn aerosol cloud. Trop. Ocean-Atmos. Newslett., 18, 14-15. [Available from Rosentiel School of Marne and At mospheric Science, 4600 Rickenbacker Causeway, Miami, FL 33149.]Thompson, P. D., 1961: Numerical Weather Analysis and Prediction. The Macmillan Co., 170 pp.