INTEGRATEDANALYSISOFPHYSICALANDBIOLOGICAL PAN-ARCTICCHANGE JAMESE.OVERLAND1,MICHAELC.SPILLANE2andNANCYN.SOREIDE1 1NOAA/PacificMarineEnvironmentalLaboratory,7600SandPointWayNE,Seattle, WA98115-6349,U.S.A. E-mail:[email protected] 2JointInstitutefortheStudyoftheAtmosphereandOceans,Box354235,UniversityofWashington, Seattle,WA98195-4235,U.S.A. Abstract. WeinvestigatetherecentlargechangesthathaveoccurredintheArcticovertheperiod of 1965–1995 through examination of 86 regionally-dispersed timeseries representing seven data types:climateindices,atmosphere,ocean,terrestrial,seaice,fisheries,andotherbiologicaldata.To ourknowledge,thisisthefirstsemi-quantitativeanalysisofArcticdatathatspansmultipledisciplines andgeographicregions.AlthoughvisualinspectionandPrincipalComponent Analysisofthedata collectionindicatethat Arcticchange iscomplex, threepatternsareevident. Thetemporal pattern of change calculated as the first Principal Component (PC1), representing 23% of the variance, has a single regime-like shift near 1989 based on a large number of time series, which include projections fromastrong stratosphericvortexinspring, theArcticOscillation, seaicedeclines in severalregions,andchangesinselectedmammal,bird,andfishpopulations. Thepatternbasedon thesecondPrincipalComponent(PC2)showsinterdecadalvariabilityovertheArcticOceanBasin ◦ northof70 N;thisvariabilityisobservedinsurfacewindfields,seaice,andoceancirculation,with themostrecentshiftnear1989.ContributionstoPC1coveralargergeographicareathanPC2,and areconsistent witharecent amplificationof theinterdecadal mode duetopolar processes suchas increased incidence of cold stratospheric temperature anomalies or internal feedbacks. Most land processes–suchassnow cover,greenness,Siberianrunoff,permafrosttemperatures–andcertain subarctic sea icerecords show a thirdpattern of alinear trend over the 30-year interval, which is qualitatively different than either PC1or PC2. These variables are from lower latitudes and often integratetheatmosphericoroceanographic influenceoverseveralseasonsincludingsummer.That morethanhalfofthedatacollectionprojectsstronglyontooneofthethreepatterns,suggeststhatthe Arcticisrespondingasacoherentsystemoverthepreviousthreedecades.However,nosingleindex or classof observations exclusively trackschange intheArctic,aconclusion that emergesfroma multivariateanalysis. 1. Introduction The Arctic has undergone significant shifts in surface temperatures over the last century (Polyakov etal.,2002) anddemonstrable environmental changes haveoc- curred over the previous three decades. These changes have made it difficult for thosewholiveandworkinthenorthtoanticipatethecourseofthesechangesorat leastdeterminetheirpotentialrange.Thereisevidencethatchangesinmidlatitudes are increasingly connected to those in the Arctic. Areal coverage of sea ice has ClimaticChange 63: 291–322,2004. ©2004KluwerAcademicPublishers. PrintedintheNetherlands. 292 JAMESE.OVERLANDETAL. diminished and sea-level pressures in the central Arctic have decreased. Warmer surface temperatures are observed in northern Europe during winter and Alaska andnorthwestCanadaduringspring.Thereisanincreaseinthefrequencyofyears with colder than normal temperature in the lowerstratosphere overhigh latitudes. Permafrost temperatures have risen in Siberia and Alaska with increased erosion. Satellite estimates of ‘greening’ have increased over both the eastern and western hemispheres, with less snow cover, longer growing seasons and changes in the character ofthe tundra. Theinfluence of warm Atlantic water in the Arctic Ocean becamemorewidespread andintenseinthe1990s,withimplications fortheupper watercolumn. Manyofthesechanges arenoted inSerrezeetal.(2000) and Dick- sonetal.(2000).Thesechangesarerobust,andmanyotherbiologicalandphysical changesaresuggested –increases incodintheBarentsSeashrimpoffofsouthern Greenland, and caribou populations in NorthAmerica, and declines and redistrib- utions of marine mammal populations, although the causes for these changes are less certain (Ottersen et al., 2001). Primary references for these changes will be providedlaterinTableII. It has been hypothesized that the present changes in the Arctic are interre- lated (Morison et al., 2001), and are associated with a rising trend in the Arctic Oscillation (AO) since the 1960s (Thompson and Wallace, 1998). Here the AO phenomenon is used to broadly describe the strengthening and increased zonality of the polar vortex as shown by the AO index and related teleconnection indices. Determining whether the covariability of these changes are coincidental or have a causallinkisofmajorimportance. Inthispaperweexaminethesechanges froma heuristic perspective, basedonexamination of86representative biotic andabiotic timeseriesfromtheArcticandsubarctic. The advantage of such a pan-Arctic study, which spans multiple scientific dis- ciplines, is that the credibility for analyzing and possibly detecting change in the Arcticisincreased byconsidering multiplelinesofevidence(ParmesanandYohe, 2003). Ideally, there would be two important interpretations of our approach. If linesofevidenceareindependent, thencombiningthisinformationintoonemetric increases thelevel ofconfidence. Alternatively, ifthere is ahigh decree ofcovari- ability between them, further investigation is needed to understand the potential causality between the records. In practice, each record has uncertainties in its measurement value and regional representativeness. Also, because Arctic change is poorly understood, each record may project only partially onto the important underlying processes. Thususe ofmultiple lines ofevidence mayprovide abetter representation ofchangethanasinglevariableorindex. There are three major difficulties with this approach. The most obvious is the use of short records. Most records in our data collection span less then 35 years. While many of the time series show a trend during this period, it is known for atmospheric temperature that the value of the trend depends on record length (Polyakov, 2002). Since only some series have long records, our interpretation is limited by the short records; this is particularly true for biology. A second issue INTEGRATEDANALYSISOFPHYSICALANDBIOLOGICALPAN-ARCTICCHANGE 293 is that short records can show trends, and thus covariably between records, for a variety of reasons. Some series can be related through nonlinear processes that are not easily detected by linear analyses, for example, lags in the data or a short physical eventthatresults inanecosystem reorganization whichpersists formany years.Thethirddifficultyisthelargeautocorrelationinseveralseries.Thisiseither explicit in averaging in a few indices or implicit in certain biological or chemical time series (CO for example). Correlation-based analyses may favor these later 2 timeseriesoverthosewithlargeinterannual variability. Thus we approach the analyses from the point of view of a screening process. We use the degree of dissimilarity of the time series as a way of reducing the dimensionality of the data collection. In the end, each time series represents its ownunderlying process. Finallyweinvestigate commonalities ofthereduced data set. 2. TheUnaamiDataCollection TheStudyofEnvironmental ArcticChange (SEARCH)isaU.S.interagency pro- gram with international connections focused on understanding recent large-scale changes in the Arctic. The SEARCH Science Plan(cid:1) has given the name Unaami, the Yup’ik word for tomorrow, to the complex of intertwined pan-Arctic changes. Althoughitappearsthatmanyofthesechangesareinterrelated, thecausalrelation between them, their feedbacks and long-term impacts arefar from certain. Tothis endwehaveselected 86representative timeseries foradatacollection forfurther investigation.(cid:1)(cid:1) Wehave chosen data thatrepresent diverse regions and sevendata types: climate indices, surface and upper atmosphere, ocean , sea ice, terrestrial, fisheries,andotherbiologicalindicators(Figure1).Thecompletesetoftimeseries islistedinAppendix A. The Unaami data collection is a representative subset of Arctic data bases. For example, time series are included for all major regions of sea-ice variability, as determined by primary authors. Data are included for different species in all major fisheries. Our goal in assembling the Unaami data collection was to have yearly coverage for at least 1975 through 1995, although many of the time series extend to 1965 and earlier (Figure 2). Our criteria of yearly sampling excluded someimportant Arcticchange examples wheremeasurements wereonly available in several years in the 1970s or early 1980s, and a few years in the 1990s, e.g., subsurface Arctic ocean temperatures, observed sea ice thickness, and tundra car- bonflux.Wehaveincluded, however,mammalpopulation datawhichhadregular, but not yearly, sampling. We have not smoothed the data by combining values intoannualmeans.Weconsiderthatassessmentoftrendswithoutsmoothinglends greater insight to the analyses. Thus patterns ofsurface and upper air temperature (cid:1) SEARCHSciencePlan,2001,http://psc.apl.washington.edu/search (cid:1)(cid:1) Unaamiwebsite,http://www.unaami.noaa.gov 294 JAMESE.OVERLANDETAL. Figure1.Websiteshowingthesevendatatypesandapproximategeographiclocationofthe86time seriesinthedatacollection.‘Unaami’istheYupikwordfor‘tomorrow’. werekept asseparate timeseries forwinterorspringtime values. Manybiological variables represent yearclassesorsummervalues. Seaiceextentdatafordifferent regionswasincluded fortheseason whichshowedthelargest changes. Most ofthe timeseries represent data from published sources. Someseries are derivable from primary sources such as the NCEP/NCAR reanalysis (Kalnay et al., 1996), and others are obtained directly from the investigator. All metadata, including contact information, is provided on the Unaami web site. Several of the time series, such as primary productivity and fisheries recruitment, have been renormalized so that they are better suited to a multivariate linear analysis (Hare andMantua,2000). INTEGRATEDANALYSISOFPHYSICALANDBIOLOGICALPAN-ARCTICCHANGE 295 Figure2.Temporalcoverageforthe86timeseriesinthedatacollection. 3. InitialResults Theprimary analysis technique isPrincipal Component Analysis (PCA),whichis used to isolate common modes of variability in the data set. PCA has three types of output: (1) principal components (PC), which give the temporal structure of theisolatedmodes;(2)eigenvectors (alsocalledEmpiricalOrthogonalFunctions), whichprovidetheloadingorimportanceofeachvariabletothatparticularPCtime series; and (3) the eigenvalues, which indicate the fraction of total data variance representedbyeachPCmode.ThefirstPCtimeseriesexplainsthegreatestamount of total variance of the combined dataset, and each successive PC explains the largestamountofresidualvariance, subjecttothemathematical constraint thatthe subsequent EOFsareorthogonal totheearlierset.PCAisnottheonlyapproachto multivariate analyses. Multidimensional scaling, forexample,focuses ondatasets that tend to cluster in groups. We have chosen PCA as a method that would em- phasizejointvariationwithinthedatacollection, withtheloadings(orprojections) representing theimportance ofindividual timeseriestotheassociated PC. As an initial screening process, we applied PCA to the correlation matrix of the86timeseries.Thiswasdoneforboththe1975–1995 and1965–1995 periods; this represents the trade-off between complete coverage versus longer records as 296 JAMESE.OVERLANDETAL. TableI PercentvarianceexplainedfromanPCAanalysisofthecorrelation matrix Mode 1 2 3 4 5 86series1975–95 29.0 11.4 8.2 6.7 5.8 86series1965–95 23.3 12.0 8.9 6.3 5.2 noted in Figure 2. The percent variance explained for each mode, based on the eigenvalues is listed in Table I; the first two PCs are selected as significant based onthemethodofNorthetal.(1982). The first principal component (PC1) for the years 1965–1995 is shown in Fig- ure 3. It can be interpreted as a regime shift over these 30 years with an increase in the magnitude of the slope near 1989. Note that the weights of the individual timeseriescontributionscanbepositiveornegativesothattheincreaseinPC1can represent either an increase or decrease in the individual contributing time series. The correlation of each series with PC1 is shown in the lower part of Figure 3. Shapesrepresentdatatypesandcolorsrepresentabsolutevaluesofthemagnitudes; a key to the representative location of all variables is provided in the appendix. Forty of the 86 series have an correlation >0.5 with PC1. It should be noted that everydatatypeandeveryregionoftheArcticarerepresentedbystrongcorrelations withPC1>0.5.TheshapeofPC1issimilarfortheshorter periodof1975–1995. Thesecondprincipalcomponent(PC2)andcorrelationmapfortheyears1965– 1995areshowninFigure4.Itshowsaminus-plus-minusstructurewithbreaksnear the mid 1970s and 1990. While many of the time series have this structure, one mustkeepinmindthatPC2isalsomathematicallyconstrained tobeorthogonal to PC1. Fourteen of the series have an ABS (r) > 0.5 with PC2. The shape of PC2 isnecessaryinunderstanding thedifferenceinusing20yearsofdata(1975–1995) and using 30years (1965–1995). BothPC1and PC2have shifts near 1989. Inclu- sionoftheadditional10yearsofdataisimportantbecauseitsuggestsinterdecadal shape in PC2 in contrast to the more linear or regime shape of PC1. Both PC1 and PC2 are consistent with a major discontinuity in the late 1980s, while their behavior inthe1970s provides acontrast between the twomodes. Thuscare must be taken in interpreting the 20-year records. Here we encounter the limitation of multivariate retrospective analysis ofArcticchange inthatfewpan-Arctic records exist before the 1960s. With roughly one and a half cycles for PC2, one cannot fix an absolute time scale, but it could be considered as part of an interdecadal or decadalmode. INTEGRATEDANALYSISOFPHYSICALANDBIOLOGICALPAN-ARCTICCHANGE 297 Figure3.(top)Thefirstprincipalcomponenttimehistory(PC1)fortheentiredatacollection.Note thesingle‘regime-like’ change near 1989. Errorbarsrepresent theestimatedinfluence ofmissing values in the data collection. (bottom) The correlation magnitude of each time series with (PC1). A location key is provided in the appendix. Note that a wide variety of data types and locations contributetoPC1. 298 JAMESE.OVERLANDETAL. Figure4.SameasFigure3,butforthesecondPrincipalComponent(PC2). A major issue for PCAis the impact of data gaps in the middle and end of the records. We address this sensitivity in three ways. First, we compared PCA for the 21-year and 31-year periods, which showed similar PC time series for mode 1 and 2 in the period of overlap. Second, we used the method of Davis (1976) to permitthecalculation ofPCtimeseries inthepresence ofdatagaps. Thismethod also provides an associated square error estimate of the uncertainty of the PC at each time value. This error is zero when all variables are present in a given year. INTEGRATEDANALYSISOFPHYSICALANDBIOLOGICALPAN-ARCTICCHANGE 299 The error bars on the PC time series in Figures 3 and 4 are greatest for the years with reduced data coverage (1965–1974) as expected, but show that extending the analysis to include the relatively less sampled earlier period is warranted. The averageuncertaintyinPCamplitudesduetomissingdataislessthan11%forPC1 and 12% for PC2. A third approach is to fill the missing data values in the data collection before the PCA analysis using a Monte Carlo approach (D. Percival, 2003,personal communication). Thismethodproduced PCtimeseriesformode1 and2similartothosepresented basedontheDavismethod. Thetimeseries thatmake majorcontributions tothe PCAarelisted inTableII and many are displayed in Figure 5. Table II lists their primary season and loca- tion, principal reference and whether they represent primarily atrend, regime-like character or interdecadal oscillation. Figure 5 plots the time series with the 1/3 largestvaluesinredandthe1/3lowestvaluesingreen;ifthetimeseriesshoweda decreaseovertimewehaveinvertedtheseriesandnotedthiswithastar(*).There is an overall shift from green to red over the 30-year period across the different datatypes. Themajorvariability isonadecadal ratherthaninterannual scale. The next sections present a discussion of the response of different data types in more detail. 4. AtmosphericandClimateIndices Because many climate indices relate directly to atmospheric circulation, we have combinedthiscategoryintoonegroupwithatmospheric timeseries.However,we haveseparatedCO andozoneintoaseparategroup,astheyrepresentatmospheric 2 chemistry, and show a different trend shape than the time series representing atmospheric circulation. Theobservations ofCO atBarrowandOzoneoverCanadashowacorrelation 2 with PC1. Stratospheric ozone levels, as represented by five Canadian stations, have decreased steadily from 1970 to the late 1990s, and now appear stabilized at a lower level due in part to the reduction in CFC use. It is estimated that it will take at least several decades for ozone to return to previous values (Weatherhead et al., 2000). Carbon dioxide as represented by the Barrow Alaska observatory recordhassteadily increased overtheperiod; CO isnowestimatedtobeatlevels 2 which are higher than at any time in the past 20 million years, and the increase continues (Pearson and Palmer, 2000). Increases of CO in models (Shindell, 2 2003) show a contribution toward cooling of the stratosphere. Ozone chemistry is understood to contribute to cooler stratospheric temperatures through positive feedback mechanisms (Randall and Wu, 1999). However, the importance of these results is controversial (Graf et al., 1998; Gillett et al., 2002). As we will note in the next paragraphs, an increase in the frequency of cold temperature anomalies and increased stratospheric winds occurred in the 1990s, which are coupled to atmospheric changes inthelowertroposphere. CO and O have linear trends and 2 3 300 JAMESE.OVERLANDETAL. TableII AsubsetoftheUnaamidatacollectionwhichrepresentsmajorArcticchange1965–1995 Variable Season Location Principalreference Trendtype Latestchange Globalchemistry CO2 BarrowTBP Peterson1986 Trend Up Stratosphericozone Canada Fergusson1998 Trend Down Atmosphereandindices Polarvortex50hPa Dec–Apr Graf2000 Regime Up Arcticoscillation Nov–Mar Thompson1998 Regime Up Zonalwind300hPa Jan–Mar NAtlantic Wang/PMEL2002PC Regime Up NorthAtlanticOscillation Dec–Mar Hurrell2001 Regime Up 200hPaairtemps(mode1) March Wang/PMEL2002PC Regime Down 925hPaairtemps(mode2) April Wang/PMEL2002PC Trend Up Surfacevorticity 2-yrmean Walsh1996 Interdecadal Up SiberianHigh Jan–Mar Savelieva2000 Interdecadal Down ArcticOceansea-levelgradient Proshutinsky1999 Interdecadal Up ‘E’(meridional)index Savelieva2000 Trend Down ‘W’(meridional)index Savelieva2000 Trend Up Zonalwind300hPa Jan–Mar NPacific Wang/PMEL2002PC Pacific/regime Up Aleutianlow Jan–Mar Trenberth1994 Pacific Up Seaice Thickness September ESiberia Holloway/AES2002PC RegimeDown Thickness September FramStrait Holloway/AES2002PC Interdecadal Down Extent Summer Arctic Parkinson1999 Regime Down Extent Spring Barents Parkinson1999 Trend Down Extent Winter Okhotsk Parkinson1999 Trend Down Duration Summer ResoluteBay Flato1996 Trend Up Extent Spring BeringSea Wyllie-Echevarria1998 Pacific Down Oceanic Watertemps January KolaPeninsula Dickson2000 Interdecadal Up SST Jan–Mar LabradorSea Salo/PMEL2002PC Regime Down SST Winter PribilofIs. Hare2000 Regime Down Transport BeringStrait Roach1995 Trend Up Terrestrial Greening Apr–Oct Eurasia Zhou2001 Trend Up Greening Apr–Oct NAmerica Zhou2001 Trend Up Snow February Eurasia RutgersClimateLab2002PC Trend Down Snow February NAmerica RutgersClimateLab2002PC Interdecadal Down Permafrosttemperature Annual Alaska Osterkamp1999 Trend Up Permafrosttemperature Annual ESiberia Romanovsky/UAF2002PC Trend Up Discharge Wateryear Siberia Savelieva2000 Trend Up Burnarea Annual NECanada Murphy2000 Regime Up Biological Carabou WAlaska Griffith/USGSPC Regime Up BlackGuillemont NAlaska Divoky/UWPC Regime Up Zooplankton NSea Planque1998 Regime Up Reddeer Post1997 Trend Up Benthicinvertebrates Bering Conners/AFSC2002PC Pacific Up Furseals Bering York2000 Pacific Down Fisheries Herring Norway StatisticsNorway1999 Regime Up Shrimp Greenland FAO2001 Trend Up Cod Barents OttersenandLoeng2000 Trend/regime Up Cod Baltic ICES/ACFM:182001 Regime Down Turbot Bering Hollowed1998 Pacific/regime Down Plaice Bering Hollowed1998 Regime Down Chumsalmon Bering Hare1999 Interdecadal Down Herring Baltic Kull1996 Interdecadal Down Redfish Barents ICES/ACFM:192001 Interdecadal Down Halibut Greenland ICES/ACFM:202001 Interdecadal Down Sockeyesalmon Bering Hare1999 Pacific Up Arrowtoothflounder Bering Hollowed1998 Interdecadal Down PC=PersonalCommunication(seewebsite).