Updating Australia's High-Quality Annual Temperature Dataset

Aust.Met. Mag. 53 (2004)15-93

Updating Australia's high-quality annual temperature dataset

Paul Della-Marta and Dean Collins National Climate Centre,Bureau of Meteorology,Australia and Karl Braganza CSIRO AtmosphericRese arch, Aspendale, Australia

(Manuscriptreceived June 2003;revised September 2003)

An updated and improved version of the Australian high-quality annual mean temperature dataset of Torok and Nicholls (1996)has been produced. This was achievedby undertaking a thorough post-1993homogeneity assessment using a number of objective and semi-objectivetechniques, by matching closed records onto continuing records, and by adding some shorter duration records in data-sparseregions. Each record has been re-assessedfor quality on the basis of recent metadata, resulting in many records being rejected from the dataset. In addition, records have been re-examinedfor possibleurban contamination using some new approaches.This update has high- lighted the need for accurate and completestation metadata. It has also demonstrated the value of at least two years of overlapping observations for major site changes to ensure the homogeneityof the climate record. A total of 133 good-quality, homogenisedrecords have been produced. A non-urban subset of 99 stationsprovides reliable calculations of Australia'sannu- al mean temperature anomalies with observation error vari- ancesbetween 15 and 25 per cent of the total variance and decorrelation length scalesgreater than the ayerageinter-station separation.

trntroduction

Torok and Nicholls (1996) produced a homogenised primary purposeof this datasetwas to enablethe 'high-quality' or datasetof annual mean maximum reliablemonitoring of climatetrends and variability and minimum temperatureseries for Australia. The at annual and decadaltime-scales. Consequently, eachstation record was correctedfor discontinuities Correspondingauthor address:D. Collins, National Climate Centre. causedby changesin sitelocation and exposure, and Bureau of Meteorology,GPO Box 1289K, Melbourne,Vic. 3001, other known data problems(Peterson Australia. et al. 1998). Email: [email protected] Suchdiscontinuities can be as large,or largerthan, 16 Australian Meteorological Magazine 53:2 June 2004 real temperaturechanges (e.g. the approximate 1"C Assessmentof station quality drop in maximum temperaturesassociated with the switch to Stevensonscreen exposure (Nicholls et al. Torok and Nicholls (1996) subjectively ranked the 1996)) and consequently confound the true long- quality of records on a scale of 1 to 6, with rank 1 term trend. Temperaturerecords for 224 observation being very good, 5 being very poor and 6 being urban. locations were homogenised, 170 of which were This update re-assessedeach station, mainly on the classified as non-urbani.e. free from the influence of basis of post-1993metadata, but also using metadata urban warming. Generally the high-quality records compiled by Torok (1996). All stationswere ranked were homogenisedfrom 1910,by which time most on the original 1 to 5 scale,but insteadof a rank 6, all stations are believedto have been equippedwith the stations were noted as either urban or non-urban, current standard instrument shelter. The dataset enabling urban stationsto also be given a quality rank. endedin 1993. Rankswere subjectivelyassigned, taking into account At the end of each year Australia's annual mean the current and historical standardof site exposure, temperature anomaly is calculated using all avail- number of site moves, number of instrument changes, able high-quality, non-urban temperature records. recent standardof observationsand amount of miss- This is routinely provided to the public via an annu- ing data. al climate summary, a media statement and the Since the development of the initial high-quality Bureau of Meteorology's internet site (e.g. datasetthere have been changesin the way station http ://www.bom. gov. aulannouncements/media- history information is recorded by the Australian releases/climate/change/20030106.shtml). The Bursau of Meteorology. Traditionally station metada- datasethas also beenused extensively in the calcu- ta have beenrecorded on paper history files. With the lation of atmospheric indicators for Australia's developmentof SitesDB (Lane 1995), an electronic 20Ot State of the Environmenl report (Collins and metadata database, the process of assessing station Della-Marta 2002). In order for such analyses to quality hasbeen streamlined.However, at present,the remain credible. the climate records used need to be information in the databaseis only reliable since the routinely checkedfor homogeneity. mid to late 1990s. Also, there is currently great varia- Since 1993, the high-quality dataset has been tion in the quality of information held in SitesDB, updatedusing data that have undergonebasic quality reflecting different levels of use and appreciation of assurance checks but no checks for homogeneity. needs throughout the Bureau. Many stationsare likely to have undergonechanges The introduction of SitesDB has resultedin paper requiring data adjustmentin this time. Also, the sta- history files generally not being used routinely. If tion exposure and quality of observations may have SitesDB is not supported adequately this will result changed since the initial assessmentof Torok and (and possibly already has resulted) in the loss of Nicholls (1996). Until this update, the number of important information. Also, most communication high-quality stations still operating has gradually about observation stations is now via email. rather declined due to site moves and closures,resulting in than formal inter-office memoranda.Such communi- gaps in the network. For the 2002 anntal summary cations often contain useful information for future only 138 stations had a full year of data available; 99 assessmentsof the climate record, but are generally of theseclassified as non-urban.The intention of this not placed on file. A concerted effort is currently update was to assessthe quality and homogeneityof underwaywithin the Bureau of Meteorologyto popu- open stations,to re-open'closed' recordsusing near- late SitesDB with historical metadatafrom paper his- by open stationsand to fill datavoids if possible. tory files and to archive station-specificemail com- Most of the homogeneitytechniques employed by munication. Torok and Nicholls (1996) have been replicated. After this latest assessment,sixteen of the origi- However, a number of new methodologies have also nal non-urban stations were consideredto have bet- been used to assesswhether a record is likely to have ter quality than the assessmentof Torok and been contaminated by urban warming, and to utilise Nicholls (1996).This was usually due to the station availablecomparison observation data in caseswhere having moved to a location with better exposure. a station has moved. The developmentof web-based However, 99 stations were ranked with lower quali- technology and greater digitisation of metadata since ty. Predominantly this has occurred at manual sta- the production of the original datasethave allowed tions, possibly due to diminishing resourcesbeing summary information, relevantanalyses and metada- used to maintain the cooperative network as ta to be conveniently combined into a webpage for reliance on the automatic weather station network each individual station.This has enabledthe effrcient increases.This has resulted in documented. but assessmentof eachrecord. uncorrected, exposure problems, less interest from Della-Martaet al.: UpdatingAustralia's annual temperature dataset 17

cooperativeobservers and the downgrading of man- ' Assessmentof recent homogeneity ual observationprograms, such as the cessationof weekendobservations. Torok (1996) provides a brief metadatahistory up to Some of the apparentdecline in quality is simply 1993 for each high-quality station. These histories due to more stringent assessmentcriteria during the were updatedfor this assessmentusing the metadata latestassessment. Some records contain many years contained in SitesDB and hardcopy station files to in which annual mean values were estimatedby identify potential datesof new discontinuities.Torok Torok and Nicholls (1996) using a referenceseries and Nicholls (1996) and Torok (1996) also used a basedon highly correlatedneighbour series. Often numberof objectivestatistical tests to help determine this amountedto more than twenty contiguousyears datesof discontinuity.Using the metadatawhere pos- of estimated data. The amount of in-filling appears sible,decisions were madeas to which discontinuities to have had no bearing on the original station rank. would be correctedto homogenisea record. Torok However, for this assessmentit was decided that (1996) providesa record of the datesat which homo- substantialperiods of estimateddata could be mis- geneity adjustmentswere made, the magnitude and leading to potential users. Generally, any records sign of adjustments,the reasonsfor adjustmentsand with around twenty or more years of estimated which testswere usedto identify them. annual mean values were ranked as poor or very In an attempt to validate previous adjustmentsit poor in the updated dataset(see Table 1). was decidedto reconstructthe unhomogenisedcom- Stations assessedas poor or very poor were not positecandidate series using raw stationdata for each removed from the original datasetin order to max- high-quality station.The raw datasetwas createdby imise the number of long-term records available merging data from the Australian Data Archive for (Torok 1996). Jones and Trewin (2000, 2002) have Meteorology (ADAM) and data compiled and digi- since shown that significant interstationcorrelations tised by Torok (1996). These data were occasionally of monthly (daily) temperatureanomalies can occur slightly different from the unhomogenised dataset for separationsup to at least700 (360) km, and up to used by Torok and Nicholls (1996), which was 1200 (570) km in somemonths. It would be reason- sourced from an early version of ADAM. In most able to expect that significant interstation correla- casesthe candidaterecords were a compositeof two tions on the annual time-scale would be equivalent or more records with different station numbers, the or greaterthan those for monthly temperatureanom- combination being chosen to achieve the highest alies since the spatial variability of the annual time- numberof contiguousyears ofrecord. scale is lower. The monthly decorrelationlength scales calculated Objective statistical test by Jones and Trewin (2000) suggest that, in most of The objective test of Torok and Nicholls (1996), Australia, the number of high-quality records avail- based on a two-phaseregression model (Easterling able is more than adequatefor calculating a reliable and Peterson1995), was replicatedas closely as pos- annual mean temperature anomaly. This redundancy sible for the update assessment.The objective test allowed the removal of all stationsranked 4 (poor) or relies on the creation of a homogeneous reference 5 (very poor) from the updateddataset. The adequacy series to create a difference series (candidate series of the updated network density is examined later. minus referenceseries). Two methods of creating a Most of the removed stations were located in referenceseries were comparedto assessthe reliabil- southeastAustralia and were ranked poorly due to ity of the objectivetest results.One used the median large data gaps in their record or poor site exposure. of interannual temperaturedifference of highly corre- The reasonsfor removing eachpoor quality record are lated neighbouring stations (hereafter known as the described in Table 1. Four stations in New South median referenceseries) (Torok and Nicholls 1996; Wales (Condobolin, Coonamble,Trangie and White Torok 1996) and the other was createdusing a dis- Clitrs) were removed from the dataset due to high tance weighted mean of highly correlated neighbour- maximum temperatures in the early part of their ing stationanomalies (hereafter known as the distance record. These data were consideredsuspect because weightedreference series) (Trewin 2001). some metadataindicated that the instrumentshelters For eachmethod an initial searchof stationswithin used were not standardStevenson screens until the six degreesof latitude and longitude and an elevation 1940s, which would bias maximum temperatures difference of no more than 300 m was performed to higher (Torok 1996). Despite intending to resurrect as find suitable reference stations. These limits were many high-quality records as possible,a total of 81 relaxed to included stations within eight degreeslati- pogr-quality records were rejected from the updated tude and longitude and 500 m elevation difference in datasetin order to producehigher quality overall. data-sparseareas if the initial searchyielded fewer than Australi an Meteorologic al Magazine 53:2 J une 2004

'able 1. Stations included in the updated annual temperature dataset.Quality rank and whether or not a station is con- sideredurbanarealsoshown.(Rankl=verygood,Rank2-goodrRank3=fair,Rank{=poor,Rank5=Y€rY poor.) Note that stations ranked 4 or 5 have been removed from the high-quality dataset and are not used for the calculation'of Australian mean temperatures. The reason(s) for the removal of these stations is shown. MD = missing data (e.g.more than 20 contiguous missing years, or only weekday or part-year observationsduring parts of the record); PE - poor exposure (e.g. screen in poor condition, partly shaded or set in bitumen during parts of the record); PO = poor quality or no overlapping data for a recent site movel pI) = poor quality or sus' pect observationsduring parts ofthe record, and; NS = suspectedto have a non-standard screenuntil the 1940s. tatrcn name Station no. Latitude Longitude . Elevation Start Qualiry Reason Urban ('s) ("E) (m) year Rank (Y/N)

a .LBANY AIRPORT 009741 -34.94 117.80 0068.0 1910 J N ,LICE SPRINGSAIRPORT 015590 -23.80 133.89 0546.0 1910 J N .MBERLEY 040004 -2t.63 152.11 0021.0 1913 2 N .RARATPRISON 089085 -37.28 142.98 0295.0 1910 J N .RMIDALE 056037 -30.53 151.61 0987.0 1910 4PE Y iTHERTONSHIRE COUNCIL 031193 -11.26 145.48 0752.0 1910 3 N JTRDPI RESEARCH 033002 -19.62 141.38 0012.0 1910 J N IAIRNSDALEAIRPORT 085219 -37.88 147.57 0049.4 19r0 J Y IALLADONIA 011017 -32.46 123.86 0148.0 1910 5MD N IALLARAT AIRPORT 089002 -31.sr 143.79 0435.2 1910 2 Y IALRANALD 049002 -34.64 143.56 0061.0 1910 5 MD, PE N |ARCALDINEPOST OFFICE 036007 -23.ss 145.29 0266.9 1913 z N IAROONPOCKET DAM 040850 -26.12 152.87 0230.0 1913 5 PO,MD N IATHURSTAG. STATION 063005 -33.43 r49.56 0713.0 1910 2 Y TEECHWORTHWOOLSHED 082131 -36.32 146.68 0330.0 l9l0 5MD N IEGA 069139 -36.61 149.82 0041.0 r9r0 4 PE,MD N TENALLA 082002 -36.55 r45.91 0169.s 1910 J Y TOLLON 044010 -28.03 t47.48 0182.9 1910 J N TOMBALAPOST OFFICE 070005 -36.9r r49.24 0705.0 7912 5MD N TOULIAAIRPORT 038003 -22.9r 139.90 0161.8 l9l0 J N }OURKEAIRPORT 048245 -30.04 r4s.95 0107.3 1910 z N IOWEN AIRPORT 0332s1 -20.02 148.22 0005.0 1910 5 N }OWRAL 068102 -34.49 r50.40 0690.0 1910 J Y }REWARRINAHOSPITAL 048015 -29.96 146.87 0l15.0 19ll 5MD N }RIDGETOWNPOST OFFICE 009510 -33.96 I 16.14 0t49.9 1910 L N }RISBANEAIRPORT 040842 -21.39 153.13 0004.0 1910 Y }ROKENHILL 04'/007 -31.98 141.41 0315.0 1910 3 Y }ROOMEAIRPORT 003003 -11.95 122.23 0007.0 1910 z N }UNDABERGAIRPORT 039128 -24.89 r52.32 0021.0 1910 5MD N }URKETOWNPOST OFFICE 029004 -11.74 139.55 0005.5 1910 5 MD, PE N }URRINJUCKDAM 013007 -35.00 148.60 0390.0 t912 4 MD, PE N ]AIRNS AIRPORT 0310r1 -16.87 145.75 0003.0 1910 z N ]APE BORDA 022801 -35.75 136.59 0143.0 1910 J N ]APE BRUNY 094198 -43.49 147.14 0059.7 r924 2 N ]APE LEEUWIN 009s18 -34.37 115.13 0013.0 1910 1 N ]APE MORETONLIGHTHOUSE 040043 -27.03 r53.4',1 0099.9 1913 z N ]APE NATURALISTE 009519 -33.s4 115.02 0109.0 1910 1 N ]APE NORTHUMBERLAND 026005 -38.06 r40.67 0005.0 1910 2 N ]APE OTWAYLIGHTHOUSE 0900rs -38.86 143.51 0082.0 1910 J N ]ARDWELL 032004 -18.26 146.02 0005.7 1910 J N ]ARNARVON AIRPORT 006011 -24.89 113.67 0004.0 1910 t N ]ASHMERE DOWNS 012022 -28.91 1r9,57 0450.0 1910 4MD N ]ASHMOREAIRPORT 09017r -38.32 141.47 0080.9 r9l0 2 N ]ASINO AIRPORT 058063 -28.88 153.05 0026.0 l9l0 5MD N ]ASTLEMAINE PRISON 088r10 -37.08 144.24 0330.0 1910 5MD N ]EDUNA 018012 -32.13 133.7| 0015.3 1910 J N 04402r -26.4r 146.26 0302.6 1910 2 N ]HARLEVILLE AIRPORT .) ]HARTERSTOWERS AIRPORT 034084 -20.04 146.21 0290.0 1910 N ]LARE HIGHSCHOOL 021131 -33.82 138.59 0395.0 1910 5PD N ]LERMONT POSTOFFICE 035019 -22.83 147.64 0267.0 1910 5PE N Della-Martaet al.: UpdatingAustralia's annual temperature dataset

Table 1. Continued.

Station name Station no. Latitude Longitude Elevation Start Quality Reason Urban ('s) ("E) (m) year Rank (y/N)

CLONCURRYAIRPORT 029141 -20.67 140.51 0186.0 1910 5 MD N -3r.49 COBAR 048027 145.83 0260.0 1910 3 N COEN AIRPORT 027006 -13.16 r43.rt 0160.5 1910 4 PD N COLLARENEBRI 048031 -29.55 148.59 0145.0 1910 s MD N CONDOBOLINRESEARCH 0s0052 -33.01 147.23 0195.0 1910 4 NS N -15.45 COOKTOWNAIRPORT 031017 t4s.r9 0005.5 1910 2 N COOMA VISITORSCENTRE 070278 -36.23 149.12 0'178.0 1910 4 PD N -31.21 COONABARABRAN 064008 149.27 0505.0 1910 4 PE N COONAMBLECOMPARISON 051010 -30.98 148.38 0180.0 1910 5 PD,NS N COOTAMUNDRAAIRPORT 073142 -34.63 148.04 0335.0 1910 5 PE N -35.99 COROWAAIRPORT 074034 146.36 0143.0 1910 5 MD N COWEAL TOWNSHIP 037010 -r9.92 138.12 0231.2 1910 3 N -33.8s COWRA AIRPORT 06509r 148.65 0300.0 1910 3 N .142.24 CROYDONTOWNSHIP 029012 -18.20 0116.3 1912 5 MD N CUNNAMULLA POSTOFFICE 044026 -28.07 145.68 0188.7 1910 3 N DALBY AIRPORT 041522 -21.t] t5t.2'7 0345.0 1910 3 Y DARWIN -12.42 AIRPORT '074128014015 130.89 0030.4 1910 3 Y DENILIQUIN POSTOFFICE -35.5s 144.95 0093.0 1910 2 N DERBYAIRPORT 003032 -17.37 123.66 0006.2 1910 3 N DONNYBROOK 009534 -33.57 115.83 0063.0 1910 4 PE N -32.22 DUBBOAIRPORT 065070 148.58 0284.0 1910 3 Y EAST SALE AIRPORT 08s072 -38.11 147.13 0004.6 1910 2 N ECHUCAAIRPORT 080015 -36.11 144.76 0096.0 1910 3 Y -23.s1 EMERALDAIRPORT 035264 148.18 0187.9 1910 5 PE N ESPERANCE 009789 -33.83 121.89 002s.0 1910 I N EYRE 011019 -32.25 126.30 0005.6 l9l0 5 MD N FORBESAIRPORT 065103 -33.36 141.92 0230.4 1910 4 PE Y GABO ISLAND LIGHTHOUSE 084016 -37.57 149.91 0015.2 1910 2 N GATTON 040082 -27.5s 152.34 0094.0 1913 4 MD N GEELONGAIRPORT 087163 -38.23 144.33 0033.4 1910 3 Y GEORGETOWNPOST OFFICE 030018 -18.29 143.55 0291.1 1910 3 N -28.80 GERALDTONAIRPORT 008051 114.70 0033.0 1910 2 N GILES METEOROLOGICALOFFICE 013017 -25.04 128.29 0598.0 1957 I N GLADSTONERADAR 039123 -23.86 15r.26 0014.5 1910 2 N GLEN INNES AIRPORT 056243 -29.68 151.69 1044.3 1910 3 N GOONDIWINDIAIRPORT 04r52r -28.52 150.33 0211.6 1910 3 N GOOSEBERRYHILL 009204 -31.94 116.05 0220.0 1910 5 PO,MD Y GOULBURN 010263 -34.72 149.74 0650.0 1910 4 MD, PD Y GRAFTONOLYMPIC POOL 058130 -29.68 152.93 0009.0 1910 3 Y -33.90 GRENFELL 073014 148.17 0410.0 1910 5 MD N GRIFFITHAIRPORT 07504r -34.25 146.07 0134.0 1916 5 PD,PO Y -30.98 GUNNEDAHPOOL 055023 r50.2s 0306.0 1910 3 Y GUYRA HOSPITAL 056229 -30.2r 1sr .68 1332.0 l9ll 5 MD N -26.18 GYMPIE 040093 152.64 0064.5 1910 4 MD Y HALLS CREEKAIRPORT 002012 -18.23 121.66 0422.0 1910 2 N HAMILTON -37.6s AIRPORT 090173 142.06 0241.1 1910 3 Y HAY 075031 -34.52 t44.85 0093.3 1910 3 N HILLSTON -33.49 AIRPORT 075032 r45.s2 0122.0 1910 s MD N HOBART 094029 -42.89 t41.33 0050.5 r9l0 3 Y HUGHENDENAIRPORT 030022 -20.82 144.23 031s.5 l9r0 3 N INNISFAIL -11.s3 032025 t46.03 0008.0 1910 3 Y INVERELL 0s6242 -29.78 151.11 0582.0 1910 2 Y ISISFORDPOST OFFICE 036026 -24.26 144.44 0205.4 1913 4 MD N JARRAHWOOD 009842 -33.80 115.66 0130.0 l9l0 2 N JERRYSPLAINS POSTOFFICE 061086 -32.50 1s0.91 0090.0 r9l0 3 N KALGOORLIE-BOULDER -30.78 AIRPORT 012038 12t.45 0365.3 1910 2 N KATANNING -33.69 COMPARISON 010579 rr7.56 0310.0 1910 3 N 80 Austral ian Meteorologic al Magazine 53:2 June 2004

Table 1. Continued.

Station name Station no. Latitude Longitude Elevation Start Qualiry Reason Urban ("s] fE) (m) year Rank (Y/N)

KATOOMBA 063039 -33.7| 150.30 1030.0 1910 4PEY KELLERBERRIN 010073 -31.62 1t].72 02s0.0 1910 2N KEMPSEY 0590t7 -31.08 152.82 0010.0 1910 4 MD,PE Y KENT TOWN 023090 -34.92 138.62 0048.0 1910 2Y KERANG 080023 -3s.73 143.92 0077.7 1910 2N KING ISLAND AIRPORT 098017 -39.88 r43.88 0036.2 1914 3N KINGSCOTEAIRPORT 022841 -35.71 137.52 000s.0 1914 4PEN LAKE CARGELLIGOAIRPORT 015039 -33.28 146.31 0169.0 19r0 5 PE,MD N LAMEROO 025509 -35.33 140.52 0099.0 1915 5PEN LAUNCESTONAIRPORT 091 104 -4r.54 141.20 0170.0 1910 2N LAVERTON 087031 -37.86 144.75 0016.0 1910 2N LAVERTONAIRPORT 012305 -28.62 14.> A.'' 0464.8 1910 4 PD,MD N LEONORA 012046 -28.88 121.33 0316.0 1910 5MDN LISMORE 058037 -28.81 r53.26 0011.0 1910 2Y LITHGOW 063224 -33.49 150.1 5 09s0.0 t913 4 PE,MD Y LONGERENONG 079028 -36.61 142.30 0091.0 1910 5MDN LONGREACHAIRPORT 036031 -23.44 144.28 0192.2 1910 IN LOW HEAD 091293 -41.06 146.79 0003.0 1910 2N MACKAY 033I l9 -21.12 149.22 0030.2 l9l0 2Y MARBLE BAR COMPARISON 004020 -21.18 1r9.75 0182.3 1910 3N MARYBOROUGHQUEENSLAND 040126 -25.52 152.11 0011.0 1910 2Y MARYBOROUGHVICTORIA 088043 -31.06 143.73 0249.3 1910 3Y MEEKAIHARRA AIRPORT 001045 -26.61 I 18.54 0517.0 1910 2N MELBOURNEREGIONAL OFFICE 086071 -37.81 144.97 003r.2 1910 3Y MENINDEE POSTOFFICE 041019 -32.39 142.42 006r.0 1910 5 PE,PD N MERREDIN 010092 -3r.48 l r 8.28 0315.0 t9t3 2N MILDURA AIRPORT 016031 -34.23 142.08 0050.0 1910 2N MILES 042rr2 -26.66 r50.18 0304.8 1910 2N MITCHELL POSTOFFICE 043020 -26.49 147.98 0336.3 l9l0 4 MD,PE N MOORA 008091 -30.64 r 16.01 0203.0 1910 5MDN MOREE AIRPORT 053l 15 -29.49 t49.8s 0213.0 r9l0 2Y MORUYA HEADS 069018 -35.91 l 50.15 0017.0 1910 IN MOUNTBARKER S.A. 023133 -35.06 r38.85 0360.0 1910 4 PE,PD Y MOUNT BARKERWA. 009581 -34.63 tlt.64 0300.0 l9l0 5 MD,PE N MOUNT GAMBIER AIRPORT 026021 -37.75 r40.77 0063.0 l9l0 IN MT GELLIBRAND 090035 -38.24 143.19 0261.0 1910 4POY MUDGEE AIRPORT 062101 -32.56 t49.62 o4t1.0 1910 3N MUNGINDI POSTOFFICE 052020 -28.98 r48.99 0160.0 1915 4MDN MURRURUNDIPOST OFFICE 0610s1 -3r.76 150.84 0466.0 1910 5 MD,PE N MUSGRAVE 028007 -14.18 t43.s0 0080.1 1910 5MDN NARRABRI WESTPOST OFFICE 053030 -30.34 r49.76 0212.0 l9l0 5PEN NARROGIN 010614 -32.93 I 17.18 0338.0 1913 5MDN NEWCASTLE 061055 -32.92 151.80 0033.0 1910 2Y NEWMAN AIRPORT 007t76 -23.42 119.80 0524.0 t916 2N NHILL 078031 -36.34 141.64 0133.0 1910 3N NORMANTONAIRPORT 029063 -11.69 14r.07 0018.4 l9l0 IN NORTHAM 01011 1 -31.64 116.61 01s5.0 l9t0 4PDN NYANG STATION 006012 -23.03 115.04 0111.0 1910 5MDN OMEO 083025 -37.10 t41.60 0685.0 l9l0 2N ONSLOW 005016 -21.64 115.1 1 0004.0 19r0 4PEN ORANGEAIRPORT COMPARISON 06323r -33.38 149.t2 0948.0 t9l0 3Y PADTHAWAYSOUTH 026100 -36.65 140.52 003s.0 1910 4MDN PALMERVILLE 028004 -16.00 144.08 0203.8 l9r0 3N PARKES 065026 -33.14 r48.16 0324.0 1910 4PEY PARRAMATTANORTH 066124 -33.79 lsl.02 0055.0 1910 4 PE,MD Y PERTHAIRPORT 009021 -31.93 115.98 00r 5.4 l910 1Y POINT PERPENDICULARLIGHT. 068034 -35.09 150.80 008s.0 l910 3N Della-Martaet al.: UpdatingAustralia's annual temperature dataset 8l

Table 1. Continued.

Station name Station no. Latitude Longitude Elevation Start Quality Reason Urban ("s) ("E) (m) year Rank (Y/N)

POLKEMMETT -36.65 079023 142.10 014t.0 1910 4 Y PORTHEDLAND AIRPORT -20.31 004032 1r8.63 0006.4 1913 2 N PORTLINCOLN -34.60 0t8192 13s.88 0008.5 1910 3 Y PORTMACQUARIE -31.44 060026 1s2.92 0007.0 1910 2 Y POSTOFFICE -31.51 QUIRINDI 0s5049 150.68 0390.0 1910 4 PE,MD N RAYVILLE PARK -33.77 02rt33 138.22 0109.1 1910 3 N RICHMONDPOST OFFICE -20.73 QLD. 030045 143.14 02t1.1 1910 1 N RICHMONDRAAFN.S.W. -33.60 06710s 150.78 0019.0 1910 2 N ROBE -31.16 026026 139.'76 0003.3 1910 2 N ROCKHAMPTONAIRPORT -23.38 039083 150.48 0010.0 1910 2 N ROEBOURNE -20.18 00403s 117.15 0012.0 1910 2 N ROMA AIRPORT -26.55 043091 148.18 0303.2 1910 3 N ROSEWORTHY :34.51 023122 r38.68 0065.0 1910 5 N ROTTNESTISLAND -32.0r 009193 115.50 0043.1 1910 3 N RUTHERGLENRESEARCH -36.11 082039 146.51 0167.6 1910 1 N SANDY CAPELIGHTHOUSE -24.13 03908s 153.21 0099.1 1910 2 N SCONE -32.06 061089 150.93 0216.0 1910 4 MD N SOUTHERNCROSS -31.23 012014 rr9.33 0355.0 1910 3 N STAWELLAIRPORT -37.07 079105 142.14 02279 1910 4 PD,PE Y STRAHANAIRPORT -42.16 091072 145.29 0020.0 r9l0 3 N STRATHALBYNRACECOURSE -35.29 024580 138.89 0062.0 1910 3 N STREAKY BAY -32.80 018079 134.21 0013.0 1926 5 PE N SWANHILL AIRPORT -3s.38 077094 143.54 0071.0 r9l0 I N SYDNEY -33.86 066062 r51.21 0039.0 l9l0 3 Y TAMWORTHAIRPORT -31.07 055325 1s0.84 0394.9 1910 3 Y TARCOOLAAIRPORT -30.71 016098 134.s8 0123.0 1922 3 N TAREE -31.90 060030 152.48 0005.0 1910 4 PE Y TE KOWAI -21.16 033047 149.12 0013.7 1910 5 MD N TENNANT CREEKAIRPORT -19.64 015135 134.18 0315;7 1911 2 N TENTERFIELD -29.0s 0s6032 152.02 0838.0 1910 3 N TIBOOBURRAPOST OFFICE -29.44 046031 142.01 0183.0 1910 3 N TRANGIE RESEARCH -31.99 STATION 051049 147.95 0215.0 r9r3 s MD, NS N WAGGAWAGGA -35.16 0t2ts0 141.46 0212.0 1910 2 N WALGETTAIRPORT -30.04 052088 148.12 0133.0 1910 3 N WANDERING -32.61 010911 116.67 027s.0 1910 2 N WANGARAITA AIRPORT -36.42 082138 146.31 0rs2.6 1910 3' Y WARRNAMBOOLAIRPORT -38.29 090186 t42.45 0070.8 1910 2 Y WELL]NGTON -32.s6 065034 148.9s 0300.0 1910 4 PE,PD N WHITE CLIFFSPOST OFFICE -30.85 046042 143.09 0151.0 1910 4 PE,NS N WILCANNIA -31.56 046043 143.37 0075.0 1910 3 N WILSONSPROMONTORY LIGHT. -39.13 085096 146.42 0088.7 1910 1 N WILUNA -26.59 013012 120.23 0521.0 1910 5 MD N WOLLONGONGUNIVERSITY -34.40 068188 150.88 0030.0 1910 4 PE Y WOOMERAAIRPORT -31.16 016001 136.80 0166.6 1950 I N WYNDHAM POSTOFFICE -15.49 001013 128.12 0011.0 1910 4 PE,PD N YAMBA PILOT STATION -29.43 058012 rs3.36 0029.0 1910 I N YASS -34.83 070091 148.91 0s20.0 1910 5 MD N YONGALA -33.03 019062 138.7s 051s.0 1910 I N YORK -31.90 01031I 116.17 0179.0 1910 2 N YOUNG AIRPORT -34.2s 073138 r48.2s 0319.6 1910 5 MD N ten potential stations to include in the referenceseries. than 10 000 (here population values used by Torok After the initial station search,the referencestation list (1996) are usedas a proxy for determininga significant was shortenedby only acceptingstations which had at influence of urbanisation on the temperaturerecord). least ten years of monthly data and a population less Time seriesfrom referencestations also had to be hish- \z AustralianMeteorological Magazine 53:2 June2004

.y corelated with the candidate station such that the ods. Both methods show a general overall warming rne-sidedsignificance (116: correlation > 0) was 95 per trend in minimum temperatureacross Australia (not rent or greater(Press et al. 1996).The minimum num- shown). However, the magnitudes of the warming rer of referencestations required to createa reference trends in the median reference series are generally ieries was one, however the reliability of such refer- greater than the trends in the distance weighted refer- lnce series was low and objective test results were enceseries. For maximum temperature,both methods :reatedwith extremecaution if the number of reference show generalwarming trendsin the centraland west- ;tations was less than ten. ern parts of the continent and distinct cooling trends The median reference series was createdby taking in New South Wales and parts of Queensland.Again, he median of the ranked interannual temperaturedif- the magnitude of the positive and negativetrends in ,erencesfrom each suitable referencestation seriesfor the median reference series is greater in magnitude :ach year. This serieswas then added to the annual than the distanceweighted refeience series. This leads nean temperature of the first year of the candidate us to remain cautiousin the use of the median refer- ieries to createthe final median referenceseries. Torok ence sedes as a basis for the objective test. Where ,1996) suggeststhat taking the median of the interan- possiblethe results of the objectivetest basedon the rual differences is more robust to outliers caused by distanceweighted referenceseries were used to help nhomogeneitiesin reference stations than taking the identify and correct inhomogeneities. nean and that the median difference series will only The objectivetest'has been found to over-estimate re significantly biased if a systematic change in the number of discontinuitiesin a difference series rbserving techniques, data processing or exposure since the nominated distribution of the test statistic, )ccurs at the sametime throughoutthe network (e.g. Fr,r_oincorrectly assumesthat multiple discontinu- he systematic change to expose thermometers in ities are independent (Lund and Reeves 2002). The itevenson screens).Torok (1996) concedesthat the critical values of the corrected underlying distribution nedian reference series overall trend and variability are typically twice those of the Fr,,r_ocritical values, :an be biasedby small differencesin the referencesta- indicating that if detecteddiscontinuities in a differ- :ion selectionprocess and that the objective test can ence series are applied withor,rtmetadata support then letect trends in the difference series by identifying a the resulting seriescould include adjustmentsassoci- ;eriesof small discontinuities.While the detectionof atedwith naturalclimate variability or as we havejust r trend in the difference seriesis desirableif the trend discussed artificial trends associatedwith biased ref-- s the result of a slow changein the exposureof acan- erenceseries. This problem is not likely to havegreat- lidate site (e.g. vegetationgrowth), it is clearly not ly affected this dataset,since only 22 per cent of lesirable if the trend in the difference series is the adjustmentsmade were based on the results of the 'esult of an artificial trend in the median reference objective test without supporting documentation. ;eries.Trewin (2001) commentsthat the medianrefer- Someunsupporteddiscontinuitiesmustbeaccepted :nce series used by Torok (1996) can have biases as real for it is known that many historical station ntroduced when converting the median interannual changes have not been documented.The metadata lifferences back into an absolute reference series by summariesof Torok (1996) reveal large gaps in sta- he accumulation of rounding errors in the interannual tion history files. There can often be many years lifferences (also well documented in Peterson and betweenfile entries,with stationphotos or site plans Sasterling(1994)). This can introducespurious trends before and after the gap revealingobvious, but undoc- n the referenceseries. Trewin (2001) usesa distance umented, changesoccurring sometime in the inter- veighted mean of highly correlated reference station vening period. momalies to create the reference series and so avoids he problems associatedwith the conversion of inter- Reference series test rnnualdifferences. The weights, W1,dta given by Eqn The referenceseries test was simply a visual exami- I where d; is the inter-station separationmeasured in nation of the difference seriesused by the objective legrees, and r1 is the interstation correlation and N is statisticaltest to detect discontinuities.Suspect look- he number of stations. ing stepjumps in the serieswere usedas evidenceof possible dates of discontinuity in a record. Analyses a ri(6-d1)-, r>0.6,c|i.6 of the number of neighbour stationsused in a refer- w,=' [ ]i=t.....1r ...1 to ,r6) ence seriesthroughout time were undertaken,as well as which stationsactually contributed data to eachref- The trend in the median reference series and dis- erence series. These provided an indication of ance weighted reference series were compared spa- whether an apparent discontinuity in a difference ially to assessthe potential biasesin the two meth- seriescould simply be due to a changein the number Della-Martaet al.: UpdatingAustralia's annual temperature dataset 83

of referencestations, or a changein which stations usedhere are classifiedas RCSs but not all, and con_ were usedfor the reference series. sequently some were moved without a period of overlap observations being undertaken. However, Diurnal temperature range test some location moves at non-RCSs had overlap Visual examination of the unhomogenisedannual observationsavailable due to chancecircumstances, mean diurnal temperaturerange (DTR) time series at or the diligence of regional observersand managers each station was also used to detect discontinuities. who recognisedthe value of overlapobservations for Occasionallythe presence of an inhomogeneitywas all long-recordstations, not just RCSs.Of the 44 sta_ more obvious in the DTR seriesdue to oppositesign tion movesthat involved a changein station number, responsesto a non-climatic influencein the individual 30 had some parallel databetween the old and new maximum and minimum temperature series. The sites. These data were examined for quality and visual examination procedure was subjective and completeness. hence a discontinuity in the DTR series was only Only 26 station moves were considered to have acceptedif supported by metadata. overlapdata of sufficientquality to calculatea reliable adjustment between the old and new temperature records.The climatologicaldifference between an old Parallel observationdata and new site can differ throughout the year, and the period of overlap rarely corresponds to calendar Since 1993 most major observationsite moves in years.Consequently the annualmean adjustment can_ Australia have involved a changein station number. not simply be calculatedusing the averageof daily Prior to this update,if a high-qualitystation moved in differences over the overlap period. Instead, the annu_ the post-1993 period, and assumeda new station al mean adjustmentfor a site move was calculated number, its entire record was removed from the using the averageof the mean monthly differences.To datasetto avoid the known bias associatedwith com- determine the mean monthly differences for each cal_ positing a new record without adjustment.One of the endar month, all available monthly differences main objectives of this update was to resurrect as betweenthe siteswere averaged(Fig. 1). many of these 'closed' stationsas possibleby match- Generally, comparison observations for longer ing the homogenisedrecord of Torok and Nicholls than five years were found to provide excellentcom_ (1996) onto a continuingrecord. parison statisticsbetween the old and new sites,e.g. Since the development of the original dataset,44 Figs 1(a) and 1(b). Comparisonslonger rhan two stationsmoved to a new location with a new station years, and sometimesbetween one and two years, number. Usually this was due to the closure of a co- were also found to be useful if completeand of good operativemanual cibservation site after the installation quality, e.g. Fig. 1(c). Poor quality comparisonslast_ of an automatic weather station (AWS) at a nearby ing lessthan two yearswere generallyfound to be of airport. A few stationswere moved with no changein limited use, e.g. Fig. 1(d). Also, the deteriorationof site number and, theoretically, should not have been the old site can sometimesresult in some of the com_ used in the calculation of Australian mean tempera- parisondata being unrepresentativeand of little value. tures until corrected for the change. The period of comparison required for meaningful Karl et al. (1995) states that one of the basic statisticsat any particularlocation depends on the dis_ requirements for sound climate monitoring is the tribution of climatological differencesbetween the undertaking of periods of parallel observation to old and new sites. enable the impact of changes in the observation An example of high quality comparison data for regime to be assessed.This has been adoptedas one maximum temperaturesrecorded at old and new of the standardpractices for stationsincluded in the observationsites at Perth is shown in Fig. 1(a). Over Global Climare Observarion System (GCOS) ten years of complete comparison data were avail_ Surface Network which requires one, preferably able, allowing reliable differencesbetween the two two, years of overlap (Daan 2002).In Australia, an sites to be calculated.Figure 1(a) is also a good overlap period preferably of two years is requested example of how the climatological difference for site moves at Reference Climate Stations between an old and new site can differ throughout (RCSs), stations specifically chosen to monitor the year, with the mean differences during June to long-term climate changes.Prior to the development Septemberbeing of reversedsign to the remainderof of the RCS network in the early 1990s, parallel the year. Despite the good quality comparison data observations for site moves were not an official the spreadof daily differencesduring each month is Bureau of Meteorology policy and were less com- still quite high, reflecting real climatological differ_ mon. Many of the high-quality temperaturestations encesbetween the two sites. E4 AustralianMeteorological Magazine 53:2 June2004

Fig. I For each calendar month, the mean (dots), inter-quartile range (shaded) and highest/lowest values (line bars) of difference between new and old observation sites for (a) Perth maximum temperature, (b) Strathalbyn minimum temperature, (c) Stawell maximum temperature and (d) Strahan minimum temperature. Based on (a) 12 years of good quality, (b) five years of good qualit5 (c) 22 months of good quality and (d) 18 months of poor quality overlap observations.Note change in vertical scalein (d).

a) (c) (New'Old) Perth MonthlyMaximum Temperature Difference StawellMonthly Maximum Temperature Difierence {New'Old)

o q1 81 o e O^ o^ o o i-1 =-1 E e $-e E ,o -5

-4 -4

March May Jury September Nov€mb€r Jiluary January March May JUly SeDtemb€t Nov€mb€r Februsry April June Augusl October Dffimber February April June August Ociobor Oeemb6r Month Month :b) (d)

StrathalbynMonthly Minimum Temperature Difference (New-Old) StrahanMonthly Minimum Temperature Ditference (New-Old)

1 oz Yc o o I Or Pr e c o gn =-l p E.2 e^ o

.@ .o

May July Seplmb€r Nov€mb€r January March May Juty Septsnber N@ember Janury Marci April Augwt Octob€r Dsffib€r Februsry April June August Octob€r Oeemb€f Februsry June Monlh Month

The adjustments calculated using overlap data Applying homogeneityadjustments were compared to the shifts determined by the objective method where possible. Generally the objective The objective technique requires a test window of method did not detect a discontinuity due to a site data ten years in length (five years before and after a move unless it was of relatively large magnitude, test point) so is unableto detecta discontinuityoccur- lbout 0.5"C or larger. Where the discontinuity was ring within the last five yearsof record.Consequently, Jetectedthere was usually good agreementin magni- for this latestassessment any post-1988homogeneity ilde of the shift. This highlights the importance of adjustment detected by the objective method, and good quality metadataand comparisondata, and sug- supported by metadata,was applied to the updated geststhat it is likely that small discontinuitiesin the record. Some adjustmentsnot supportedby metadata recordshave not been detectedby the objectivetech- were also made. Where a discontinuity was evident nique. For site moves where no parallel data were from the reference series or DTR test, but not detect- available the shift value calculated by the objective ed by the objectivemethod, the magnitudeof adjust- techniquewas used. If this was not availablean esti- ment was estimated. An example of differences mated value was used based on subjective techniques. between the candidate and homoseneousreference Della-Martaet al.: UpdatingAustralia's annual temperature dataset 85

seriesfor Launcestonmaximum temperatureseries is Fig.2 (a) Differences between candidate and refer- shown in Fig. 2(a), alongwith the serieshomogeneity ence series (line with open circles), and the cor- adjustmentsapplied. The conesponding unadjusted responding series of applied homogeneity and adjusted temperatureseries are shown in Fig. adjustments (bold line), and (b) unadjusted (thin line) 2(b). For the posr1993assessment a total of 108new and adjusted (bold tine with open circles) annual mean maximum temperature adjustments(48 to maximum temperatureseries and series for Launceston. No adjustments were 60 to minimum temperatureseries) were made across required since 1940. the network. Overall, the updateddataset includes an (a) averageof approximatelysix adjustmentsper record since1910. The decision of whether or not to correct for a potential inhomogeneity is often a subjective one. Supporting metadata evidence of a discontinuity makesthat decisioneasier but is not alwaysavailable. Consequently,it was often impossible to reproduce the exact homogeneity adjustmentsof Torok and Nicholls (1996). Slightly different rechniques,reference stationsand sourcedatacan apparentlyproduce different results.Rather than correctthe entire dataset with new adjustments,the pre-1994 homogeneity adjustmentsdetermined by Torok (1996) were generally re-appliedfor the updatedseries, except for about 30 adjustmentsin which an obvious error had been made, such as the wrong sign of adjustmentor an incorrect year of adjustment. A few records were completely reassesseddue to significant differences between the objective test results based on the two types of referenceseries used.

Assessmentof potential urban tw warming

Torok and Nicholls (1996)used a simple definition of perature trends calculated using the actual station data an urban station, classifying any location with a popu- and an interpolatedtrend value from a spatialanalysis lation greater than 10 000 as urban. However, some basedon the non-urbanstations 'urban' definedby Torok and stations located near, but not within, a large Nicholls (1996). A significantly greater rise in the population centre, such as at an airport or lighthouse, minimum temperature at the station compared to the 'non-urban' may not necessarily have experienced a significant background (using a two-sided Kendall urban warming signal,parlicularly if the station was re- Tau test(Press et al. 1996))was assumedto be indica- located to the airport many decadesago. Torok et al. tive of an urban warming signal. However, this (2001) show thateven relatively small towns with pop- approachassumes that the stationsoriginally classi- ulations less than 10 000 can have a detectableurban fied as non-urban have not been affected bv some heat island signal during the right synoptic conditions. urbanisation. However, suchconditions are only prevalentfor a small Other information usedin the assessmentof urban- fraction of the year (Torok et al. 2001). The influence isation included the dateat which the town exceeded of urbanisationis also complicatedby poor exposureat a populationof 10 000 (ABS 1996,1986,1981,I976, many observationsites, such as being locatedin or near l97I),land-use codesdefined in SitesDBand the sta- a bituminised carpark. Consequently, it remains tion location relative to the town centre. Ultimately a unclear how much urban warming is really evident at subjectivedecision was made for each station as to the annual time-scalefor small towns. whether it was likely to have been influenced by This study attempted a number of approachesto urbanisation. Some stations originally classified as determine whether a station was likely to have been urban were re-classifiedas non-urbanand vice-versa, affected by urban warming. One of the key tools used with 39 stationsbeing classifiedas urbanin this latest was a statistical test of the difference between tem- assessment(Table 1). 86 AustralianMeteorological Magazine 53:2 June2004

The updated Australian high-quality Fig.3 Updated high-quality temperature network showing non-urban (solid circles) and urban annual temperature network (solid triangles) station locations. A total of31 ceasedtemperature records ofTorok and Nicholls (1996) have been re-opened by matching them onto a continuing record. Also, parts of inland ,S:"f Australia and Tasmaniaare now bettersampled due to the inclusion of five homogenisedrecords located at Giles, Woomera, Tennant Creek, Cape Bruny and Tarcoola (Fig. 3). These records are of shorter duration than other seriesand hencewere not included in ;rll4 t I ,/I the original dataset. 1".1 ! al -- Unfortunately sixteen of the original records were ri r \\. 1 found to have closed with potential \'t no candidatesta- 'll tion available to continue the record. Also, due to a i.l '..^r' decline or reassessmentof data quality, 81 stations +: were removedfrom the dataset.Of the remaining 133 r:.,-* recordsranked fair (rank 3) to very good (rank 1), 34 were classifiedas being urban, leaving 99 non-urban recordson which to calculateAustralia's annual mean maximum and minimum temperature.This is equalto the numberavailable prior to the update.However, the- removal of poor quality records, the application of post-1993homogeneity adjustments and the inclusion of additional stations in data-sparseregions means Fig.4 196l-1990 inter-station correlations (expres- that overall the datasethas been greatly improved. sed as Vo)for annual mean (a) maximum and The stationsincluded in the updatedAustralian annu- (b) minimum temperatures as a function of al mean temperaturedataset are listed in Table 1, and separation distance. Exponentially damped the distribution of the high-quality subset (records cosine functions are fitted where the station separation is lessthan 3000 with rank 7,2 or 3) shownin Fig. 3. (a) km. tm s s Adequacy of the updated annual 70 temperature network m 8so 2 9o In order to determine the adequacy of the updated net- q iEs

G work to resolve interannual climate variability the de- 8s correlation (maximum significant inter-stationcorrela- t0 tion) length scale of annual de-trended temperature 0 {0 anomalieswas calculatedusing the method of Jonesand -20 Trewin (2000), based on the theory of Daley (1993). -30 m Figure 4 shows all inter-station correlations over the STATIONSEPAMION (kn) 196I-1990 period for annual maximum and minimum (b) temperatureanomalies as a function of station separation. The cor:relationsshow a generaldecline with inter- tm s station separation up to around 2000 km, then tend to s remain around 0.2 or increaseslightly before becoming 70 scatteredat separationsbetween 3000 and 4000 km. s 8so Plots of the angle (0-n) between stationsagainst z {Pro inter-stationdistance indicate that the higher correla- iE 30 t tions tend to have a north-southorientation at separa- 820 tions betweenzero and 2000 km, then tend to become 10 0 more northwest-southeast at separations between -t0 2000and 3000 km (not shown).The orientationof the -20 '30 inter-stationcorelations largely reflects the station 2000

ST TION SEPARATIOT'I(km) Della-Martaet al.: UpdatingAustralia's annual temperature dataset 87

distribution(Fig. 3). However,some of the increasein the theoreticalobservation errors and compare them maximum temperature (FiS. correlations a@D and,the to the total variability of annual maximum and mini_ taperingdecline in minimum temperaturecorrelations mum temperature using the techniques of Daley at separations berween2000 and 3000 km (Fig. 4(b)), (1993).The study has shownthat the y intercept (sta_ could be due to the anisotropic northwest-southeast tion separation equals zero) of the exponentially nature of Australian temperatureanomalies (Jones dampedcosine function plotted in Fig. 4 can be used and Trewin 2000; Seaman1982). to approximate the observation error variance. Exponentially dampedcosine functions were only Theoretically, instruments that have infinitesimally f,rttedto the inter-station correlation data for separa- small separationshould be samplingthe sameair tem_ tions up to 3000 km sincethe number of inter-station perature and hence their correlation should be 1.0. correlationsdiminishes rapidly beyond this distance. However, due to instrument, observer and exposure Using a crjtical conelation (the value of 0.54 one- erors this never occurs in practice. Assuming that the sided 99.9 per cent significant correlationfor a non- homogenisedstation records in the datasetare free autocorrelated series of length 30 (Jones and Trewin from systematicbiases, the inter-stationcorrelations 2000)) the fitted curves suggest that for all points are isotropic-in space,and the remaining observation within 745 (735) km of a high-quality station, an errors are random and not correlatedin time or space, interpolated annual maximum (minimum) tempera- then the observationerror varianceis given by Eqn 3. ture anomalyvalue will be a reasonableestimate. The R, is the theoreticalcorrelation at an inter-stationdis- inter-stationcorrelations shown in Fig. 4 are basedon tance of zero and E? = Rz 02, where o'2 is the total de-trended data and therefore not inflated by time variance of the station observationswhich. in this serieshaving similar'trends. case,is basedon the 196I-1990 climatology (seeFig. Thesede-correlation length scaleswere compared 5 for plots of o). to the averageinter-station separation.Koch et al. (1983) calculatea theoreticalaverage station separa- J tion given by Eqn 2 that is weighted such that if the stationdistribution is highly clusteredits value will be greater than the simple averageof the minimum dis- Using Eqn 3, the averagestandard deviation of obser_ t^ tance from one station to another. vatronal error lEoz for annual maximum and mini_ mum temperature is estimated to be 0.25'C and Ln,= 4tt2111+ utr\t(u - l)l z 0.28'C respectively.These representapproximately 40 per cent (16 per cent) and 46per cent (22 per cent) In Eqn 2, A representsthe land surface area of of the all-Australianaverage maximum and minimum Australia (krnz)and M is the number of stationsin the annual temperature standard deviations (variance) network. Using this equationthe averagestation sep- shown in Fig. 5. As expected,the avsragestandard aration of the updatedhigh-quality ahnual tempera- deviations of observational enors !Eo2 for annual ture datasetis 316 km, comparedto the simple aver- meanmaximum and minimum temperatures(0.25"C; age inter-station separationof 163 km, indicating 0.28"C) are lower than those for mean monthly someclustering of stations.Both thesevalues are well (0.45"C;0.42"C) and daily (3.16"C;2.gg"C) anom_ below the de-correlation length scales calculated alies obtained using the same method by Jones and above using this dataset,as well as the monthly de- Trewin (2000; 2002). The results above can be inter_ correlation length scales determined by Jones and preted as,the Australian averageobservation error in Trewin (2000),suggesring thar rhe high-qualiry annu- each annual, monthly and daily station observation, al temperature network is adequate to resolve the respectively.This network is consideredto be ade_ interannualvariability of annual mean temperatures. quate since the observational errors are less than half So despitesome apparentlylarge gaps remaining in the observed standard deviation of the annual maxi_ the network the updated distribution is able to provide mum and minimum temperature. adequatecoverage of the country for the purposes of Della-Marta (2002) shows similar independent calculatingannual mean temperatureanomalies. estimatesof annual observationerror based on the summationof uncertaintiesassociated with bias cor_ rections and measurement emor with root mean Observational errors in the updated squarederror (rmse) of 0.26"C and0.24"C for annual network maximum and minimum temperatlue, respectively. These figures are not exactly comparablesince the Another approach to assessingthe adequacyof the method of Daley (1993) assumesthat any bias in a network to resolveinterannual variability is to look at station record has been removed and that the error AustralianMeteorological Magazine 53:2 June2004

tig.5 196l-1990 standard deviation of annual mean Accuracy of the objective analysis (a) maximum and (b) minimum temperature anomalies. The non-urban annual temperature records ranked 1 to 3 (99 stations detailed in Table 1) represent the high-quality subsetused to createall-Australian annual mean temperatureanomalies. A thorough investi- gation of the theoretical analysis elTors at each grid point can be achievedthrough the optimal interpolation techniquesof Daley (1993). However, a cross- validation approach was adopted here because the most influential factor on the accuracyof an analysis is the station density (Jonesand Trewin 2000) and hence this approachwill yield an upper limit to the analysiserror. Daley showsthat the analysiseror is also dependenton the accuracyofthe observationand climatological (background)error covariance structures to their true structures.Seaman (1983) showsa good example of how the theoretical analysis error increasesas a function of the misspecificationof the decorrelation length and observation error. Initially station anomalies(with respectto 1961- 90 normals) were gridded using Statistical Interpolation (SI) given a fitted dampedcosine function (Fig. 4) to explain the background covariance strlrcture (Jones and Trewin 2001; Daley 1993). Individual stationswere then removed from the analysis and the difference between the observedstation temperatureanomaly and interpolated station temperature anomaly was used to calculate a rmse due to interpolationat the stationlocation. The spatial plots (not shown) of these erors (hereafter known as interpolation errors) indicate that generally maximum temperatureanomalies have lower interpolation errors than minimum temperature'The averagerms interpolation errors are 0.3loC and 0.34'C for maximum and minimum temperature,respectively. Figure 6 showsthe ratio of rms interpolation errors to the observedstandard deviation expressed as a per- centage.This ratio shows, averagedover Australia, that interpolation errors are approximately 50 per cent and 58 per cent of the standarddeviation for maximum and minimum temperaturerespectively, a similar result to the 45 per cent determined by Jones and Trewin (2000) using monthly temperatures.The inter- expressedhere are a combination of associatedwith this bias removal is not explicitly polation elrors and analysiserror and by comparing taken into account, whereas Della-Marta (2002) observationerror the theoretical magnitude of the removesthe bias in stationrecords but retainsan esti- these figures to detailed in the last section it is clear mate of the error of the homogenisationprocess and observationerrors of interpolation erors are due to explicitly takes these into account. In reality, the that the majority rather than analysiselror. process of homogenisation will not remove a bias observationerror in Fig. 6 denotesregions in which the completely and so the observationsused to calculate The shading errors to observed standard the theoreticalobservation effor (given by Eqn 3) will ratio of rms interpolation than 0.5. For maximum temperatures contain some componentthat is due to the remaining deviationis less (Fig. be seen that there are consistently bias that will manifest itself as a change in the corre- 6(a)) it can errors along the east coast of lation structure. higher interpolation Della-Martaet al.: UpdatingAustralia's annual temperature dataset 89

Fig.6 Ratio (Vo) of rms error of cross-validated The noise-to-signalratios shownin Fig. 6 are only observations and observed standard 'true' deviation an indication of the analysiserrors since inter- for annual (a) maximum and (b) minimum polation error is only assessedat station locations. temperature using data from 1961-1990. These are likely to be an overestimateof the analysis error since, in reality, the datasetcontains a station where the cross-validationis taking place. In order to get a reliable estimate of the total analysis error the interpolation effor at each grid-point needs to be examined.This was consideredbeyond the scopeof this paper.

Annual mean temperature trends

Torok and Nicholls (1996) calculated Australian annualmean temperatureanomalies using a Theissen polygon weighted-average technique (Lavery et al. 1992). For the updateddataset this was done using the successivecorrection scheme used operationally to produce gridded analysesof temperatureand rainfall (Jonesand Weymouth 1997).Simple averagesof the interpolated grid values were used to' produce Australian annual mean maximum and minimum temperaturetime series (Fig. 7). These show the general rise in temperatureduring the 20th century which has been previously documented (eg. Torok and Nicholls 1996; Collins and Della-Marra 1999). The similarity to previous results suggests that all- Australian mean temperature anomalies based on a network ofabout 100 stationsare robust to changesin someof the individual stationswithin that network, as well as changes in the averaging technique. Linear regression over 1910 to 2002 reveals trends of 0.6+0.3'C/(100 years) in the maximum temperature series and 1.7+0.3'C/(100years) in the minimum temperature series, again consistent with previous studies.The uncertaintiesshown here are basedon the standard error of the linear residuals (Collins and Della-Marta 1999). The general rise in Australian temperaturesduring Australia and the west coast of Tasmania,probably the 20th century is consistent with temperaturetrends due to the large gradients of maximum temperature measured by surface instrumentation throughout the anomaly associatedwith broadscale topographicfea- globe. IPCC (2001) statesthat the global averagesur- tures such as the Great Dividing Range. Another face temperature increasedby 0.6+0.2"C over the region of high noise-to-signal ratio occurs in the 20th century. This global warming trend is supported northwest of Western Australia due to low station by independent data records such as those from density and high maximum temperature gradients weatherballoons, satellites, sea-ice and glacial extent, associatedwith sea-breeze (Jones effects and Trewin and is at least panly attributable to human activities 2000). Minimum temperature interpolation errors (PCC 2001). (Fig. 6(b)) generally show Iarge noise-to-signalrarios Throughout most of the country updated post- where the station density is low. Topographic influ- 1910 trend values (Fig. 8) have changedlittle com- ences on error calculations for minimum temperature pared with those of the original dataset. All of the are not as well defined as for maximum temperatures, country displays a warming trend in minimum tem- except in the southeast. Land-sea proximity effects peratures, the largest trends being in the northeast. are likely to be of greater influence. Increases in maximum temperatures have been 90 AustralianMeteorological Magazine 53:2 June2004

Fig.7 Time series of Australian annual mean maxi- Fig.8 Maps of 1910-2002linear trend in Australian mum (solid line) and minimum (dashed line) annual mean (a) maximum and (b) minimum temperatures basedon the updated non'urban temperatures ('C/100 years) based on the network (linear trends (1910-2002) also updated, non-urban network. shown). Anomalies are with respect to the 196l-1990 normal.

-1.5 @ '1910 1920 F4 \"/

strongest in the westem two-thirds of the country, while some parts of New South Wales and Queensland have actually recorded weak cooling trends since 1910. However, the magnitude and extent of the cooling in central and northern New South Wales is reduced in the updated datasetcompared to previous published analyses based on the original dataset(e.g. Collins and Della-Marta 2002). This is due to the removal of many poor quality records in the region. Most New South Wales stations, and some Queensland stations, were missing digitised data between 1957 and 1964 due to data processing problems (Trewin 2001). In inland New South Wales the annual mean, unhomogenised maximum temperatures prior to 1957 were generally greater than post- 1964 and consequently there is a decrease in the median reference series for most stations over this period. Further analysesusing only thoserecords with complete 195'7-1964data continuedto suggestsome confounding variable in the verification of this cool- maximum temperature cooling in the region since ing trend is the fact that some of the high-quality 1910. The distance weighted reference series of (New South Wales) stationscompiled by Torok and Trewin (2001) also displayed cooling through New Nicholls (1996) were not exposed in a standard South Wales and southern Queensland(not shown). Stevensonscreen until the 1940s(Torok 1996;Parker The cooling is also consistent with an increase in 1994). Where sufficient evidence (metadata) existed, annual total rainfall in the regipn (Collins and Della- these stations were removed from this dataset.Given Marta 2002). Consequently, some maximum temper- this documentednon-standard exposure at some can- ature cooling is likely to have occurred through parts didatestations it is possiblethat other stations,includ- of New South Wales and Queenslandsince 1910. A ing those used in the calculation of reference series, Della-Martaet al.; UpdatingAustralia's annual temperature dataset 'gt

might contain a negativebias in their overall trend, areas (predominately the northern and western parts which in turn could influence the outcome of the of the continent) the median referenceseries some- homogenisationprocess. A more detailed investiga- times displayed discontinuities associated with tion of the available metadataat referencestations and changesin the numberof referencestations available. the digitisation of missing 1957-1964 dara are Extreme caution was used when making undocu- required to provide a clearer understanding of the mentedchanges based on the medianreference series extent and magnitudeof this cooling. in theseareas. However, it is possiblethat recordsin remote location have not beenfully homogenised. The difficulties associatedwith developing truly Discussionand conclusions homogeneousreference series highlight the importance of good quality station metadatato aid in the The updated datasetof 99 good quality, non-urban identification of discontinuities.This requiresclimate temperaturerecords has been found to provide a reli- data users to maintain strong links with observation able measure of Australian annual mean tempera- network- managers and to continue to promote the tures.This supportsthe decision to include approxi- need for accurate,complete and accessiblestation mately 100 mainland stationsin Australia's RCS net- metadata.Technological developments must be used work. However,such a network doesnot meet all the to ensurethat the vast amountsof electronicmetada- observationneeds for climate applications and ser- ta that are now available are capturedfor the needsof vices. It is too coarseto fully describelocal-scale cli- future generations. mate or resolve trends in variables with high spatial The imperfection of statistical homogeneity tech- variability, such as rainfall. Nevertheless, a network niques also underscoresthe valueofparallel observa- of about 100 high-quality stations does provide a tions for major changesin the observationregime. backboneon which to basedenser networks suitable When artificial discontinuities are small relative to for other climate pulposes. natural variability, objectivestatistical techniques are Despite the suitability of the dataset for narional often unableto detectthe shift. In such casesthe only and regional-scaleanalyses, any individual station real way to determine the shift is with the aid of com- record within the datasetshould still be treatedwith plete and accurate comparisondata. The period of caution. The subjectivity inherently involved in the comparisonrequired is dependenton the broadscale homogeneityprocess means that two different adjust- climate and the microclimatic differencesof the loca- ment schemeswill not necessarilyresult in the same tions. In this study a good understandingof the cli- homogeneity adjustmentsbeing calculated for indi- matological impact of a site move was generallyonly vidual records.However, if the overall biasesof the achieved with at least two years of good quality over- different approachesare neutral then spatial averages lap, supportingthe Bureauof Meteorology'spolicy of should be highly consistent acrossa large number of at least one, preferably two, years of overlapping homogenisedrecords. Also, even though a subjective records for location moves at important climate obser- assessmentof the likelihood of urban contamination vation sites. within these records has been made. further work It is also vital that comparisonsbetween instru- needsto be undertakento better understand the rela- ment types continue to be undertaken in order to tive contributionof urban warming to the temperature ensure the homogeneity of the climate record. The recordsof small Australiantowns. updateddataset remains limited to the post-1910peri- Torok (1996) and Trewin (2001) both note bias od becausebefore this time the use of Stevenson problems associated with using median reference screenswas not consistentacross Australia (Nicholls seriesbased on interannualtemperature differences. et al. 1996).Reference series prior to 1910are unreli- This techniquewas comparedto the distanceweight- able becauseof the systematicchange in exposure ed referenceseries of Trewin (2001) by analysingthe across the network at approximately the same time. difference in trend in the reference series for each Nevertheless,already published exposure compar- candidatestation. The trendswere found to be consis- isons and possiblemodern comparisonsbetween his- tent in sign but not in magnitude,with trendsin the torical exposure types may eventually allow the median referenceseries often greater than those of the extensionof some homogeneousannual temperature distance weighted reference series. The total influ- records backwards in time (Nicholls et al. 1996). enceof this bias in the pre-1993homogeneity adjusr However, at most, about 60 unevenly-distributedsta- ments is small since the majority of adjustmentsare tions have sufficient data back to 1900, and only verified by documentation.Both reference seriesin about 37 stationsback to 1890, suggestingthat reli- data sparseareas were sometimesunreliable and led able all-Australian averageswill remain restricted to to misleadingresults from the objective test. In these the post-1910period. 92 Australian Meteorological Magazine 53:2 June 2004

Greater scope exists to extend monthly and sea- Collins, D.A. and Della-Marta,P.M. 2002. Atmosphericindicators sonal trend analyses further back in time than cur- for the Stateof the EnvironmentReport 2001. TechnicalReport No. 74, Bur.Met., Australia. rently possible. The records in the datasetupdated Daan,H. 2002.Manual on theGCOS surface and upper-air networks: herehave been adjustedat the annual time-scaleand GSN and CUAN. GCOS-73, WMO/TD No. 1106, WMO. lonsequently are inappropriatefor analysesat intra- Geneva. annualtime-scales. However, the datesof adjustment Daley, R. 1993.Atmospheric data analysis. Cambridge University andmetadata summaries generated could easily form Press,NY, U.S.A.,457pp. Della-Marta, P.M. 2002. Measuring uncertainty in the Australian lhe basis for a high-quality dataset corrected at the annualtemperature record. AMOS 2002: 9th National Australian monthly time-scale. A high-quality temperature Meteorological and OceanographicSociety Conference, 18-20 Jataset corrected at the daily time-scale has been February2002, University of Melbourne,p.25. leveloped by Trewin (2001) but generally only Easterling,D.R. and Peterson,T.C. 1995. A new method for detect- ing undocumenteddiscontinuities in climatologicaltime series. :xtends from around 1957 due to a lack of digitised Int. J. Climatol.. 15.369-77. laily data prior to this time. It is hoped that addition- IPCC 2001. Climate Change 2001: The Scientific Basis: il digitised data provided by the CLIMARC project Contribtftionof WorkingGroup I to the Third AssessntentReport iClarkson et al. 2001) will ultimately enable the of the IntergovernmentalPanel on Climate Change (IPCC). CambridgeUniversity Press,UK, l000pp. :xtensionof the high-quality daily recordsbackwards Jones,D.A. andTrewin, B.C. 2000.The spatialstructure of monrhly n time and provide a good coverage of corrected, temperatureanomalies over Australia.Aust. Met. Mag., 49,261- :entury-scale daily temperature records across 76. Australia.This would enablelonger monthly and sea- Jones,D.A. andTrewin, B.C. 2002.On theadequacy of digitisedhis- ;onal trend analyses than currently available. torical Australian daily temperaturedata for climate monitoring. Aust.Met Mag.,5 1, 23'7-50. Flowever, this will require significant resources and Jones,D.A. andWeymouth , G.T. 1997. An Australianmonthly rain- :onsequently the annual temperaturedataset updated fall dataset.Technical Report No. 70.Bun Met., Australia. rere will continue to provide the basis for century- Karl, T.R., Derr, V.E., Easterling,D.R., Folland, C.K., Hofmann, ;cale climate changeanalyses for sometime. D.J., Levitus, S., Nicholls, N., Parker,D.E. and Withee, G.W 1995.Critical issuesfor long-termclimate monitoring. Clirnatic Change,31,185-221. Koch, S.E., DesJardins,M. and Kocin, PJ. 1983. An interacrive Acknowledgments Barnesobjective map analysisscheme for usewith satelliteand conventionaldata. Jnl Clim. appl. Met., 22, 1487-503. fhe updateof station metadatasummaries would not Lane, J. 1995.Sites Database: Functional Requirements.Bureau of rave been possible without funding receivedthrough Meteorologyinternal report. Bur. Met., Australia. .he Bureau of Meteorology's special project bids Lavery,B., Kariko, A. and Nicholls, N. 1992.A historicalrainfall dataset for Australia.Aust. Met. Mag., 40,33-9. )rocess.The authorswish to thank Dr David Jonesfor Lund, R. and Reeves,L 2OO2.Detection of undocumentedchange- tssistancein calculating inter-stationcorrelations and points: a revision ofthe two-phaseregression model. Jnl climate, rseful discussionson observationand analysiserrors, I 5,2547-54. md Dr Blair Trewin for discussionson historicalther- Nicholls,N., Tapp,R., Burrows,K. andRichards, D. 1996.Historical thermometerexposures in Australia.Int. Clirnatol., nometer exposuresand homogenisationtechniques. J. 16,705-rc. Parker,D.E. 1994.Effects of changingexposure of thermometersat lhe authors are also grateful to Mr Neil Plummer and land stations.Int. J. Climatol.,14, l-31. )r Blair Trewin for internally reviewing a draft man- Peterson,T.C. and Easterling,D.R. 1994.Creation of homogeneous rscript, and to three external reviewers whose com- composite climatological referenceseries. lrt J. Climatol., 14, nents were useful in shaping the final form of this 6'7l-9. Peterson, )aper. T.C., Easterling,D.R., Karl, T.R., Groisman,P., Nicholls, N., Plummer,N., Torok, S.J.,Auer, I., Boehm, R., Gulletr,D., Vincent,L., Heino,R., Tuomenvirta,H., Mestre,O., Szentimrey, T., Salinger,M.J., Forland, E., Hanssen-Bauer,I., Alexandersson, H., Jones,P.D. and Parker, D.E. 1998.Homogeneity adjustments References of In Situ atmosphericclimate data:a review.Int. J. Climatol., 18, 1493-517. \BS 1996. Selected characteristicsfor urban centres 2016.0. Press,W.H., Teukolsky,S.A., Vetterling,W.T. and Flannery, Australian Bureau of Statistics,Canbena. B.P. 1996.In: Numerical Recipesin Fortran 77: The Art of Scientific \BS 1986, 1981, 1976, 1971. Censusof populationand housing. Computing, 2nd Ed. Cambridge University Press, New York, AustralianBureau of Statistics,Canberra. 630-639pp. llarkson, N.M., Trewin, 8.C., Jones, D.A., PIummer, N., Seaman,R.S. 1982.A systematicdescription of the spatialvariabili- Hutchinson,R.L. and Wong, K. 2001. Extendingthe comput- ty of geopotentialand temperature in theAustralian region. Aasr. erisedAustralian climate archivesto unlock our climate history - Met MeLg.,30, 133-41. the CLIMARC project. 14th Australia New Zealand Climate Seaman,R.S. 1983.Objective analysis accuracies of statisticalinter- Forum,Darwi4 September18-21, 2001, p.49. polation and successivecorrection schemes.Aasl. Met. Mag., 3l , lollins, D.A. and P.M. 1999.Annual Della-Marta, climatesummary 225-40. 1998:Australia's warmest year on record.Aust. Met. Mag., 48, Torok, S.J. 1996.The developmentof a high qualiry 273-83. historicalrem- perature data base for Australia. PhD Thesis, School of Earth Sciences,University of Melbourne,Australia.