Ecology and Evolution. 2017;1–15.  | 1www.ecolevol.org

Received:20June2017  |  Revised:31August2017  |  Accepted:2September2017DOI:10.1002/ece3.3499

O R I G I N A L R E S E A R C H

Ad hoc instrumentation methods in ecological studies produce highly biased temperature measurements

Adam J. Terando1,2  | Elsa Youngsteadt3 | Emily K. Meineke4  | Sara G. Prado5

ThisisanopenaccessarticleunderthetermsoftheCreativeCommonsAttributionLicense,whichpermitsuse,distributionandreproductioninanymedium,providedtheoriginalworkisproperlycited.©2017TheAuthors.Ecology and EvolutionpublishedbyJohnWiley&SonsLtd.

1DepartmentofInteriorSoutheastClimateScienceCenter,USGeologicalSurvey,Raleigh,NC,USA2DepartmentofAppliedEcology,NorthCarolinaStateUniversity,Raleigh,NC,USA3DepartmentofEntomologyandPlantPathology,NorthCarolinaStateUniversity,Raleigh,NC,USA4DepartmentofOrganismicandEvolutionaryBiology,HarvardUniversityHerbaria,Cambridge,MA,USA5NorthCarolinaCooperativeFishandWildlifeResearchUnit,DepartmentofAppliedEcology,NorthCarolinaStateUniversity,Raleigh,NC,USA

CorrespondenceAdamJ.Terando,DepartmentofInteriorSoutheastClimateScienceCenter,USGeologicalSurvey,Raleigh,NC,USA.Email:aterando@usgs.gov

AbstractInlightofglobalclimatechange,ecologicalstudiesincreasinglyaddresseffectsoftem-peratureonorganisms andecosystems.Tomeasure air temperature at biologicallyrelevantscalesinthefield,ecologistsoftenusesmall,portabletemperaturesensors.Sensorsmustbeshieldedfromsolarradiationtoprovideaccuratetemperaturemeas-urements,butourreviewof18yearsofecologicalliteratureindicatesthatshieldingpracticesvaryacrossstudies(whenreportedatall),andthatecologistsofteninventandconstructadhocradiationshieldswithouttestingtheirefficacy.Weperformedtwo field experiments to examine the accuracy of temperature observations fromthreecommonlyusedportabledataloggers(HOBOPro,HOBOPendant,andiButtonhygrochron)housedinmanufacturedGillshieldsoradhoc,custom-fabricatedshieldsconstructed fromeverydaymaterials such as plastic cups.We installed this sensorarray(fivereplicatesof11sensor-shieldcombinations)atweatherstationslocatedinopenandforestedsites.HOBOProsensorswithGillshieldswerethemostaccuratedevices,withameanabsoluteerrorof0.2°Crelativetoweatherstationsateachsite.Errorinadhocshieldtreatmentsrangedfrom0.8to3.0°C,withthelargesterrorsattheopensite.Wethendeployedonereplicateofeachsensor-shieldcombinationatfivesitesthatvariedintheamountofurbanimpervioussurfacecover,whichpresentsafurthershieldingchallenge.Biasinsensorspairedwithadhocshieldsincreasedbyupto0.7°Cforevery10%increaseinimpervioussurface.Ourresultsindicatethat,duetovariableshieldingpractices,theecologicalliteraturelikelyincludeshighlybiasedtemperaturedatathatcannotbecompareddirectlyacrossstudies.Ifleftunaddressed,theseerrorswillhindereffortstopredictbiologicalresponsestoclimatechange.Wecallforgreaterstandardizationinhowtemperaturedataarerecordedinthefield,han-dledinanalyses,andreportedinpublications.

K E Y W O R D S

airtemperature,climatechange,datalogger,HOBO,iButton,radiationshield

1  | INTRODUCTION

Global surface temperatures have increased an average of 0.85°Csince1880(Hartmannetal.,2013),resultinginprofoundchangesin

phenology (Anderson, Inouye,McKinney, Colautti, &Mitchell-Olds,2012;Waltheretal.,2002),abundances(Parmesan,2006),anddistri-butions(Chen,Hill,Ohlemüller,Roy,&Thomas,2011;Moritz&Agudo,2013) of species worldwide. Yet, large uncertainties remain about

2  |     TERANDO ET Al.

the ecological consequences of recent warming. The constructionof accurate, mechanistic models of biological sensitivity to climatechangewillbeacriticalpartofmanagementeffortstopreventglobalbiodiversity loss (Urbanetal.,2016).Thesemodels requireaccurateandpreciseestimatesoftemperatureatthescaleofrelevantbiologicalprocesses.

Permanent, stationary weather stations are currently the goldstandardformonitoringairtemperaturesinthefield(Diamondetal.,2014; Forister& Shapiro, 2003;Marra, Francis,Mulvihill, &Moore,2005; Rundel, Graham, Allen, Fisher, & Harmon, 2009). However,theseinstrumentsaresparselydistributedandareoftensitedinflat,openareas.Incontrast,biologicalprocesses,andthereforeterrestrialfieldecologystudies,aremorelikelytooperateatfinerspatialscales.At these scales, local environmental variationmore strongly affectsairtemperature(Fridley,2009;Potter,Woods,&Pincebourde,2013),which in turn affects numerous biological and ecological processes(Chenetal., 1999). In response to theneed for thesemicroclimaticdata,ecologistshaveincreasinglyturnedtosmallandinexpensiveen-vironmentalsensors(alsoknownasdataloggers)thatcanbequicklydeployed and simultaneously record observations at high densitiesinthefield.Thesedevicescapturethelocalvariancestructureofairtemperature,improvingecologists’abilitytomakeinferencesaboutitseffectonthebiotaof interest. It isworthnoting thatother thermalparameters, such as operative temperature or habitat surface tem-peratures,maybe themostbiologically relevantmeasures forsomeresearchquestions(Bakken,1992;Huey,Peterson,Arnold,&Porter,1989).Wefocushereonair temperaturebecause ithasa longhis-toricalrecord,iswidelyused,andhaspotentialtobeconsistentlyde-ployedinwaysthatallowcomparisonbetweenstudiesconductedindifferenthabitats.

Temperaturesensorsvary inaccuracy,precision,andprice,whichcanmakechoosingamongthemdifficult.Inadditiontotheseinternaldifferences, sensors are sensitive to solar radiation,which can resultinsignificantbiasesinrecordedobservationsduringperiodsofdirectsunlightandhightemperatures(Holden,Klene,Keefe,&Moisen,2013;Hubbart, Link,Campbell,&Cobos,2005).Radiation shields canbuf-fertheseinaccuracies(Holdenetal.,2013;Hubbart,2011)butarenotavailableformanyinexpensivesensorsandaremarketedasanoptionalpurchasewithmoreexpensiveones.Asacost-effectivesolution,ecol-ogistshavedevelopedseveraltypesofradiationshieldsusinginexpen-sive everyday materials. These custom-fabricated shields seeminglyobviatetheneedtopurchaseexpensivemanufacturedshields,allowingecologiststotakefulladvantageofthelow-costtemperaturesensors.

Andyet,weareawareofonlyasmallnumberofstudieswhereintheefficacyofcustom-fabricatedradiationshieldsforusewith low-costdataloggersisassessedthroughsystematiccomparisontoman-ufacturedshieldsorpermanentweatherstations(Ashcroft&Gollan,2013; Cheung, Levermore, & Watkins, 2010; Holden etal., 2013;Hubbart, 2011; Hubbart etal., 2005; Lundquist & Huggett, 2008;Tarara&Hoheisel,2007).Thesestudies,noneofwhichareinecology-focusedpublications,havegenerallyconcluded that sensorshousedincustom-fabricatedshieldsprovideacceptableaccuracy,withmeanbiasesof<1°Crelativetothechosenstandard(i.e.,aweatherstation

orasensorhousedinamanufacturedshield)butresultsvaryacrossenvironments.Forexample,ashieldconstructedoftwomodifiedfun-nelsperformednearlyaswellasmanufacturedshieldsinagreenhouse(Hubbart,2011)butyieldedabiasofupto8°Cinfieldtests(Holdenetal.,2013).Suchdifferencesamongcustom-fabricatedshieldscastdoubt on the accuracy and comparability of temperature data re-portedinstudiesthatrelyondifferent(andoftenuntested)methods.Moreover,giventhecostadvantagesoftheleastexpensivetempera-ture sensors, these errors could rapidly proliferate if standardizedmethods tominimizebiasesarenotdevelopedandadopted in fieldecology,significantlyhinderingourabilitytoaccuratelypredictspeciesresponsesandsensitivitiestoclimatechange.

Here,weassesscurrenttrendsintheuseandshieldingofportabletemperaturesensorsbysampling18yearsofecologicalliterature.Wethenpresenttheresultsofanexperimentdesignedtotesttheaccu-racyofthemostcommonlyusedsmall,portabletemperaturesensorsacrossthreeenvironmentalsettingswherefieldecologyisoftencar-riedout:openfields,closed-canopytemperateforests,andurbanizedareas.Thefirsttwositeswerecolocatedwithpermanentweathersta-tionsthatincludedhigh-qualitytemperaturesensors,whiletheurbansitesspannedagradientofimpervioussurfacecoverthatrevealhowobservationbiasescanvaryacrosshabitatsinatypicalfieldstudy.Weappliedmultipletreatmentcombinationsconsistingofmanufacturedand custom-fabricated radiation shields.Our goalwas to provide amuch-neededadvanceinthedevelopmentofstandardizedmethodsforaccuratelymeasuringandmonitoringairtemperature,whenusinginexpensivesensorsinfieldecologyandglobalchangestudies.

2  | MATERIALS AND METHODS

2.1 | Assessment of current practice in ecology

Weconducteda literature search todeterminewhether theuseofsmall, portable temperature sensors in ecology has become morecommonoverthepast18years.WefocusedourreviewonHOBOs(Onset Computer Corporation, Bourne, MA) and iButtons (MaximIntegrated,SanJose,CA)as,anecdotally, thesearecommonlyusedsensorsinfieldecology.Although“HOBOs”includeavarietyofport-ableenvironmental sensors,weused thebroadsearch for “HOBO”becausemodelnamesandnumbersarenotconsistentlyreportedintheliterature(seetheSupplementaryInformationforcompletelistofHOBOslistedinthepapersexamined).Wethenassessedasubsetoftherecoveredpaperstodeterminehowecologistswereusingthesesensors,and,inthecaseofairtemperaturemeasurements,whetherandhowsensorswereshieldedfromdirectanddiffusesolarradiation.BecauseaWebofSciencesearch returned fewrelevant results forthenamesofcommonsensors(“iButton”and“HOBO”),weselected20journalstosearchdirectly.Toidentifytargetjournals,weusedtheISIJournalCitationReportforthecategory“ecology”andconsideredalljournalswitha5-yearimpactfactor>3.5.Wethenexcludedjour-nalsthatpublishprimarilyreviews,thescopeofwhichdidnotincludefieldecology,andthosethatyieldednosearchresultsfor“iButton”or“HOBO.”The20journalsusedarelistedinTableS1.

     |  3TERANDO ET Al.

ToassesshowthefrequencyofiButtonandHOBOusehaschangedover time, we used each journal’s ownwebsite search function tosearchseparatelyforkeywords“iButton”and“HOBO.”Wescannedtheresultstoconfirmrelevance,forexample, that“HOBO”referredto an environmental sensor andnot a genetic element, and that allresultswere original research contributions and not review articles.Werecordedthenumberofpaperspublishedusingeachsensortypeineachyearfrom1998through2015.Wechose1998asthestartingyearbecauseitwasthefirstyearofpublicationoronlinearchivingofrelevantjournalssuchasEcology Letters,Diversity and Distributions,andEcosystems.Toavoidinflatingthenumberofstudiesin2015,weex-cludedpapersthatwerepublishedonlinein2015pendingassignmenttoa2016issue.WeperformedallsearchesinFebruary2016.

ToexaminethedetailsofhowecologistsuseandshieldiButtonsandHOBOs,weexaminedupto25papersperjournalfromthe“iBut-ton”searches.Thisresulted inasampleof170papers,representing100%ofstudies inthedesignatedtimeperiod inall journalsexceptOecologia,fromwhichwerandomlyselectedasubsetof25ofthe40papersavailable.Asthe“HOBO”searchreturnedalmostthreetimesasmany records compared to the iButton search,we randomly se-lectedasubsetofpapersfromeach journal toequal thenumberofiButtonpapersfromthesamejournal.Hence,wealsoexaminedatotalof170papersthatmentionedHOBOsensors(1–25perjournal).SixpapersweresharedbetweentheiButtonandHOBOgroups;thetotalnumber of papers examinedwas 334 (see Supporting Information).Foreachpaper,werecordedsensoridentitiesandenvironmentalpa-rametersmeasured(e.g.,soiltemperature,airtemperature,andwatertemperature). For those that measured air temperature,we furtherrecordedanydetailsorcitationsabouthowthesensorwasprotectedfromsolarradiation.

2.2 | Field experiment: temperature sensors and radiation shields

To assess the accuracy of commonly used environmental sensors,we conducted a field experiment using a total of twelve tempera-turesensor-radiation-shieldcombinations tocompare topermanentweatherstations.Weusedthreetypesoftemperaturesensors:iBut-ton(DS1923Hygrochron,MaximIntegrated,SanJose,CA),HOBOPendant (UA-001-08, Onset Computer Corporation, Bourne, MA),and HOBO Pro (U23-001 Pro v2, Onset Computer Corporation;TableS2).TheHOBOProisaself-containeddataloggerwithanat-tachedthermistor-basedtemperaturesensor;theHOBOPendant isasmallerandlower-costthermistor-baseddatalogger,andtheiBut-tonhygrochronisalow-costdataloggerwithasilicon-basedinternaltemperaturesensor.Weassignedeachsensormodeltoanunshieldedtreatmentandatleastone(uptoseven)shieldedtreatments(Figure1).

Specifically,weassignedHOBOProsensorstoonlyoneshieldedtreatment, the manufacturer-recommended M-RSA naturally ven-tilated multiplate solar radiation shield (also known as a “Gill”shield, (Gill, 1979); Figure1a).We also subjected HOBO Pendantsto only one shielded treatment, the custom-fabricated Radshieldof Holden etal. (2013) (Figure1b). We subjected iButtons to the

seven custom-fabricated shield treatments shown in Figure1 (nomanufacturer-recommendedradiationshieldexistsfortheiButtons).These include three previously described shields and four originalvariations.Thepreviouslydescribedshields include (i) theRadshieldofHoldenetal.(2013)(Figure1b);(ii)theAlternativeradiationshieldof Hubbart (2011) (Figure1c); and (iii) a translucent plastic cup, asdescribedbyMeineke,Dunn, Sexton, andFrank (2013) andCarper,Adler,Warren,and Irwin (2014) (Figure1d).Allwereconstructed tothebestofourabilities,basedonourunderstandingofthemethodsandmaterialsused.However,minordeviationsinourattemptedrep-licationoftheauthors’methodsmayaffectanyreportedbiases.Thefirst of the original variationswas a smaller radiation shield similartothatofHoldenetal. (2013)withthedimensionsreducedby50%(Figure1e).Wealsotestedthreevariationsoftheplasticcupdesignthatincorporated(i)ventilation(Figure1f),(ii)ventilationandamorereflective,whiteoutercup(Figure1g),and(iii)ventilation,whiteoutercup, and abasal foil shield toprotect the sensor from radiation re-flected from the ground (Figure1h). Details of shieldmaterials anddimensions are included in Fig. S1. iButtons in all treatmentsweremounted using iButton wall mounts (DS9093S, Maxim Integrated)withtwisttiesorcableties.WetestedmoreiButtontreatmentsthanHOBOtreatmentsbecause iButtonsaresmaller, lessexpensive,andlessconspicuousthaneitherHOBOsensor,oftenmakingthemmoreattractiveforlarge-scalestudies,andinareaswheretheftorestheticsareanissue.ManyofthetestediButtonshieldsweretoosmalltoac-commodateHOBOsensors.

2.3 | Sensor accuracy in exposed and forested locations

To examine the accuracy of unshielded and differently shieldedtemperaturesensors,weused60sensors,comprisingfivereplicatesof each of the 12 sensor-shield combinations (Table S3). All sen-sorswereprogrammedtorecordtemperaturesynchronouslyevery30min(onthehourandhalf-hour)andthenweredeployedtorecordtemperaturesattwositesthatcontainpermanentweatherstationswithcalibratedtemperaturesensors.Thefirstweatherstation(re-ferredtoasLakeWheeler),partoftheNorthCarolinaEnvironmentandClimateObservingNetwork(ECONet),islocatedinafullyex-posedrural locationnearRaleigh,NC(35.728°N,78.680°W).ThisstationisoutfittedwithaVAISALAplatinumresistanceairtempera-turesensor(HMP45C)mountedinsideawind-aspiratedmultiplateradiationshield (Fig.S2a).Air temperaturewas loggedat this siteforthe16-dayperiodcovering6–21August2015 intypicalsum-merconditionsforthesoutheasternUnitedStates.Wethenplacedthe60 sensors in a rural forested location from21 to28August2015,ataRemoteAssessmentofForestEcosystemStress(RAFES)network fixed weather station (referred to as Duke Forest) nearDurham,NC(35.98°N,−79.09°W;Fig.S2b),whichalsousesanatu-rally aspiratedmultiplate radiation shield. Sensorswere randomlyplaced on 1–2 meter-long 25mm×51mm wood boards, whichwerepassedthroughtheinstrumenttowerateachstation,placingthe sensors at the same height as the permanentwind-aspirated

4  |     TERANDO ET Al.

temperature sensors, approximately 2m above the ground. Thefixed weather stations recorded temperature every minute (LakeWheeler)oreveryhour(DukeForest),andweobtainedthesedataforthetimeperiodsthatoursensorswereinstalledateachstation.

2.4 | Variation in sensor performance along an impervious surface gradient

To determine how sensor accuracy varied across field conditions,we placed sensors at five sites along an impervious surface gradi-ent,whichallowedustoevaluatetheeffectofvariationinupward-directedradiationonsensoraccuracy.Ateachsite,weselectedafocaltreeandsuspended12sensors,onepersensor/shieldcombination,onabranch2–3maboveground.Sensorscontinuedtorecordsyn-chronouslyeveryhalf-hourandwereon site2–9September2015.Wemeasuredimperviousgroundcoverwithin100mofthefocaltreeusingArcMapversion10.3.1(ESRI,Redlands,CA),witha1mresolu-tionimpervioussurfacemapofRaleigh,NC,obtainedfromtheWake

County GIS Map Services website (http://www.wakegov.com/gis/services/pages/data.aspx). The impervious surface sites included anear-urbanforestedsite(0%impervioussurface),asuburbanresiden-tialbackyard(20%),anurbanresidentialfrontyard(31%),astreet-sidelawn(41%),andaparkinglot(46%).

2.5 | Data analysis

All statistical analyseswere conducted usingR version3.3 (RCoreTeam,2016).Weused twoerrormetrics toassessoverallbiasandaccuracyofthesensor/shieldtreatmentsinsunnyandshadedcondi-tions.First,wecalculatedtheaveragebiasofeachsensortreatmentinrelationtothetwopermanentweatherstations:

where si is the recorded temperature for the sensor/shield com-bination for observation i, oi is the corresponding weather station

bias=1

n

n∑

i=1

(si−oi)

F IGURE  1 Custom-fabricatedradiationshieldstestedinthefield.(a–d)areradiationshieldspreviouslyusedinpublishedresearchpapers.(a)Amanufacturedradiationshield,alsoreferredtoasa“gillshield”(OnsetComputerCorp.,Bourne,Massachusetts,partMRSA).(b,c)Custom-fabricatedradiationshieldstestedbyHoldenetal.(2013)andHubbart(2011),respectively.(d)Acustom-fabricatedshieldusedinafieldstudy(Carperetal.,2014;Meinekeetal.,2013).(e–h)areradiationshieldscreatedfortestinginthisstudy.(e)Asmallerversionof(b).(f)Amodificationof(d)withholestoallowairflow.(g)Amodificationof(c)usingalargerwhitecup.(h)Amodificationof(g)withashieldplacedbelowthesensor.ConstructiondetailsareprovidedintheFig.S1

(a) (b)

(c) (d)

(e) (f)

(g) (h)

     |  5TERANDO ET Al.

observation,andnisthetotalnumberofrecordedobservations.Wealsocalculatedthemeanabsoluteerror(MAE)toestimatetheoverallexpectederrorforeachsensor/shieldcombination:

wherethesymbolsarethesameasinthebiasequation.BoththebiasandtheMAEvalueswerecalculatedseparatelyfortheperiodsfrom6a.m.to8p.m.(LST)and8p.m.to6a.m.tohighlighttheeffectsofsolarradiationonsensorreadings.TheerrormetricswerecalculatedforeachtreatmentreplicateandthenaveragedtoobtaintheoverallbiasorMAEvalue.

WeobservedthattheDukeForestpermanentweatherstationmayitselfbemiscalibratedorexposedtositeconditionsthatappeartocausetemperaturerecordingsthatarebiasedlowinthemorninghoursaroundandaftersunrise (seeFigure3).Whilewecouldnotdetermine the exact cause of these anomalies, it did appear thatthebiaswasconsistentacrossourexperimentalperiod.Therefore,we calculated an additional “Adjusted”MAE for the sensor treat-ment data associatedwith the Duke Forest site. To calculate theadjusted MAE, we subtracted the daytime MAE of the shieldedHOBO Pro sensor at the LakeWheeler site (the sensorwith thelowest overallMAE values) from theMAE of the same sensor attheDukeForestsite.Wethensubtractedthisvalue,0.25°C, fromtheMAEofeachsensor/shieldcombinationattheDukeForestsiteto obtain an estimated daytimeMAE.The rationale for using thisadjustmentisthatweassumethatanexperimentperformedusingapermanentweatherstationofsimilarqualityas towhat isavail-ableattheLakeWheelersiteshouldresultinsimilarMAEvaluesforthemostaccuratesensor/shieldtreatment(i.e.,theshieldedHOBOPro).Therefore,theadjustedMAEfortheshieldedHOBOProattheDukeForestsiteequalstheMAEfortheshieldedHOBOProattheLakeWheelersite.

For the impervioussurfacegradientexperiment,weused lin-ear regression to estimate the effect of percent impervious sur-face(thepredictor)onrecordedtemperatures.Generalizedlinearregression models were fit independently for each hour of thedayusing theR functionglm (RCoreTeam,2016)withassumedGaussianerrors.

Wealsousedlinearregressiontotesttheextenttowhichsolarradiation,windspeed,and the interactionbetween these twoco-variates,canexplaintheobservedbiasesinthesensor/shieldcom-binations. For this analysis,weused2m solar radiation andwindspeedobservationsfromtheLakeWheelersite,whichisequivalenttooursensorheights.Nosolarradiationobservationswererecordedfor theDuke Forest site, sowe limited our analysis to the sunnyandopenLakeWheelersite(wheretheeffectswereexpectedtobelarger).Wetestedfivemodels(Table1)thatrepresentedvariationson the hypothesis that high amounts of solar radiation combinedwith low wind speeds would result in the largest sensor biases.OnceagainweusedtheglmfunctioninRtoestimatetheregressionmodel, andAkaike InformationCriterion (AIC)was used to evalu-atemodel fit and parsimony. Finally, the predicted bias given the

interactionbetweenwindspeedandsolarradiationlevelsisplottedusingtheeffectspackageinR(Fox,2003).

Our initial analysis of the recorded temperatures at the LakeWheelerandDukeForestsitessuggestedthatseveraloftheiButtontreatmentswererecordingnearlyidenticaltemperatures(cf.Figure3).AnANOVAtestwasconductedandtheresults(notshown)indicatedno statistically significant differences between the four plastic cuptreatmentsateitheroftheweatherstationlocations,norweretherestatistically significant differences between the original “Radshield”and ourmodified “Small Radshield.” Based on these results, for theregressionanalysisofsolarradiationandwindspeedeffects(andforthe conditional quantile plots in Figure 4),we pooled the recordedtemperaturesfromthesegroups,increasingthenumberofreplicates,sothattheoriginalsevenshieldedtreatmentsappliedtotheiButtonsensorswerereducedtothreegroups:CUPS(includesthe“Cup,”“Cupventilated,” “Cup,ventilated&sheltered,”and“Cup,ventilated,shel-tered,shieldedbeneath”treatments),Alternativeradiationshield(onetreatmentonly),andRadshields(“Radshield,”“SmallRadshield”).

All data associated with this research project and manuscriptare publicly available and can be found at https://doi.org/10.5066/f7b56hpw.Pleasecontactthefirstauthorforanyquestionsconcern-ingthedataormetadata.

3  | RESULTS

3.1 | Current practice in ecology

Oursearchof18yearsofliterature(1998–2015)in20targetecologyjournalsidentifiedatotalof185papersthatusediButtonsand539papers that usedHOBOdata sensors.Useof these small, portablesensorshasincreasedovertime,from0iButtonand5HOBOpapersin1998to34iButtonand72HOBOpapersin2015(Figure2a).Amongthe334studiesthatweexaminedingreaterdetail,werecorded417sensor applications (some studies recorded multiple environmentalparameters,andsixusedbothiButtonsandHOBOs,resultinginmul-tipleapplicationsperpaper.)Aboutone-thirdoftheapplicationsusedsensorstomeasureandrecordairtemperature.Othersusedthemtorecordparameterssuchasbody,soil,surface,orwatertemperature(TableS4).

Amongthe138papersthatusedthesensorstorecordairtempera-ture,nearlyhalfdidnotreportradiationshieldingpractices(Figure2b).Evenifbestpracticeswereactuallyusedinthesestudies,theywere

MAE=1

n

n∑

i=1

∣ si−oi ∣

TABLE  1 ListofmodelsconsideredintheregressionanalysisoftheeffectsofsolarradiationandwindspeedonairtemperaturesensorbiasrelativetotheLakeWheelerweatherstation

Model # Model description

1 Solarradiation+error

2 Model1+inverse(windspeed)

3 Model1+log(inverse[windspeed])

4 Model2+solarradiation×inverse(windspeed)

5 Model3+solarradiation×log(inverse[windspeed])

6  |     TERANDO ET Al.

notreportedinthemethodsnorillustratedinafigure.Seventeenper-cent apparentlydidnot require shieldingbecause sensorswerede-ployedatnightorindoors,orrecordedtemperatureminima.Only35%ofpapersmentionedshielding,andamongthissubsetof48papers,only 10% provided either product information for a manufacturedshieldorvalidationofacustom-fabricatedshield.Inthiscase,valida-tioncouldincludeadditionaldatathatinsomewayassesstheefficacyoftheshield,oracitationthatincludedsuchdata.Theremaining90%ofthepapersthatmentionedshieldingfellintothreecategories:theyusedcustom-fabricatedshieldswithoutvalidation(44%),providedtoolittleinformationtounderstandhowsensorswereshielded(33%),ordeployedsensorsinnaturalshade(13%).

3.2 | Sensor accuracy under different shield types

Our comparison of different sensor and radiation shield combina-tions showed large biases across all the custom-fabricated shieldsexceptforthepreviouslypublishedHoldenetal.(2013)methodandourmodificationofit(Table2).MeanbiaseswerelargestattheLakeWheelersiteandrangedfromaverysmallpositivebiasof0.05°Cfor

theHOBOProsensorwiththemanufacturedradiationshield(sensorshowninFigure1a),toa3.39°CpositivebiasfortheunshieldedHOBOPendantsensor.Thelargestbiasforashieldedsensorwas3.03°CfortheAlternativeradiationshield (seen inFigure1c),whilethe largestbias among theunpublishedmethods for custom-fabricated shieldswas2.84°C for theventilated and shelteredplastic cup (Figure1g).Thesmallestbiasamongthecustom-fabricatedmethodswas0.75°CfortheRadshielddesign(Figure1b)describedbyHoldenetal.(2013).Our smaller version of this design (Figure1e) had a similar bias of0.81°C.TheresultsfortheDukeForestsitefollowedasimilarpatternbutwithsmallerpositivebiases.Evenatthisforestedsite,alliButtonandHOBOPendantsensor/shieldcombinations(includingnoshields),withtheexceptionoftheRadshieldandSmallRadshielddesigns,hadpositivebiases>1°Cduringtheday.

ThedaytimeMAEvaluesformostofthesensor/shieldcombina-tionsweresimilartothebiasvaluesandhadthesamerank-orderattheLakeWheelersite(Table3).Atthissite,thelargestabsolutediffer-encebetweenthebiasandtheMAEforasensorwas0.16°CfortheshieldedHOBOProsensor(equaltoanMAEof0.21°C).Thisindicatesthatfortherestofthesensor/shieldcombinations,mostoftheMAEisexplainedbythepositivebiaserrors,whiletheshieldedHOBOProisaccuratetoapproximately0.2°Cwithasmallbiasof0.05°C.Incon-trast,thenighttimeMAEvaluesaresimilarforallsensor/shieldcom-binations,rangingfrom0.19°CfortheunshieldediButtonto0.30fortheunshieldedHOBOProattheLakeWheelersite(TableS5).

We observed a similar rank-order but with smaller (adjusted)MAE values for the Duke Forest site. The shaded site conditions

F IGURE  2 UseandreportedradiationshieldingpracticesofiButtonandHOBOtemperaturesensorsinecologicalfieldstudies.Despitetheincreasinguseofsuchsensorsovertime(a),airtemperaturedatacollectedwiththesedevicesaremarkedbyunevendeploymentandreportingofmethods(b)(n=138papersthatmeasuredairtemperature)

TABLE  2 Meandaytime(6a.m.–8p.m.LST)biasforeachsensorandshieldtreatmentattheopenLakeWheelerandforestedDukeForestsites

Sensor Treatment

Daytime bias (°C)

Open site Forested site

HOBOPro Manufactured(Gill)shield

0.05 0.34

iButton Radshield 0.75 0.93

iButton SmallRadshield 0.81 0.93

HOBOPendant Radshield 0.92 0.90

HOBOPro Noshield 1.05 0.72

iButton Noshield 2.17 1.49

iButton Cup 2.44 1.56

iButton Cup,ventilated,sheltered,andshieldedbeneath

2.57 1.80

iButton Cup,ventilated 2.60 1.74

iButton Cup,ventilatedandsheltered

2.84 1.72

iButton Alternativeradiationshield

3.03 1.71

HOBOPendant Noshield 3.39 1.51

Sensor/shieldcombinationsarerank-orderedfromlowesttohighestbiasattheLakeWheelersite.

     |  7TERANDO ET Al.

likelyreducethebiasesanderrorsthatresultfromdirectsolarradi-ationexposure.However,evenintheseforestedconditions,exceptfortheshieldedHOBOProsensor,allofthesensor/shieldcombina-tionshaveadjustedMAEvaluesabove0.5°C,andonlythemanufac-turedGillShieldandRadshielddesignshadMAEvalues thatwere<1°C.

These large biases and errors can be seen in the diurnal tem-peratureobservationsateachsite(Figure3).Forsomesensor/shieldcombinations, averagemidday (between1100and1700hours LST)readingsweremorethan5°Cwarmer,particularlyatthesunnyLakeWheelersite.Thenon-RadshieldiButtontreatmentsshowparticularlylargedaytimebiases(Figure3a,b).Andconsistentwithexpectations,theunshieldedsensorsinallcasesandatbothsiteshadaveragemid-daybiasesofatleast1°C.Overnighttemperaturereadingswereverysimilartotheobservedweatherstationvalues,furthersuggestingthatmany custom-fabricated shield types are not adequately mitigatingthe effects of solar radiation on the accuracy of inexpensive tem-perature sensors.Notably, the shieldedHOBOPro sensor recordedalmost exactly the same temperatures as the LakeWheeler perma-nentweatherstation(Figure3e).Overall,themajorityofthedaytime(06:00–20:00hoursLST)observationsforallunshieldedsensorsandcustom-fabricated shielded sensors had biases >1°C,with the four“Cup” treatments and the Alternative radiation shield designs ap-proaching100%ofmiddayobservationsthatwereatleast1°Cwarmerthanthepermanentweatherstations(Fig.S3).

TheeffectsofsolarradiationonsensorbiasareevenstrongerathighertemperaturesasseenintheconditionalquantileplotsfortheLakeWheelerresultsinFigure4.Notonlyistheaveragebiasoftheunshielded and custom-fabricated shield treatments highest during

thedaytimewhensolarradiationisstrongest,butthevarianceofthebiasincreasessubstantiallyatthehighesttemperatures.Inparticular,the unshielded iButtons (Figure4a), the unshieldedHOBOPendant(Figure4e), the combined iButtonCUPS treatments (Figure4b), andtheiButtonAlternativeradiationshield(Figure4c)allrecordincreas-inglyextremetemperaturesinconjunctionwiththewarmestweatherstation observations. For some shield/sensor combinations (i.e.,Figure4b,c,e)the10thandeventhe25thquantilesextendwellabove40°Cwhen theweather station air temperature readings approach35°C. In contrast, thebiases for theRadshield treatmentsaremoremuted (Figure4d,f),andtherecordedtemperatures for theshieldedHOBO Pro are nearly identical to the permanent weather stationobservations.

3.3 | Variation in sensor performance along an impervious surface gradient

Asexpected,duetourbanheatislandeffectsallsensorsrecordedincreasing temperatureswith increased impervious surface cover,with stronger effects during daylight hours (Figure5). However,thevariationinobservedairtemperatureduringthedaytimehours(Figure5b)wasalsoaffectedbythetypeofsensorandshieldused.Forexample,theregressioncoefficients(withstandarderrors)thatestimatetheeffectofimpervioussurfacecoveronairtemperatureobservationsrecordedat3a.m.alloverlapeachother(Fig.S4).Butthecoefficientsforthesamemodelfittothe3p.m.datashowbe-tweentwoandsevennonoverlappingestimatesmeasuredagainstallpairwisecombinationsofsensors/shields(Fig.S4).Modelresultsindicatethat impervioussurfacecoverhadthesmallestafternoon

Sensor Treatment

Daytime MAE (°C)

Open site Forested siteAdjusted forested site

HOBOPro Manufactured(Gill)shield 0.21 0.47 0.21

iButton Radshield 0.75 0.96 0.70

iButton SmallRadshield 0.81 0.98 0.73

HOBOPendant Radshield 0.92 0.93 0.67

HOBOPro Noshield 1.13 0.83 0.57

iButton Noshield 2.18 1.56 1.31

iButton Cup 2.47 1.63 1.37

iButton Cup,ventilated,sheltered,andshieldedbeneath

2.58 1.83 1.58

iButton Cup,ventilated 2.62 1.80 1.55

iButton Cup,ventilatedandsheltered

2.86 1.76 1.51

iButton Alternativeradiationshield 3.04 1.78 1.52

HOBOPendant Noshield 3.40 1.58 1.32

An“Adjusted”MAEwascalculatedfortheforestedsite,bycalculatingthebiasbetweentheHOBOProGill shield and theRAFES station, and subtracting that bias from theoriginal “Forested site”MAE.Sensor-shieldcombinationsareorderedfromsmallest(best)tolargestMAEattheopensite.

TABLE  3 Daytimemeanabsoluteerror(MAE)ofthecomparisonoftemperaturesrecordedbyeachsensor-shieldcombinationtotheweatherstationstandardatopenandforestedsites

8  |     TERANDO ET Al.

F IGURE  3 Hourlyaveragetemperaturesrecordedbytheweatherstationsensor(solidblacklines)andeachcombinationsensor/treatmentattheopenLakeWheelersite(a,c,e)andtheforestedDukeForestsite(b,d,f).(a,b)ShowresultsfortheiButtonshieldtreatments,(c,d)fortheHoboPendanttreatments,and(e,f)fortheHoboProtreatments.Measurementswererecordedfrom6–21and21–28August,2015attheLakeWheelerandDukeForestsites,respectively

0 5 10 15 20

2025

3035 Wx Station Sensor

UnshieldedRadshieldSmall RadshieldAlt Radiation ShieldCup (Fig. 1d)Cup (Fig. 1f)Cup (Fig. 1g)Cup (Fig. 1h)

Time of Day

(a)

0 5 10 15 20

2025

3035 Wx Station Sensor

UnshieldedRadshieldSmall RadshieldAlt Radiation ShieldCup (Fig. 1d)Cup (Fig. 1f)Cup (Fig. 1g)Cup (Fig. 1h)

Time of Day

(b)

0 5 10 15 20

2025

3035 Wx Station Sensor

UnshieldedRadshield

Time of Day

(d)

0 5 10 15 20

2025

3035 Wx Station Sensor

UnshieldedGill Shield

Time of Day

(f)

0 5 10 15 20

2025

3035 Wx Station Sensor

UnshieldedRadshield

Time of Day

(c)

0 5 10 15 20

2025

3035 Wx Station Sensor

UnshieldedGill Shield

Time of Day

(e)

     |  9TERANDO ET Al.

15 20 25 30 35 40 45

1520

2530

3540

45

15 20 25 30 35 40 45

1520

2530

3540

45

15 20 25 30 35 40 45

1520

2530

3540

45

15 20 25 30 35 40 45

1520

2530

3540

45

15 20 25 30 35 40 45

1520

2530

3540

45

15 20 25 30 35 40 45

1520

2530

3540

45

15 20 25 30 35 40 45

1520

2530

3540

45

(b) iButton Cups

(c) iButton Alt Radiation Shield (d) iButton Radshield

(e) HOBO Pendant Unshielded (f) HOBO Pendant Radshield

(g) HOBO Pro Unshielded (h) HOBO Pro Gill Shield

15 20 25 30 35 40 45

1520

2530

3540

45

Median25th/75th Quantiles10th/90th Quantiles

(a) iButton Unshielded

F IGURE  4 Conditionalquantileplotsforeachsensor/treatmentcombinationattheLakeWheelersite.Coloredlinesrepresentthe50th(red),25thand75th(green),and10thand90th(blue)quantilesoftherecordedsensortemperaturesrelativetotherecordedtemperatureoftheweatherstation

10  |     TERANDO ET Al.

effectontheshieldedHOBOProsensor(0.05°Cper%impervioussurface),whilethelargesteffectwasmorethandoublefortheiBut-tonwith the custom-fabricated cupdepicted in Figure1f (0.12°Cper%impervioussurface).

Figure5c shows all hourly differences between the estimatedregression coefficients for each sensor/shield combination and theshieldedHOBOProsensor(thebestperformingsensorbasedonthere-sultsfromtheLakeWheelerandDukeForestexperiments).Differencesare inunitsofdegreecelsiusper10%increase in impervioussurfacecover. Rapid increases in sensors differences, particularly among thecustom-fabricatedshieldsattachedtoinexpensivesensors,areseeninthemorninghourswithaninitialpeakaroundlocalsolarnoon;followedbya lateafternoonpeakdifferencethat is likely relatedtomaximumdaytimeheatingandupwardheatflux.Thesedifferencessuggestthatforsomecustom-fabricatedsensor/shieldcombinations,themeanaf-ternoonrecordedtemperaturesoversiteswith~50%impervioussur-facecovercouldbeover3°CwarmerthantheshieldedHOBOPro.

3.4 | Effects of solar radiation and wind speed on sensor performance

To better understand which environmental conditions are mostconducive to creating large biases in the sensor/shield combina-tions,wetestedfivelinearregressionmodelswithadditiveandmul-tiplicativecombinationsofsolarradiationand(inverse)windspeedusing theLakeWheelerair temperaturebias results (asdescribedin theMaterials andMethods section).Models 4 and 5, the twomodels that included an interaction termbetween solar radiationandinversewindspeed,hadthelowestAICscoresforalleightsen-sor/shieldcombinations(Table4).TheresultsfromModel5,whichincluded as a predictor the log-transform of inverse wind speed,are shown inFigures6 and7, and the results fromModel4withtheabsolutewindspeedareshowninFigsS5andS6.Theseplotsshowthepredictedbiasforagivenwindspeedunderthreelevelsof solar radiation (i.e., themean and±one standard deviation) at

F IGURE  5 Temperaturevs.impervioussurfacecoverat03:00hours(a)and15:00hours(b)withtheleastsquaresfitforeachsensor/shieldcombination.Asimpervioussurfacecoverincreases,biasedshieldsamplifytheeffectonrecordedtemperatures.(c)Showsthedifferencebetweentheestimatedslopes(changeintemperatureper10%increaseinimpervioussurfacecover)ofeachsensorandtheshieldedHOBOPro(thebestperformingfieldsensor)forallhoursoftheday.Orangedotsthereforeactasareferenceline,astheyrepresentthedifferencebetweentheshieldedHOBOProanditself(i.e.zero)

(a)

20

25

30

35

0 10 20 30 40 50% Impervious Surface Cover

0300

hr

Tem

p (°

C)

Sensor/Shield TypeHOBO Pendant RadshieldHOBO Pendant UnshieldedHOBO Pro Gill ShieldHOBO Pro UnshieldediButton Alt Radiation ShieldiButton Cup (Fig. 1d)iButton Cup (Fig. 1f)iButton Cup (Fig. 1g)iButton Cup (Fig. 1h)iButton RadshieldiButton Small RadshieldiButton Unshielded

(b)

20

25

30

35

0 10 20 30 40 50% Impervious Surface Cover

1500

hr

Tem

p (°

C)

(c)

0

5

10

15

20

0.0 0.2 0.4 0.6

(Sensor/Shield HOBO Pro Gill Shield)

Tim

e of

Day

     |  11TERANDO ET Al.

theLakeWheelersite.Becauseofthelog-transformoftheinversewindspeedpredictor,Model5ismoreconservativethanModel4(intermsofthepredictedbiasatlowwindspeeds),althoughgiventhe very similarAIC values, neither can be ruled out as implausi-ble,andothertechniquessuchasBayesianmodelaveragingcouldbe used to incorporate information from the full model set (e.g.,Terandoetal.,2016).

Overall,theresultsinFigures6and7indicatethatlargebiasescould result under conditions of lowwind speeds and high solarradiation formostcombinationsofcustom-fabricatedshieldswithinexpensive temperature sensors. Indeed, for some treatmentsthe predicted lowwind speed/high solar radiation biases exceed10°C (e.g., Figure6c). In contrast, the predicted shielded HOBO

Prosensorbiasesarelowthroughouttherangeofwindspeedsandsolarradiationvalues.Thenonlinearnatureoftheseresultsfortheinexpensiveshield/sensorcombinationssuggeststhatsimple,con-stantbias-correctionmethodsmaynotbeadequatetocontrol forthiserror.

4  | DISCUSSION

Ourresultsdefinitivelyshowthattheaccuracyofreportedtempera-tureobservationsinfieldecologyishighlysensitivetothechoiceofin-struments,materials,andmethodsusedtocollectthedata.Combinedwith our review of the literature, this reveals that the quality of

Model # Model descriptionFrequency of lowest AIC score (max = 8)

R2 average and (range)

1 Solarradiation+error 0 0.49(0.05,0.63)

2 Model1+inverse(windspeed) 0 0.51(0.09,0.64)

3 Model1+log(inverse[windspeed]) 0 0.52(0.09,0.66)

4 Model2+solarradia-tion×inverse(windspeed)

4 0.54(0.11,0.71)

5 Model3+solarradia-tion×log(inverse[windspeed])

4 0.54(0.09,0.71)

ResultsaresummarizedintermsofthefrequencythateachtestedmodelhadthelowestAICvalueforatreatment(n=8sensor/shieldcombinations),andthemeanandrangeoftheadjustedR2valueacrossthosetreatments.

TABLE  4 Resultsofregressionanalysisofeffectofsolarradiationandwindspeedonairtemperaturesensorbias

F IGURE  6 PredicteddaytimeiButtonsensorbiasresultingfromtheinteractionofsolarradiationandtheinverseandlog-transformedwindspeed.HeavyblacklinerepresentsthepredictedbiasforthemeandaytimesolarradiationexperiencedovertheexperimentperiodattheLakeWheelersite.Greylinesshowpredictedbiasatonestandarddeviationaboveandbelowthemeansolarradiation.Resultsaredisplayedwithwindspeedbacktransformedtotheoriginalunitsofms- 1

0

3

6

9

12

0 1 2 3 4

0

3

6

9

12

0 1 2 3 4

0

3

6

9

12

0 1 2 3 4

0

3

6

9

12

0 1 2 3 4

Wind Speed (ms–1) Wind Speed (ms–1)

Wind Speed (ms–1) Wind Speed (ms–1)

(b) iButton Cups

(d) iButton Radshield(c) iButton Alt Radiation Shield

(a) iButton Unshielded

75 Wm–2

353 Wm–2

631 Wm–2

12  |     TERANDO ET Al.

reportedtemperaturedatalikelyvarieswidelyacrossecologicalstud-ies.Inaddition,wefoundthattocollectaccurateairtemperaturedata,ecologistsmustalwaysusehigh-qualityradiationshields;withinoursubsetofshieldtypes,themanufacturedGillshieldwasmosteffec-tivewhilethetwovariationsontheRadshieldperformedbestamongthe custom-fabricated shields. Many custom-fabricated shields didnot prevent biases frombeing introduced by solar radiation and insomecases,resultedinlargerbiasesthantheoriginalunshieldedsen-sors.Therefore,suchadhocmethodsmustbetestedbeforeuseinthefield.HOBOProtemperaturesensorswerebyfarthemostaccuratesensorsacrosshabitats.Thesealsohappentobethemostexpensiveinstrumentsinourtestset(at~5–40timesthecostoftheinexpen-sivesensors),andtherefore,itisunrealistictoexpecttheiruseinallstudies.

Rather,wecautionecologiststounderstandthebiasesassociatedwiththeirsensorsandradiationshields,throughtheirownexperimentsorthroughtheliterature,andtakethesebiasesintoaccountwhende-signingexperimentsandanalyses.Forexample,ifmeasuringmaximumtemperaturesduringthedayisanobjectiveofagivenstudy,investinginhigherqualitysensorsandshieldsmaybenecessary,particularlyinexposedenvironments.WefoundthatforthesunnyLakeWheelersite,inexpensiveandimproperlyshieldedsensorshadlargepositivebiasesintherecordedmaximumdailytemperatures(Fig.S7).Onlythedataloggerswith theRadshield treatmentshadmeanbiases<2°C,whilemeanbiasesassociatedwithothershieldswere>5°C.Conversely,nosensors hadmeanminimum temperature biases >0.5°C, suggesting

thatstudiesthatfocusonmeasuringnighttimetemperaturesmaynotrequireinvestmentinthemostexpensivesensors.Regardless,aware-nessofbiasesintroducedbythechoiceofsensor-shieldcombinationwouldresultinmoreaccurateairtemperaturedata,andthusdatathataremorereadilycomparableacrossstudies.

Temperature readings were variable across all environmentstested.Compared toweather stations, sensors in theopensitehadhigher temperature recordings for a longer period of time than thesensors in the forested location.The strong and possibly nonlinearinteractionbetweenlowwindspeedsandhighsolarradiationattheopen site is likely to lead to significant biases that could extend toeventhebestperformingradiationshieldswhenappliedtolow-costsensors(e.g.,Figure6d).Underlowwindspeedandhighsolarradia-tionconditions, theaccuracyof thesedata loggers,especiallywhenimproperlyshielded,islikelyreducedduetoheatingofboththesen-sorand shieldhousing.Thesecombinedheatingeffects are seen inFig.S7,wheretherecordedmaximumtemperaturesofsomeshieldediButtonswere0.27–1.91°ChigherthantheunshieldediButton,whilethetwoRadshields loweredtheestimated iButtonbiasby3.06and3.28°C. Intheabsenceofwindspeedsthatcanefficientlytransportheatedairmoleculesawayfromthesensor,thenear-sensorairtem-peraturewillriseaboveambientconditions.

Solarheatingoftheinexpensivesensors/shieldsisreducedattheforestedsite,andsotherecordedtemperatureswillbemoresimilarto thesurroundingnear-surfaceenvironment (Lundquist&Huggett,2008). However,while trees did dampen some of the temperature

F IGURE  7 SameasFigure6butfortheHOBOPendantandHOBOProsensors

0

3

6

9

12

0 1 2 3 4

0

3

6

9

12

0 1 2 3 4

0

3

6

9

12

0 1 2 3 4

0

3

6

9

12

0 1 2 3 4

Wind Speed (ms–1) Wind Speed (ms–1)

Wind Speed (ms–1) Wind Speed (ms–1)

(c) HOBO Pro Unshielded (d) HOBO Pro Gill Shield

(a) HOBO Pendant Unshielded (b) HOBO Pendant Radshield

75 Wm–2

353 Wm–2

631 Wm–2

     |  13TERANDO ET Al.

extremes,ourdataindicatethatnaturalshade(althoughafrequentlyreportedshieldingmethodintheliterature)doesnotcompletelymit-igatesolarradiationeffects.Furthermore,ourimpervioussurfaceex-perimentresultsshowthattemperaturesensorsareaffectednotonlybydirect,incomingsolarradiation,butalsobyupward-directedradia-tionfluxesfromthesurface.Thisislikelytointroduceanothersourceofbiasinstudysiteslocatedinurbanareasandotherenvironmentswhereradiativeheatingofimproperlyshieldedsensorscanoccurduetoreflectivesurfacesorlowsunangles,suchasinhigh-latitudefieldsitesorareaswithhighalbedovalues(Huwaldetal.,2009).

Temperatures recorded by the inexpensive sensors usingRadshieldshadthelowestbiasesamongthecustom-fabricatedshieldsandwereclosesttothoseoftheshieldedHOBOProintheimpervi-oussurfacegradientexperiment.Thissuggeststhatthisdesign (andourslightmodificationof it) iscurrentlyoneof thebestperformingcustom-fabricatedshieldsintheliteratureandmayallowfortheuseofinexpensivetemperaturesensorsacrossarangeoffutureecologicalstudies.However,eventheRadshielddesignsresultedintemperaturebiasesthatwere>0.5°C,andothercustom-fabricatedshieldsmayper-formsimilarly.Forexample,radiationshieldsconstructedoutofPVCcapshavebeenusedwithiButtonsinpriorstudies,andthereportedbiaseswere1°Corless(Ashcroft&Gollan,2011),whilea“homemadeGillshield”testedbyTararaandHoheisel(2007)alsoperformedrea-sonably well. Finally, nighttime temperatures in both experimentswere least biased and so unshielded sensors mayworkwell whenthereislittleornodirectradiativeheating.

Ourfieldresultsillustratethatdesigningasmall,effective,custom-fabricated radiation shield is not straightforward, and our literaturereviewsuggeststhatthischallengemaybeunder-appreciatedintheecological literature. Field ecologists, as a community, do appear tounderstand that solar radiationaffects theaccuracyof air tempera-turemeasurements.Morethanone-thirdofthepapersweexaminedclearlyacknowledgedtheneedforshielding,butthedescriptionsoftheconstructeddevicesoftenprecludedanevaluationoftheiraccu-racy in the field.All sevenofour iButton shielddesigns,whichhadmanyelementsincommonwithothercustom-fabricatedshieldsmen-tionedintheliterature,yieldedbiaseddaytimetemperaturemeasure-ments.Biasedtemperaturedataneednotinvalidateobservedpatternsor conclusions in the studies that contain them; for example, thesemeasurementsmaystillcorrectlyarraysitesonanaxisfromcoolertowarmer(Figure5),buttheactualtemperaturevaluesrecordedarenotlikelytobecomparableamongstudies.Abouthalfofthepapersweexamineddidnotmentionshieldingatall,butwesuspectthatatleastsomeoftheseauthorsdidnotreportthetypeofradiationshieldsusedbecausetheytookthemforgranted,ratherthanbecausetheydidnotuse them. When manufacturer-recommended shields are availableandhabituallydeployed,theymayappeartobea“packagedeal”withthesensorsthemselves.Eventhisoptimistic interpretationpointstoawidespreadneed formore thorough reporting to improve repeat-abilityandensurethatbestpracticesareunderstoodbystudentsandreaders.

Together, lackofaccuracyandstandardization indatacollectionacrossecologicalstudieslimitstheutilityofthereportedtemperature

data. We fully acknowledge that our experiment was conductedunderarelativelynarrowrangeofenvironmentalandtemporalcondi-tions,andthattheexpectedtemperaturebiaseswillvarybylocationandclimate.Yet,morecarefulconsiderationofbiases>1°C is likelywarranted.Forexample,biasesof2°Corgreater (foundinsevenofourtwelvesensor/shieldtreatments)equatetotheprojectedannualmeantemperatureincreasesovermostofNorthAmericabythemid-dleofthe21stcenturyduetoanthropogenicclimatechange(Collinsetal.,2013).Astheclimatewarms,havingaccuratetemperatureob-servationswillbecriticalforunderstandingthecascadingeffectsofglobalclimatechangeatsmallermicroclimaticscales,andforevalu-atingtheattendantexposureimpactsonorganismsinhabitingtheseenvirons.

We therefore call for an increased awareness among field ecol-ogists of the biases theymay introducewhen using small, portabletemperaturesensorswithoutradiationshields,orwhenusingadhocmethods to construct radiation shields. The increased use of suchsensors over the past two decades demonstrates that theymeet ademandforcollectingclimatedataonaspatialscalerelevanttotheorganisms’ecologistsstudy.However,thereisaneedformorewide-spreadadoptionofbestpractices in theiruse.Weseeseveralwaysforward depending on research goals; each of these relies on clearreportingofmethodsandintentions:

• Invest in high-quality aspirated radiation shields, and high-quality (ex-pensive) sensors when needed—Whenactualtemperaturesareofin-terestforcomparisontootherdatasources,manufacturedshields,orthoroughlyvalidatedcustom-fabricatedshieldsthatallowfor(i)thefreeflowofairaroundthesensor,(ii)minimalsensorexposuretosolarradiation,and(iii)minimalradiationabsorptionbytheshield(Huwaldetal.,2009;Richardsonetal.,1999),shouldbeusedandtheir specifications clearly reported. In situationswherebiases inexcessof0.5°Cormoreareunacceptable,higherquality sensorsandshieldsmaybenecessary.Morefrequentuseofpublishedcal-ibrationmethodsbeforedeployment (suchas sensorwaterbathsasinToohey,Neal,andSolin(2014)andMauger,Shaftel,Trammell,Geist,andBogan(2015))couldalsoincreaseconfidenceinthere-sultsfromlow-costdataloggers.

• Provide clear disclaimers about data use and applicability—In somecases,itmaybehelpfultoexplicitlyacknowledgethatsensorspro-videarelativetemperaturemeasurewithinthecontextofthestudybutarenotmeantforuseincomparisontootherstudies.

• Consider local landscape effects on temperature measurement—Asillustrated inour impervioussurfacecoverexperiment,solarradi-ationeffectsonsomesensorscouldbeexacerbatedbyadditionalsurfaceheating.Assuch,employingthebestavailablematerialsandmethodsbecomesincreasinglyimportantinareasthatarelikelytoexperiencetheseconditions.

• Consider alternative sensor technologies—The iButton is a com-pact, inexpensive data logger with a silicon-based temperaturesensor,making itanattractiveoption inmanyfieldecologystud-ies. However, other established temperature sensors, such asthermocouples or thermistors (which are used in theHOBOPro

14  |     TERANDO ET Al.

data logger),arealsoavailableforusewithcompactdata loggers.Ecologistsshouldconsiderthesewhenevaluatingthetradeoffsbe-tweentime, labor,andcostofdeployinganytypeoftemperaturesensorinthefield.

Iftemperaturedatawerecollectedusingstandard,reliablemeth-ods,theycouldbeanalyzedacrossstudies,increasingecologists’abil-itytoinferclimatesensitivityandexposurerisks.Carefulapplicationofthesemethodswouldhelptofullyrealizetheopportunitypresentedbytheavailabilityofinexpensivesensors;potentiallytransformingthefieldofglobalchangebiologybyallowingunprecedentedcomparisonsofbioticresponsesacrossphylogenetic,spatial,andtemporalscales.As fieldecologyproceeds in thecontextof rapidanthropogeniccli-mate change, researchers increasinglyplace theirwork in a thermalcontext.This trendhasgreatpotential to improveunderstandingoforganismalandecosystemresponsestochangingtemperaturesworld-wide.Toadvancealongthispath,ecologistsmustensurethatpotentialbiasesintheirdataareminimizedorclearlyconveyed.Therefore,wearguethatstandardizingclimatedatacollectionmethodsinecologyisacriticalgoalinordertomakesignificantadvancesinunderstandingtheeffectsofglobalchange.

ACKNOWLEDGMENTS

WethankRyanBoyles,GlennCatts,SaraChilds,JamesClark,andSallyThigpenforfacilitatingaccesstostudysitesorweatherstationdata.JaimeCollazo,StevenFrank,andEricaHenryprovidedthedatalog-gersandradiationshields.AccesstostudysiteswasapprovedbytheDukeForestTeachingandResearchLaboratory,NorthCarolinaStateUniversity Research Forests, theCity of Raleigh Parks, Recreation,and Cultural Resources Department, and the North Carolina StateClimateOffice.Anyuseof trade, firm,orproductnames is forde-scriptivepurposesonlyanddoesnotimplyendorsementbytheU.S.Government.

CONFLICT OF INTEREST

Nonedeclared.

AUTHORS CONTRIBUTION

A.J.T.designedandperformedexperiments,analyzeddata,andwrotethepaper;E.Y.designedandperformedexperiments,conductedtheliteraturereview,andwrotethepaper;E.K.M.andS.G.P.designedandperformedexperimentsandwrotethepaper.

ORCID

Adam J. Terando http://orcid.org/0000-0002-9280-043X

Emily K. Meineke http://orcid.org/0000-0002-5416-4233

REFERENCES

Anderson, J. T., Inouye, D. W., McKinney, A. M., Colautti, R. I., &Mitchell-Olds,T.(2012).Phenotypicplasticityandadaptiveevolutioncontribute to advancing flowering phenology in response to climatechange.Proceedings of the Royal Society of London. Series B, Biological Sciences,279,3843–3852.

Ashcroft,M.B.,&Gollan,J.R.(2011).Fine-resolution(25m)topoclimaticgrids of near-surface (5 cm) extreme temperatures and humiditiesacrossvarioushabitats ina large(200×300km)anddiverseregion.International Journal of Climatology,32,2134–2148.

Ashcroft,M.B.,&Gollan,J.R.(2013).Thesensitivityoftopoclimaticmod-elstofine-scalemicroclimaticvariabilityandtherelevanceforecologi-calstudies.Theoretical and Applied Climatology,114,281–289.

Bakken, G. S. (1992). Measurement and application of operative andstandard operative temperatures in ecology. American Zoologist, 32,194–216.

Carper,A. L.,Adler, L. S.,Warren, P. S., & Irwin, R. E. (2014). Effects ofsuburbanizationonforestbeecommunities.Environmental Entomology,43,253–262.

Chen, I.-C.,Hill,J.K.,Ohlemüller,R.,Roy,D.B.,&Thomas,C.D. (2011).Rapid range shifts of species associatedwith high levels of climatewarming.Science,333,1024–1026.

Chen,J.,Saunders,S.C.,Crow,T.R.,Naiman,R.J.,Brosofske,K.D.,Mroz,G.D.,…Franklin, J. F. (1999).Microclimate in forest ecosystemandlandscapeecology.BioScience,49,288–297.

Cheung,H.K., Levermore,G.J.,&Watkins,R. (2010).A lowcost,easilyfabricated radiation shield for temperature measurements to moni-tordrybulbair temperature inbuiltupurbanareas.Building Services Engineering Research and Technology,4,371–380.

Collins, M., Knutti, R., Arblaster, J., Dufresne, J.-L., Fichefet, T.,Friedlingstein, P., …Wehner, M. (2013). Long-term climate change:Projections, commitments and irreversibility. InT. F. Stocker,D.Qin,G.-K.Plattner,M.Tignor,S.K.Allen,J.Boschung,A.Nauels,Y.Xia,V.Bex,&P.M.Midgley(Eds.),Climate change 2013: The physical science basis.ContributionofWorkinggroupItothefifthAssessmentReportoftheIntergovernmentalPanelonClimateChange(pp.1029–1136).Cambridge, United Kingdom and New York, NY, USA: CambridgeUniversityPress.

Diamond, S. E., Cayton, H., Wepprich, T., Jenkins, C. N., Dunn, R. R.,Haddad,N.M.,&Ries,L.(2014).Unexpectedphenologicalresponsesofbutterflies to the interactionofurbanizationandgeographic tem-perature.Ecology,95,2613–2621.

Forister, M. L., & Shapiro, A. M. (2003). Climatic trends and advancingspringflightofbutterfliesinlowlandCalifornia.Global Change Biology,9,1130–1135.

Fox,J.(2003).EffectdisplaysinRforgeneralisedlinearmodels.Journal of Statistical Software,8,1–27.

Fridley,J.D.(2009).Downscalingclimateovercomplexterrain:Highfines-cale(<1000m)spatialvariationofnear-groundtemperaturesinamon-tane forested landscape (GreatSmokyMountains). Journal of Applied Meteorology and Climatology,48,1033–1049.

Gill,G.C. (1979).Development of a small rugged radiation shield for air temperature measurements on drifting buoys.Report toNOAADataBuoyOfficeforDevelopmentContract#01-7-038-827(IF).BaySt.Louis,MS.

Hartmann, D. L., Klein Tank, A. M. G., Rusticucci, M., Alexander, L. V.,Brönnimann, S., Charabi, Y., … Zhai, P. M. (2013). Observations:Atmosphere and surface. In T. F. Stocker, D. Qin, G.-K. Plattner,M.Tignor,S.K.Allen,J.Boschung,A.Nauels,Y.Xia,V.Bex&P.M.Midgley (Eds.), Climate change 2013: The physical science basis. ContributionofWorkingGroupItotheFifthAssessmentreportofthe

     |  15TERANDO ET Al.

intergovernmentalpanelonclimatechange(p.1535).Cambridge,UKandNewYork,NY,USA:CambridgeUniversityPress.

Holden,Z.A.,Klene,A.E.,Keefe,R.F.,&Moisen,G.G.(2013).Designandevaluationofaninexpensiveradiationshieldformonitoringsurfaceairtemperatures.Agricultural and Forest Meteorology,180,281–286.

Hubbart, J. A. (2011). An inexpensive alternative solar radiation shieldfor ambient air temperature micro-sensors. Journal of Natural & Environmental Sciences,2,9–14.

Hubbart,J.,Link,T.,Campbell,C.,&Cobos,D.(2005).Evaluationofalow-costtemperaturemeasurementsystemforenvironmentalapplications.Hydrological Processes,19,1517–1523.

Huey,R.B.,Peterson,C.R.,Arnold,S.J.,&Porter,W.P.(1989).Hotrocksandnot-so-hot rocks:Retreat-site selectionbygarter snakes and itsthermalconsequences.Ecology,70,931–944.

Huwald, H., Higgins, C. W., Boldi, M.-O., Bou-Zeid, E., Lehning, M., &Parlange,M.B.(2009).Albedoeffectonradiativeerrorsinairtempera-turemeasurements.Water Resources Research,45,W08431.

Lundquist, J. D., & Huggett, B. (2008). Evergreen trees as inexpensiveradiation shields for temperature sensors.Water Resources Research,44,W00D04.

Marra,P.P.,Francis,C.M.,Mulvihill,R.S.,&Moore,F.R.(2005).Thein-fluence of climate on the timing and rate of spring bird migration.Oecologia,142,307–315.

Mauger,S.,Shaftel,R.,Trammell,E.J.,Geist,M.,&Bogan,D.(2015).StreamtemperaturedatacollectionstandardsforAlaska:Minimumstandardstogeneratedatausefulforregional-scaleanalyses.Journal of Hydrology,4,431–438.

Meineke,E.K.,Dunn,R.R.,Sexton,J.O.,&Frank,S.D.(2013).Urbanwarm-ingdrivesinsectpestabundanceonstreettrees.PLoS ONE,8,e59687.

Moritz,C.,&Agudo,R.(2013).Thefutureofspeciesunderclimate.Natural Systems in Changing Climates,904,504–508.

Parmesan,C.(2006).Ecologicalandevolutionaryresponsestorecentcli-matechange.The Annual Review of Ecology, Evolution, and Systematics,37,637–671.

Potter,K.A.,Woods,H.A.,&Pincebourde,S.(2013).Microclimaticchal-lengesinglobalchangebiology.Global Change Biology,19,2932–2939.

RCoreTeam(2016).R: A language and environment for statistical computing. Vienna,Austria:RFoundationforStatisticalComputing.URLhttps://www.R-project.org/

Richardson,S.J.,Brock,F.V.,Semmer,S.R.,&Jirak,C.(1999).Minimizinger-rorsassociatedwithmultiplateradiationshields.Journal of Atmospheric and Oceanic Technology,16,1862–1872.

Rundel, P.W.,Graham, E.A.,Allen,M. F., Fisher, J. C., &Harmon,T. C.(2009). Environmental sensor networks in ecological research. New Phytologist,182,589–607.

Tarara, J. M., & Hoheisel, G.-A. (2007). Low-cost shielding to minimizeradiation errors of temperature sensors in the field.HortScience,42,1372–1379.

Terando,A.J.,Reich,B.,Pacifici,K.,Costanza,J.,McKerrow,A.,&Collazo,J.A.(2016).Uncertaintyquantificationandpropagationforprojectionsofextremesinmonthlyareaburnedunderclimatechange.InK.Riley,P.Webley&M.Thomspon(Eds.),Natural hazard uncertainty assessment: Modeling and decision support (pp.245–256).Hoboken,NJ:AmericanGeophysicalUnion.

Toohey,R.C.,Neal,E.G.,&Solin,G.L.(2014).Guidelines for the collection of continuous stream water-temperature data in Alaska.U.S.GeologicalSurveyOpen-FileReport,2014–1182,34.

Urban,M.C.,Bocedi,G.,Hendry,A.P.,Mihoub,J.-B.,Pe’er,G.,Singer,A.,…Travis,J.M.J.(2016).Improvingtheforecastforbiodiversityunderclimatechange.Science,353,1113

Walther,G.R.,Post,E.,Convey,P.,Menzel,A.,Parmesan,C.,Beebee,T.J.C.,…Bairlein,F.(2002).Ecologicalresponsestorecentclimatechange.Nature,416,389–395.

SUPPORTING INFORMATION

Additional Supporting Information may be found online in thesupportinginformationtabforthisarticle.

How to cite this article:TerandoAJ,YoungsteadtE,MeinekeEK,PradoSG.Adhocinstrumentationmethodsinecologicalstudiesproducehighlybiasedtemperaturemeasurements.Ecol Evol. 2017;00:1–15. https://doi.org/10.1002/ece3.3499