maxent原版英文说明.docx

maxent原版英文说明.docx

《maxent原版英文说明.docx》由会员分享，可在线阅读，更多相关《maxent原版英文说明.docx（15页珍藏版）》请在冰豆网上搜索。

maxent原版英文说明

maxent原版英文说明

RESM575SpatialAnalysis

Spring2010

Lab6MaximumEntropy

Assigned:

MondayMarch1

Due:

MondayMarch8

20points

ThislabexercisewasprimarilywrittenbyStevenPhillips,MiroDudikandRobSchapire,withsupportfromAT&TLabs-Research,PrincetonUniversity,andtheCenterforBiodiversityandConservation,AmericanMuseumofNaturalHistory.Thislabexerciseisbasedontheirpaperanddata:

StevenJ.Phillips,RobertP.Anderson,RobertE.Schapire.

Maximumentropymodelingofspeciesgeographicdistributions.

EcologicalModelling,Vol190/3-4pp231-259,2006.

MygoalistogiveyouabasicintroductiontouseoftheMaxEntprogramformaximumentropymodelingofspecies’geographicdistributions.

TheenvironmentaldataconsistofclimaticandelevationaldataforSouthAmerica,togetherwithapotentialvegetationlayer.ThesamplespeciestheauthorsusedwillbeBradypusvariegatus,thebrown-throatedthree-toedsloth.

NOTEontheMaxentsoftware

Thesoftwareconsistsofajarfile,maxent.jar,whichcanbeusedonanycomputerrunningJavaversion1.4orlater.Itcanbedownloaded,alongwithassociatedliterature,fromwww.cs.princeton.edu/~schapire/maxent.IfyouareusingMicrosoftWindows（asweassumehere）,youshouldalsodownloadthefilemaxent.bat,andsaveitinthesamedirectoryasmaxent.jar.Thewebsitehasafilecalled“readme.txt”,whichcontainsinstructionsforinstallingtheprogramonyourcomputer.

Thesoftwarehasalreadybeendownloadedandinstalledonthemachinesin317Percival.

ðFirstgototheclasswebsiteanddownloadthemaxent-tutorial-data.zipfile.Extractittothec:

/tempfolderwhichwillcreateac:

/temp/tutorial-datadirectory.

ðFindthemaxentdirectoryonthec:

/driveofyourcomputerandsimplyclickonthefilemaxent.bat.Thefollowingscreenwillappear:

Toperformarun,youneedtosupplyafilecontainingpresencelocalities（“samples”）,adirectorycontainingenvironmentalvariables,andanoutputdirectory.Inourcase,thepresencelocalitiesareinthefile“c:

\temp\tutorial-data\samples\bradypus.csv”,theenvironmentallayersareinthedirectory“layers”,andtheoutputsaregoingtogointhedirectory“outputs”.Youcanentertheselocationsbyhand,orbrowseforthem.Whilebrowsingfortheenvironmentalvariables,rememberthatyouarelookingforthedirectorythatcontainsthem–youdon’tneedtobrowsedowntothefilesinthedirectory.AfterenteringorbrowsingforthefilesforBradypus,theprogramlookslikethis:

Thefile“samples\bradypus.csv”containsthepresencelocalitiesin.csvformat.Thefirstfewlinesareasfollows:

species,longitude,latitude

bradypus_variegatus,-65.4,-10.3833

bradypus_variegatus,-65.3833,-10.3833

bradypus_variegatus,-65.1333,-16.8

bradypus_variegatus,-63.6667,-17.45

bradypus_variegatus,-63.85,-17.4

Therecanbemultiplespeciesinthesamesamplesfile,inwhichcasemorespecieswouldappearinthepanel,alongwithBradypus.Othercoordinatesystemscanbeused,otherthanlatitudeandlongitude,aslongasthesamplesfileandenvironmentallayersusethesamecoordinatesystem.The“x”coordinateshouldcomebeforethe“y”coordinateinthesamplesfile.

Thedirectory“layers”containsanumberofasciirastergrids（inESRI’s.ascformat）,eachofwhichdescribesanenvironmentalvariable.Thegridsmustallhavethesamegeographicboundsandcellsize.MAKESUREYOURASCIIFILESHAVETHE.ascEXTENSION!

!

!

!

!

!

!

!

!

Oneofourvariables,“ecoreg”,isacategoricalvariabledescribingpotentialvegetationclasses.Youmusttelltheprogramwhichvariablesarecategorical,ashasbeendoneinthepictureabove.

Doingarun

Simplypressthe“Run”button.Aprogressmonitordescribesthestepsbeingtaken.Aftertheenvironmentallayersareloadedandsomeinitializationisdone,progresstowardstrainingofthemaxentmodelisshownlikethis:

The“gain”startsat0andincreasestowardsanasymptoteduringtherun.Maxentisamaximum-likelihoodmethod,andwhatitisgeneratingisaprobabilitydistributionoverpixelsinthegrid.Notethatitisn’tcalculating“probabilityofoccurrence”–itsprobabilitiesaretypicallyverysmallvalues,astheymustsumto1overthewholegrid.Thegainisameasureofthelikelihoodofthesamples;forexample,ifthegainis2,itmeansthattheaveragesamplelikelihoodisexp

（2）≈7.4timeshigherthanthatofarandombackgroundpixel.Theuniformdistributionhasgain0,soyoucaninterpretthegainasrepresentinghowmuchbetterthedistributionfitsthesamplepointsthantheuniformdistributiondoes.Thegainiscloselyrelatedto“deviance”,asusedinstatistics.

Therunproducesanumberofoutputfiles,ofwhichthemostimportantisanhtmlfilecalled“bradypus.html”.Partofthisfilegivespointerstotheotheroutputs,likethis:

Lookingataprediction

Toseewhatother（moreinteresting）contenttherecanbeinc:

\temp\tutorial-data\outpus\bradpus_variegatus.html,wewillturnonacoupleofoptionsandrerunthemodel.Pressthe“Makepicturesofpredictions”button,thenclickon“Settings”,andtype“25”inthe“Randomtestpercentage”entry.Lastly,pressthe“Run”buttonagain.Youmayhavetosay“ReplaceAll”forthisnewrun.Aftertheruncompletes,thefilebradypus.htmlcontainsthispicture:

Theimageusescolorstoshowpredictionstrength,withredindicatingstrongpredictionofsuitableconditionsforthespecies,yellowindicatingweakpredictionofsuitableconditions,andblueindicatingveryunsuitableconditions.ForBradypus,weseestrongpredictionthroughmostoflowlandCentralAmerica,wetlowlandareasofnorthwesternSouthAmerica,theAmazonbasin,Caribeanislands,andmuchoftheAtlanticforestsinsouth-easternBrazil.Thefilepointedtoisanimagefile（.png）thatyoucanjustclickon（inWindows）oropeninmostimageprocessingsoftware.

Thetestpointsarearandomsampletakenfromthespeciespresencelocalities.Testdatacanalternativelybeprovidedinaseparatefile,bytypingthenameofa“Testsamplefile”intheSettingspanel.Thetestsamplefilecanhavetestlocalitiesformultiplespecies.

Statisticalanalysis

The“25”weenteredfor“randomtestpercentage”toldtheprogramtorandomlysetaside25%ofthesamplerecordsfortesting.Thisallowstheprogramtodosomesimplestatisticalanalysis.Itplots（testingandtraining）omissionagainstthreshold,andpredictedareaagainstthreshold,aswellasthereceiveroperatingcurveshowbelow.TheareaundertheROCcurve（AUC）isshownhere,andiftestdataareavailable,thestandarderroroftheAUConthetestdataisgivenlateroninthewebpage.

Asecondkindofstatisticalanalysisthatisautomaticallydoneiftestdataareavailableisatestofthestatisticalsignificanceoftheprediction,usingabinomialtestofomission.ForBradypus,thisgives:

Whichvariablesmatter?

Togetasenseofwhichvariablesaremostimportantinthemodel,wecanrunajackknifetest,byselectingthe“Dojackknifetomeasurevariableimportant”checkbox.Whenwepressthe“Run”buttonagain,anumberofmodelsgetcreated.Eachvariableisexcludedinturn,andamodelcreatedwiththeremainingvariables.Thenamodeliscreatedusingeachvariableinisolation.Inaddition,amodeliscreatedusingallvariables,asbefore.Theresultsofthejackknifeappearinthe“bradypus.html”filesinthreebarcharts,andthefirstoftheseisshownbelow.

WeseethatifMaxentusesonlypre6190_l1（averageJanuaryrainfall）itachievesalmostnogain,sothatvariableisnot（byitself）agoodpredictorofthedistributionofBradypus.Ontheotherhand,Octoberrainfall（pre6190_l10）isamuchbetterpredictor.Turningtothelighterbluebars,itappearsthatnovariablehasalotofusefulinformationthatisnotalreadycontainedintheothers,asomittingeachoneinturndidnotdecreasethetraininggainmuch.

Thebradypus_variegatus.htmlfilehastwomorejackknifeplots,usingtestgainandAUCinplaceoftraininggain.Thisallowstheimportanceofeachvariabletobemeasurebothintermsofthemodelfitontrainingdata,anditspredictiveabilityontestdata.

Howdoesthepredictiondependonthevariables?

Nowpressthe“Createresponsecurves”,deselectthejackknifeoption,andrerunthemodel.Thisresultsinthefollowingsectionbeingaddedtothe“bradypus_variegatus.html”file:

Eachofthethumbnailimagescanbeclickedontogetamoredetailedplot.Lookingatfrs6190_ann,weseethattheresponseishighestforfrs6190_ann=0,andisfairlyhighforvaluesoffrs6190_annbelowabout75.Beyondthatpoint,theresponsedropsoffsharply,reaching-50atthetopofthevariable’srange.

Sowhatdothevaluesonthey-axismean?

Themaxentmodelisanexponentialmodel,whichmeansthattheprobabilityassignedtoapixelisproportionaltotheexponentialofsomeadditivecombinationofthevariables.Theresponsecurveaboveshowsthecontributionoffrs6190_anntotheexponent.Adifferenceof50intheexponentishuge,sotheplotforfrs6190_annshowsaverystrongdropinpredictedsuitabilityforlargevaluesofthevariable.

Onatechnicalnote,ifwearemodelinginteractionsbetweenvariables（byusingproductfeatures）asweareforBradypushere,thentheresponsecurveforonevariablewilldependonthesettingsofothervariables.Inthiscase,theresponsecurvesgeneratedbytheprogramhaveallothervariablessettotheirmeanonthesetofpresencelocalities.

Notealsothatiftheenvironmentalvariablesarecorrelated,astheyarehere,theresponsecurvescanbemisleading.Iftwocloselycorrelatedvariableshavestrongresponsecurvesthatarenearopposites