1

EcoMapper:GenerativeModelingforClimate-AwareSatelliteImagery

MuhammedGoktepe*1AmirHosseinShamseddin*1ErencanUysal1JavierMuineloMonteagudo1

LukasDrees2AysimToker3SentholdAsseng4MaltevonBloh4

Abstract

SatelliteimageryisessentialforEarthobserva-tion,enablingapplicationslikecropyieldpre-diction,environmentalmonitoring,andclimatechangeassessment.However,integratingsatelliteimagerywithclimatedataremainsachallenge,limitingitsutilityforforecastingandscenarioanalysis.Weintroduceanoveldatasetof2.9mil-lionSentinel-2imagesspanning15landcovertypeswithcorrespondingclimaterecords,form-ingthefoundationfortwosatelliteimagegenera-tionapproachesusingfine-tunedStableDiffusion

3models.Thefirstisatext-to-imagegenerationmodelthatusestextualpromptswithclimateandlandcoverdetailstoproducerealisticsyntheticim-ageryforspecificregions.ThesecondleveragesControlNetformulti-conditionalimagegenera-tion,preservingspatialstructureswhilemappingclimatedataorgeneratingtime-seriestosimulatelandscapeevolution.Bycombiningsyntheticim-agegenerationwithclimateandlandcoverdata,ourworkadvancesgenerativemodelinginremotesensing,offeringrealisticinputsforenvironmen-talforecastingandnewpossibilitiesforclimateadaptationandgeospatialanalysis.

1.Introduction

Satelliteimageryisacriticalresourceforavarietyofre-searchandindustrialapplications,rangingfromenviron-

mentalmonitoringandclimateresearch(

deAra.

joetal,

*Equalcontribution1TechnicalUniversityofMunich,SchoolofComputation,InformationandTechnology,Germany2UniversityofZurich,DepartmentofMathematicalModelingandMachineLearning,EcoVisionLab,Switzerland3TechnicalUniversityofMunich,SchoolofComputation,InformationandTechnology,Dy-namicVisionandLearningGroup,Germany4TechnicalUniversityofMunich,SchoolofLifeSciences,DepartmentofLifeScienceEngineering,HEFWorldAgriculturalSystemsCenter,ChairofDigitalAgriculture,Germany.Correspondenceto:MaltevonBloh<malte.von.bloh@tum.de>.

Proceedingsofthe42ndInternationalConferenceonMachineLearning,Vancouver,Canada.PMLR267,2025.Copyright2025bytheauthor(s).

2025

;

Hussainetal.

,

2022

)toagriculture(

Khanaletal.

,

2020

),archaeology(

Mar.

n-Buznetal,

2021

),resourceexploration(

Shirmardetal.

,

2022

),andhumandevelop-mentmapping(

Burkeetal.

,

2021b

).Itprovidesvaluablespatialandtemporaldata,oftenservingastheprimaryorsupplementarysourceforpredictionandanalysismodels,includingcropyieldforecasting(

Khakietal.

,

2021

;

Mu-

rugananthametal.

,

2022

;

vonBlohetal.

,

2023

),forestrysurveillance(

Fassnachtetal.

,

2024

),anddisastermanage-ment(

Burkeetal.

,

2021a

).Inmanyregionsoftheworld,satellitedataistheonlyfeasiblemeansofacquiringnearreal-timeinformationaboutenvironmentalconditions.

However,theuseofremotesensingishamperedbysig-nificantchallenges,includingcloudcoverage,atmosphericdistortions,andtemporalresolutionconstraints(

Dubovik

etal.

,

2021

).Cloudcoverage,inparticular,renderssatelliteimageryunusableforlargepartsoftheworld,disruptingsatelliteobservationsincloud-proneareasfordaysorevenweeks,dependingontheseason(

Kingetal.

,

2013

;

Mercury

etal.

,

2012

).Theseoperationalchallengesnotonlyhinderreal-timemonitoringbutalsoraiseacriticalconceptualgap:theneedtointegratesatelliteimagerywithfutureclimatescenariostoenhancepredictionaccuracy.Whilevariousdatasetssupportmachinelearningapplicationstoaddressthesechallenges,mostaretask-specificorregionallycon-strained,limitingtheirgeneralizability(

Clasenetal.

,

2024

;

Christieetal.

,

2018

;

Schneideretal.

,

2023

;

VanEttenetal.

,

2018

).Toaddressthis,weintroduceacomprehensivere-motesensingdataset—oneofthelargesttodate—featuringover2.9millionSentinel-2RGBimageslinkedwithcli-matedata,enablingmorerobustandscalableapplicationsacrossdiverseenvironmentalconditions.Recentadvancesinmultimodalfoundationmodelsforremotesensinghavesignificantlyimprovedgenerationofsyntheticsatelliteim-ageryacrossdomains(

Hongetal.

,

2024

;

Khannaetal.

,

2024

;

Liuetal.

,

2024a

;

Yuetal.

,

2024b

).Butakeygapremainsingenerativemodelscapableofproducinglocation-specificsatelliteimageryconditionedonfutureclimaticconditions.Thislimitationhinderspredictiveapplicationssuchasseasonalcropyieldforecastingandtheassessmentofclimatechangeimpactsonlandcover(

Iizumietal.

,

2018

;

Zachowetal.

,

2024

).Thesemodelsrelyondetailedweatherandclimateprojectionstoimprovetheiraccuracy,butoften

EcoMapper:GenerativeModelingforClimate-AwareSatelliteImagery

2

Figure1.Ourframeworkenablesmulti-conditional(text+image)satelliteimagegenerationusingStableDiffusion3(

Esseretal.

,

2024

)andControlNet(

Zhangetal.

,

2023

).Onthetext-inputside,themodeltakesdetailedspatialandclimaticpromptembeddingscreatedbyCLIPandT5textencoders.Ontheimage-inputside,ControlNetisfine-tunedtoprocessanimageforspatialguidance.Bothispassedtopre-trainedStable-Diffusion3,whichgeneratesalocation-andclimatespecificsatelliteimage,alignedwithactualspatialcharacteristics.

dependonhistoricalorapproximatedsatellitedataintheabsenceofintegrationwithclimatescenarios.Thisresultsinaqualitativegapbetweenpredictedenvironmentalcondi-tionsandavailableimagery,complicatingtherepresentationofdynamicchangesinlandcoverorvegetationstates,re-ducingthepredictivepowerofthesemodels(

Ebeletal.

,

2020

;

Jozdanietal.

,

2022

).Thesechallengesunderscoretheneedforsyntheticsatelliteimagerytoenhancedatasetsandproviderealisticprojectionsforfutureconditions.Inthispaper,weintroduceanovelapproachtogeneratesatelliteimagesconditionedongeographic-climatepromptsusingStableDiffusion3.OurmethodenablesthesimulationofhowweatherandclimateaffectEarth’ssurface-generatingsyntheticimagesthatcansupportforecastingmodels(e.g.,cropyieldpredictionorlandcoverclassification),visualizeclimatechangemodelsundervariousscenarioassumptions,andfillobservationalgapsinregionsaffectedbypersistentcloudcover.Theapproachisgloballyapplicableandgener-atesrealisticimageswith10-meterspatialresolutionacrossdiversevegetationtypes(e.g.,cropland,broadleafforests,savannas),usinginformationaboutlocation,landcovertype,andclimateconditions.Weproposetwoinnovations:

1.Atext-to-imagegenerationmodelthatleveragesStableDiffusion3withclimatepromptengineering.

2.Amulti-conditional(text+image)frameworkutilizingControlNet,whichpreservesspatialfeaturesanden-ablesthegenerationoftimeseries.

Fig.

1

illustratestheconcept.Tosupportthisresearch,we

curatedadatasetofover2.9millionRGBsatelliteimagesfrom104,424locationsworldwide,sourcedfromSentinel-2(

Druschetal.

,

2012

).ThisdatasetspansthewholeEarthcat-egorizedin15vegetationzonesandeightyearsofhistoricaldata.Together,thesecontributionsadvancetheapplicationofgenerativemodelsinremotesensingandoffernovelsolu-tionsforavarietyofenvironmentalmonitoringchallenges.

2.RelatedWork

2.1.DiffusionModels

Diffusionmodelsareapowerfulclassofgenerativemodels,achievingstate-of-the-artperformanceinhigh-qualityimagesynthesisacrossdiversedomains(

Hoetal.

,

2020

;

Khanna

etal.

,

2024

;

Zhangetal.

,

2023

).TheyoperatebyiterativelydenoisingrandomGaussiannoisethroughalearnedreverse-diffusionprocess,producingrealisticsamples.Thesemod-elsexcelintaskslikeimagesuper-resolution,inpainting,anddomainadaptation(

Manvietal.

,

2024

;

Tokeretal.

,

2024

),andhavebeeneffectivelyappliedtosatelliteim-agerysynthesisbyleveragingcontrolimagesandtextualpromptstomaintainspatialstructureandstylisticfidelity(

Sastryetal.

,

2024

).Foundationalworkby

Sohl-Dickstein

etal.

(

2015

)introduceddiffusionprobabilisticmodels,laterrefinedby

Hoetal.

(

2020

)withasimplifieddenoisingob-jective.Score-basedgenerativemodeling(

Songetal.

,

2021

)andclassifier-freeguidance(

Dhariwal&Nichol

,

2021

)haveenhancedtheiradaptabilityandoutputquality.

EcoMapper:GenerativeModelingforClimate-AwareSatelliteImagery

3

2.2.SatelliteImageGeneration

Inremotesensing,imagegenerationhasalsobeenenhancedbydiffusionmodels,withmodelslikeStableDiffusion(SD)fine-tunedtogeneratesatelliteimagesfromtextualdescrip-tions(

Liuetal.

,

2024b

).

Khannaetal.

(

2024

)enrichedintheir“DiffusionSat”theSDmodel’sinputbyincorporatinggeo-locationandsamplingtimeasprompts,enablingthegenerationofhigh-qualitysatelliteimagestailoredtospe-cificgeographicandtemporalconditions.Theirexperimentsfocusedongeneratingdiverselandcovertypes,includingur-banareas,croplands,andforests,demonstratingthemodel’sabilitytocapturefine-graineddetailssuchasbuildingstruc-turesandvegetationpatterns.

Diffusionmodelshavealsobeenwidelyadoptedforimage-to-imagegeneration,producingrealisticsatelliteimagesfromguidinginputssuchasmaps,semanticlayouts,andmulti-modaldata.

Espinosa&Crowley

(

2023

)successfullygeneratedsatelliteimagesconditionedonhistoricalmaps,focusingonurbanandrurallandscapestosimulatehistori-calandfuturelandusechanges.

Tangetal.

(

2024

)refinedthegenerationprocessbyintegratingbothglobal(e.g.,tex-tualdescriptions)andlocal(e.g.,depthmaps,segmentationmasks)controlinformation,expandingthescopeofsatelliteimagesynthesis.

Satelliteimagedatasetsrangefromsmalltask-specificcol-lectionstolargegeneral-purposesets,withunlabeledremotesensingdataoftenexceedingonemillionimages,whilela-beleddatasetsaretypicallysmallerinsize.Reben(

Clasen

etal.

,

2024

)andEarthView(

Velazquezetal.

,

2025

)supportgeneral-purposeandself-supervisedlearning,withEarth-Viewcomprisingover15terapixelsofmulti-sourceimagery.fMoW(

Christieetal.

,

2018

)andEuroCrops(

Schneider

etal.

,

2023

)providelabeleddata,targetingfunctionallanduseandharmonizedcroptypes.Multimodalandcloud-robustdatasetssuchasSEN12MS-CR-TS(

Ebeletal.

,

2022

)andDiffCR(

Zouetal.

,

2024

)offerpairedradar-opticaltimeseriesforcloudremoval.Incontrast,MetaEarth(

Yuetal.

,

2024b

)presentsagenerativefoundationmodeltrainedonmulti-resolutionimageryforlarge-scaleimagesynthesis.Incontrast,climate-integrateddatasetshavesofarbeensmallerandhighlyspecializedforresearchapplications(

Nathetal.

,

2024

;

Requena-Mesaetal.

,

2021

).

2.3.IntegratingClimateData

Whenprocessingdatasetswithhigh-dimensionalclimatevariableslikecyclonedistribution,cloudcover,andwatervapor,diffusionmodelshaveproveneffective(

Liuetal.

,

2024b

).

Hatanakaetal.

(

2023

)usedcascadeddiffusionmod-elstogeneratehigh-resolutioncloudcoverimages,while

Nathetal.

(

2024

)employedmulti-stagediffusionframe-worksforprecipitationandcycloneforecasting.

Gaoetal.

(

2023

)and

Leinonenetal.

(

2023

)unifiedprecipitationnow-

castingwithinsinglediffusionmodels,achievingstate-of-the-artresultsbycapturingcomplexspatiotemporalrelation-ships.DiffCastfrom

Yuetal.

(

2024a

)outperformspreviousworksintheCriticalSuccessIndex(CSI)-whichmeasuresthefractionofcorrectlypredictedprecipitationeventsrela-tivetoallobservedorpredictedevents-by15.59%.

3.Preliminaries

3.1.Diffusion

DiffusionmodelsaimtoproduceimagesbyreversingastochasticGaussiannoisingprocess.Givenaninputimage,anoisyinputxtiscreatedbyaddingGaussiannoise:

xt=√1—βt·xt-1+√βt·?,?~N(0,I).(1)

Theparametersβtdenotethenoisevarianceateachtimestept,thehigherthetthemoretheaddednoise.Theaimofthediffusionmodelistodenoisearandomsampleusinganeuralnetworktopredictitsnoisevectors,effectivelylearningthemappingxt}→xt-1.ThisresultsinastochasticgenerativemodelMthatmapsfromapredefinednoisedistributionN(0,I)togeneratedimages.

M(z)=x0,z~N(0,I).(2)

Latentdiffusionmodelsareatypeofgenerativemodelthatappliesthediffusionprocesswithinalower-dimensionalla-tentspace,whichisobtainedbyencodingtheinputimages.Thisreducescomputationalcostcomparedtotraditionaldiffusionmodels.Adecoderthenreconstructstheimagefromthelatentspace.Conditionalarchitectures,includingStableDiffusion2,guidethedenoisingprocessbyincor-poratinganadditionalsignalC(·),suchastextorimages.Thisenablesthegenerationofmeaningfuloutputsthatalignwiththecontrolsignalandthespecifictaskforwhichthemodelwastrained.

M(z,C(z))=x0,z~N(0,I).(3)

3.2.ControlNet

ControlNet(

Zhangetal.

,

2023

)enhancesdiffusionmodelsbyintegratingexplicitspatialcontrolintothegenerativeprocess.LetF(x)denoteaneuralnetworkblockfromtheoriginalarchitecture,whichhasfrozenparametersinternally(omittedhereforbrevity).Thenewcontrolblockmodifiesitsoutputbyadding

y=F(x)+Z(F(x+Z(c;θ1);θcopy);θ2),(4)whereθcopyisatrainablecopyoftheoriginalparameters,andθ1andθ2aretheparametersofzeromodules(e.g.

convolutionlayersinitializedtozero).Thisdesignensuresthatthecontrolblockhasminimalinitialeffectonthemainblock,functioningsimilarlytoaskipconnection.

EcoMapper:GenerativeModelingforClimate-AwareSatelliteImagery

4

4.MaterialandMethods

4.1.TheEcoMapperDataset

WepresenttheEcoMapperdatasetthatcombinesSentinel-2RGBimageswithcorrespondingmetadata,assembledusingGoogleEarthEngine,SentinelHubandNASAPower(

Dr-

uschetal.

,

2012

;

Milcinskietal.

,

2019

;

NASA

,

2025

;

Phan

etal.

,

2020

).

4.1.1.SATELLITEIMAGERY

Thedatasetincludes104,424uniquegeographiclocations,randomlysampledfrom15distinctlandcoverclasses(

Phan

etal.

,

2020

)excludingwaterbodiesasshowninFigure

2

.Foreachlocation,weselectedonemonthlyobservationforatwoyearperiodbasedontheleastcloudyday,resultinginasequenceof24imagesperlocation.Thetwoyearsofob-servationarerandomlydistributedbetween2017and2022.Thetestsetconsistsof5,500uniquegeographiclocations,eachmonitoredmonthlyovera96-monthperiodfrom2017to2024.Thisensuressufficientspatialandtemporalinde-pendenceintheevaluation,enablingrobustassessmentofthemodel’sgeneralizationacrossdiverseregionsandunseenclimateconditions.Withaspatialcoverageof~26.21km2perobservationtheoveralldatasetcovers~2,704,000km2,accountingfor~2.05%ofEarth’sterrestrialarea.Anex-cerptofthedatasetispublishedintheGithubrepository,thefulldatasetisavailableattheuniversitiesservers.Formoredatasetdetailswereferto

A.1

.

4.1.2.CLIMATEDATA

Eachsampledlocationisenrichedwithmetadata,includinggeographiclocation(longitudeandlatitude),observationdate(monthandyear),landcovertype,andcloudcoverage(in%).Weincorporatedaveragemonthlytemperature,solarradiation,andtotalprecipitationfromNASAPower(

NASA

,

2025

),asthesefactorsmainlydrivevegetationgrowth,en-ergyavailabilityandwaterbalance,whichinturninfluenceagriculturalconditions,forestry,biodiversity,andlandcover(

PielkeSretal.

,

2011

).

4.2.GenerativeModels

Ourgoalistosynthesizesatelliteimageryconditionedongeographicandclimatemetadata,enablingrealisticpro-jectionsofenvironmentalconditions.Toachievethis,weleveragestate-of-the-artgenerativemodelsfortwokeytasks:text-to-imagegenerationandmulti-conditionalimagegen-eration.

Fortext-to-imagegeneration,weemploygenerationmodelsthatsynthesizesatelliteimagesbasedonstructuredmetadataprompts.Additionally,weintroduceamulti-conditionalgenerationapproachusingaControlNet-enhancedmodel,

Figure2.The104,424locationsweresampledgloballyacross15landcovertypes,providingarepresentativedistributionofEarth’slandsurface.Grasslandsandsparselyvegetatedregionsdomi-nate,followedbyforestedareasandcroplands,withadditionalcategoriessummarizedinto”O(jiān)thers”(mixed-,evergreenneedle-leafforest,permanentwetlands,cropland/naturalmosaics,urban,closedshrublandanddecidiuousneedleleafforest).Eachlocationincludesatimeseriesof24months(training)or96months(test-ing).

whichpreservesthespatialstructureofaninputimagewhilemappingclimate-inducedvariationsontoit.Byleveragingourdataset,wedemonstratehowenvironmentalchangescanbevisuallyrepresentedbymodifyingclimatemetadatainthegenerationprocess.

Weevaluatetwogenerativemodelsfortheirabilitytointe-grateclimatemetadataintosatelliteimagesynthesis:

StableDiffusion3(SD3)from

Esseretal.

(

2024

)-Amul-timodallatentdiffusionmodelincorporatingCLIPandT5textencodersforflexiblepromptconditioning.Wefine-tuneSD3usingourdatasettogeneraterealisticsatelliteimageryconditionedongeographic,climatic,andtemporalmetadata.Akeychallengeistherepresentationofcontinu-ousclimatevariables,whichweaddressthroughstructuredpromptengineering.DiffusionSatfrom

Khannaetal.

(

2024

)

-aspecializedfoundationmodelforsatelliteimagery,ex-tendingStableDiffusion2withadedicatedmetadataem-beddinglayerfornumericalconditioning.Thisarchitectureencodeskeyspatialandtemporalattributes,includinglati-

EcoMapper:GenerativeModelingforClimate-AwareSatelliteImagery

5

tude,longitude,timestamp,groundsamplingdistance,andcloudcover.Unlikegenericdiffusionmodels,DiffusionSatisexplicitlydesignedforremotesensingtasks,includingsuper-resolution,inpainting,andtemporalprediction.

4.2.1.TEXT-TO-IMAGEGENERATION

WecomparemultipleconfigurationsofStableDiffusion3andDiffusionSat,withandwithoutfine-tuning,toassesstheircapacityforclimate-awaresatelliteimagesynthesis.ThebaseSD3model,leveragingaT5encoder,allowsfullclimatepromptconditioning.ThebaseDiffusionSatmodel,limitedtoonetextencoderwith77tokens,wasonlycondi-tionedonmonth,year,cloudiness,andlandcovertypeduetoitspredefinedmetadataembeddingstructure.ToenableDiffusionSatfine-tuningwithadditionalclimatemetadata,wemodifieditsSatUNetarchitecturebyintroducing10metadatalayers.WeinitializedthisadaptednetworkwithpretrainedDiffusionSatweightswhilerandomlyinitializingthenewlayers,followedbyfull-modelfine-tuning.Forafaircomparison,bothmodelsweretrainedat512×512res-olution,whichalignswithDiffusionSat’soriginaltrainingsetup.Additionally,SD3,whichsupportshigherresolu-tions,wastestedinafine-tunedexperimentat1024×1024resolution.Insummaryweevaluate:

1.Baselinemodels:Bothmodelswereevaluatedwithoutfine-tuningat512×512resolution.

2.Fine-tunedmodels:Bothmodelswerefine-tuned(-FT)withclimatemetadataat512×512resolution.

3.High-resolutionSD3:SD3-FT-HRwasfine-tunedwithclimatemetadataat1024×1024resolution.

4.2.2.CLIMATE-AWARESENSITIVITYANALYSIS

ToassessthesensitivityoftheSD3-FT-HRmodeltocli-matevariablesandensurethatperformancegainsstemfrommeaningfulclimateconditioningratherthanspuriouscor-relationswithmonth,landcover,orlocationweperformatargetedsensitivityanalysis.Weevaluatethemodel’sabilitytoincorporateclimateeffectsintosatelliteimagegenerationbytestingitunderextremeconditions,rangingfromdrybo-real(cold,dry)tohumidtropical(hot,wet)climates.Thesevariationsspanmultiplelandcovertypesandregions,en-ablingustodeterminewhetherthemodelcapturesgenuineclimateinfluencesormerelyexploitsdatasetcorrelations.

4.2.3.MULTI-CONDITIONALIMAGEGENERATION

Forthetaskofmulti-conditional(text+image)imagegen-eration,weutilizeafine-tunedStableDiffusion3modelenhancedwithLoRA(Low-RankAdaptation).Thismodel,trainedata512×512resolution,servesasafoundationalpriorforgeneratinghigh-qualityandcontextuallyrelevantoutputs.Toconditiontheimagegenerationprocessonboth

spatialstructureandclimatedynamics,weincorporateadual-conditioningmechanismusingControlNet.Control-NetextendsStableDiffusionbyintroducingtrainableneurallayersthatguidethedenoisingprocessusinganexternalcontrolsignal.Inourapproach,twocriticalconditioningsignalsareused:Satelliteimageryfrompreviousmonthsasacontrolsignalthatpreservesthespatialstructureofthegeneratedimage,ensuringthatlandforms,urbanlayouts,andothergeographicalfeaturesremainintact.Thisalsoen-ablesthemodeltoincorporatetemporalchangesovertime,reflectingreal-worldenvironmentalshifts.ClimatePrompts:Atextualconditioningmechanismthatspecifiesclimaticandatmosphericconditionsunderwhichthesatelliteimageshouldbegenerated.Bycombiningthesetwoconditioningfactors,themodeliscapableofgeneratingrealisticsatelliteimagesthatintegrateclimatevariationswhilemaintainingspatialconsistency.Thismethodsupportstime-seriesgener-ation,allowingthesimulationoflandscapeevolutionunderchangingclimateconditions.

4.3.PromptStructure

Wedesigntwotypesofpromptstoeffectivelyconditionsatelliteimagegeneration:aspatialprompt,whichencodesmetadata,andaclimatepromptextendingitwithenviron-mentaldetails.BothpromptsleveragethetextencodersofStableDiffusion3,withspatialinformationprocessedbyCLIPandclimatedatahandledbytheT5encoder.

1.SpatialPrompt:Capturesfundamentalmetadata,in-cludinglandcovertype,location,date,andcloudiness.Thisensuresthatthegeneratedimagesalignwiththegeographicandtemporalcontext.

2.ClimatePrompt:Extendsthespatialpromptbyincor-poratingmonthlyclimatevariablestemperature,pre-cipitation,andsolarradiationprovidingadditionalen-vironmentalconditioningforimagegeneration.

Thestructuredpromptfollowstheformat:“Asatelliteimageof[landcovertype]in[location]on[date].Theaveragetemperaturewas[temperature],with[precipitation]and[so-larradiation].”Thisformatensuresthegeneratedimagesremaincontextuallyandenvironmentallyaccuratebyinte-gratingbothspatialandclimaticfactors.Forexample:“AsatelliteimageofcroplandsinNorthernCape,SouthAfrica,onOctober2019.Theaveragetemperatureoverthelastmonthwas20°C,withanaverageprecipitationof0mmandanaveragedailysolarradiationof25W/m2.”Thisstructuredpromptingframeworkenableseffectivecondi-tioningacrossspatialandenvironmentaldimensions.Ad-ditionaldetailsonthepromptingstrategyandcomparativeresultsfordifferentpromptformulationsareprovidedinAppendix

A.3.5

.

EcoMapper:GenerativeModelingforClimate-AwareSatelliteImagery

6

4.3.1.METRICS

Toevaluatethequality,diversity,andperceptualfidelityofthegeneratedsatelliteimages,weusefiveestablishedmet-rics:FID,LPIPS,SSIM,PSNR,andCLIPScore.FIDandLPIPSassessrealismandperceptualsimilarity,whileSSIMandPSNRmeasurestructuralconsistencyandreconstruc-tionquality.CLIPScoreevaluatestext-imagealignment.Adetaileddescriptionisprovidedin

A.2

.

5.Results

5.1.Text-to-ImageGeneration

Wetestourmodelson5,500worldwidelocationsacrossalllandcovertypes,usingeightyearsofsatellitedatawithmonthlyobservations.AsshowninTab.

1

,thebaselinemod-elsfromSD3andDiffusionSathadthelowestevaluationscores.However,DiffusionSatdemonstratedsuperiorper-formanceoverSD3,showingadvantagesfromthesatellite-specificpretraining.Fine-tuningsignificantlyimprovedbothmodelsacrossallmetrics:SD3-FTachievedhigherCLIP,SSIM,andPSNRscores,whileDiffusionSat-FTexcelledinFIDandLPIPS.Thebest-performingmodelintermsofFIDwasSD3-FT-HR,whichproducedthehighest-resolutionimages.

ThequalitativeresultsinFig.

3

demonstratethecapabil-ityofourmodelsingeneratingrealisticsatelliteimagesconditionedongeographicandlandtypemetadata.Inpar-ticular,themodelsexcelatcapturingthestructuredpatternsofcroplands/grasslands,suchasthoseseeninKazakhstanandParaguay,whereregularfieldpatternsarefaithfullyreproduced.Allthreemodelseffectivelypreservetheessen-tialstructureoftheselandscapes.InmountainousregionslikeCanada,themodelssuccessfullycapturethedistinctfeaturesofsnow-coveredterrainsandrockysurfaces.Whilebotharchitectureshandlesnowcoveragewell,SD3-FT-HRexcelsatpreservingthesharpcontrastsbetweensnowandrockformations,providingfinerdetailcomparedtoDiffu-sionSatinthiscontext.IngrasslandregionsofParaguay,themodelsrepresenttheexpansive,flatterrainwithsparsevegetationeffectively,capturingthehomogeneousstruc-turetypicalofgrasslands.Allmodelsmanagetorepresentthebroadnessoftheseregions,thoughSD3modelsshowaslightimprovementincapturingthesubtlevariationsinvegetationdensity.ForwetlandareasDiffusionSatcapturesthewaterpresencewithhighfidelity,whileSD3alsoeffec-tivelyrepresentsthewetland’sdynamicstructure,withbothmodelsexcellinginmaintainingth