ESA UNCLASSIFIED For Official Use - PDF document

ESA UNCLASSIFIED For Official UseESA UN
ESA UNCLASSIFIED For Official UseESA UNCLASSIFIED For Official UseThe Sen2Cor and MAJA cloud masks and classification productsMagdalena MainKnorn, Jerome Louis, Olivier Hagolle, Uwe MüllerWilm, Kevin AlonsoOutlineSen2CorMAJAComparison and Validation procedure Results and ExamplesConclusionsSEN2COR ProcessorAtmosphericcorrectionprocessordeveloped by Telespazio behalf of ESAtemporal processor: corrects singledate Sentinel2 L1C TopAtmosphere (TOA) products from the effects of

the atmosphere and delivers a L2A Botto
the atmosphere and delivers a L2A BottomAtmosphere (BOA) reflectanceproductAdditional outputs: Aerosol Optical Thickness (AOT) map, Water Vapour(WV) map and Scene Classification (SCL) map with Quality Indicators for cloud and snow probabilities, in JPEG 2000 image formatL2A data production on the systematicbasisoverEurope withdisseminationthrough the Copernicus Open Access HubsinceMay 2017 Processor can be freely downloaded (new Version 2.5) and used for processing L1C imagesM

AJA ProcessorAtmosphericcorrectionproces
AJA ProcessorAtmosphericcorrectionprocessordeveloped by CNES, DLR and CESBIOMAJA uses a combination of multispectral and multitemporal (Sentinel2 time series) criteria to detect invalid pixels, and to estimate AOT before correcting foratmospheric effects Multitemporal algorithm based on the assumption that surface reflectance tends to change slower in time than cloud coverCorrects Sentinel2 L1C TOA products from the effects of the atmosphere and delivers a L2A BOA reflectancepro

ductsAdditional outputs: Aerosol Optical
ductsAdditional outputs: Aerosol Optical Thickness (AOT), Water Vapour(WV), and set of masks for clouds, cloud shadows, topographic shadows, snow and waterL2A data production (7M kmCNES, �60 000 L2A productsfreelyavailable Processor (V1.0) isavailableexecutablecodeforlinuxRedHatCentOShttpslogiciels.cnes.fr/en/content/majaSen2Cor and MAJA Masking and ClassificationMAJASEN2CORAlgorithmMultitemporal andmultispectralMonotemporal andmultispectralClassesclouds (+cirrus), clo

ud shadows, topographic shadows, water a
ud shadows, topographic shadows, water and snowsaturateddefectivedarkareacloudshadowsvegetation, nonvegetatedwaterunclassifiedcloudmedium probabilitycloudhigh probabilitythincirrussnowMethodsthresholds, reflectance temporal variation; tests of the local temporal correlation; band ration and indicesthresholds on L1C spectral bands;Band ratios and indicesOutputs8 bit set of binary masksScene Classification(SCL) mapQuality Indicators for cloud and snow probabilitiesL2A spatialresol

ution240m (time*2 resolution20m, 60mDila
ution240m (time*2 resolution20m, 60mDilationTotal L2A Processing time~20min~30minValidationand Comparison procedureAllocation key to consolidate Sen2Cor and MAJA classes for comparison:Comparison of MAJA V2 and Sen2Cor 2.5 Validation testsites (MAJAbased)cloud cover: 373%good representativeness SiteDateFloresta(Brazil)21LWK2016/06/01 Davos(Switzerland)32TNS2016/09/09 Davos(Switzerland)32TNS2016/05/22 Hyytiala(Finland)35VLJ2016/05/23 Ispra(Italy)32TNR2016/04/22 Mong

u(Zambia)34LGH2016/10/03 Mongu(Zambia)
u(Zambia)34LGH2016/10/03 Mongu(Zambia)34LGH2016/10/23 Ouarzazate(Morocco) 29RPQ2016/04/17 Railroadall(USA)11SNC2016/05/09 Railroadall(USA)11SNC2016/06/28 SedeBoker(Israel)36RXV2016/05/06 Validation and Comparison procedureValidation stepsStratified random samplingPixel/area labelling (visually) based on:S2 bands2, 8A6 & 12118Aspectralprofiles, QIs, L1C cirrusband, Google dataarchCreation of the validation image (reference image)Creation of the corresponding subset

of the classification imageReportedconfu
of the classification imageReportedconfusionmatrixuser‘sproducer‘saccuraciescommissionandomissionerrorsoverallaccuracy Validation procedure limitationStratifiedrandomsamplingisbasedon MAJA masksitcanbiasvalidationresultswhereSen2Cor classificationfailedReference samplessetisinbalancedsemiproportional the classareaotal numberof sampelsoverlandis~416 000 (~23% of total referencesampelsbut forsomeproductsnot sufficientenoughDefinition of thincirrusandvisualinterpretatio

nissubjectiveInternal testthe subjectivi
nissubjectiveInternal testthe subjectivity of the validation method (4 validatingpersons, 2 productsrevealedquitestableresultsst.devclassificationaccuracy~56%). More xtended tests are required.In preparation classification validation protocol for L2A to assure validation quality and comparability between algorithmsL1C (8AMAJASEN2CORPreliminary Validation ResultsAll classescomparison: Cloud detectionworksverygoodforbothprocessorsCloud shadowsclassificationshowslowperformanc

eTopographicshadowsareverychallengingcla
eTopographicshadowsareverychallengingclassifyDetectionof clearlandandwaterpixels for11 productsseemslittlebithigherSen2Cor but… itcouldbiasedthe numberof validatedsamplesSnow iswellrecognizedbut overestimatedin someplacesMAJA V2SEN2COR 2.5 commissionomissioncommission omissionClouds 2,9 6,6 5,0 3,7 Cloud shadows18,9 44,0 2,9 60,0 Water 18,9 29,2 13,6 11,5 Snow 23,2 1,4 20,9 7,2 Topographicshadows86,2 90,4 97,8 57,4 L

and 34,3 21,1 16,2 9,8 OA com
and 34,3 21,1 16,2 9,8 OA complete90,2 ( 92,7 ( Pixel validatedaverage166685 Total validatedpixels1833536 Preliminary Validation ResultsValid andinvalid pixels comparison:Issuesrelatedstratificationbasedon MAJA maskingInvalid pixels detectioniscorrectforbothprocessorsTwopoorresults of MAJA related to decision on thin cloud thresholdOverall accuracyforbothprocessorsishigh(�94andmayimprovewiththe increasingnumberof validatedproductsandfurtherproces

sorevolutionOverall accuracyTestsiteMAJ
sorevolutionOverall accuracyTestsiteMAJA SEN2CORCommentAlta Floresta97,8 96,8 Low numberlandreferencesHyytiala95,9 98,7 Davos92,4 94,7 Davos281,5 90,6 ThincirrusconfusionOuarzazate99,6 93,9 StratificationbiasMongu85,2 96,2 ThincirrusconfusionMongu299,3 95,4 Ispra93,0 95,9 SedeBoker96,3 97,3 Stratification biasRailroad Valley95,8 92,5 StratificationbiasRailroadValley 295,9 92,2 Low numberlandreferencesValidation ExamplesMon

guZambia) 03.10.2016; L1C_T34LGH_A006696
guZambia) 03.10.2016; L1C_T34LGH_A006696_20161003T084305L1C (8AMAJASEN2CORCirrusB10Definition of thincirrusandvisualinterpretationissubjective& challengingOmissionisforsomeapplicationsmorecriticalthancommissionerrorAccordingthresholdused, MAJA showscommissionandSen2Cor omissionerrorsin thincloudsdetectionL1C (3CirrusB10Validation ExamplesQuarzazateMorocco) 17.04.2016; L1C_T29RPQ_A004281_20160417T111159L1C (8AMAJASEN2CORCirrusB10cloudhereSen2Cor misclassifiesbrightsur

facecompactcloudHere, MAJA provideshighe
facecompactcloudHere, MAJA provideshigheraccuracythanSen2CorValidation ExamplesQuarzazateMorocco) 17.04.2016; L1C_T29RPQ_A004281_20160417T111159L1C (8AMAJASEN2CORCirrusB10WhichspatialresolutionisneededforusersownapplicationsMAJA snow mask is only indicative (other snow product by CNES available)Validation ExamplesSedeBoker(Israel) 06.05.2016; L1C_T36RXV_A004551_20160506T083117L1C (3MAJASEN2CORLarge amount of false clouds detected by Sen2Cor in bright and urban areasL1

C (3MAJASEN2COROmission of clouds over
C (3MAJASEN2COROmission of clouds over water by MAJAConclusionsMAJA andSen2Cor providehigh qualitycloudmasksbothproductsarefreelyavailableandusersareinvitedtestandcomparethoseowntestsites!ThinirrusclouddefinitionandlabelingischallengingBothprocessorsarelessaccuratein cloudshadowsandtopographicshadowsclassificationAdvantage of MAJA multitemporal algorithmparticularlyforarid andsemiarid locationsZupancet al.: MAJA outperforms singlescene algorithms in terms of land misclassific

ationas clouds; Very bright surfaces are
ationas clouds; Very bright surfaces are problematic for Sen2CorDisadventageof MAJA concerninglowspatialresolutionof maskingpixel+ dilation=480mSen2Cor couldperformstill betterifdilationof cloudmaskwouldappliedplannedin the followingversionsBuilding a representative reference data set is a huge work … validation protocol for L2A to assure validation qualityneeded. ESA UNCLASSIFIED For Official UseMany thanks for your attentionReferencesHagolle, O. et al. (in prepa

rationValidation of Sentinel2 cloud mask
rationValidation of Sentinel2 cloud masks obtained from MAJA software and comparison with Sen2cor cloud masks”Louis, J., et al. (2017), Evolutions of Cloud Screening and Scene Classification in Level2A atmospheric correction processor Sen2Cor”, RAQRS V Symposium, 18.22.Sep. 2017, ValenciaZupanc, A. et al. (2017), „Improving Cloud Detection with MachineLearning” (available: https://medium.com/sentinelhub/improvingclouddetectionwithmachinelearningc09dc5d7cf13)