Landcover and crop type classification with intra-annual times series of sentinel-2 and machine learning at central Portugal

i I zá Alejand a He nández Sequei a LANDCOVER AND CROP TYPE CLASSIFICATION wi h in a-annual imes se ies o sen inel-2 and machine lea ning a Cen al Po ugal ii LANDCOVER AND CROP TYPE CLASSIFICATION wi h in a-annual imes se ies o sen inel-2 and machine lea ning a Cen al Po ugal Disse a ion supe ised by: Ph.D. Má io Síl io Rochinha de And ade Cae ano Associa e P o esso , No a In o ma ion Managemen School (NOVA IMS) Uni e si y o No a, Lisbon, Po ugal Ph.D. Filibe o Pla Bañón P o esso , Ins i u e o New Imaging Technologies (INIT) Uni e si y o Jaume I (UJI) Cas ellón, Spain Ph.D. Hugo Alexand e Gomes da Cos a In i ed Lec u e , No a In o ma ion Managemen School (NOVA IMS) Uni e si y o No a, Lisbon, Po ugal Feb ua y 24, 2020 iii DECLARATION OF ORIGINALITY I decla e ha he wo k desc ibed in his documen is my own and no om someone else. All he assis ance I ha e ecei ed om o he people is duly acknowledged and all he sou ces (published o no published) a e e e enced. This wo k has no been p e iously e alua ed o submi ed o NOVA In o ma ion Managemen School o elsewhe e. Lisboa, Feb ua y 24, 2020 I zá Alejand a He nández Sequei a [ he signed o iginal has been a chi ed by he NOVA IMS se ices] i ACKNOWLEDGMENTS Fi s , I would like o hank my supe iso Má io Cae ano o allowing Geo ech mas e s uden s o de elop esea ch in he Di eção-Ge al do Te i ó io (DGT); o isualizing he p ojec and encou aging me o ake his challenge. Addi ionally, I would like o hank my co-supe iso Hugo Cos a and (my un-o icial co-supe iso ) Ped o Bene ides om he Di isão de In o mação Geog á ica (DIG) o hei help ul guidance and all he cons an eedback p o ided o his wo k; and o my co-supe iso Filibe o Plá o his en husiasm on he hesis opic and cons uc i e sugges ions p o ided. Special hanks o William Ma ínez, because his applied coding knowledge o geospa ial in o ma ion was essen ial o de elop many o he sc ip s. I lea ned a lo om hei insigh s in he emo e sensing and land co e mapping ield; wo king alongside hem was a g ea p o essional expe ience. Also, I exp ess my g a i ude owa ds he E asmus Mundus P og am o inancing s uden s om all o e he wo ld o pu sue hei academic de elopmen wi h he Mas e in Geospa ial Technologies. The mul i-cul u al en i onmen we a e su ounded by makes us mo e ole an and open owa ds o he s. Many hanks o Ma co Painho, Ch is oph B ox, and Joaquín Hue a o coo dina ing his p og am h ough he h ee uni e si ies. Special hanks o Joel Sil a om NOVA IMS o he codes he p o ided du ing he emo e sensing cou se as hey we e a undamen al inpu o his wo k. Addi ionally, I am g a e ul o my GeoTech amily, which is an en husias ic and alen ed g oup o people. I was encou aged by he quali y o hei wo k in e e y p ojec hey deli e ed. Especially du ing he Geomundus Con e ence ha we o ganized, ha is an excellen memo y o ou s anding eamwo k. Special hanks o he wonde ul people ha suppo ed me un il he e y end o he mas e Ca los, F ancisco, Mihail, and Vicen e; i was a pleasu e sha ing many Lisbon memo ies wi h you. Finally, I canno hank enough my amily in Nica agua; hey a e he eal s eng h behind e e y p ojec I achie e. This hesis was de eloped unde he amewo k o he p ojec oRESTER (PCIF/SSI/0102/2017) unded by Fundação pa a a Ciência e Tecnologia LANDCOVER AND CROP TYPE CLASSIFICATION wi h in a-annual imes se ies o sen inel-2 and machine lea ning a Cen al Po ugal ABSTRACT Land co e and c op ype mapping ha e bene i ed om a daily e isi ing pe iod o senso s such as MODIS, SPOT-VGT, NOAA-AVHRR ha con ains long ime-se ies a chi e. Howe e , hey ha e low accu acy in an A ea o In e es (ROI) due o hei coa se spa ial esolu ion (i.e., pixel size > 250m). The Cope nicus Sen inel-2 mission om he Eu opean Spa ial Agency (ESA) p o ides ee da a access o Sen inel 2-A(S2a) and B (S2b). This sa elli e cons ella ion gua an ees a high empo al (5-day e isi cycle) and high spa ial esolu ion (10m), allowing equen upda es on land co e p oduc s h ough supe ised classi ica ion. Ne e heless, his equi es aining samples ha a e adi ionally collec ed manually ia ieldwo k o image in e p e a ion. This hesis aims o implemen an au oma ic wo k low o classi y land co e and c op ypes a 10m esolu ion in cen al Po ugal using exis ing da abases, in a-annual ime se ies o S2a and S2b, and Random Fo es , a supe ised machine lea ning algo i hm. The ag icul u al classes such as empo a y and pe manen c ops as well as ag icul u al g asslands we e ex ac ed om he Po uguese Land Pa cel Iden i ica ion Sys em (LPIS) o he Ins i u o de Financiamen o da Ag icul u a e Pescas (IFAP); land co e classes like u ban, o es and wa e we e ained om he Ca a de Ocupação do Solo (COS) ha is he na ional Land Use and Land Co e (LULC) map o Po ugal; and las ly, he bu ned a eas a e iden i ied om he co esponding na ional map o he Ins i u o da Conse ação da Na u eza e das Flo es as (ICNF). Also, a se o p ep ocessing s eps we e de ined based on he implemen a ion o ancilla y da a allowing o a oid he inclusion o mislabeled pixels o he classi ie . Mislabeling o pixels can occu due o e o s in digi aliza ion, gene aliza ion, and di e ences in he Minimum Mapping Uni (MMU) be ween da ase s. An inne bu e was applied o all da ase s o educe bo de o e lap among classes; he mask om he ICNF was applied o emo e bu ned a eas, and NDVI ule based on Landsa 8 allowed o e ase ecen clea -cu s in he o es . Also, he Cope nicus High-Resolu ion Laye s (HRL) da ase s om 2015 (la es a ailable), namely Dominan Lea Type (DLT) and T ee Co e Densi y (TCD) a e used o dis inguish i be ween o es wi h mo e han 60% co e age (coni e ous and b oadlea ) such as Holm Oak and S one Pine and be ween 10 and 60% (coni e ous) o ins ance Open Ma i ime Pine. Nex , empo al gap- illed mon hly composi es we e c ea ed o he ag icul u al pe iod in Po ugal, anging om Oc obe 2017 ill Sep embe 2018. The composi es p o ided da a ee o missing alues in opposi ion o single da e acquisi ion images. Finally, a pixel-based app oach classi ica ion was ca ied ou in he “Tejo and Sado” egion o Po ugal using Random Fo es (RF). The esul ing map achie es a 76% o e all accu acy o 31 classes (17 land co e and 14 c op ypes). The RF algo i hm cap u ed he mos ele an ea u es o he classi ica ion om he cloud- ee composi es, mainly du ing he sp ing and summe and in he bands on he Red Edge, NIR and SWIR. O e all, he classi ica ion was mo e success ul on he i iga ed empo a y c ops whe eas he g asslands p esen ed he mos complexi y o classi y as hey we e con used wi h o he ain ed c ops and bu ned a eas. ii KEYWORDS Sen inel-2 Land Co e Mapping C op ype Mapping Random Fo es Au oma ic Sample Ex ac ion In a-annual ime se ies Po ugal iii ACRONYMS CAP – Common Ag icul u al Policy COS – Ca a de Uso e Ocupação do Solo (Po uguese o LCLU) DGT – Di eção-Ge al do Te i ó io - Gene al Di ec o a e o Te i o ial Managemen FCT - Fundação pa a a Ciência e Tecnologia - Founda ion o Science and Technology. oRESTER – Rede de senso es combinada com modelação da p opagação do ogo in eg ado num sis ema de apoio à decisão pa a o comba e a incêndios lo es ais - Da a usion o senso ne wo ks and i e sp ead modelling o decision suppo in o es i e supp ession. EU – Eu opean Union HRL – High-Resolu ion Laye s ICNF - Ins i u o da Conse ação da Na u eza e das Flo es as - Ins i u e o Na u e Conse a ion and Fo es s IFAP – Ins i u o de Financiamen o da Ag icul u a e Pescas - Ag icul u e and Fishe ies Financing Ins i u e INSPIRE– In as uc u e o spa ial in o ma ion in Eu ope IPSTERS - IPSen inel Te es ial Enhanced Recogni ion Sys em LCLU – Land Co e Land Use LPIS - Land Pa cel Iden i ica ion Sys em NCPA – Na ional Con ol and Paying Agency OTS – On-The-Spo SCAPE FIRE - A sus ainable landSCAPE planning model o u al FIREs p e en ion ix INDEX OF THE TEXT ACKNOWLEDGMENTS ......................................................................................................... i ABSTRACT .................................................................................................................................... KEYWORDS ............................................................................................................................... ii ACRONYMS ............................................................................................................................... iii INDEX OF TABLES ................................................................................................................... x INDEX OF FIGURES ............................................................................................................... xi 1 INTRODUCTION .............................................................................................................. 1 1.1 Backg ound .................................................................................................................... 1 1.2 P oblem S a emen and Mo i a ion ............................................................................ 1 1.3 Resea ch Ques ion ......................................................................................................... 4 1.4 Aim .................................................................................................................................. 4 1.5 Objec i es ....................................................................................................................... 5 1.6 Thesis s uc u e .............................................................................................................. 5 2 LITERATURE REVIEW ................................................................................................... 6 2.1 Land co e mapping ..................................................................................................... 6 2.2 C op ype mapping ........................................................................................................ 7 2.3 Sen inel-2 ime se ies ..................................................................................................... 8 2.3.1 Time se ies image y ............................................................................................... 8 2.3.2 In a-annual and In e -annual ............................................................................. 8 2.3.3 Gap- illed image composi es ............................................................................... 8 2.4 Machine Lea ning – S a is ical Lea ning .................................................................... 9 2.4.1 Machine Lea ning P ocess ................................................................................... 9 2.4.2 Supe ised and Unsupe ised Lea ning ........................................................... 11 2.4.3 Reg ession, Classi ica ion, and Clus e ing ....................................................... 11 2.4.4 Random Fo es (RF) ........................................................................................... 11 3 METHODOLOGY ........................................................................................................... 14 3.1 S udy a ea ...................................................................................................................... 14 3.2 Da a ............................................................................................................................... 15 3.2.1 Ancilla y da a ....................................................................................................... 16 3.2.2 Remo e Sensing Da a ......................................................................................... 21 3.3 Me hods ........................................................................................................................ 23 3.3.1 So wa e and de ice speci ica ions ................................................................... 25 3.3.2 P ep ocessing o he e e ence da ase s ........................................................... 25 3.3.3 P ep ocessing o he in a-annual ime se ies o Sen inel 2 .......................... 31 3.3.4 Supe ised lea ning ............................................................................................. 33 4 1.3 Resea ch Ques ion When pe o ming supe ised image classi ica ion, se e al algo i hms can be applied, and di e en da a sou ces can be u ilized o aining he classi ie . Depending on he numbe o a ge classes, he o e all accu acy o he map luc ua es. The mo e classes a e added o he classi ica ion, he highe he p obabili y o misclassi ica ions and, he e o e, he educ ion in he abili y o he classi ie o map he classes accu a ely. To c ea e an au oma ic map, he main ask is o gene a e a s able wo k low o classi ying sa elli e image y in a ep oducible way. De eloping he p esen esea ch a DGT and he a ailabili y o ime se ies o Sen inel-2 and he up o da e ancilla y da ase s (COS 2018, IFAP pa cels 2018 and ICNF bu ned a eas 2018), h ee esea ch ques ions a e p oposed: 1. How accu a e is i o classi y 31 classes o land co e and c op ypes a 10m esolu ion? 2. Which a e he mos impo an ea u es/ a iables o conside when using in a- annual ime se ies? 3. When pe o ming au oma ic sample ex ac ion, can a se o p e-p ocessing ules allow us o ex ac spec al signa u es o he classes sui able o image classi ica ion? 1.4 Aim This in es iga ion p ojec aims o gene a e an au oma ic land co e and c op ype map in as e o ma using in si u and up- o-da e da a, sa elli e image y, and machine lea ning algo i hms. The au oma iza ion me hod elies on he e ie al o he spec al signa u es o he land co e and c op ypes om in a-annual ime-se ies image y o Sen inel-2 a he pixel le el aking ad an age o he a ailabili y om COS 2018 land co e da ase , and IFAP 2018 moni o ed ag icul u al pa cels as well as ICNF 2018 bu ned a eas. These a eas will se e as aining and es ing inpu o he Random Fo es classi ie , allowing o implemen he supe ised machine lea ning me hod o land co e and c op ype classi ica ion. 5 1.5 Objec i es • Con ibu e o an au oma ic supe ised classi ica ion wo k low o p oduce land- co e and c op ype maps. • Re iew he exis ing s a e-o - he-a in machine lea ning and mul i- empo al op ical image y o classi ying land co e and c op ype a eas. • Classi y land co e and ag icul u al a eas using he Random Fo es algo i hm and he ea u es ex ac ed om he aining da ase s and Sen inel-2 ime se ies. 1.6 Thesis s uc u e • The li e a u e e iew p esen s he co e concep s o he de elopmen o he esea ch ocusing on he s a e-o - he-a in land co e and c op ype mapping, use o sen inel-2 in a-annual ime se ies, and andom o es classi ie . • The me hodology comp ehensi ely desc ibes he s udy a ea ha is he ocus o his esea ch, he p ep ocessing o he p ima y da ase s, he p ac ical s eps o he sample selec ion and spec al signa u e ex ac ion, he selec ion o he bes pa ame e s o he andom o es model and he challenges o aining a machine- lea ning algo i hm o classi y la ge s udy a eas and gene a e a inal map. • Resul s and discussion desc ibe he esul s o he classi ica ion and con ex ualize he goodness-o - i based on accu acy me ics. C i ically analyze he esul s and ela e hem o li e a u e. • Conclusions, limi a ions, and ecommenda ions: summa y o he esea ch, p esen he main indings and con ibu ion as well as he limi a ions and ecommenda ions o u u e s eps in he au oma ic annual classi ica ion o land co e and c op ype mapping. 6 2 LITERATURE REVIEW This chap e ocus on ou sec ions ha will allow con ex ualizing he amewo k o his s udy. Sec ion 2.1 is dedica ed o he p ima y conside a ions o land co e mapping, whe eas sec ion 2.2 p esen s he cu en ag icul u al moni o ing sys ems based on emo ely sensed da a and how hese e o s bene i om exis ing da ase s such as he Land Pa cel In o ma ion Sys em (LPIS). Then, he concep o ime se ies o sa elli e image y is in oduced in sec ion 2.3, co e ing he di e ences be ween in a-annual and in e -annual ime se ies and he need o p oduce gap- illed composi es be o e using he da a. Finally, sec ion 2.4 comp ises he p inciples o Machine Lea ning, also called S a is ical Lea ning, and how hese algo i hms a e di e en om he mos commonly employed in emo e sensing (i.e., Maximum Likelihood). I also desc ibes he adi ional machine lea ning wo k low and i s key cons i uen s, he di e ences be ween supe ised and unsupe ised lea ning as well as eg ession, classi ica ion and clus e ing. A las , i in oduces he algo i hm ha will be implemen ed h ough he hesis ha is Random Fo es and pu i in o he con ex o emo e sensing and image classi ica ion. 2.1 Land co e mapping Land co e analyses ha e e ol ed om s udying a small geog aphic egion a a de e mined pe iod o global s udies using smalle spa ial esolu ion and highe empo al pe iods [18]. I emains, howe e , an in ica e p ocess, and in supe ised land co e app oaches, he c i ical componen is he a ailabili y o aining da a (g ound u h o e e ence da a) o he signa u e gene a ion [5]. Acco ding o he me a-analysis on supe ised pixel-based land-co e image classi ica ion [19] ha compa ed 266 a icles be ween 1998 and 2012 he mos ele an ea u es conside ed when pe o ming a classi ica ion p ocess we e ex u e, ancilla y da a (e.g., opog aphic, ac i e senso s such as ada o LiDAR and passi e senso s), mul i- ime image y (e.g., usion o images o he same a ea cap u ed a di e en imes), mul i-angle image y, image p e-p ocessing (e.g., adiome ic co ec ion, a mosphe ic co ec ion, pan sha pening, and geome ic co ec ions), spec al indices (e.g., NDVI -a i hme ic combina ions o di e en spec al bands) and ea u e ex ac ion (e.g., dimensionali y educ ion). Howe e , aining da a collec ion is delica e in la ge ju isdic ions and o e emo e a eas [3]. Ongoing esea ch on he ex ac ion o au oma ic aining da a includes he majo i y ule app oach in polygon le el sou ce maps [16]. 7 2.2 C op ype mapping When using Ea h Obse a ion da a o moni o ing ag icul u e, he e a e ecognized amewo ks such as he GEOGLAM ini ia i e. Cu en ly, se e al main global and egional scale ag icul u al moni o ing sys ems a e in place, some o hem a e he Global In o ma ion And Ea ly Wa ning Sys em (GIEWS), he Famine Ea ly Wa ning Sys ems Ne wo k (FEWS NET) and he C op Wa ch o China [20]. The ESA “Sen inel 2 o Ag icul u e” (Sen2Ag i) ha s a ed in 2015 aims o c ea e ope a ional c op ypes maps and dynamic c opland masks [21] ha a e equi ed as inpu o global ag icul u al moni o ing sys ems. Di e ences ha e been es ablished be ween c opland maps, c op calenda s, c opping in ensi y, c op ype, g owing calenda , c op condi ion indica o s, and c op yield [20]. In c op ype classi ica ion, he classi ie s yielding he bes pe o mances a e Random Fo es , ollowed by he g adien boos ed ees and hen SVM [22]. RF has also been implemen ed in bina y ope a ions (c opland/non-c opland) sys ems [23]. Ne e heless, he key o di e en ia ing indi idual c op ypes is he a ailabili y o empo al in o ma ion [24]. Fo he calib a ion and co ela ion o he spec al signal o he a ious c ops, in o ma ion a a pa cel-le el is also a c ucial elemen [24]. Se e al s udies, including he Sen2-Ag i sys em, ha e epo ed he use o he Land Pa cel Iden i ica ion Sys em (LPIS) o ex ac samples om ag icul u al pa cels [6], [17], [24]. The LPIS is an IT sys em based on ae ial pho og aphs o ag icul u al pa cels employed o check paymen s made unde he Common Ag icul u al Policy (CAP) o app oxima ely 45.5 billion eu o in 2015 [25]. In Po ugal, he Ins i u o de Financiamen o da Ag icul u a e Pescas (IFAP) ensu es he inancing, implemen a ion, and con ol mechanisms o he measu es de ined a he na ional le el in ag icul u e and ishe ies. I ac s as Na ional Con ol and Paying Agency (NCPA) designa ed by he Eu opean Union (EU) unde he Common Ag icul u al Policy (CAP) and is esponsible o he adminis a ion and con ol o he subsidies in his sec o . Fo applying o inancial suppo , a me s a e equi ed o submi an applica ion o he NCPA and decla e he p ecise loca ion and a ea o he ag icul u al pa cels and he c op ype [26]. Fo his, landowne s use an online Geog aphic In o ma ion Sys em (GIS) o digi ize hei pa cels on o hopho os o e y high- esolu ion sa elli e image y [24]. The NCPA con ols a leas 5% o he decla a ions by pe o ming an On-The-Spo (OTS) check, penalizing he a me s ha submi ed inco ec in o ma ion [26]. 8 2.3 Sen inel-2 ime se ies 2.3.1 Time se ies image y The coa se (i.e., pixel size > 250 m) o medium esolu ion op ical ins umen s on boa d o SPOT-Vege a ion, MODIS, and PROBA-V ha e a daily e isi cycle, global co e age, and long- e m a chi e [6]. This da a can be exploi ed in long ime-se ies esea ch a egional o global scales, bu o en su e om low local accu acy in land co e p oduc s [27] and high mix u es o c op ypes [28]. The e isi ing pe iod o 16 days o he Landsa 8 sa elli e o he U.S. Geological Su ey (USGS) allows us o desc ibe spa ial de ails o land co e bu canno cap u e changes in c op phenology and g ow h due o low empo al epea cycles and equen cloud con amina ion [28]. Sen inel-2A (S2a) sa elli e o he Eu opean Spa ial Agency (ESA) p o ides a e isi ime o 10 days, and he Sen inel-2B (S2B) ensu es a 5-days e isi ime allowing he collec ion o high-quali y spa ial and empo al da a [6]. Mo e gene ally, ime se ies algo i hms ha e eme ged o e he las decade ha can exploi dense, mul i-senso ime se ies o de i e imp o ed land co e classi ica ions [29]. Recen coun y-scale s udies ha e demons a ed he added alue o mul i-senso ime se ies om Landsa and Sen inel-2 o di e en ia e c op ypes and g asslands [24]. 2.3.2 In a-annual and In e -annual In he e iew o ime se ies analysis o Land Co e mapping [27] Gómez e al. (2016) poin ed ou he empo al ele ance o images collec ed o e in e als in he same yea (in a-annual) o o e some yea s (in e -annual). The in a-annual image y allows moni o ing he sub le di e ences and a ia ions o e he g owing pe iod by calcula ing an a e aged phenology while he in e -annual image y allows o compu e a unique spec al p o ile ha makes mo e isible when ab up changes occu in he land co e . Mapping landco e is complex, ime-se ies spec al da a (in a-annual o phenology and in e - annual o land co e dynamics) p o ide mo e in o ma ion o inc easing he classi ica ion. Fo a speci ic class o in e es (e.g., c op ype mapping), i is necessa y o inco po a e he knowledge o o he unde lying p ocesses (e.g., phenology, dis u bance, succession) [27]. 2.3.3 Gap- illed image composi es The a ailabili y o ope a ional imaging sa elli es ha co e s all lands equen ly, such as Landsa 8 and Sen inel-2, p o iding ee and open access o hese da a has p omp ed new applica ions based on ime se ies o images co e ing as e i o ies [3], [5], [24]. 9 Howe e , when p ocessing op ical sa elli e images on land su aces, he de ec ion o cloud and cloud shadows is one o he i s issues. Clouds can equen ly be mis aken wi h b igh landscapes, semi- anspa en clouds obse ed e lec ance con ains a mix u e o cloud and land signals, and cloud shadows can be con used wi h wa e pixels, bu n a eas o opog aphic shadows [30]. As o now, cloud and cloud shadow masking algo i hms o Landsa 8 and Sen inel 2 include he MAJA algo i hm by he F ench Space Agency (CNES), Sen2Co , om he Eu opean Space Agency (ESA) and FMask o he Uni ed S a es Geological Su ey (USGS). When in eg a ing empo al ime se ies o image y, mos app oaches ollow a bes - pixel selec ion s a egy ha allows exploi ing all he image y a ailable [24]. The Bes A ailable Pixel (BAP) enables he compu a ion o pe iodic image composi es ee o haze, clouds, o shadows o e la ge a eas [31]. Whi e, J. C. e al. (2014) [32] p oposed h ee unique ypes o pixel-based image composi es: annual (single-yea ) composi es, mul i-yea composi es, and p oxy- alue composi es. Whe ein, De ou ny, P. e al. (2019) [6] gene a es mon hly composi es using a weigh ed a e age algo i hm, ha a e ages cloud- ee su ace e lec ance alues o e he gi en pe iod. In e pola ion o e su ace e lec ance alues o ill missing alues due o he p esence o clouds and clouds shadows has also been implemen ed o ope a ional sys ems [3]. 2.4 Machine Lea ning – S a is ical Lea ning Wi h he apid g ow h o “Big da a,” machine lea ning, also e e ed o as s a is ical lea ning, a b oad se o ools o analyzing and unde s anding he da a eme ged. Se e al models can be buil and equi e a se o inpu da a o p edic o es ima e ou pu da a [33]. An ad an age o Machine-lea ning algo i hms is ha hey do no make assump ions abou he da a dis ibu ion (i.e., non-pa ame ic), can handle da a o high dimensionali y, and can e icien ly classi y emo ely sensed image y [34]. 2.4.1 Machine Lea ning P ocess A adi ional machine lea ning wo k low (Figu e 1) equi es key cons i uen s: da a collec ion, ea u e enginee ing (cleaning and ea u e selec ion), model lea ning ( aining, alida ing and es ing), and model e alua ion [35]. 10 Figu e 1 Machine Lea ning Wo k low The collec ed aw da a may be noisy, incomple e, o inconsis en , and be o e using he da a as inpu in he model, i is equi ed o p e-p ocess i by emo ing e o s and ou lie s and ill missing alues. O he asks include he in eg a ion o mul iple da ase s and ans o m he da a in o an app op ia e o ma , so i is eadable depending on he ool deployed o pe o m he machine lea ning p ocess. Fea u e selec ion and ex ac ion a e u ilized o educe dimensionali y in oluminous da a, allowing o emo e i ele an o edundan ea u es ha p omo e o e - i ing and o educe compu a ional equi emen s. Some echniques o dimension educ ion include en opy, Fou ie ans o m, and P incipal Componen Analysis (PCA) [35]. The aining da ase is implemen ed o each he model how o es ima e he unc ion ha will be able o p edic ou pu o any new obse a ion [33]. A alida ion da ase is applied o choose a sui able a chi ec u e o he model. I he a chi ec u e is p e- selec ed, he e is no need o a alida ion se . Finally, he es ing da ase allows he model o i e a e and une he di e en pa ame e s un il he model is eady o be deployed. The main decomposi ions o he da ase a e 60/20/20% i aining, alida ion, and es da ase s o 70/30% i alida ion is no equi ed [35]. The aining da a selec ion is ele an , because la ge, and accu a e aining da ase s esul in inc eased classi ica ion accu acy. I is being sugges ed ha he minimum numbe o aining samples should be en o p e e ably 100 imes he numbe o a iables [34]. The e alua ion ocusses on he p edic i e e icacy o he model and on he compu a ional equi emen s ( aining and es ing ime) o i s applica ion [36]. A high bias e e s o a simple ML model ha poo ly maps he ela ions be ween ea u es and ou comes (unde - i ing) while a high a iance implies an ML model ha i s he aining da a bu does no gene alize well o p edic new da a (o e - i ing) [35]. Techniques o expe imen al algo i hm e alua ion include boo s ap sampling, c oss- alida ion, and holdou e alua ion [36]. 11 2.4.2 Supe ised and Unsupe ised Lea ning Supe ised lea ning uses labeled aining da ase s o c ea e models, and ypically, his app oach is used o sol e classi ica ion and eg ession p oblems [35]. The e o e, he algo i hm uses pa e ns o p edic he alues o he labeled da a on addi ional unlabeled da a, and by compa ing he ac ual ou pu wi h he co ec ou pu , i inds he e o s, lea ns, and modi ies he model acco dingly. Unsupe ised lea ning uses unlabeled aining da ase s o c ea e models ha can disc imina e be ween pa e ns in he da a. This app oach is mos sui ed o clus e ing p oblems [35]. The algo i hm explo es he da a and inds pa e ns o g ouping oge he alues based on hei ea u es. 2.4.3 Reg ession, Classi ica ion, and Clus e ing In da a clus e ing, he aim is o pa i ion objec s in o g oups such ha simila objec s a e g ouped while dissimila objec s a e g ouped sepa a ely. Ca ego ical clus e ing iews he da a as a se o a wo-dimensional ma ix o da a objec s and a ibu es (a se o disc e e alues ha a e no compa able) and a emp s o pa i ion he se o objec s in o g oups wi h simila a ibu es [36]. Well-known clus e ing algo i hms a e K-means and Kohonen Sel -O ganizing Maps (SOM). Whe eas in classi ica ion and eg ession p oblems, he goal is o map a se o new inpu da a o a se o disc e e o con inuous- alued ou pu s [35]. Some classi ica ion algo i hms include Decision T ees (DT), Neu al Ne wo ks (NN), K-Nea es Neighbo s (k-NN), Bayesian Ne wo ks (BN), and Suppo Vec o Machines (SVM). The e a e also ways o combining hem in o ensemble classi ie s such as boos ing, bagging, and he ensemble DT - Random Fo es (RF). While known eg ession algo i hms a e mainly linea models such as Leas Squa es ha include speci ic echniques such as OLS, MaxEn , Logis ic Reg ession [35], [36], LASSO Reg ession, SVM and Mul i a ia e Reg ession algo i hm. 2.4.4 Random Fo es (RF) Random Fo es algo i hm speci ica ions o classi ica ion The RF classi ie is an ensemble classi ie ha uses mul iple Classi ica ion and Reg ession T ees (CART) and combines hei ou pu s o make a p edic ion, ea ing hem as a “commi ee” o decision-make s [36], [37]. I combines he Bagging algo i hm o educe a iance by he andom selec ion o samples and he Random Subspace me hod o educe bias by he andom selec ion o he ea u es employed a each spli [36]. When ope a ed o classi ica ion, each ee “ o es” o a class and hen classi y using he “majo i y o e” o he o es [38]. Findings o Random Fo es is ha i does no o e i as mo e 12 ees a e added, i is ela i ely obus o ou lie s and noise, gi es use ul in e nal es ima es o e o , s eng h, co ela ion and a iable impo ance and is easily pa allelized [37]. 1. Fo b = 1 o B (n° o ees in he o es ): a. D aw a boo s ap sample Z* o he size N om he aining da a. b. G ow a andom- o es ee 𝑇𝑇𝑏𝑏 o he boo s apped da a, by ecu si ely epea ing he ollowing s eps o each e minal node o he ee, un il he minimum node size nmin is eached. i. Selec m a iables a andom om he p a iables. ii. Pick he bes a iable/spli -poin among he m. iii. Spli he node in o wo daugh e nodes. 2. Ou pu he ensemble o ees {𝑇𝑇𝑏𝑏}1 𝐵𝐵. To make a p edic ion a a new poin x: Classi ica ion: Le Ĉ𝑏𝑏(𝑥𝑥) be he class p edic ion o he b h andom- o es ee. Then Ĉ𝑏𝑏𝑟𝑟𝑟𝑟 𝐵𝐵(𝑥𝑥) = majo i y o e {Ĉ𝑏𝑏(𝑥𝑥)}1 𝐵𝐵. Table 1 Algo i hm: Random Fo es o Classi ica ion [37], [38] The i s s ep in he RF algo i hm in Table 1 (a) is o ex ac a “boo s ap sample” om he aining da ase ; boo s apping allows o selec he same sample mo e han once and include i in he subse da ase while o he samples may no be selec ed a all. Bagging (ac onym de i ed om Boo s ap AGG ega ING) allows ha each membe o he ensemble is cons uc ed om a di e en aining se , each da ase being a boo s ap sample om he o iginal [36]. Abou wo- hi ds o he samples (in-bag samples) a e used o ain he ees wi h he emaining one hi d (ou -o - he-bag) a e employed in an in e nal c oss- alida ion echnique o pe o mance es ima ion o he model [37], [39]. Then, each ee is g own using samples om he boo s apped da ase (b); howe e , i will selec a andom a iable m om he ull se o a iables p a ailable and pick he bes one o he op spli [38]. The andom subspace p inciple is o inc ease di e si y be ween membe s o he ensemble by es ic ing classi ie s o wo k on di e en andom subse s o he ull ea u e space [36]. This p ocedu e is epea ed o he numbe o ees in he o es ; bagging seems o enhance accu acy when andom samples a e u ilized, his is also he case when using a single andomly chosen inpu a iable o spli on a each node [37]. 13 Random o es classi ica ion o emo e sensing da ase s ML algo i hms ha e use -de ined pa ame e s ha may imp o e classi ica ion accu acy when unning pa ame e op imiza ion. One o he bene i s o he RF algo i hm is ha i is conside ed easy o op imize in compa ison o mo e complex models such as ANN. Thei s abili y conce ning he choice o pa ame e s makes hem excellen candida es o ope a ional p ocessing chains, yielding classi ica ion accu acies as high as mo e sophis ica ed algo i hms such as SVM bu wi h much lowe compu a ional complexi y [3]. RF has been exploi ed in ime se ies analysis o c ea ing mul i- empo al cloud mask o Sen inel-2 image y [30]; in supe ised classi ica ion o p oducing land co e maps a a coun y scale o F ance [3], as well as c op ype and land co e maps o Ge many [24] and in global ope a ional sys ems such as Sen2Ag i o c op ype maps [6]. RF algo i hm equi es only wo use -de ined pa ame e s: he numbe o Decision T ees in he ensemble and he numbe o andom a iables a each node [34]. RF is compu a ionally e icien and does no o e i [39]; he numbe o ees does no impac accu acy as long as i is la ge enough, being 500 a e y conse a i e alue [37]. The es ima ed e o a e can be plo ed o each ensemble size o de e mine when he pe o mance s abilizes [34], [37]. Ano he ad an age o RF is ha he algo i hm i sel gene a es addi ional in o ma ion [37]. The ou -o -bag (OOB) e o p o ides an unbiased es ima e o gene aliza ion e o and esembles he e o es ima e ob ained by N- old c oss- alida ion [38]. Also, he Va iable Impo ance (VI) es ima ion anks he a iables based on he p edic i e capabili ies o disc imina ing be ween he a ge classes [37], [39]. The VI has been exploi ed in emo e sensing o educe he numbe o dimensions o hype spec al da a (i.e., he con ibu ion o bands), o iden i y ele an ancilla y da a (i.e., opog aphy) and o selec he sui able season o classi y a ge classes [34], [39]. This allows add essing he challenges o mi iga ing he Hughes phenomenon (i.e., he cu se o dimensionali y) ha occu s when he numbe o a iables is much la ge han he numbe o aining samples [39]. The main d awback is ha i is sensi i e o sampling design when imbalanced da a is used; he inal classi ica ion will unde -p edic he mino i y class [34]. To educe misclassi ica ion, his sensi i i y o sampling design needs o be conside ed by ensu ing ha aining and es ing a e independen , es ablish balance and ep esen a i i y o each class, and ha e an ex ensi e aining sample o deal wi h he numbe o da a dimensions [39]. 20 Table 2 Nomencla u e o Land Co e and C op ype 21 3.2.2 Remo e Sensing Da a Sen inel 2 The Cope nicus Sen inel-2 mission is a cons ella ion o wo pola -o bi ing sa elli es ha ope a e simul aneously, phased a 180° o each o he a a mean al i ude o 768 km. This allows a and high e isi ime (10 days o S2-A and 5 days o S2A/B), and i s wide swa h wid h (290 km) p o ides a high co e age [43] being ideal o he p oposed s udy. The image y is acqui ed by he Mul ispec al Ins umen (MSI) on-boa d Sen inel-2 and con ains 13 spec al bands om Visible/Nea In a ed (VNIR) o Sho Wa e In a ed (SWIR) and comes in h ee spa ial esolu ions (10, 20 and 60m) as seen in Table 3. Band Spa ial esolu ion (m) Cen al wa eleng h(nm) Bandwid h (nm) Pu pose B01 60 443 20 Ae osol de ec ion B02 10 490 65 Blue B03 10 560 35 G een B04 10 665 30 Red B05 20 705 15 Red Edge B06 20 740 15 Red Edge B07 20 783 20 Red Edge B08 10 842 115 Nea In a ed (NIR) B08A 20 865 20 NIR B09 60 945 20 Wa e apo B10 60 1375 30 Ci us B11 20 1610 90 Snow/ice/cloud disc imina ion (SWIR) B12 20 2190 180 Snow/ice/cloud disc imina ion (SWIR) Table 3 Speci ica ions o he Sen inel-2 bands The Sen inel-2 da a was ob ained om he F ench Theia Land Da a Cen e (THEIA). The da a is in he Coo dina e Re e ence Sys em (CRS) o Uni e sal T ans e se Me ca o (UTM) Zone 29N, and i is iled in he Mili a y G id Re e ence Sys em (MGRS) allowing all he images o ha e he same size (100x100 km2) and a code (e.g., a ile in Po ugal is T29SND). 22 The biogeog aphical egion 214 “Tejo and Sado” is co e ed by ou iles o Sen inel-2 ha co espond o 29SNB, 29SNC, 29SND, and 29SPD. The classi ica ion will be done o he s a a 214 ha is wi hin he iles 29SNC and 29SND. Mos o he s udy a ea is comp ised wi hin he same o bi ; howe e , a sligh co ne in he 29SND ile has swa h o e lap wi h he adjacen o bi , accoun ing o mo e image y collec ed wi hin he same pe iod han ile 29SNC. Al hough he e- isi a ion pe iod is he same, he da es in he collec ion o he o bi s a y om le o igh in adjacen land. Figu e 3 Sen inel-2 o bi , swa h, and illing o he s udy a ea and images acqui ed o Oc obe 2017 in he ile 29SND. O hopho os o Con inen al Po ugal Fo all he isualiza ions, he o hopho os a ailable as Web Map Se ice (WMS) om DGT we e used. The image y has a spa ial esolu ion o 25 cm, and i is a ailable o Con inen al Po ugal. 23 3.3 Me hods The p oposed me hodology co esponds o an au oma ic supe ised classi ica ion p ocedu e using he andom o es classi ie , in a-annual ime-se ies o Sen inel-2, and il e ed auxilia y da a o ex ac he labels o land co e and c op ypes au oma ically (Figu e 4). Fi s , he e e ence da ase s (COS2018 and IFAP 2018) a e eclassi ied using he nomencla u e om, hen a se o p ep ocessing ules is applied o he da ase s o emo e he pixels ha do no ma ch he class label (3.3.2). Nex , sec ion 3.3.3 illus a es he p ep ocessing o he Sen inel-2 in a-annual ime se ies. Ini ially, he image y is downloaded o he pe iod o Oc obe 2017 o Sep embe 2018; a mask o emo e clouds and cloud shadows a e applied, and all he bands a e esampled o 10m. La e , i e spec al indices a e calcula ed, and all he image y is agg ega ed o mon hly composi es. The po en ial missing alues (pixels wi h no da a du ing a mon h) a e illed using linea in e pola ion in ime o ensu e con inui y o in o ma ion du ing he pe iod. A e wa d, he supe ised lea ning p ocedu e is p esen ed in sec ion 3.3.4. This sec ion s a s wi h he au oma ic ex ac ion o samples by class om he p e-p ocessed da ase s. Two independen sample da ase s a e acqui ed, one o aining and one o es ing wi h a ying pe cen ages 80/20 o 75/25 depending on he numbe o pixels a ailable. Fo each sample, he spec al signa u es a e e ie ed a he pixel le el om all he bands o he composi es and he spec al indices. Then, a g id sea ch is used o de e mine he bes hype pa ame e s o he RF; he models a e i o he aining da ase and assessed wi h 10-spli c oss- alida ion. A las , he pe o mance o he bes model is quan i ied ollowing he me ics in (3.3.5) based on he p edic ed labels in he es ing da ase . Then, he model is applied o unlabeled da a allowing o gene a e he inal map o he biogeog aphical egion. 24 Figu e 4 Flowcha o he au oma ic p oduc ion o a land co e and c op ype map in cen al Po ugal 25 3.3.1 So wa e and de ice speci ica ions The da a p e-p ep ocessing was done using he so wa e A cGIS P o 2.4.0 om ESRI. A geop ocessing wo k low was de eloped wi h he isual p og amming language o Model Builde ha can la e be expo ed as a py hon sc ip . The sample ex ac ion, model aining, and classi ica ion we e done using Anaconda Dis ibu ion ha is an open-sou ce pla o m o pe o m da a science and machine lea ning. The e sion ins alled co esponds o Anaconda3-4.4.0-Windows- x86_64 ha con ains Py hon (3.5.4) and he equi ed lib a ies such as NumPy (1.13.1) [44] and Pandas (0.20.3) [45] o da a s uc u es, Seabo n and Ma plo lib (2.0) [46] o da a isualiza ion and Sciki -Lea n [47] o conduc ing machine lea ning analysis since i includes he andom o es classi ica ion algo i hm. O he lib a ies ins alled comp ise GeoPandas o i s spa ial unc ionali y wi h geospa ial da a and he Geospa ial Da a Abs ac ion Lib a y (GDAL), which is a ansla o lib a y o as e and ec o geospa ial da a o ma s [48]. The ea u e ex ac ion, classi ica ion, and he elabo a ion o he inal map we e done using he compu e s o DGT. The compu e s ha e an ins alled RAM o 64.0 GB wi h a p ocesso In el (R) Xeon (R) Gold 6140 CPU @ 2.3GHz 2.29 GHz. Fo all he o he p ocedu es, a pe sonal compu e was used, wi h a p ocesso In el (R) Co e (TM) i7- 7500U CPU @ 2.70 GHz 2.90 GHz and ins alled RAM o 8.00 GB. 3.3.2 P ep ocessing o he e e ence da ase s The wo k low o p e-p ocessing he e e ence da ase s is de ailed in Figu e 5. Fi s , he IFAP 2018 da ase was eclassi ied om 175 c op ypes o 14, and a bu e o - 40m was applied. Nex , om he 83 classes a ailable in COS 2018 da ase , a o al o 15 we e ex ac ed; likewise, an inne bu e was used. Then, he emaining polygons we e c ossed wi h he auxilia y da a; his includes he ICNF bu ned a eas 2015-2018, NDVI ale s o clea cu s 2015-2018, and HRL laye s 2015 (DLT and TCD). Finally, i IFAP had o e lapping a eas wi h COS, hese we e emo ed om he la e ; he inal da ase comp ises IFAP 2018, COS 2018 and ICNF 2018. As he IFAP con olled pa cels we e used o es ing while he na ional pa cel egis y was used o aining, bo h da ase s we e kep spa ially independen . This is no he case o COS no ICNF, being he whole da ase used bo h o aining and es ing. 26 Figu e 5 Da a p ep ocessing wo k low o e e ence da ase s C op ype da ase (IFAP 2018) F om he IFAP da ase , he en mos abundan empo a y c ops (5 ain ed and 5 i iga ed), h ee pe manen c ops, and he ag icul u al g asslands we e selec ed o he analysis o a o al o 14 c op ypes classes. An Explo a o y Spa ial Da a Analysis (ESDA) was pe o med o iden i y he en mos abundan empo a y c ops in he s udy a ea o he classi ica ion (Figu e 6); hese co esponds o maize (24,012ha), ice (21,595 ha), oma o (12,742ha), yeg ass (5,472ha), oa meal (4,163ha), whea (2,723ha), so ghum (2,104ha), ba ley (1,844ha), lupin (1,762ha) and po a o (1,748ha). As o he pe manen c ops, oli e ees, ineya ds, and o cha ds we e conside ed due o hei impo ance in Po ugal’s ag icul u e. The o cha d class combines 17 ypes o ees om igs and o anges o walnu s and hazelnu s. Figu e 6 A ea co e ed by he 10 mos abundan c ops in hec a es 27 The numbe o pa cels is ele an ; he mo e pa cels a e dis ibu ed wi hin he s udy a ea, he mo e ep esen a i i y is possible o ob ain. Acco ding o he dis ibu ion o he pa cels (Figu e 7), mos o hem a e less han 10 ha. The a e age a ea o all he classes is 3.35 ha, being he oma o pa cels wi h he highe mean a ea (5.38ha) and oa meal he lowe mean a ea (2.35ha). Figu e 7 Dis ibu ion o he ee main c op pa cels by a ea (ha) An in e se bu e o -40 m was pe o med o he o iginal pa cels o a oid selec ing pixels o which he spec al signa u e does no ma ch he class label, as can be seen in Figu e 8. Du ing he bu e ing p ocess, i can occu ha he smalles pa cels a e emo ed om he da ase , educing he numbe o a ailable pixels o aining. No o he c ossing wi h ancilla y da a was applied o his da ase , aining (na ional pa cel egis y) and es ing (con olled pa cels) we e kep independen . This was ensu ed by pe o ming an in e sec ion be ween he da ase s and emo ing he con olled pa cels om he na ional pa cel egis y gua an eeing ha all he polygons a e spa ially disjoin ed. Figu e 8 IFAP pa cels p e-p ocessing: in e se bu e (- 40m) 28 Land Co e da ase (COS 2018) A o al o 16 LC classes we e de i ed om he 83 classes a ailable in COS 2018, and a -40 m bu e was applied o he emaining polygons. Ye , i is c i ical o emphasize ha he MMU o COS (1ha) en ails a educ ion in de ail o be e model he eali y; and many imes, i equi es he gene aliza ion o polygons. This means ha a eas smalle han 1ha (pa hs, edi ica ions and o he objec s) will be agg ega ed wi h he p edominan class up o 25% o he o al a ea o he polygon [7]. Classi ica ion a he pixel le el o Sen inel- 2 con empla es a 10 m MMU; he e o e, some p e-p ocessing s eps a e equi ed o p e en he selec ion o pixels wi h spec al in o ma ion ha misma ch he class label inhe i ed om COS which has a la ge MMU (and po en ial hema ic e o s). The i s s ep was o in e sec all COS polygons wi h he ICNF bu ned mask o he yea s 2015- 2018. The mask allowed o c ea e holes in he polygons by elimina ing he sco ched a eas; consequen ly, no au oma ic sample will be ex ac ed om hese a eas. A o al o 392 ha in 2015, 2565 in 2016, 6002 ha in 2017, and 99 ha in 2018 we e emo ed om he da ase . In Figu e 9, i is possible o ecognize a blackened a ea inside a Ma i ime Pine class in 2016 and inside a Eucalyp us class in 2017. S ill, he bu ned mask does no co e he polygon ex ensi ely, as i can be app ecia ed in he Ma i ime pine whe e wo holes emained in he polygon co esponding o edi ica ions and in Eucalyp us whe e he mask does no co e he o al ex en o he a ea. The o es ype mos educed in he a ea a e applying by he bu ned mask co espond mainly o co k oak (263 772 ha) and eucalyp us (80 323 ha). In Po ugal, he eucalyp us o es is indus ially g own o supply pulp ibe o he pape indus y, al hough hey a e highly lammable [49]. 29 Figu e 9 COS2018 polygon o e laid wi h he bu ned mask, Ma i ime Pine OBJECTID: 493848 and Eucalyp us OBJECTID: 415864. Scale 1:7,000 (1) and 1:25,000 (2). The second s ep applies only o he o es a eas since aining and es ing canno be sampled om o es cu s. NDVI di e encing echniques allow disc imina ing be ween eal changes and seasonal o in e -annual a iabili y o o es s [50]. This echnique has been implemen ed in Po ugal [51] o de ec ege a ion loss ha occu ed be ween 2015-2018 in o es s, so-called NDVI ale s. The o es polygons we e c ossed wi h he NDVI ale s mask o emo e he a eas whe e he e ha e been changes, and hence he class label o COS does no co espond o he pixel spec al signa u e. A e applying he mask, he mos a ec ed o es s a e eucalyp us wi h 13785 ha educed, ollowed by 7076 ha in s one pine and 3498 ha in ma i ime pine. This o es agmen a ion (i.e., b eaking o la ge, con iguous o es ed a eas in o smalle pieces o a o es ) is due o some ex en o oad cons uc ion, i es, logging and con e sion o ag icul u e. In he case o o es plan a ions like eucalyp us, clea -cu s a e pa o he o es managemen cycle; as a new o es is expec ed o ollow, he land use emains a o es [51]. Howe e , in a s ic ly land co e map de i ed om supe ised classi ica ion, hese changes in ege a ion can esul in misclassi ica ions when implemen ing he model and he e o e equi e o be emo ed om sample ex ac ion. Figu e 10 NDVI ale s in a Eucalyp us plan a ion (COS 2018 OBJECTID: 382944). Scale 1:80,000 (1) and 1:6,000 (2). The inal s ep was o c oss he o es a eas, and he sh ublands wi h he High- Resolu ion Laye s (HRL) masks c ea ed ollowing he ules in Table 4. The Dominan Lea Type (DLT) allows sepa a ing b oadlea o coni e ous majo i y, while he T ee Co e Densi y (TCD) anges om 0 o 100%. Fo co k oak, holm oak, o he b oadlea , and eucalyp us, i is equi ed ha hey co espond o b oadlea wi h a ee co e highe han 60% o he pixel in he HLR. Fo coni e ous, i mo e han 60% o he pixel is co e ed by 36 Figu e 17 Au oma ic sample ex ac ion o COS 2018 (Co k Oak OBJECTID: 382944). Scale 1:20,000 (1) and 1:5,000 (2) and (3). The iles 29SND and 29SNC ha e o e lapping a eas; o a oid ha he same sample ex ac ea u es om bo h iles, he samples we e di ided. The p io i y was gi en o ile 29SND as i con ains mo e images a ailable o he pe iod; he aining and es ing da ase s we e c opped o he whole ex en whe eas, o ile 29SNC, he o e lapping a ea was no conside ed (Figu e 18). The g een do s indica e he samples ha will ex ac spec al in o ma ion om ile 29SND, and he blue ones will e ie e he in o ma ion om ile 29SNC. Figu e 18 T aining and es ing samples o he biogeog aphical egion di ided be ween iles 29SND and 29SNC. 37 Ex ac ion o he spec al ea u es Be o e inpu ing he da a in o he model, he aining and es ing da ase mus con ain he equi ed ea u es/ a iables. The in o ma ion ha is p o ided o he classi ie co esponds o he su ace e lec ance alues o he Sen inel-2 composi es and he spec al indices de i ed. A o al o 180 ea u es a e ex ac ed, equi alen o 10 bands and i e spec al indices o each o he 12 mon hs (Oc obe 2017 o Sep embe 2018), as shown in Figu e 19. This p ocess was done wi h py hon, adap ing he code s-u il in sec ion 7.1. The samples we e used as a mask o ex ac he spec al in o ma ion a he pixel le el om all he image y. The da a e ie ed is sa ed as an a ay o 180 ea u es; all he a ays we e con e ed o a pandas Da aF ame, a wo-dimensional abula da a s uc u e in .cs o ma . Figu e 19 Ex ac ion o he spec al ea u es o aining and es ing da ase s The esul ing aining Da aF ame con ains a o al o 115,880 samples wi h 180 ea u es ex ac ed (Figu e 20), while he es ing Da aF ame con ains 29,150 samples. The ows ep esen he samples used o ex ac he da a a he pixel le el, and he columns con ain he ea u es e ie ed om he Sen inel-2 image y. The ‘CLASS’ column con ains he labels co esponding o he land co e and c op ype a ge classes; howe e , he RF model does no accep s ing a iables and he e o e, he nume ical codes ‘LV4’ a e he ones used as inpu in he classi ica ion. The o al ime o ea u e ex ac ion was app oxima ely 2 hou s using he compu e s a DGT. 38 Figu e 20 Da a ame o he aining da ase con aining he 115,880 samples wi h 180 ex ac ed ea u es One o he bene i s o he Da aF ame abula s uc u e is ha i allows que ies and a i hme ic ope a ions along bo h ows and columns. Fo isualiza ion pu poses, he alues o only one mon h we e acqui ed; in his case, Oc obe 2018 (Figu e 21). F om he ea u es ob ained, we can app ecia e ha he e a e high co ela ions be ween he isible spec um (b2, b3, b4) and he ed-edge bands and NIR o ege a ion disc imina ion (b6 o b8a) and be ween he bands 11 and 12 used o snow/ice and cloud disc imina ion (SWIR). The bands 1 o ae osol de ec ion and 9 o wa e apo we e no included in his analysis as hey would no p o ide in o ma ion abou su ace e lec ance’s alues 39 Figu e 21 Co ela ion be ween spec al signa u es o bands and indices o Oc obe 2017 Model uning and aining Model e alua ion is a equi ed ask in classi ica ion using machine lea ning algo i hms; a adi ional way o alida e he pe o mances o classi ica ion is o use pa o he a ailable samples o aining and ano he o alida ion [30]. The 10- old c oss- alida ion me hod [59] allowed o andomly spli he aining da ase in o 10 pa s and c ea e 10 alida ion expe imen s using each ime a di e en pa o alida e and 9 o he pa s o aining he classi ie . This was done o each o he hype pa ame e s in he G id Sea ch Table 8. The RF model was buil using he open-sou ce Sciki -lea n lib a y o machine lea ning ha is a ailable o deploy using sklea n.ensemble.RandomFo es Classi ie [47]. The sklea n.model_selec ion.G idSea chCV allowed es ing a ious hype pa ame e s combina ions o RF o achie e he op imal classi ica ion pe o mance o he aining da ase ; a summa y is p o ided in Table 8. 40 Algo i hm Use -de ined pa ame e s Values Random o es C i e ion Gini Numbe o ees 100, 200, 300, 400, 500 Maximum dep h None, 2, 16 Table 8 Use -de ined pa ame e s o G id Sea ch The Gini c i e ion is conside ed o ex ac he a iable impo ance, and he numbe o es ima o s is es ed om 100 ha is he de aul alue in Sciki lea n o 500, which is he ecommended alue o RF [39]. Al hough RF does no o e i , i is possible no o p une he ees; ne e heless, he es is conside ed om ‘none’ un il 16 spli s. By lea ing he max ea u es pa ame e in ‘au o,’ he so wa e will conside he numbe o ea u es a each spli (m) o be �p (being p he o al numbe o a iables); his he de aul o RF [38]. The aining da ase (80% o al da a o classes wi h 5000 samples and 75% o classes below 5000 samples) will be used o ain he model pa ame e and adjus he pa ame e s using 10- old c oss- alida ion. The bes pe o ming model was selec ed by he anking achie ed in he aining da ase , and he 10 c oss- alida ions esul s o ha model a e p esen ed in Table 10. A o al o 500 ees wi hou p uning ou pe o med he o he models es ed; his is consis en wi h mos li e a u e on RF epo ing ha he e o s abilizes be o e eaching 500 ees and ha is ecommended o le each ee o e i un il he node eaches pu i y [34], [37], [39]. A o al o 15 models we e es ed using 10 co es and 32 GB o RAM o he DGT se e s, he o al aining ime ook 93 minu es, o 100 ees as es ima o he a e age aining ime is 2 minu es while o 500 ees i akes 10 min (Table 9). Mean accu acy Numbe o ees Max dep h 100 200 300 400 500 None 5 4 2 3 1 4 15 14 11 13 12 16 10 9 6 8 7 Mean i ime (min) 2 4 6 8 10 Table 9 Ranking o he hype pa ame e s g id using 10- old c oss- alida ion 41 1 2 3 4 5 6 7 8 9 10 Mean S d 0.609 0.78 0.737 0.703 0.740 0.753 0.693 0.64 0.63 0.57 0.687 0.065 Table 10 C oss- alida ion esul s o he bes model (10 olds, mean and s anda d de ia ion) 3.3.5 Accu acy Assessmen o he model pe o mance and map p oduc ion This inal subsec ion ocuses on alida ing he pe o mance o he model and classi ying bo h iles o p oduce a inal map in as e o ma ; he biogeog aphical egion is used as a mask o ex ac he ROI (Figu e 22). Figu e 22 Accu acy assessmen and map p oduc ion wo k low Valida ion When using classi ica ion models in emo e sensing, i is equi ed o quan i y he numbe o imes he model p edic ions ma ch he eali y being modeled [36]. Accu acy assessmen compa es he pixels o polygons om a map classi ied using ML algo i hms o a e e ence es da ase (g ound u h) o which he labels a e known [3]. A con usion ma ix summa izes he classi ica ion pe o mance, whe e he ow en ies a e he ac ual classes ( e e ence da a), and he column en ies con ain he numbe o pixels p edic ed by he classi ie belonging o he column class. In a wo-class p oblem, he con usion ma ix is a wo-dimensional ma ix, one designa ed he posi i e class and he o he he nega i e class (Table 11). T ue posi i es (TP) a e he posi i e samples co ec ly classi ied, and False posi i es (FP) a e a nega i e class inco ec ly classi ied as posi i e. Whe eas T ue nega i es 42 (TN) a e he nega i e samples co ec ly classi ied, and False Nega i es (FN) a e a posi i e class inco ec ly classi ied as nega i e [36]. Assigned Class To al popula ion Posi i e Nega i e Ac ual class Posi i e T ue Posi i e (TP) False Nega i e (FN) Nega i e False Posi i e (FP) T ue Nega i e (TN) Table 11 Bina y con usion ma ix F om he bina y con usion ma ix, se e al indices can be calcula ed o each class, and he a e age o all he alues ac oss classes se e o he mul iclass pu poses. The e alua ion me ics conside ed in his s udy a e based on he py hon implemen a ion o he me ics is a ailable in he Sciki -lea n documen a ion [47]. A summa y is shown below: • Accu acy (sklea n.me ics.accu acy_sco e): he numbe o co ec ly classi ied samples/ o al numbe o samples. The E o a e is di ec ly ela ed o he accu acy, being e o a e = 1.0 – accu acy. • Use ’s accu acy (sklea n.me ics.p ecision_sco e): a a io be ween ue posi i es/ o al numbe o posi i es p edic ed ( 𝑇𝑇𝑇𝑇 𝑇𝑇𝑇𝑇 +𝐹𝐹𝑇𝑇). I is he abili y o he classi ie no o label as posi i e a sample ha is nega i e. • P oduce s accu acy (sklea n.me ics. ecall_sco e): a a io be ween ue posi i es/ o al numbe o ac ual posi i es ( 𝑇𝑇𝑇𝑇 𝑇𝑇𝑇𝑇 +𝐹𝐹𝐹𝐹). I is he abili y o he classi ie o ind all he posi i e samples. • F1-sco e: ha monic mean o p ecision and ecall (2 𝑇𝑇 𝑥𝑥 𝑅𝑅 𝑇𝑇 + 𝑅𝑅). The Kappa coe icien is he p opo ion o ag eemen a e he chance ag eemen is emo ed [60]; howe e , i is no epo ed in his s udy. Kappa can p o ide in o ma ion on assessing he pe o mance o a classi ie , bu i does no p o ide in o ma ion on assessing a map because i is no possible o iden i y ac ual pixels classi ied co ec ly by andom chance hence andom classi ica ion is no a ealis ic al e na i e o c ea e a map [18]. Se e al au ho s ha e explained he unsui abili y o he kappa coe icien in accu acy assessmen o image classi ica ion and encou age esea che s o p o ide mo e s aigh o wa d me ics such as es ima es o pe -class accu acy and con usion ma ices [18], [61], [62]. Aside om he accu acy me ics, he main misclassi ied classes will unde go a isual in e p e a ion o unde s anding he con usion in he classi ie . 43 Map p oduc ion Fo he map p oduc ion, he ained RF model was applied o each unclassi ied pixel assigning he land co e o c op ype ha go he mos o es in he ensemble. The i s a emp o make he map was o ex ac he ea u es pe ile, each ile con aining 10980 x 10980 pixels (app ox. 120 million pixels); howe e , i p o ed o be compu a ionally demanding. The nex app oach was o classi y subsamples o iles con aining 2.5 million pixels; he me hod wo ked; howe e , i was no e icien as his would ha e equi ed o classi y 48 subsamples o 2 hou s each (app ox. 4 ull compu ing days). Las ly, wi h he help o Ped o Bene ides and Hugo Cos a a DGT, i was possible o implemen a classi ica ion app oach using mul ico e p ocessing (18 co es/36 h eads). Each ile was classi ied independen ly wi h he same ained model; he o al compu ing ime was 4 hou s (2 hou s pe ile). 44 4 RESULTS AND DISCUSSION In his sec ion, he bes hype pa ame e s selec ed a e i ed o he aining da a hen he ained model is used o classi y he es ing da a allowing o ex ac he essen ial a iables in he classi ica ion 4.1. The inal map is p esen ed in sec ion 4.2, and i is e alua ed based on he me ics p oposed in sec ion 3.3.5. Then, a pa icula emphasis on c op phenology ollows (4.2.2) ha will compa e he pe o mance o he use o ime se ies o speci ic band e lec ance’s based on he c op calenda o Po ugal. And inally, a isual assessmen o he map (4.2.3) is necessa y o comp ehend he ex en o he accu a e classi ica ions bu also he misclassi ica ions ela ed o he p e-p ocessing s eps. 4.1 Va iable impo ance A e choosing he bes pa ame e , he andom RF was ained acco dingly wi h 500 ees and no p uning; a andom s a e o 101 was used o ensu e ep oducibili y. The model was applied o he es ing da a o e ie e he a iable impo ance, om he 180 a iables used in he model. The 10 mos in o ma i e a iables and he 10 leas in o ma i e a iables a e displayed in Table 12, along wi h hei sco es. The a iables a e anked om 0 o 1, meaning ha he close hey a e o 1, he mo e in o ma ion hey p o ided du ing he spli o he decision ees. The mos c i ical ea u es co espond o he mon hs o sp ing (Ap il) and summe (June and Augus ), whe eas he leas impo an is du ing au umn (No embe and Decembe ) and win e (Janua y and Feb ua y). While he mos in luen ial bands a e in he Red edge (b5, b6, and b7), NIR (b8a) and SWIR (b11 and b12) whe ein he leas impo an a e in he isible spec um (b2, b3, and b4) and he NIR (b8). This co ela es wi h he a ailabili y o spec al in o ma ion, some o he mon hs a e en i ely cloud- ee o all he mainland Po ugal (Oc obe 2017, No embe 2017 and Augus 2018), bu he e a e also some c i ical pe iods whe e he numbe o missing alues is signi ican (e.g., Feb ua y, Ap il o July) [52]. This is he main eason why he mon hly composi es a e in e pola ed in ime, as some o he mon hs can ha e pixels wi h la ge spans o missing da a. Also, he Medi e anean ype o clima e in Po ugal is cha ac e ized by wa m and d y summe s and cool and we win e s [63]; he ege a ion u ilizes he p ecipi a ion ha is accumula ed om No embe o Ap il du ing sp ing and summe o i s pho osyn he ic ac i i ies. The pho osyn he ic phenology is cap u ed by he di e en bands in he MSI, 45 mainly in he Red Edge and NIR, whe eas he SWIR allows pene a ing hin clouds o mois u e disc imina ion on soils. The high co ela ion be ween bands 8 and 8a in luences he in o ma ion gain in he spli selec ion, educing he impo ance o he band 8 and assigning mo e weigh o band 8a. The inclusion o he i e spec al indices appea s o p o ide a mino in o ma ion gain in he classi ica ion, making hem no a p edominan a iable [19]. 10 mos impo an a iables in he model JUNb8a AGOb11 JUNb11 AGOb5 JUNb6 APRb8a JUNb7 ABRb7 AGOb8a JUNb12 0.01146 0.01060 0.01053 0.00998 0.00924 0.00913 0.00882 0.00869 0.00851 0.00849 10 leas ele an a iables in he model JANb3 DICb8a FEBb4 DICb6 JANb8 DICb7 FEBb2 FEBb8 NOVb8 DICb8 0.00297 0.00294 0.00286 0.00276 0.00270 0.00267 0.00265 0.00253 0.00239 0.00228 Table 12 Ex ac ion o he 10 mos impo an ea u e in he classi ica ion and he 10 leas ele an o he selec ed model 4.2 Land Co e and C op Type Classi ica ion The land co e and c op ype map in as e o ma (10m pixel size) is p esen ed in (Figu e 23) based on he me hodology p oposed in sec ion 3. The map con ains 31 classes, om which 14 co espond o ag icul u al classes om whea o ag icul u al g asslands (in he legend), i also includes bu ned a eas and 16 land co e classes. The quali y o he map will be assessed quan i a i ely (4.2.1) based on he accu acy me ics om sec ion 3.3.5 and discussed using li e a u e. 52 Class co ec ly p edic ed on he map The i s example co esponds o whe e “Open Ma i ime Pine” (PA=78) is equal o “Open Ma i ime Pine”, and i illus a es when he labels allow classi ying he class co ec ly. In Figu e 25, he COS 2018 polygon co esponding o he CODE 3121 “Flo es as de Pinhei o b a o” was eclassi ied o Ma i ime Pine. Acco ding o he COS guidelines [7], he polygon con ains 75% o mo e o he o al a ea co e ed by o es . Whe ein, i con ains a egula ne wo k o se ice oads inside he polygon ha can ha e he same p obabili y o being selec ed as “Ma i ime Pine”. Ne e heless, a e he applica ion o he p e- p ocessing s eps in sec ion 3.3.2, he a ea was classi ied as “Open Ma i ime Pine Fo es ” because i con ains mo e han 10% and less han 60% o coni e ous ee co e acco ding o he HRL. When including he samples o his class, nei he he se ices oads no logged a eas a e aken in o conside a ion o ea u e ex ac ion, educing he possible misclassi ica ions. As i is possible o isualize, he inal classi ica ion coincides wi h he class; howe e , he se ice oad ne wo k is classi ied as Ba esoil and U ban in some a eas, being Ba esoil mo e app op ia e. Figu e 25 COS 2018 (OBJECTID: 491011) polygon p e and pos -p ocessed compa ison o p edic ions o he class Open Ma i ime Pine Fo es . Scale: 1:30.000 Nex , i is possible o app ecia e Whe e “Holm Oak” (PA=94) is equal o “Holm Oak”; he COS 2018 polygon is labeled as a pu e o es o Holm Oak. The HRL mask allowed o alloca e aining and es ing poin s only whe e he ee co e densi y was highe han 60%. Howe e , when o e laying he IFAP 2018 da ase , as his da ase is p io i ized o e 53 COS, i ends up emo ing some Holm Oak a eas in a o o ag icul u al g asslands. The inal classi ica ion is a combina ion o Holm Oak and ag icul u al g asslands o he polygon. Figu e 26 COS 2018 (OBJECTID: 382944) polygon p e and pos -p ocessed compa ison o p edic ions o he class Holm Oak. Scale: 1:10.000 Class inco ec ly p edic ed on he map In he con usion ma ix, one o he classes wi h he lowes accu acy is O cha ds; in his example Whe e “O cha ds” (PA=35) is equal o “Na u al g assland” i is possible o app ecia e a classi ica ion issue ela ed o plan a ions. This class was assigned 201 imes by commission e o o na u al g asslands. One o he di icul ies in classi ying his class is ha i con ains 17 ypes o ees anging om ci us o almonds. Also, o cha ds a e usually plan ed in 2m sepa a ion, meaning ha he su ace e lec ance alues cap u ed co espond o a mix u e o he c op and soil as he MMU o he Sen inel-2 is 10m. In he i s squa e o Figu e 27, i is possible o isualize ha hal o he polygon con ains mo e ege a ion in a ows a he soil le el, his can co espond o c eeping ege a ion (i.e., close o he g ound). The inal classi ica ion dic a ed ha he polygon is conside ed as oli e ees and na u al g assland; no hing was classi ied in o he class hey genuinely belong. 54 Figu e 27 IFAP 2018 (OSAID: 4410598) polygon p e and pos -p ocessed compa ison o p edic ions o he class O cha ds. Scale: 1:6000 Tile ansi ions The aim was o classi y he biogeog aphic egion co esponding o he s a a 214 in Figu e 2. Though his a ea was co e ed by wo sepa a e Sen inel-2 iles, which could en ain discon inui ies in hei limi s [3]. Ye , he app oach was o au oma ically ex ac he samples o he whole s udy a ea and e ie ing he ea u es om bo h iles. Hence, he classi ie con ained he in o ma ion o he o e all s a a, allowing adjacen pixels in he bo de s o be assigned in he same class as i can be app ecia ed in Figu e 28 ha po ays he Land Co e and C op Type map in h ee loca ions (a). The i s loca ion (b) co esponds o he Tejo es ua y, p ese ing he con inui y o he i e and we lands. The o he wo loca ions ha a e in he bo de o San a ém and Sé ubal (c) and nea É o a (d) kep he con inui y o classes such as o es s and ice ields along he iles. 55 Figu e 28 (a) Land Co e and C op Type in as e o ma wi h h ee loca ions on he bo de o he Sen inel-2 iles 29SND (uppe ) and 29SNC (lowe ), (b) Tejo es ua y (c) bo de o San a ém and Sé ubal, (d) loca ion nea É o a. 56 5 CONCLUSIONS Up- o-da e land co e and c op ype in o ma ion play an essen ial ole in comme cial and en i onmen al moni o ing and planning. Fo i s upda ing, hey ha e bene i ed om emo e sensing image y a a na ional, con inen al and global le el. Howe e , many challenges emain o p oduce accu a e and imely land co e and c op ype maps. This hesis ocused on he use o in a-annual composi es o Sen inel-2, supe ised classi ica ion wi h andom o es , and au oma ic sample ex ac ion based on a p e-p ocessing se o ules. The o e all accu acy o 76% was achie ed o 31 land co e and c op ype classes. The use o mon hly composi es o L2 Sen inel-2 da a allowed ha ing cloud- ee da a in con as o single acquisi ions ha ha e missing alues due o cloud co e o cloud shadows. Also, since he classi ica ion is done a he pixel le el, ha ing missing da a would a ec he spec al signa u e ex ac ion and incomple ely cha ac e ize he classes wi h missing da a. Likewise, he composi es ep esen an excellen oppo uni y o dimensionali y educ ion as he numbe o ea u es would co espond o 10 bands pe mon h. Fo single acquisi ions, each acquisi ion would con ain 10 bands, and he numbe o ea u es would inc ease based on acquisi ions du ing he pe iod. The Random Fo es classi ie equi ed ew hype pa ame e s o une as opposed o o he classi ie s and p o ed o be compu a ionally e icien as i was possible o pa allelize i (mul i-co e p ocessing) o classi y he whole a ea. Also, i allowed ex ac ing he mos impo an ea u es du ing he classi ica ion. As expec ed, he mos ele an ea u es om he ime se ies co espond o he sp ing and summe mon hs and he bands on he Red Edge (b5, b6, b7), NIR (b8a) and SWIR (b11 and b12). The inclusion o spec al indices sligh ly imp o ed he accu acy bu was no a p edominan a iable. One o he pu poses o his esea ch was o es i a p e-de ined se o ules could emo e possible sou ces o misclassi ica ion, allowing us o ex ac samples o aining and es ing au oma ically. This would pe mi he classi ie o adequa ely cha ac e ize he spec al signa u e o each class and make an accu a e p edic ion. The da a sou ces (IFAP 2018 and COS 2018) hemsel es a e a p oduc o isual in e p e a ion o high- esolu ion image y; in he case o he LPIS, he yea ly upda e o he p oduc and he MMU o a pa cel allowed o cha ac e ize he ypes o c ops. Ne e heless, he ag icul u al g assland class co e age was o e -op imis ic in his da ase and some imes would mask ou o es a eas causing se e al mix-ups wi hin classes. 57 Rega ding he il e ing ules, he applica ion o he bu ned mask allowed o emo e om he da ase he a eas ha igni ed wi h wild i es. Though, in some cases, he bu n mask includes build-up a eas and wa e leading o con usion wi hin classes. The e o e, he bu n mask can bene i om a se o p e-p ocessing ules be o e sample ex ac ion, such as build- up, canno be pa o he bu ned mask and nei he wa e . The ollowing il e was he NDVI ale s; hese we e p oduced o he yea 2015-2018 om Landsa 8 images a 30m esolu ion, con aining an omission e o o 33% [51]. This implica es ha some changes a e no de ec ed. Also, he di e ence in pixel size be ween sa elli es educes he p ecision in he de ec ion o hese a eas; he same app oach ye implemen ing Sen inel-2 image y migh imp o e he iden i ica ion o clea -cu s o his s udy. Las ly, he HRL ules educed he numbe o samples a ailable pe class d ama ically. By emo ing many o he o es pixels, he spec al signa u e was no p ecisely cha ac e ized, and he model could no classi y he whole a ea accu a ely. Fo es in he Po uguese landscape is no as dense as ees a e spa sely dis ibu ed in space; he e o e, a dec ease in he ee co e densi y is encou aged o de ec ing he o es ypes. 5.1 Limi a ions and Recommenda ions The limi a ions o he s udy can be summa ized in he equi emen o an independen e i ica ion da ase and conside a ions o inc easing accu acy; hen, ecommenda ions a e p o ided o po en ial enhancemen o he me hodology. Fo he IFAP 2018, he a ailabili y o an independen da ase ensu ed a p ope e i ica ion o he classi ica ion o he ag icul u al classes. Howe e , in he case o land co e , he una ailabili y o a e i ied es ing da ase o COS2018 aised h ee main issues. Fi s , he es ing da ase unde wen he same p e-p ocessing as he aining da ase ; in consequence, many o he po en ial es ing pixels we e emo ed. Also, he aining and es ing polygons we e no spa ially disjoin , meaning ha pixels coming om he same polygon can be used as aining and es ing. This can induce posi i e alues in he accu acies as he spec al signa u e can be simila o bo h da ase s [3]. The inal conce n o no ha ing a alida ed da ase is ha he model can co ec ly p edic a class; howe e , i he class is inco ec ly labeled, he class accu acy dec eases. The adi ional me hod o alida ion is isual inspec ion. Howe e , his equi es knowledge o he landscape o iden i y he di e en classes co ec ly. In e ms o inc easing accu acy, his s udy can bene i om a educ ion in he numbe o classes. I he ocus is in c opland a eas, a bina y c opland mask [23] can allow 58 es ic ing he classi ica ion; his app oach has been implemen ed in ope a ional sys ems [3]. Fo land co e classes, he use o he mos de ailed le el in he hie a chical nomencla u e (i.e., le el 4 in he nomencla u e in sec ion 7.3) is use ul o he pu pose o ege a ion cha ac e iza ion o i e modeling. Howe e , his de ailed nomencla u e adds noise o he classi ica ion. A educ ion om 16 o he 11 classes (i.e., le el 2 in he nomencla u e in sec ion 7.3) p o ides eliable in o ma ion and can inc ease he classi ica ion accu acy. Fo he elabo a ion o he inal map, in o de o educe he sal - and-peppe e ec , i is possible he implemen a ion objec segmen a ion algo i hms whe e pixels can be agg ega ed in homogeneous bounda ies [17], [66] howe e his app oach equi es mo e complex analysis. A las , Random Fo es has been used o classi y hype spec al da ase s [39], demons a ing i s capabili ies o deal wi h an inc easing numbe o dimensions. Mo e ea u es can be added o his model o imp o e he accu acy, some o hem a e he spec al, empo al me ics o desc ibe he dis ibu ion o a spec al band o index o e a speci ic pe iod [67] and ex u e me ics [19]. 59 6 BIBLIOGRAPHIC REFERENCES [1] GAETANO, R. e al. A Two-B anch CNN A chi ec u e o Land Co e Classi ica ion o PAN and MS Image y. Remo e Sensing, 2018, 10(11), 1746. Re ie ed om: h ps://doi.o g/10.3390/ s10111746 [2] LATHAM, J. e al. Global Land Co e SHARE (GLC-SHARE) - da abase Be a-Release Ve sion 1.0-2014. FAO: Rome, 2014. [3] INGLADA, J. e al. Ope a ional High Resolu ion Land Co e Map P oduc ion a he Coun y Scale Using Sa elli e Image Time Se ies. Remo e Sensing, 2017, 9(1), 95. Re ie ed om: h ps://doi.o g/10.3390/ s9010095 [4] WULDER, M. A. e al. Land co e 2.0. In e na ional Jou nal o Remo e Sensing, 2018, 39(12), 4254–4284. Re ei ed om: h ps://doi.o g/10.1080/01431161.2018.1452075. [5] HERMOSILLA, T. e al. Dis u bance-In o med Annual Land Co e Classi ica ion Maps o Canada’s Fo es ed Ecosys ems o a 29-Yea Landsa Time Se ies. Canadian Jou nal o Remo e Sensing, 2018, 44(1), 67–87. Re ie ed om: h ps://doi.o g/10.1080/07038992.2018.1437719 [6] DEFOURNY, P. e al. Nea eal- ime ag icul u e moni o ing a na ional scale a pa cel esolu ion: Pe o mance assessmen o he Sen2-Ag i au oma ed sys em in a ious c opping sys ems a ound he wo ld. Remo e Sensing o En i onmen , 2019, 221, 551–568. Re ie ed om: h ps://doi.o g/10.1016/j. se.2018.11.007 [7] DIREÇÃO-GERAL DO TERRITÓRIO (DGT). Especi icações écnicas da Ca a de Uso e Ocupação do Solo (COS) de Po ugal Con inen al pa a 2018. Rela ó io Técnico. 2019. [8] TURCO, M. e al. Clima e d i e s o he 2017 de as a ing i es in Po ugal. Scien i ic Repo s, 2019, 9, 13886. Re ie ed om: h ps://doi.o g/10.1038/s41598-019- 50281-2 [9] WISNER, B.; ADAMS, J. En i onmen al heal h in eme gencies and disas e s. Wo ld Heal h O ganiza ion, 2002. [10] BODROZIC, L.; MARASOVIC, J.; STIPANICEV, D. Fi e modeling in o es i e managemen . CEEPUS Sp ing School, 2005. [11] JAHDI, R. e al. E alua ing i e modelling sys ems in ecen wild i es o he Goles an Na ional Pa k, I an. Fo es y: An In e na ional Jou nal o Fo es Resea ch, 2016, 89(2), 136–149. Re ie ed om: h ps://doi.o g/10.1093/ o es y/cp 045 [12] UNINOVA; NOVAIMS; DGT. IPSTERS (IPSen inel Te es ial Enhanced Recogni ion 60 Sys em). 2019. URL: h ps://www.ca3-unino a.o g/p ojec _ips e s [Accessed: 26- Jan-2020]. [13] UNINOVA; NOVA IMS; UA; ISA; DGT; IT. FoRESTER (Da a usion o senso ne wo ks and i e sp ead modelling o decision suppo in o es i e supp ession). 2019. URL: h ps://www.ca3-unino a.o g/p ojec _ o es e [Accessed: 26-Jan-2020]. [14] LÜDTKE, D. Land co e mapping wi h andom o es using in a-annual sen inel 2 da a in cen al Po ugal: a compa a i e analysis. NIMS - MSc Disse a ions Geospa ial Technologies (E asmus-Mundus), 2018. URL: h ps:// un.unl.p /handle/10362/33648 [Accessed: 15-Feb-2020]. [15] BLANCO-MARTÍNEZ, W.A. In a-Annual land co e mapping: Au oma ic aining sample ex ac ion om old maps o in a-annual land co e mapping a cen al o Po ugal. NIMS - MSc Disse a ions Geospa ial Technologies (E asmus-Mundus), 2019. URL: h ps:// un.unl.p /handle/10362/63946 [Accessed: 15-Feb-2020]. [16] PARIS, C.; BRUZZONE, L.; FERNÁNDEZ-PRIETO, D. A No el App oach o he Unsupe ised Upda e o Land-Co e Maps by Classi ica ion o Time Se ies o Mul ispec al Images. IEEE T ansac ions on Geoscience and Remo e Sensing, 2019, 57(7). 4259-4277. Re ie ed om: h ps://doi.o g/10.1109/TGRS.2018.2890404 [17] COSTA, H. e al. Land Co e Mapping om Remo ely Sensed and Auxilia y Da a o Ha monized O icial S a is ics. In e na ional Jou nal o Geo-In o ma ion, 2018, 7(4), 157. Re ei ed om: h ps://doi.o g/10.3390/ijgi7040157. [18] STEHMAN, S. V.; FOODY, G. M. Key issues in igo ous accu acy assessmen o land co e p oduc s. Remo e Sensing o En i onmen , 2019, 231, 111199. Re ie ed om: h ps://doi.o g/10.1016/j. se.2019.05.018 [19] KHATAMI, R.; MOUNTRAKIS, G.; STEHMAN, S. V. A me a-analysis o emo e sensing esea ch on supe ised pixel-based land-co e image classi ica ion p ocesses: Gene al guidelines o p ac i ione s and u u e esea ch. Remo e Sensing o En i onmen , 2016, 177, 89–100. Re ie ed om: h ps://doi.o g/10.1016/j. se.2016.02.028 [20] FRITZ, S. e al. A compa ison o global ag icul u al moni o ing sys ems and cu en gaps. Ag icul u al Sys ems, 2019, 168, 258–272. Re ie ed om: h ps://doi.o g/10.1016/j.agsy.2018.05.010 [21] EUROPEAN SPACE AGENCY (ESA). Sen2-Ag i. 2018. URL: h p://www.esa- sen2ag i.o g/ope a ional-sys em/sys em-desc ip ion/ [Accessed: 08-Sep-2019]. [22] INGLADA, J. e al. Assessmen o an Ope a ional Sys em o C op Type Map P oduc ion Using High Tempo al and Spa ial Resolu ion Sa elli e Op ical Image y. Remo e Sensing, 61 2015, 7(9), 12356–12379. Re ie ed om: h ps://doi.o g/10.3390/ s70912356 [23] VALERO S. e al. P oduc ion o a Dynamic C opland Mask by P ocessing Remo e Sensing Image Se ies a High Tempo al and Spa ial Resolu ions. Remo e Sensing, 2016, 8(1), 55. Re ie ed om: h ps://doi.o g/10.3390/ s8010055 [24] GRIFFITHS, P.; NENDEL, C.; HOSTERT, P. In a-annual e lec ance composi es om Sen inel-2 and Landsa o na ional-scale c op and land co e mapping. Remo e Sensing o En i onmen , 2019, 220, 135–151. Re ie ed om: h ps://doi.o g/10.1016/j. se.2018.10.031 [25] EUROPEAN COURT OF AUDITORS (ECA). The Land Pa cel Iden i ica ion Sys em A use ul ool o de e mine he eligibili y o ag icul u al land – bu i s managemen could be u he imp o ed. Special epo no. 25, 2016. URL: h ps://www.eca.eu opa.eu/en/Pages/DocI em.aspx?did=38180 [Accessed: 08-Sep-2019]. [26] SCHMEDTMANN, J.; CAMPAGNOLO, M. L. Reliable c op iden i ica ion wi h sa elli e image y in he con ex o Common Ag icul u e Policy subsidy con ol. Remo e Sensing, 2015, 7(7), 9325–9346. Re ie ed om: h ps://doi.o g/10.3390/ s70709325 [27] GÓMEZ, C.; WHITE, J. C.; WULDER, M. A. Op ical emo ely sensed ime se ies da a o land co e classi ica ion: A e iew. ISPRS Jou nal o Pho og amme y and Remo e Sensing, 2016, 116, 55–72. Re ie ed om: h ps://doi.o g/10.1016/j.isp sjp s.2016.03.008 [28] SONG, Q. e al. In-Season C op Mapping wi h GF-1/WFV Da aby Combining Objec - Based Image Analysisand Random Fo es . Remo e Sensing, 2017, 9(11), 1184. Re ei ed om: h ps://doi.o g/10.3390/ s9111184. [29] ZHU, Z.; WOODCOCK, C. E. Con inuous change de ec ion and classi ica ion o land co e using all a ailable Landsa da a. Remo e Sensing o En i onmen , 2014, 144, 152–171. Re ei ed om: h ps://doi.o g/10.1016/j. se.2014.01.011. [30] BAETENS, L.; DESJARDINS, C.; HAGOLLE, O. Valida ion o Cope nicus Sen inel- 2 Cloud Masks Ob ained om MAJA, Sen2Co , and FMask P ocesso s Using Re e ence Cloud Masks Gene a ed wi h a Supe ised Ac i e Lea ning P ocedu e. Remo e Sensing, 2019 11(4), 433. Re ie ed om: h ps://doi.o g/10.3390/ s11040433 [31] HERMOSILLA, T. e al. An in eg a ed Landsa ime se ies p o ocol o change de ec ion and gene a ion o annual gap- ee su ace e lec ance composi es. Remo e Sensing o En i onmen , 2015, 158, 220–234. Re ie ed om: h ps://doi.o g/10.1016/j. se.2014.11.005 [32] WHITE, J. C. e al. Pixel-Based Image Composi ing o La ge-A ea Dense Time 68 9.1.1.1 Wa e na u al and a i icial eshwa e su aces, oceans and su aces, and coas al lagoons and i e mou hs. COS 2018 (5111, 5121, 5122, 5123, 5124, 5125, 521, 522); 9999. Bu ned A eas a eas ha bu ned in 2018 and de ec ed by he ICNF. ICNF (2018) 69 7.3 RGB colo amp o he Land Co e and C op Type Classes The ollowing colo amp is a combina ion o he CLC 2018 RGB ha can be ound in he Eu opean En i onmen Agency (EEA) websi e and he C opScape RGB a ailable on he websi e o he Uni ed S a es Depa men o Ag icul u e - Na ional Ag icul u al S a is ics Se ice (USDA-NASS). LV1 LV2 LV3 LV4 RGB CODE 1.Build up (BUI) 1.1 Build up (BUI) 1.1.1 Build up (BUI) 1.1.1.1 Build up (BUI) 255-000-00 1111 2.Ag icul u e (AGR) 2.1 Tempo a y c ops (TCO) 2.1.1 Rain ed empo a y c ops (RAI) 2.1.1.1 Whea (WHE) 168-112-0 2111 2.1.1.2 Ba ley (BAR) 226-0-127 2112 2.1.1.3 Oa meal (OAT) 161-88-137 2113 2.1.1.4 Ryeg ass (RYE) 174-1-126 2114 2.1.1.5 Lupin (LUP) 255-255-168 2115 2.1.2 I iga ed empo a y c ops (IRR) 2.1.2.1 Maize (MAI) 255-212-0 2121 2.1.2.2 So ghum (SOR) 255-158-15 2122 2.1.2.3 Rice (RIC) 0-38-115 2123 2.1.2.4 Toma o (TOM) 255-255-0 2124 2.1.2.5 Po a o (POT) 115-38-0 2125 2.2 Pe manen c ops (PCO) 2.2.1 Vineya ds (VIN) 2.2.1.1 Vineya ds (VIN) 230-128-000 2211 2.2.2 O cha ds (ORC) 2.2.2.1 O cha ds (ORC) 242-166-077 2221 2.2.3 Oli e T ees (OLI) 2.2.3.1 Oli e T ees (OLI) 230-166-000 2231 3.G assland (GRA) 3.1 G assland (GRA) 3.1.1 Ag icul u al g assland (AGR) 3.1.1.1 Ag icul u al g assland (AGR) 255-230-077 3111 3.1.2 Na u al g assland (NAT) 3.1.2.1 Na u al g assland (NAT) 230-230-000 3121 70 5.Fo es s (FOR) 5.1 B oadlea o es (BOF) 5.1.1 Co k oak o es (COR) 5.1.1.1 Co k oak o es (COR) 128-255-000 5111 5.1.2 Holm oak o es (HOL) 5.1.2.1 Holm oak o es (HOL) 000-166-000 5121 5.1.3 Eucalyp us o es (EUC) 5.1.3.1 Eucalyp us o es (EUC) 077-255-000 5131 5.1.4 O he b oadlea o es (OBL) 5.1.4.1 O he b oadlea o es (OBL) 077-200-0 5141 5.2 Coni e ous o es (COF) 5.2.1 Ma i ime pine o es (MAR) 5.2.1.1 Closed Ma i ime pine o es (CMA) 166-255-128 5211 5.2.1.2 Open ma i ime pine o es (OMA) 204-242-077 5212 5.2.2 S one pine o es (STO) 5.2.2.1 S one pine o es (STO) 166-230-077 5221 5.2.3 O he coni e ous o es (OCO) 5.2.3.1 O he coni e ous o es (OCO) 166-242-000 5231 6. Sh ubland (SHR) 6.1 Sh ubland (SHR) 6.1.1 Sh ubland (SHR) 6.1.1.1 Sh ubland (SHR) 242-204-166 6111 7. Open spaces wi h li le o no ege a ion (OPE) 7.1 Open spaces wi h li le o no ege a ion (OPE) 7.1.1 Ba esoil (BSL) 7.1.1.1 Ba esoil (BSL) 230-230-230 7111 7.1.2 Ba e Rock (BRK) 7.1.2.1 Ba e Rock (BRK) 204-204-204 7121 7.1.3 Spa se ege a ion (SPA) 7.1.3.1 Spa se ege a ion (SPA) 204-255-204 7131 8.We lands (WET) 8.1 We lands (WET) 8.1.1 We lands (WET) 8.1.1.1 We lands (WET) 166-166-255 8111 9. Wa e (WAT) 9.1 Wa e (WAT) 9.1.1 Wa e (WAT) 9.1.1.1 Wa e (WAT) 000-204-242 9111 9. Bu ned a eas (BUR) 9.9 Bu ned a eas (BUR) 9.9.9 Bu ned a eas (BUR) 9.9.9.9 Bu ned a eas (BUR) 000-000-000 9999 71 7.4 C op Calenda This co esponds o he c op calenda o he moni o ed IFAP pa cels, each calenda is de ined based on he Po uguese ag icul u al cycle (Oc obe o Sep embe ). No e: Fo o cha ds, i co esponds o an a e age o 17 di e en ypes o ees. Pe iod OCT NOV DEC JAN FEB MAR APR MAY JUN JUL AUG SEP Tempo a y Rain ed (au umn/win e ) Whea Ba ley Flooding Oa meal Seed Ryeg ass Ge mina ion Lupin Tille ing I iga ed (sp ing/summe ) Maize Flowe ing So ghum F ui Rice Ripening Toma o Ha es Po a o Pe manen Vineya ds P uning O cha ds Ha es Oli e T ees 72 7.5 Con usion ma ix LV4 BUI WHE BAR OAT RYE LUP MAI SOR RIC TOM POT VIN ORC OLI AGR NAT COR HOL EUC OBL CMA OMA STO OCO SHR BSL BRK SPA WET WAT BUR BUI 880 1 0 0 0 0 0 0 0 0 0 15 6 9 7 18 1 0 0 2 0 2 0 0 7 25 0 5 2 0 20 WHE 1 667 10 223 10 15 0 0 0 0 3 0 1 1 6 13 0 0 0 0 0 0 0 0 0 1 0 26 0 0 23 BAR 0 226 464 0 100 0 0 27 0 0 0 1 0 3 1 7 0 0 0 0 0 0 0 0 21 0 0 6 0 0 0 OAT 0 47 27 492 65 129 0 0 1 0 0 6 0 16 207 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 9 RYE 0 2 0 80 619 75 0 68 0 0 3 3 10 3 32 26 0 0 0 0 0 0 0 0 0 0 0 2 0 0 77 LUP 4 0 0 122 198 552 0 0 0 0 108 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 16 MAI 0 0 0 0 2 0 990 1 0 3 4 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 SOR 0 0 2 18 8 6 0 637 2 36 0 126 16 73 46 24 1 0 0 0 0 0 0 0 0 0 0 0 0 0 5 RIC 0 0 0 0 0 0 10 5 985 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 TOM 0 0 0 0 0 0 0 2 0 998 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 POT 5 0 28 0 0 4 0 0 0 31 845 1 28 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 58 VIN 8 0 0 0 0 0 0 0 0 0 5 944 4 3 0 12 0 0 0 0 0 0 0 0 1 17 0 1 0 0 5 ORC 0 0 0 1 25 0 0 1 0 1 59 43 350 48 42 201 11 0 0 171 0 0 0 0 1 0 0 0 45 0 1 OLI 0 0 11 18 20 24 0 6 0 0 54 33 21 460 174 89 0 3 1 1 0 4 3 0 23 7 0 1 0 0 47 AGR 2 3 0 13 11 34 0 6 0 0 0 14 3 41 720 37 45 1 2 5 0 15 15 0 15 3 0 3 1 0 11 NAT 14 1 3 13 20 16 1 2 0 0 0 18 5 26 75 679 7 1 2 11 0 6 0 0 65 17 0 3 4 0 11 COR 0 0 0 0 0 2 0 0 0 0 0 10 4 3 19 2 700 37 32 92 30 18 10 0 19 1 0 0 2 6 13 HOL 0 0 0 0 0 0 0 0 0 0 0 0 0 0 5 0 29 1063 3 28 1 1 1 0 3 0 0 0 1 0 1 EUC 1 0 0 0 0 1 0 0 0 0 0 1 2 3 2 0 78 10 821 28 22 15 6 0 8 2 0 0 0 0 0 OBL 0 0 0 0 0 0 0 0 1 0 0 0 2 1 4 3 93 31 11 808 12 1 11 0 13 0 0 0 7 0 2 CMA 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2 1 58 5 28 19 751 69 60 0 3 0 0 0 1 1 2 OMA 6 0 0 0 0 2 0 0 0 0 0 0 0 2 12 0 16 3 5 2 113 781 25 0 21 5 0 5 0 0 2 STO 0 0 0 0 0 2 0 0 0 0 0 0 0 3 3 0 16 8 12 20 64 24 843 0 3 0 0 0 1 0 1 73 OCO 3 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2 3 1 4 9 1 20 50 0 0 0 0 0 0 0 SHR 23 1 0 4 2 4 0 0 0 0 0 9 8 16 47 42 18 3 2 28 3 25 1 0 634 69 0 40 1 2 18 BSL 36 0 0 1 0 3 0 1 0 0 1 11 7 7 14 14 4 0 0 4 0 14 0 0 22 804 2 27 5 9 14 BRK 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 60 4 0 0 0 SPA 11 0 0 0 0 0 0 0 0 1 0 1 0 0 8 3 5 0 0 1 0 5 1 0 12 8 1 942 0 0 1 WET 7 0 0 0 0 1 0 0 0 0 0 2 1 2 4 7 3 1 1 10 0 0 0 0 4 2 0 0 944 11 0 WAT 1 0 0 1 0 1 0 0 0 0 0 0 0 0 3 6 0 1 0 0 0 0 0 0 2 11 0 1 24 948 1 BUR 17 0 0 6 7 13 0 3 0 0 0 8 9 37 54 35 10 0 7 8 1 14 4 0 24 7 0 4 7 2 723 LV4 BUI WHE BAR OAT RYE LUP MAI SOR RIC TOM POT VIN ORC OLI AGR NAT COR HOL EUC OBL CMA OMA STO OCO SHR BSL BRK SPA WET WAT BUR 74 LANDCOVER AND CROP TYPE CLASSIFICATION Using in a -annual imes se ies o sen inel-2 and machine lea ning a cen al Po ugal 2020 I zá Alejand a He nández Sequei a Guia pa a a o ma ação de eses Ve são 4.0 Janei o 2006