Mapping Cropland Abandonment in the Aral Sea Basin with MODIS Time Series

24 Pages·2017·3.97 MB·English

remote sensing
Article
Mapping Cropland Abandonment in the Aral Sea
Basin with MODIS Time Series
FabianLöw1,2,* ID,AlexanderV.Prishchepov3,4,5,FrançoisWaldner6,7,OlenaDubovyk8,
AkmalAkramkhanov1 ID,ChandrashekharBiradar1andJohnP.A.Lamers8
1 InternationalCentreforAgriculturalResearchinDryAreas(ICARDA),11431Cairo,Egypt;
[email protected](A.A.);[email protected](C.B.)
2 MapTailorGeospatialConsulting,53113Bonn,Germany
3 DepartmentofGeosciencesandNaturalResourceManagement(IGN),UniversityofCopenhagen,
1165København,Denmark;[email protected]
4 LeibnizInstituteofAgriculturalDevelopmentinTransitionEconomies(IAMO),
06120Halle(Saale),Germany
5 InstituteofEnvironmentalSciences,KazanFederalUniversity,420008Kazan,Russia
6 EarthandLifeInstitute-Environment,UniversitéCatholiquedeLouvain,2CroixduSud,
1348Louvain-la-Neuve,Belgium;[email protected]
7 CSIROAgriculture&Food,306CarmodyRoad,StLucia,QLD4067,Australia
8 DepartmentofGeography,Rheinische-Friedrich-Wilhelms-Universität,53113Bonn,Germany;
[email protected](O.D.);[email protected](J.P.A.L.)
* Correspondence:[email protected]
Received:18November2017;Accepted:16January2018;Published:23January2018
Abstract: Croplandabandonmentisgloballywidespreadandhasstrongrepercussionsforregional
foodsecurityandtheenvironment. Statisticssuggestthatoneofthehotspotsofabandonedcropland
islocatedinthedrylandsoftheAralSeaBasin(ASB),whichcoverspartsofpost-SovietCentralAsia,
AfghanistanandIran. Todate,theexactspatialandtemporalextentsofabandonedcroplandremain
unclear,whichhampersland-useplanning. Abandonedlandisapotentiallyvaluableresourcefor
alternativelanduses. Here,wemappedtheabandonedcroplandinthedrylandsoftheASBwitha
timeseriesoftheNormalizedDifferenceVegetationIndex(NDVI)fromtheModerateResolution
ImagingSpectroradiometer(MODIS)from2003–2016. Toovercometherestrictedabilityofasingle
classifiertoaccuratelymapland-useclassesacrosslargeareasandagro-environmentalgradients,
“stratum-specific”classifierswerecalibratedandclassificationresultswerefusedbasedonalocally
weighteddecisionfusionapproach. Next,theagro-ecologicalsuitabilityofabandonedcroplandareas
was evaluated. The stratum-specific classification approach yielded an overall accuracy of 0.879,
whichwassignificantlymoreaccurate(p<0.05)thana“global”classificationwithoutstratification,
whichhadanaccuracyof0.811. In2016,theclassificationresultsshowedthat13%(1.15Mha)of
theobservedirrigatedcroplandintheASBwasidle(abandoned). Croplandabandonmentoccurred
mostly in the Amudarya and Syrdarya downstream regions and was associated with degraded
landandareaspronetowaterstress. Despitethealmosttwofoldpopulationgrowthandincreasing
food demand in the ASB area from 1990 to 2016, abandoned cropland was also located in areas
withhighsuitabilityforfarming. Themapofabandonedcroplandareasprovidesanovelbasisfor
assessingthecausesleadingtoabandonedcroplandintheASB.Thiscontributestoassessingthe
suitabilityofabandonedcroplandforfoodorbioenergyproduction,carbonstorage,orassessingthe
environmentaltrade-offsandsocialconstraintsofrecultivation.
Keywords: abandonedcropland;AralSeaBasin;changedetection;landuse;decisionfusion;MODIS
RemoteSens.2018,10,159;doi:10.3390/rs10020159 www.mdpi.com/journal/remotesensing
RemoteSens.2018,10,159 2of24
1. Introduction
Agricultural production must sustainably increase to meet the growing food demand while
preservingecosystemservicesandbiodiversity[1,2]. Giventhe(global)limitsofcroplandexpansion,
approximately80%ofthisincreasemustcomefromintensificationsuchastheexpansionofirrigated
cropproduction[3–5],whichalreadycontributestoapproximately44%ofglobalcropproduction[1].
OneoftheregionswhereagriculturallandsarescarceisthetransboundaryAralSeaBasin(ASB),
whichcoverspartsofpost-SovietCentralAsia(Uzbekistan,Kazakhstan,Kyrgyzstan,Tajikistanand
Turkmenistan), Afghanistan and Iran. This region experienced precipitous population growth in
thementionedcountries,increasingfrom61millionpeoplein1990to104millionby2016;therural
populationisstronglydependentonthedomesticagriculturalproduction[6]. Despiteincreasingfood
demand,officialstatisticsandexemplarycasestudieswithsatelliteimagerysuggestthatagricultural
landabandonmentiscommoninthisregionandmostlikelyoccursondegraded,irrigatedagricultural
lands[7–10]. Duetoeconomicrestructuring,agriculturallandabandonmentprimarilyoccursinthe
downstreamAmudaryaandSyrdaryaRiverbasinsresemblingSovietland-uselegacies. Additionally,
large tracts of cropland in Afghanistan that partly belong to the ASB were left fallow, for reasons
includingconflictsandwar[11]. Currently,theloomingscarcityofwaterresourcesandongoingland
degradation[12,13]mayfurtherenhanceabandonment,causingadversesocio-economicconsequences,
includingreducedincomeorincreasedfoodinsecurity[14,15].
A better understanding of the spatial and temporal patterns of abandonment and idle land
production potential is important to better assess the drivers of land-use change for developing
plausible land-use policies [16–18] and assess the impacts on the carbon cycle [19] as well as the
trade-offsbetweentherecultivationortheprovisionofecosystemservices[20]. Arevitalizationof
currentlyabandonedlandcouldbecomeparticularlyplausibleinareaswheresignificantinvestments
were made during the Soviet era to establish an irrigation and drainage infrastructure [21,22].
The potential for recultivation depends, however, on the biophysical properties of soils and the
landsuitabilityitself[23,24].
The main problem is that agricultural statistics for the ASB are often outdated or of doubtful
quality and little knowledge exists about the spatial extent of agricultural land abandonment in
theASB.Remotesensingisawell-knownalternativetoassesslarge-scaleland-usechange. Much
progress has been made in mapping land-use/land-cover changes (LULCCs) in drylands using
geographicinformationsystems(GIS)andremotesensing,suchaswith30-mmultiannualimagery
fromLandsat[25,26],high-resolutionRapidEyeorSentinel[7,27,28]and250-mModerateResolution
ImagingSpectroradiometer(MODIS)data[29,30]. MODISdatafitwellforassessingandmapping
LULCC and crop types [31] or land abandonment over large regions in a regular manner [32–35].
Additionally, since fallow periods can be part of the typical crop rotation cycle, it is important to
assessseveralconsecutiveyearstodeterminewhetherafieldhaseffectivelybeenabandonedorif
itisawaitingfutureuse[7]. Usingregular,consecutiveimagetimeseriessuchasthoseprovidedby
MODIShelpsavoidmisclassificationoftemporarilyfallow-croprotation[7,29].
Machine-learningclassifierswerefoundtobeparticularlyusefultoovercomethecomplexity
related to the accurate separation of spectrally similar classes, such as abandoned cropland and
cultivated cropland in drylands [36], particularly at the early stages of abandonment [7,29,33,37].
Vegetation recovery on abandoned irrigated fields in the drylands of the ASB generally follows a
certainpathway. Itstartswiththerecoveryofannualandmultiyearherbaceousspecies[8,38,39]and
gradually,perennialwoodyspeciessuchasshrubs(e.g.,blacksaxaul)andsometrees[7]canestablish.
Dependingontimeandhydrologicalandsoilproperties,bareareaswithoutanyvegetationchange
couldalsobeobserved[8,24]. Abandoned, formerlyirrigatedcroplandindrylandsmaytherefore
represent a set of multimodal distributions of reflectance for different wavelengths recorded with
opticalsatelliteimagery. Apreviousstudypointedtotheneedtohavevariousinputdatasetsand
non-parametricmachinelearningalgorithmstomapabandonedcropland[29].
RemoteSens.2018,10,159 3of24
Todate, LULCCapproacheshavereliedprimarilyonsingleclassificationmethods; thus, low
accuracies were often found for LULCC in drylands [36] due to difficulties in mapping spectrally
complexclassesacrosslargeareaswithdifferentagro-environmentalsettings[40]. Thephysiography
andtypeofplantsuccessioninabandonedareasposechallengestoaccuratelymappingabandoned
croplandacrosslargeterritoriesovertime[7,29,30].Croplandabandonmentcanalsobeassociatedwith
bothnegative(degradationofvegetation)andpositive(vegetationrecovery)vegetationtrends[41,42].
Thus,despitethegeneralsuitabilityof“global”methodsforlandcovermapping[43,44],theyhaveless
accurateperformancesthanlocallycalibratedmodelsforlandcovermapping[45–48]andcapturing
accuratelyalldivergingtrajectoriesofland-coverchange. Similarly,croprotationpracticesmaydiffer
fromcountrytocountry,reflectingdifferentland-usepolicies,regionalfoodsecurityprogramsand
marketconditions. Itmayposeanadditionalchallengeinthegeneralizationofspectralfeaturesfora
singleclassificationacrossalargearea[36]. Atthesametime,afusionofmachine-learningclassifiers
basedonstrata(e.g.,soiltypes,landforms,administrativeboundariesandfurtherstratum-specific
classifiers)mayboostclassificationaccuracy,particularlywhereoneclassifierisnotabletoaccurately
separateclasses[36]. However,mapscreatedwithdifferentclassificationmethodsbutwithsimilar
accuraciesmayyieldspatialdisagreementofclassifiedpatterns,i.e.,errorsarerarelyequallydistributed
inamap[49–51]. Thus,amajorchallengeformappinglarge-scale,abandonedcroplandistoachieve
spatialcontinuityandconsistencyinthefinalmap.
The overarching objective of this study was to develop a method that can map the various
trajectories (spectral signatures) of cropland abandonment and to create the first map of cropland
abandonmentacrosstheASB.Specifically,we(i)testedwhetheradecisionfusionyieldedastatistically
significantdifference(p<0.05)comparedtosingleclassificationmethods,(ii)allocatedthehotspots
ofabandonedcroplandandstablecultivatedcroplandsand(iii)relatedthepatternsofabandoned
cropland to crop suitability. To achieve our goals, we combined the principle of random feature
selectionwithstratifiedclassifiersinanovelfashiontocreateaspatiallyconsistentmapofcropland
abandonmentintheASB.Westratifiedthestudyregionbasedontheadministrativeboundariesand
stratum-specificclassifierswerecalibratedtomapabandonedcroplandduringthe2003–2016period
basedonMODISNormalizedDifferenceVegetationIndex(NDVI)timeseries.
2. MaterialsandMethods
2.1. StudyArea
ThestudyareacoveredtheagriculturalareasoftheASB,avasttransboundaryriverbasinatthe
heartoftheEurasiancontinent[52]. Itspreadsover1.76millionkm2andencompassesthesouthern
partofKazakhstan,Turkmenistan,Uzbekistan,Kyrgyzstan,TajikistanandsmallpartsofAfghanistan
andIranintheTedzhen/MurghabBasin. Allofthesecountries,exceptAfghanistanandIran,were
oncepartoftheformerSovietUnion(CentralAsia)[53]. Theclimateintheirrigatedregionsofthe
ASB is mostly dry-arid continental with 100–250 mm of precipitation per year, which mainly falls
duringthewinter(December-February). Precipitationinthemountainscanexceed1,000mm. Because
of the aridity, agriculture in the study area is fully dependent on a dense irrigation and drainage
network[13], whichwasextensivelydevelopedduringtheSovietera[54]. Therefore, agricultural
landispredominantlylocateddownstreamoftheAmudaryaandSyrdaryaRivers. Morethanhalfthe
meanannualrunoffintheASB,whichisapproximately114km3,isgeneratedinTajikistanandalmost
one-quarterisgeneratedinKyrgyzstan[52]. Alargeshareofthefreshwaterintheseriversisfedinto
irrigationsystemsof~8.1–8.5Mhaofcropland[55,56]. Today,irrigatedproductionisdominatedby
cottonandwheatproduction[57]inareassuchasinTurkmenistan[58]andUzbekistan[59]. Riceisan
importantcropinareassuchassouthKazakhstan[7,60]. TheaveragefieldsizesthroughouttheASB
rangefrom2.19ha(Karakalpakstan)to6.74ha(Fergana)[40].
The post-Soviet countries within the ASB have undergone large transitions in their economy
andagriculturalproductionafterindependencein1991(Table1). DuringtheSovietera,thetypical
RemoteSens.2018,10,159 4of24
crop rotation in Uzbekistan, for instance, was three years of alfalfa followed by six years or more
of cotton. The crop rotation changed after the independence in 1991; the share of alfalfa and
especially cotton decreased in favor of winter wheat [61,62]. In southern Kazakhstan, the official
recommendations [38,61,63,64] may vary following the soil quality [59,65]. However, in southern
Kazakhstan,riceandalfalfaformadistinctrotationpattern,whichdiffersfromUzbekistan: aftertwo
yearsReomfotrei Sceensc. u20lt1i8v, 1a0t,i xo FnO,Rfi PeEldERs RaErVeItEeWm porarilysetasidefromcropproduction(mainlyrice)a4n odf 2a5l falfa
orotherlegumesarecultivatedforuptothreeconsecutiveyearsforsoilregeneration(Table2). Official
statisctuilctsiviantidoinc,a ftieedldas agrrea tdeumaplodrearcilliyn eseot faslaidned fruosme icnroipr rpigroadteudctifioenl d(msabiunltya rlsicoe)a adnedc arlefaaslfea oofr iortrhigera tion
legumes are cultivated for up to three consecutive years for soil regeneration (Table 2). Official
intensityintheremainingcultivatedfields[10].
statistics indicated a gradual decline of land use in irrigated fields but also a decrease of irrigation
2.2. Dinetfiennsitiitoyn ino fthCer orepmlaanidniAngba cnudltoinvmateendt fields [10].
2A.2b. aDnedfinointieodn cofr oCprolaplnadndi sAdbaenfidnoendmehnetr eas“croplandpermanentlywithoutmanagement,”i.e.,land
thathasnotbeenused(sownbutnotcropped)foraperiodlongerthanthefallowperiodspracticed
Abandoned cropland is defined here as “cropland permanently without management,” i.e. land
under the typical crop rotations in the region (Table 2, usually four or more years). After the
that has not been used (sown but not cropped) for a period longer than the fallow periods practiced
succeusnsdioern tohfea tgyrpicicuallt ucrraolpp rrootdatuiocntiso nin, sthhreu bresgaionnd (gTraabslsee s2,e nucsuroaallcyh feoduor nora bmaonrdeo yneeadrsfi).e lAdfst—er otrhefi elds
remasiuncecdessbioanre ofa angdricduelvtuoriadl porfodvuegcteiotant,i sohnrudbus eantdo gsraalsisneisz eanticornoa(cFhiegdu orne a1b)a.ndDouneedt ofiealdws—atoerr fisehldosr tage
(e.g.,rdemroauingehdt sb)aorre faanrdm deervso’idde ocfis vioegnest,aitniotenr dmuiett teod sfaalilnloizwatpioenr i(oFdigsumrea 1y).b Deuleo ntog ear wthataenr suhnodrteargteh (ee.cgu.,r rent
manadgroeumgehntst) porra fcatricmeesr[s2’ 6d]ebciustiosnusc, hinfiteerlmdsittdeod nfaoltlonwe cpeesrsiaordilsy mmaeye btet hloenggievre nthdane fiunnidtieorn thoef acubrarnednto ned.
Abanmdaonnaegdemfieenldt spraarceticchesa r[2a6c]t ebruizt esudcbhy fiealdtysp dioc anloat nndecreescsoarginlyi zmaebelte tvhee ggeivtaetni odnefsinuicticoens soifo anb,asntadrotninedg. with
Abandoned fields are characterized by a typical and recognizable vegetation succession, starting with
therecoveryofannualandmultiyearherbaceousspecies[8,38,39],perennialwoodyspeciessuchas
the recovery of annual and multiyear herbaceous species [8,38,39], perennial woody species such as
shrubs(e.g.,blacksaxaul)andsometrees[7]. Whenclassifyingafewconsecutiveyears, previous
shrubs (e.g., black saxaul) and some trees [7]. When classifying a few consecutive years, previous
studieshavedemonstratedthatintermittentfallowperiods,e.g.,aspartofagriculturalmanagement,
studies have demonstrated that intermittent fallow periods, e.g., as part of agricultural management,
orareasleftfallowduetowatershortage,canobscureland-usetrajectoriesofabandonedfieldsthatin
or areas left fallow due to water shortage, can obscure land-use trajectories of abandoned fields that
turnleadtomisclassifications[7,29,30].
in turn lead to misclassifications [7,29,30].

Figure 1. Drylands in the Aral Sea Basin and its irrigated agricultural land. The photographs from
Figure1.DrylandsintheAralSeaBasinanditsirrigatedagriculturalland.Thephotographsfrom2011
2011 highlight examples of abandoned fields: open (A) and dense (B) shrubland in Karakalpakstan,
highlightexamplesofabandonedfields:open(A)anddense(B)shrublandinKarakalpakstan,bare
bare soil (C) and dense shrubland mixed with herbaceous vegetation (D) in Kyzyl-Orda, Kazakhstan.
soil(C)anddenseshrublandmixedwithherbaceousvegetation(D)inKyzyl-Orda,Kazakhstan.
RemoteSens.2018,10,159 5of24
Table1.Summaryofthetransitionapproachesacrossthepost-SovietcountrieswithintheAralSeaBasin.
LegalAttitudeto Agriculture,Value ShareofRural
PotentialPrivate Allocation
Country PrivatizationStrategy* Transferability RelevantLegislation* Added(%GDP) Population
Ownershipafter1990* Strategy*
after1990* in2014** in2014**
PresidentialDecreeonLand
Kazakhstan Householdplotsonly None Shares Userights 4.7 47
Reform,Feb.1994
PresidentialDecreeonDeepening
LandandAgrarianReform,
Kyrgyzstan Allland Distribution/conversion Shares Moratorium Feb.1994Referendum,June1998; 17.1 67
PresidentialDecreeonPrivate
LandOwnership,Oct.1998
Landcode,Dec.1996;
Tajikistan None None Shares Userights 27.2 73
amended1999
None;virginland
Turkmenistan Allland Leasehold None Constitution,May1992 Nodata 50
tofarmers
Uzbekistan None None Leasehold None None 18.8 64
Afghanistan Complexpatternoflandmanagementandtenure,shapedbyconflict 23.5 74
Note:*Adaptedwithpermissionfrom[66].**WorldBank.
Table2.TypicalcroprotationsforselectedareasintheAralSeaBasin.(Shownisonecompletecycleofeithertherecommendedorthemandatorycroppingscheme).
Province Year-1 Year-2 Year-3 Year-4 Year-5 Year-6 Year-7 Year-8 Source
Kazakhstan
Kazalinsk Rice Rice Fallow Rice Rice Alfalfa Alfalfa Alfalfa [63]
Kyzyl-Orda Rice Rice Alfalfa Alfalfa Alfalfa [60]
Uzbekistan
Countrywide Cotton* Cotton* Cotton Wheat** Wheat** [61,64]
Karakalpakstan*** Wheat/Alfalfa Alfalfa Alfalfa Cotton Cotton Cotton Cotton [38]
*2–3yearsofcottonarerecommended [61,64]
**1–2yearsofwheatarerecommended,insteadofwheatalsosummercropssuchasmungbean,soybean,maize,sunflowerorvegetablesarecultivated [38]
***ExpertrecommendationforKarakalpakstan;insteadofcottonalsoricecouldbecultivated [38]
RemoteSens.2018,10,159 6of24
Remote Sens. 2017, 9, x FOR PEER REVIEW 6 of 25
2.3. SatelliteDataandPreprocessing
2.3. Satellite Data and Preprocessing
TTiimmee sseerriieess ooff tthhee NNDDVVII ffrroomm tthhee TTeerrrraa aanndd AAqquuaa MMOODDIISS 225500--mm iinnssttrruummeennttss wweerree ddoowwnnllooaaddeedd
((1166--ddaayy LL33 GGlloobbaall CCoolllleeccttiioonn VV000066,, MMOODD1133QQ11 aanndd MMYYDD1133QQ11)) ffrroomm 22000033 ttoo 22001166 [[6677]].. EEaacchh MMOODDIISS
ttiillee pprroovviiddeess wwiiddee ssppaattiiaall ccoovveerraaggee ((22333300 kkmm)),, aa ddeennssee ffrreeqquueennccyy ooff oobbsseerrvvaattiioonnss aanndd aa lloonngg--tteerrmm
aarrcchhiivvee [[88,,2299]].. FFoouurr 22333300 kkmm ×× 22333300 kkmm MMOODDIISS ttiilleess ((hh2222vv0044,, hh2233vv0044,, hh2222vv0055 aanndd hh2233vv0055)) wweerree
nneecceessssaarryy ttoo ccoovveerr tthhee ssttuuddyy aarreeaa ((FFiigguurree 11)).. OOvveerraallll,, 4466 iimmaaggeess ffrroomm bbootthh tthhee TTeerrrraa aanndd AAqquuaa
ppllaattffoorrmmss wweerree aavvaaiillaabbllee ffoorr eeaacchh yyeeaarr aanndd ssttaacckkeeddi innttoo1 144a annnnuuaallt tiimmees seerriieess..
SSeevveerraall pprreepprroocceessssiinngg sstteeppss rreedduucceedd tthhee eeffffeeccttss ooff rreessiidduuaall cclloouuddss aanndd sshhaaddoowwss,, dduusstt,, aaeerroossoollss,,
ooffff--nnaaddiirr vviieewwiinngg oorr llooww ssuunn zzeenniitthh aanngglleess.. FFiirrsstt,, wwee eexxcclluuddeedd ppiixxeellss flflaaggggeedd aass nnoo ddaattaa aanndd eexxcclluuddeedd
ssnnooww//iiccee oorr cclloouuddss iinn tthhee MMOODD1133QQ11 ppiixxeell rreelliiaabbiilliittyy llaayyeerr pprriioorr ttoo fifilltteerriinngg bbaasseedd oonn MMOODDIISS qquuaalliittyy
aassssuurraannccee iinnffoorrmmaattiioonn.. OOnnllyy ppiixxeellss llaabbeelleedd ““ggoooodd ddaattaa”” oorr ““mmaarrggiinnaall ddaattaa”” wweerree rreettaaiinneedd.. SSeeccoonndd,,
tthhee NNDDVVII ttiimmee sseerriieess wweerree ssmmooootthheedd uussiinngg tthhee SSaavviittzzkkyy--GGoollaayy aapppprrooaacchh[ [6688]]..
SSiinnccee aa ggaaiinn ooff ccllaassssiifificcaattiioonn aaccccuurraaccyy wwaass oobbsseerrvveedd wwhheenn ccoommbbiinniinngg pphheennoollooggiiccaall mmeettrriiccss wwiitthh
rraaww ttiimmee--sseerriieess ddaattaa [[2299,,6699]],, sseevveerraall mmeettrriiccss wweerree ccoommppuutteedd wwiitthh TTIIMMEESSAATT aanndd sseerrvveedd aass pprreeddiiccttoorr
vvaarriiaabblleess ffoorr tthhee ffoolllloowwiinngg ccllaassssiifificcaattiioonnss ((FFiigguurree 22)):: ((aa)) aa mmeeddiiaann NNDDVVII vvaalluuee ooff tthhee ggrroowwiinngg sseeaassoonn,,
((bb)) 2255%% aanndd 7755%% qquuaarrttiilleess ooff tthhee NNDDVVII vvaalluueess ooff tthhee ggrroowwiinngg sseeaassoonn,, ((cc)) tthhee aammpplliittuuddee ooff tthhee ggrroowwiinngg
sseeaassoonn bbeetwtweeenenm maxaixmimumuman danmdi nmiminuimmuNmD VNID,(VdI),t h(de)s ttahned asrtdanddeavrida tidoenvoiaftNioDnV oIfo vNeDrVthIe ogvroewr itnhge
sgeraoswoningan sdea(seo)nt haneda r(ee)a thuen dareerat uhnedceurr tvhee (cAuUrvCe) (AfrUomC) tfhreomst athrte tsotatrht etoe tnhde eonfdth oef tsheea ssoenas(oFnig (Fuirgeu2r)e.
T2h).e Tghreo wgrionwgisnega ssoenaswoans wdeafis ndeedfinasedst aarst isntagrotinngJu olina nJudliaayn9 d1a(yA p91ri l(A1sptr,iol r1Mst,a orcrh M31asrtcho n31lesat pony elaerasp)
aynedaresn) danindg eonndJinugli aonn dJuaylia3n0 4d(aOy c3t0o4b e(rO3c1tostb,eorr 3O1sctt,o obre rO3c0tothbeinr 3le0athp iyne aleras)pb yaesaerds)o bnatsheedi onnit itahler einsuitlitasl
frreosmultTsI fMroEmSA TTIM[6E8S].AT [68].

Figure 2. The normalized difference vegetation index (NDVI) phenological metrics for the entire
Figure 2. The normalized difference vegetation index (NDVI) phenological metrics for the entire
growing season. The phenological metrics are statistical descriptors of the NDVI trajectory of a pixel.
growingseason.ThephenologicalmetricsarestatisticaldescriptorsoftheNDVItrajectoryofapixel.
2.4. Training and Testing Data
2.4. TrainingandTestingData
Reference pixels for algorithm training (calibration) and testing (accuracy assessment) were
Reference pixels for algorithm training (calibration) and testing (accuracy assessment) were
randomly selected across the study area (Figure 3). In Google EarthTM, high-resolution images from
randomlyselectedacrossthestudyarea(Figure3). InGoogleEarthTM,high-resolutionimagesfrom
2016 clearly showed typical indicators of land abandonment, such as advanced shrub encroachment,
2016clearlyshowedtypicalindicatorsoflandabandonment,suchasadvancedshrubencroachment,
which supported the labeling process (Figure 4).
whichsupportedthelabelingprocess(Figure4).
RemoteSens.2018,10,159 7of24
Remote Sens. 2018, 10, x FOR PEER REVIEW 7 of 25
Remote Sens. 2018, 10, x FOR PEER REVIEW 7 of 25

FigurFeig3u.reL 3o. cLaotcioatniono foft rtraaiinniinngg ((rreedd) )anadn dvalvidaalitdioant i(ognree(ng)r epeixne)lsp foixre dlestefromridnientge rambainndinogneadb vasn. adcotinvee d vs.
Figure 3. Location of training (red) and validation (green) pixels for determining abandoned vs. active
activecrcorpolpalnadn. d.
cropland.

Figure 4. Abandoned (left) and active cropland (right) MODIS pixels (red squares). The graphs show
Figure 4. Abandoned (left) and active cropland (right) MODIS pixels (red squares). The graphs show
Figurtehe4 .mAebana nndoromnaeldiz(elde fdti)ffaenrednacec tvievgeectarotiponla inndde(xr i(gNhDt)VMI) oOf DalIlS acptiivxee lasn(dr eadbasnqduoanreeds )r.eTfehreengcrea ppixheslss how
the mean normalized difference vegetation index (NDVI) of all active and abandoned reference pixels
themeannormalizeddifferencevegetationindex(NDVI)ofallactiveandabandonedreferencepixels
(dashedsignatures)andstandarddeviations(dottedlines).BoldlinesrepresenttheNDVIsignatures

o fthetwoselectedreferencepixels(Source:GoogleEarthTM2016).
RemoteSens.2018,10,159 8of24
ToassigntheclassesfortheselectedMODISpixels,weusedthefollowingsteps. First,forseveral
regions(Karakalpakstan,Kyzyl-OrdaandFergana),non-differentialGPSgrounddataofabandoned
fieldswereavailablefordifferentyears(2008,2009and2014)[7,8,70–72]. Thisinformationwasused
jointlywithhigh-resolutionimagesavailableviaGoogleEarthTMandMODISNDVItime-seriesprofiles
toassistintheassignmentofclassesfortrainingdata. Phenologicalsignaturesofabandonedfields
werecharacterizedbysmoother,bell-shaped,temporalNDVIsignatures,withNDVIvaluesgenerally
below0.2–0.3duetolowbiomassproductionforsemi-naturalsuccessionvegetationindrylands[7].
Incontrast,activecroplandresultedinNDVItemporalprofileswithsubstantiallysmallergrowing
seasonNDVIintegralsandahigherNDVIvalueduringthepeakofvegetationgrowththanthoseof
abandonedplots(Figure2)[73]. UsingGoogleEarthTMtocreateandvalidatemapsofabandonment
andotherlandcoverwasalsoreportedinotherstudies[29,30,35,74,75].
Only pixels with a single dominant land-cover class of at least 85% abundance in the Google
EarthTMimageryfrom2016(activeorabandonedirrigatedcropland)wereretained. Abufferofhalf
aMODISpixelwasconsideredtoaccountforthepossibleeffectsoftheMODISlarge-pointspread
function[76,77]. Toreducethespatialautocorrelation,aminimumdistanceof1.5kmbetweenpixels
wasensured.
In total, 7960 reference samples (pixels) of “abandoned vs. active cropland” were randomly
selected,withapprox.50%foreachclass.Approximatelyone-halfofthereferencepixelswasrandomly
splitandusedfortheaccuracyassessment(4030),whiletheotherhalf(3930)wasusedastrainingdata
tocalibrate/classifyactiveandabandonedcropland.
2.5. AncillaryData
Anexistingcroplandmapforthestudyarea[71]wasusedtomaskoutareaswherecropland
abandonmentdidnotoccur(e.g., unmanaged, naturallandcover). Themapdepictsthecropland
extent of the region of interest for the period before 2004, which coincided with the onset of the
analysis [71]. The overall accuracy of the cropland map, measured with an independent, random
sampleof5185croplandvs. non-croplandtestpixels,was89.8%,whereastheproducer’saccuracyand
user’saccuracyfortheclass“cropland”were92.1%and88.2%,respectively. Theproducer’saccuracy
and user’s accuracy for the class “non-cropland” were 87.5% and 91.6%, respectively. The area of
cropland(includingboth,usedandunused)inthatmapamountsto9.1Mha,similartootherstudies
andstatistics[52,78].
AdministrativeboundariesfromtheGADMdatabaseofGlobalAdministrativeAreas,version2.5
(http://www.gadm.org/), provided the basis for the stratification (Section 3.1). This database
subdividesthestudyareaintoprovinces,referredtoasoblastsinRussian. Oblastsareprovince-level
administrativeunits,equivalenttotheNUTS-3levelintheEuropeanUnion. Intotal,39provincesthat
correspondtothe“oblast”administrativelevel2intheformerSovietUnioncoverthestudyregion.
TheGlobalAgro-EcologicalZones(GAEZ)dataset,version3.0.,whichisbasedontheHarmonized
WorldSoilDatabase(HWSD)andclimatedatafor1961–1990[79,80],wastakenfromtheFoodand
AgricultureOrganization(FAO).Itprovidesadescriptionoftheenvironmentalcharacteristicsofa
regionandthecropsuitabilityindexforseveralcroptypes(cotton,wheatandrice,whicharethemajor
croptypesintheASB).
2.6. MappingAbandonedCropland
The mapping of cropland abandonment involved two stages (Figure 5): (i) a per stratum
classification of abandoned cropland following a stratified classifier and (ii) a fusion of the most
accuratestratum-specificresultsbasedonpixel-levelclassmemberships.
RemoteSens.2018,10,159 9of24
Remote Sens. 2018, 10, x FOR PEER REVIEW 9 of 25

Figure 5. Scheme showing the workflow for mapping abandoned cropland using the stratified
Figure5.Schemeshowingtheworkflowformappingabandonedcroplandusingthestratifiedclassifier
classifier (Section 3.1). The skewed rectangles represent input and output data sets, whereas ellipses
(Section3.1).Theskewedrectanglesrepresentinputandoutputdatasets,whereasellipsesrepresent
represent analytical steps.
analyticalsteps.
Because of the agro-ecological gradients and the variety of management practices in the ASB,
Becauseoftheagro-ecologicalgradientsandthevarietyofmanagementpracticesintheASB,
spectral signatures varied spatially and could lower the recognition ability by a single global
spectralsignaturesvariedspatiallyandcouldlowertherecognitionabilitybyasingleglobalclassifier.
classifier. The study area, therefore, was stratified according to administrative boundaries (Section
Thestudyarea,therefore,wasstratifiedaccordingtoadministrativeboundaries(Section2.5),assuming
2.5), assuming that they would either (i) tend to separate irrigation systems and thus follow the agro-
that they would either (i) tend to separate irrigation systems and thus follow the agro-ecological
ecological boundaries in some cases, or in others, (ii) provide finer spatial units of environmental
boundaries in some cases, or in others, (ii) provide finer spatial units of environmental zonation.
zonation. A buffer zone of 100 km was assumed to minimize boundary artifacts due to the stratum-
A buffer zone of 100 km was assumed to minimize boundary artifacts due to the stratum-specific
specific training. A non-stratified classifier was used as a benchmark, i.e., all training pixels were
training. A non-stratified classifier was used as a benchmark, i.e., all training pixels were used to
used to calibrate a global classifier model.
calibrateaglobalclassifiermodel.
The classification approach (for both, with and without stratification) consisted of mapping
The classification approach (for both, with and without stratification) consisted of mapping
abandoned cropland by using two classifier algorithms, random forest (RF) [81] and supported vector
abandonedcroplandbyusingtwoclassifieralgorithms,randomforest(RF)[81]andsupportedvector
machines (SVM) [36]. In previous studies, both algorithms demonstrated a good ability to create
machines (SVM) [36]. In previous studies, both algorithms demonstrated a good ability to create
accurate LULCC maps [36,82–85], specifically abandonment maps [19,29,30,33]. RF is an ensemble of
accurateLULCCmaps[36,82–85],specificallyabandonmentmaps[19,29,30,33]. RFisanensembleof
decision trees (DTs) that were trained based on random, bootstrapped samples of the training data,
decisiontrees(DTs)thatweretrainedbasedonrandom,bootstrappedsamplesofthetrainingdata,
which gave this algorithms its name [81]. DTs are non-parametric, hierarchical classifiers that predict
whichgavethisalgorithmsitsname[81]. DTsarenon-parametric,hierarchicalclassifiersthatpredict
class membership by recursively partitioning datasets into increasingly homogeneous, mutually
class membership by recursively partitioning datasets into increasingly homogeneous, mutually
exclusive subsets via a branched system of data splits [86]. In contrast to other classifier algorithms,
exclusivesubsetsviaabranchedsystemofdatasplits[86]. Incontrasttootherclassifieralgorithms,
which use the whole feature space at once and make a single membership decision, SVMs [87] only
whichusethewholefeaturespaceatonceandmakeasinglemembershipdecision,SVMs[87]only
require the most informative samples to make the class decision [88] and they are relatively
requirethemostinformativesamplestomaketheclassdecision[88]andtheyarerelativelyinsensitive
insensitive to high-dimensional datasets [89]. SVMs are based on the notion of separating classes into
tohigh-dimensionaldatasets[89]. SVMsarebasedonthenotionofseparatingclassesintoahigher
a higher dimensional features (Hilbert) space by fitting an optimal separating hyperplane (OSH)
dimensionalfeatures(Hilbert)spacebyfittinganoptimalseparatinghyperplane(OSH)betweenthem,
between them, focusing on those training samples that lie at the edge of the class distributions, the
focusingonthosetrainingsamplesthatlieattheedgeoftheclassdistributions,theso-calledsupport
so-called support vectors [90]. Despite their high accuracies, RF and SVM might result in
vectors[90]. Despitetheirhighaccuracies,RFandSVMmightresultincomplementaryresults,which
complementary results, which can be overcome by decision fusion [7,36,91,92].
canbeovercomebydecisionfusion[7,36,91,92].
During the classification process, subsets of features were randomly generated from the full
Duringtheclassificationprocess,subsetsoffeatureswererandomlygeneratedfromthefullinput
input data set, i.e., 50% of the 448 features (6 annual NDVI metrics plus 26 annual NDVI values per
dataset,i.e.,50%ofthe448features(6annualNDVImetricsplus26annualNDVIvaluespergrowing
growing season from 2003–2016, see Section 2.3).
seasonfrom2003–2016,seeSection2.3).
The random generation was repeated 10 times (10 per RF and 10 per SVM), which resulted in 20
The random generation was repeated 10 times (10 per RF and 10 per SVM), which resulted
maps of abandoned vs. active cropland. These 20 maps were fused at the per stratum level, resulting
in 20 maps of abandoned vs. active cropland. These 20 maps were fused at the per stratum level,
in one map per stratum, similar to Löw et al. [36]. The selection of input features (we tested 10%, 20%,
r…es,u 1l0ti0n%g)i nanodn enummabpepr eorf sittreartautimon,ssi (mwiela tretsoteLdö 1w0,e 2t0a,l .5[03 a6n].dT 1h0e0s ietleercattiioonnso)f winapsu atsfseeastsuerde se(mwpeirtiecsatleldy
and confirmed by similar studies using random feature selection [36,93].
The fusion considers the probabilistic a-posteriori values estimating class membership at the
pixel level by the RF [94] and the SVM [95]. Depending on the algorithm, the way these ‘‘softened’’
RemoteSens.2018,10,159 10of24
10%,20%,... ,100%)andnumberofiterations(wetested10,20,50and100iterations)wasassessed
empiricallyandconfirmedbysimilarstudiesusingrandomfeatureselection[36,93].
Thefusionconsiderstheprobabilistica-posteriorivaluesestimatingclassmembershipatthepixel
levelbytheRF[94]andtheSVM[95]. Dependingonthealgorithm,thewaythese“softened”outputs
arecalculateddifferfromeachother: intheRFframework,itisdefinedasthenumberoftreesinthe
RFensemblevotingforthefinalclass[94],inSVMclassification,itisbasedonthedistancesofthe
samplestotheOSHinthefeaturespace[95,96]. Previousstudiesfoundthathigh(low)a-posteriori
valuesfrombothalgorithmsarecorrelatedwithcorrectly(incorrectly)classifiedtestpixelsandare
comparablewitheachother[94–96]. Further,thedecisionfusionassessesthereliabilityofeachmap
basedonitsper-classaccuracyandaccordingtothemethodof[36].
ThenumberoftreesintheRFwassetto500becauseahighernumberdidnotincreaseaccuracyor
thenumberofrandomsplitvariablestothesquarerootofthenumberofinputvariables[83]. Training
oftheSVMincludeschoosingthekernelparameterγandtheregularizationparameterC[97],which
wasdoneusingasystematicgridsearchin2-DspacethatisspannedbyγandC,usingathreefold
cross-validation. Therangeofγwas[0.00125,10];therangeofCwasfinallysetto[1,200]. Thewidely
usedradialbasisfunction(RBF)kernelwasselectedinthisstudysincelinearorpolynomialkernels
weretestedbutresultedinloweraccuracies(notreportedhere).
Thedelineationofa100km-bufferaroundthestrataresultedinanoverlapofthestratum-specific
maps. Therefore,themapswerefusedbyusingtheper-pixelclassmembershipsfromtheRFandSVM
algorithms. Foreachpixelintheoverlappingregionsofthestrata,severalpossibleclassoutputswere
combinedbyassigningweightstotheclassdecisionofeachmethodinproportiontoitscorresponding
classification accuracy per stratum. The decision of a method t is defined as d ∈ {0,1}, with
t,j
t =1,...,Tandclassj =1,...,C,whereTisthenumberofmethodsorclassifiersandCisthenumber
ofclasses(here: 2). Iftchoosesclassω ,thend =1and0otherwise[92]. Thefinalclassificationis
j t,j
thendeterminedbythefollowing:
∑ω d = maxC ∑ω d (1)
t t,J J=1 t t,j
t=1 t=1
that is, if the total weighted vote received by ω is higher than the total weighted vote received
j
by any other class. Weights ω to t are defined by the overall classification accuracy of t. In areas
t
withoutoverlap,afusionwasnotpossibleandthefinalclasscorrespondedtothestratum-specific
classificationavailable.
2.7. AccuracyAssessment
Theaccuracyofthemapswassystematicallyassessedwithanindependentsubsetofthereference
dataset (Section 2.4). For each active vs. abandoned map, a confusion matrix was calculated
at the province level [98,99] and included the overall accuracy, user’s accuracy and producer’s
accuracy. Asrecommended[100],theoverallaccuracymeasureswerereportedusing95%confidence
intervals[101]. Theconfidenceinterval(CI)ofthedifference(inequality)inaccuracyvaluesbetween
twoclassifieralgorithmsisgivenasfollows:
p1−p0±zα/2SEp1−p0 (2)
whereSEp1−p0 isthestandarderrorofthedifferencebetweentwoestimatedproportionswithz=1.96
andα=0.05.Thevaluesp andp aretheproportionsofcorrectlyclassifiedtestpixelsoftwoclassifiers
1 0
undercomparison. Inaddition,receiveroperatingcharacteristic(ROC)curvesandthecorresponding
AUC have been calculated; the AUC is an increasingly used accuracy metric in machine-learning
and data mining [102–104]. The AUC ranges from 0–100%, with 100% representing an error-free
classification. AsarandomclassificationyieldsanAUCof50%,norealisticclassificationshouldhave