Detection and Differentiation between Laurel Wilt Disease, Phytophthora Disease, and Salinity

13 Pages·2016·0.96 MB·English

agriculture
Article
Detection and Differentiation between Laurel Wilt
Disease, Phytophthora Disease, and Salinity Damage
Using a Hyperspectral Sensing Technique
JaafarAbdulridha,RezaEhsani*andAnadeCastro
CitrusResearchandEducationCenter,UniversityofFlorida,700ExperimentStationRd.,LakeAlfred,
FL33850,USA;[email protected](J.A.);[email protected](A.d.C.)
* Correspondence:ehsani@ufl.edu;Tel.:[phone];Fax:[phone]
AcademicEditor:RitabanDutta
Received:1August2016;Accepted:9October2016;Published:27October2016
Abstract: Laurelwilt(Lw)isafataldisease. Itisavascularpathogenandisconsideredamajorthreat
totheavocadoindustryinFlorida. ManyofthesymptomsofLwresemblethosethatarecaused
byotherdiseasesorstressfactors. Inthisstudy,thebestwavelengthswithwhichtodiscriminate
plants affected by Lw from stress factors were determined and classified. Visible-near infrared
(400–950nm)spectraldatafromhealthytreesandthosewithLw,Phytophthora,orsalinitydamage
werecollectedusingahandheldspectroradiometer. Thetotalnumberofwavelengthswasaveraged
intworanges: 10nmand40nm. Threeclassificationmethods,stepwisediscriminant(STEPDISC)
analysis,multilayerperceptron(MLP),andradialbasisfunction(RBF),wereappliedintheearlystage
ofLwinfestation. TheclassificationresultsobtainedforMLP,withpercentaccuracyofclassification
as high as 98% were better than STEPDISC and RBF. The MLP neural network selected certain
wavelengthsthatwerecrucialforcorrectlyclassifyinghealthytreesfromthosewithstresstrees. The
resultsshowedthatthereweresufficientspectraldifferencesbetweenlaurelwilt,healthytrees,and
treesthathaveotherdiseases;therefore,aremotesensingtechniquecoulddiagnoseLwintheearly
stageofinfestation.
Keywords: Laurel wilt; spectroradiometer; hyperspectral classification; remote sensing;
multilayerperceptron
1. Introduction
Avocado is the second most important tropical fruit crop in Florida after citrus [1] valued at
tens of millions of dollars per year [2,3]. The avocado crop is valuable to the state economy, and
the avocado industry brings in a substantial amount of “new dollars” to the state, resulting in an
overalleconomicimpactofcloseto$100millionperannum[3]. However,therevenuefromavocado
inFloridahasbeengreatlyreduced(by50%)duetotheeffectsoflaurelwilt(Lw)disease[3,4]. Lw
diseasehascausedseriouslossesinfruitqualityandquantity,resultinginreducedsales,aswellas
anincreaseinagro-industrialwaste,pesticidecosts,andmanagementexpenses[4]. Lwdiseasehas
beenreportedasamajorthreattothecommercialproductionofavocadoinFlorida[5,6]. Onevector
ofLwdiseaseistheredbayambrosiabeetle[3,4,7,8]. Theredbayambrosiabeetle,Xyleborusglabratus
Eichhoff (Coleoptera,Curculionidae,Scolytinae)isassociatedwithfungalsymbiontssuchasRaffaelea
lauricola[9,10],afungusthatkillsthetreebyblockingwaterflowtothecanopy[11].
ThesecretionofredbayambrosiabeetlesalivahelpstospreadR.lauricolafromtreetotree[12].
Inaddition, R.lauricolagrowsrapidlyinsidethewoodenstems[5,13]. Thefungireducethetree’s
abilitytotransfernutrientsandwatertothebranchesandleaves. Generally,thediseasecausesseveral
changes to the plant, such as the change in color of the leaves from green to a red-purple brown,
Agriculture2016,6,56;doi:10.3390/agriculture6040056 www.mdpi.com/journal/agriculture
Agriculture2016,6,56 2of13
theundersideofthebarkbecomesblack,andsmallporesandholesforminsidetheoutermostlayer
ofbark[14]. Allofthesesymptomsmayindicatethepresenceoftheredbayambrosiabeetleinthe
tree [9,14]. Many of the disease symptoms are similar to those caused by other diseases such as
Phytophthorarootrot(Phytophthoracinnamomi)orfactorssuchaslightningdamage,freezedamage,or
droughtstress[15],makingvisualdetectionofLwdifficult. Moreover,thedevelopmentofexternal
symptomssignifiescolonizationofthehost[16]. Itmaynotbepossibletomanagethediseaseonce
plantsdisplayexternalsymptoms[17]. Therefore,anadvancedandrapidmethodfordetectingLw
thatcandistinguishthesebioticandabioticstressesisneeded[18]. Currently, theonlymethodto
reducetheeffectofLwisbycompleteremovaloftheinfectedtree,includingtheroot,sothatvectors
cannot transfer the disease to healthy trees [2,10,15]. Lw disease has spread very quickly in just a
fewyearsthroughFloridaandotherstates[16], soitisnecessarytofindarapidmethodtodetect
thediseaseinatimelymannertoatleastreducethespreadoftheredbayambrosiabeetle. Thereare
severalmethodsthatcandetectthedisease,suchasscoutingandpolymerasechainreaction(PCR)[19],
buttheyarecostlyandtimeconsuming;therefore,itwouldbehelpfultodevelopafasterandmore
rapidmethodofdetectingLwdisease[20]. Spectralreflectanceisamethodthatisconsideredtobe
rapidandnon-destructive[21]. Spectralreflectancemethodsdependonreflectedoremittedradiation
fromdifferentbodies,soeachmaterialhasadifferentspectralsignature[22,23]. Forplantsignatures,
thespectralreflectanceeitherincreasesordecreasesdependingonphysical(external)factorssuchas
chlorophyllandpigmentconcentrationwhichareusedasindicatorsoftheplantconditioninthevisible
range(450–760nm)[24–26]. Therearealsootherfactorssuchasphysiologicalstructure,conditionof
thecellwall,watercontent,surfaceroughness,andstomaactivitywhichwillaffectlightpenetration
throughtheleafstructure,sothespectralsignatureinthenearinfrared(NIR)rangeof760to2500nm
isusedtoindicatethesefactors[27,28].
ItispossibletouseVisible-NearInfrared(V-NIR)techniquestodistinguishhealthyfromunhealthy
leaves[26,29–31]. Maetal.[32]usedavisibleremotesensorwithdifferentmulti-spectralwavelengths
to detect citrus greening. Some studies confirmed that spectrometric methods are more effective,
moreaccurate,lesstime-consuming,andnondestructivecomparedtoDNAanalysismethods[33,34].
FrankandMenz[35]utilizedspectraldataatdifferentwavelengthstomonitorthepowderymildew
diseaseofleafrustatdifferentstagesofdevelopment. Fungicidetreatmentwasthenappliedusing
precisionagriculturetodeterminetheinteractionofthecropwiththefungicide.Resultswerecompared
withchemicalanalysisandshowedanincreaseinearlydiseasedetectionfrom56%to88.6%,which
was considered an acceptable result for early disease detection. The goals of the present study
were to: (i) determine the hyperspectral mean reflectance curves of Lw infested trees at different
stagesofdevelopment;(ii)selecttheoptimalspectralbandsfordiscriminationofdamageduetoLw,
Phytophthorarootrot,andsalinity.
2. ExperimentalSection
2.1. HostPlantsandInoculationinGreenhouse
ExperimentswereconductedinMiami-DadeCountyunderanindoorcontrolledenvironmentat
theUniversityofFlorida’sTropicalResearchandEducationCenter(TREC)inHomestead,FL.Avocado
leaveswereobtainedfromthe“Simmonds”varietyofavocadotreeswhichweregrowninpotsin
greenhousetrialswheredifferenttreatmentswereusedtoinducethecertainsymptomsandproduce
certainfactors.
Avocadotreeswereoneyearoldandalmost1mtall. Thepre-experimentmethodologyforeach
treatmentisexplainedasfollows:
Laurelwilt(Lw): TenplantswereinoculatedwithLwatapproximately5cmabovethesoillevel
by drilling four small holes around the circumference of the trunk. Each hole received 25 µL of a
R.lauricolasporedilutionataconcentrationof30,000colonyformingunits(CFU)/mLforatotalof
3000CFU/plant. Holesweresealedwithparaffinwrap. FortheearlyinoculatedstageofLw,new
Agriculture 2016, 6, 56 3 of 14
Agriculture2016,6,56 3of13
shoots were wilted and the leaf color changed from a dark oily green to light green on most leaves
14 days after inoculation (DAI), while others turned fully yellow. In that way, leaves were obtained
s1h4o DotAsIw foerr eeawrliylt esdtaagne danthde 4l5e aDfAcoI lfoorr clahtaen sgtaegdef.r omadarkoilygreentolightgreenonmostleaves
14daSyaslianftiteyr i(nSolnc)u: lOatnioen li(tDerA oI)f, swahlti lseooluthtieorns twuarns eadppfulilelyd ytoel leoawch. Itnrethe.a Ttwhea ys,allet asvoelustwioenr,e aotb 3ta6i nge/Ld,
1w4aDs AsiImfoilraer airnl ycostnacgeenatrnadtio4n5 DtoA tIhfaotr olfa ttehest asegae .water from the east coast of Florida. Leaves showed
someS ablrinoiwtyn(inSlgn )s:yOmnpetolimtesr oafftsearl tsesovleunt idonayws,a asnadp pwlieerde ttoaekaecnh ftorre ee.aTrlhye sstaalgtes.o Tluhteionn, ,aanto3t6hegr/ lLit,ewr aosf
stihme islaera inwcaotenrc elintetrra wtioans taoptphliaetdo fotnh eexspeaerwimateenrtafrl odmayth 1e7e. aFsutrctohaesrt, obfroFwlonriidnag. sLyemavpetsomshso wweedres foumlley
bdreovwelnoipnegds, ysmimpitloamr tsoa lfetaervesse vaefnfecdtaeyds ,bayn Ldww aefrteerta 2k1e ndafoyrs.e Tarhleyns,t alegaef. sTahmepnl,easn roetphreersleintetirnogf sthaleinsietay
wlaatete srtaligteer wwearse aspelpelciteeddo. nexperimentalday17. Further,browningsymptomswerefullydeveloped,
similaPrhtyotolpeahvtheosraaf freocotet drobty (PLrwr): aTfteenr p21ladntasy sw.eTreh einno,cluealaftseadm bpyl eisnfreecptirnegse 1n0ti npgotssa wlinitihty 6l agtreamstas goef
wwehreeaste sleecetde dc.olonized with Phytophthora cinnamomi. After 30 days, very few early symptoms of Prr
appePahreydto ipnh tthheo rfaorrmoo torf otth(eP yrre)l:loTweninpgl aonf tssowmeer eleianvoecsu, laatnedd tbhyosine fleecatvinegs 1re0ppreostesnwtiinthg 6eagrrlyam stsaogfe
wwheerea ttasekeedn.c Aolfotenri zseixd mwoitnhthPsh, yltaotpeh stthaograe cwinans acmoonmsiid.eAreftde rfo3r0 Pdhayytso,pvhetrhyofreaw roeoatr lryots yanmdp lteoamvseso wfPerrre
aopbptaeianreedd. in theformoftheyellowingofsomeleaves,andthoseleavesrepresentingearlystagewere
takenH. Aeaftlethrys i(xHm):o Tnhthess,e lpatleansttsa gweewrea susceodn saids ethreed cfoonrtrPohly atnodph wtheorera grrooowtnro itna fnudlll seuavne. swereobtained.
HDeuarlitnhgy t(hHe) :eTahrelys egprloawnttsh wsteargeeu, stehdea psltahnetsc osnhtorowleadn dfewwe rseygmropwtonmisn ofuf llLswu na.nd Prr, so some
leaveDs udriindg ntohte teuarrnly tgortoawllyth ysetlalgowe,,t heesppelcainatlslys htohwoseed ifnefwecsteydm pwtiotmh sPorrf. LHwoawnedvPerr,r ,dsuorsionmg ethleea vlaetse
dgirdownotht tsutrangeto tthaell yplyaenltlso wsh,oewspeedc imallaynyth soysmeipntfoemctes dofw Litwh Panrrd. HSlonw, seuvcehr, adsu driinscgotlhoeraltaitoeng arnodw tnhecsrtaogtiec
tlheeavpelas ndtespshenodwiendg monan tyhes ysmevpetroimtys ooff tLhwe satnrdesSsl nfa,cstuocrh. Fasigduirseco 1lo srhaotiwons atensdtende carvooticcaldeaov leesavdeesp ednudriinngg
othnetshee tsweov esrtiatgyeosf. thestressfactor. Figure1showstestedavocadoleavesduringthesetwostages.

FFiigguurree 11.. AAvvooccaaddoo lleeaavveess iinn ddiiffffeerreenntt ssttrreessss ssttaaggeess:: hheeaalltthhyy,, llaauurreell wwiilltt ((LLww)),, PPhhyyttoopphhtthhoorraa rroooott rroott
((PPrrrr)),, aanndd SSaalliinniittyy ((SSllnn))..
Agriculture2016,6,56 4of13
Agriculture 2016, 6, 56 5 of 14
2.22..S2.p SecptercatlraDl aDtaatCa oClolellcetcitoinon
FoFrtoyrtlye alevaevsews wereeres asmamppleleddf rforomme eaacchh ooff tthhee ccoonnttrrooll aanndd ttrreeaatteedd pplalanntst.s .MMulutilstcisacnasn wsewree rteaktaekn en
at different positions in order to reduce the variability of the spectral device and the leaves. Five
at different positions in order to reduce the variability of the spectral device and the leaves. Five
reflectance spectra per each leaf were collected at two stages of symptom development: early and
reflectancespectrapereachleafwerecollectedattwostagesofsymptomdevelopment: earlyandlate,
late, depending on the symptomatic appearance. A handheld spectroradiometer (SVC HR-1024i™)
dependingonthesymptomaticappearance. Ahandheldspectroradiometer(SVCHR-1024i™)(Spectra
Vis(tSapCecotorap eVraisttiao nC,oPoopuegrhaktieoenp, sPieo,uNgYh,kUeeSpAsi)ew, NithY,a U4◦SAfie) ldw-iothf- vai e4w° ofipetlidc-aolfl-evnieswa nodpaticsaple cletnrasl aranndg ea of
spectral range of 350 to 2500 nm was used. Two portable halogen work lamps (500 W) were used as
350to2500nmwasused. Twoportablehalogenworklamps(500W)wereusedasthelightsource,and
the light source, and the reference reflectance spectra were acquired using a white reflectance
thereferencereflectancespectrawereacquiredusingawhitereflectancereferencepanel(Spectralon
reference panel (Spectralon Reflectance Target, CSTM-STR-99-100; Spectra Vista Cooperation) [1]
ReflectanceTarget,CSTM-STR-99-100;SpectraVistaCooperation)[1](Figure2). Spectraldatawere
(Figure 2). Spectral data were selected in the visible and near infrared domain from 350 to 950 nm.
selectedinthevisibleandnearinfrareddomainfrom350to950nm. Spectraldatawereaveraged
Spectral data were averaged for both stages to 10 nm and 40 nm based on previous studies
for both stages to 10 nm and 40 nm based on previous studies [36,37]. The purpose of averaging
[36,37]The purpose of averaging data to 10 and 40 nm is to reduce the numerous wavelengths and
datato10and40nmistoreducethenumerouswavelengthsandalsoreducenoiseandinterference
also reduce noise and interference between wavebands. Additionally, waveband filters for 10 and
bet4w0 enemn waraev eabvaanildabs.leA idn dtihtieo nmaallryk,ewt aavnedb aanred rfielatseornsafbolry1 p0raicnedd 4f0orn mfutaurree asveanisloarb ledeivnetlohpemmeanrtk. et
andThaerreefroeraes,o tnhaeb nluympbriecre odf fworavfuelteunrgetshesn fsoor r4d56e vnemlo wpmase dnrto.pTpheedr etfoo 4r9e ,atnhde 1n5u wmabveerleonfgwthasv feolre n10g tahnsdf or
45640n mnmw, arsesdpreocptpiveedlyt.o S4a9mapnlde 1s5izwe afovre lsepnegctthraslf orerf1le0catanndce4 0dnamta ,wreesrpee actpivpelileyd. Sbaemtwpeleens iz1e96f–o2r0s0p efoctrr al
refleeacctha ncacteegdoartya. wereappliedbetween196–200foreachcategory.

FiFgiugruer2e .2L. aLbaobroartaotoryrys seet-tu-uppi inncclluuddiinngg ssppeeccttrroorraaddiioommeetteerr aanndd hhaalologgene nliglihgth stosuorucrecse. s.
2.3. Data Analysis and Classification
2.3. DataAnalysisandClassification
SPSS software (SPSS Statistics ver.13.0, IBM Corporation, Aramonk, NY, USA) was utilized in
SPSSsoftware(SPSSStatisticsver.13.0,IBMCorporation,Aramonk,NY,USA)wasutilizedinthis
this study for spectral analysis. The classifications were between H, LW, Prr, and Sln based on the
studyforspectralanalysis. TheclassificationswerebetweenH,LW,Prr,andSlnbasedonthe10nm
10 nm and 40 nm averaged reflectance data. The classifications were conducted in two stages of
and 40 nm averaged reflectance data. The classifications were conducted in two stages of disease
disease development: early and late. The analyses were performed independently for the early and
development: earlyandlate. Theanalyseswereperformedindependentlyfortheearlyandlatestage
late stage spectral data, as well as in a data set composed of all of the reflectance data in order to
spectraldata,aswellasinadatasetcomposedofallofthereflectancedatainordertoselectthebest
select the best bands for each development stage.
bandsforeachdevelopmentstage.
2.3.1. Stepwise Discriminant Analysis
2.3.1. StepwiseDiscriminantAnalysis
Discriminant analysis permits the setting up of a predictive model of group membership based
Discriminantanalysispermitsthesettingupofapredictivemodelofgroupmembershipbased
on characteristics observed in each case [35]. This method determines significant differences
oncharacteristicsobservedineachcase[35]. Thismethoddeterminessignificantdifferencesbetween
between variables so that repetitive variables can be eliminated [38]. The stepwise discriminant
varainaabllyessiss o(StThEaPtDreISpCe)t itpivroecvedauriraeb lceosmcbainnebse feolriwmairnda tseedle[c3ti8o]n. Tanhde sbtaecpkwwiasredd eislicmriimnaintiaonnt oafn athlyes is
(STvEaPriDabISleCs); fporrowceadrdu rseelceocmtiobnin ise semfoprwloayredds feolre cthtieo ninacnludsiboanc kofw aa vrdareialibmlei,n aantido nbaocfktwhearvda reilaimbliensa;tfioornw isa rd
seluecsteido nfoisr etmhep loreymedovfoalr tohfe vinarcilaubslieosn noof alovnagriearb lsei,gannifdicbanact kiwn atrhdee lmimodineal ti[o3n8]i.s Aus eWdiflkosr’sth elarmembdoav al
ofdviasrtiraibbuletisonn owlaosn gpeerrfsoirgmneifid ctaon dteintertmheinme othdee lsi[g3n8i]f.icAanWcei lokfs ’esaclahm dbisdcarimdiisntarnibt uftuinocntiwona;s thpee rlfoowrmere d
totdheet eWrmilkins’es tlhamesbidgan ivfiacluane,c tehoe fgeraecahterd tihscer simpeicntaranlt dfuifnfecrteinotnia;ttihoen lboewtwereethne gWrouilpkss ’[s3l9a]m. Tbhduas,v aatl euaec,ht he
gresatetepr, tthhee svpaerciatrbalel dthifafet rmenintiiamtiioznedb tehtwe oeevnergarllo uWpislk[s3’9s ]l.aTmhbudsa, avtaeluaceh wsatesp e,nttheerevda r[3ia5b].l eTthhea dtamtain wimerize ed
theraonvdeoramllizWedil kasn’ds lsaempabrdaatevda ilnuteo wtwaso einndteerpeedn[d3e5n].t Tpharetsd: aotnaew weares ruasneddo tmo idzeevdelaonpd asnedp acroantsetdruicntt othtew o
independentparts: onewasusedtodevelopandconstructthemodelandthesecondwasusedto
validatetheaccuracy. Thedatasetwasdividedintotwoparts: 30%fortrainingand70%forvalidation.
Agriculture 2016, 6, 56 6 of 14
model and the second was used to validate the accuracy. The data set was divided into two parts:
Agri3c0u%ltu rfeo2r0 t1r6a,i6n,i5n6g and 70% for validation. 5of13
2.3.2. Neural Networks
2.3.2. NeuralNetworks
Artificial neural networks operate by machine learning, in that inputs and outputs are given to
Artificial neural networks operate by machine learning, in that inputs and outputs are given
the network one at a time and the network incrementally improves a model to approximate the
tothenetworkoneatatimeandthenetworkincrementallyimprovesamodeltoapproximatethe
input/output function [40]. Once the neural network has learned to carry out the desired function,
input/outputfunction[40]. Oncetheneuralnetworkhaslearnedtocarryoutthedesiredfunction,
the input values can be entered and the neural network will calculate the output [35]. Two neural
thenientwpuotrkvsa wlueerse cuasnedb: emeunltteilraeydera pnedrctehpetrnoenu (rMalLnPe) tawnodr rkadwiailll bcaaslicsu fluantectitohne (oRuBtFp)u. t[35]. Twoneural
networkTshwe eMreLuPs ende:umraul lmtiloadyeelr pise rac efputlrlyo nc(oMnnLePc)teadn dmrualdtiilaalybera sfiesefdu-nfocrtiwoanr(dR BsuFp).ervised learning
neTtwheorMk LtrPainneeudr ablym a obdaeclki-spraofpualglyatcioonnn aelcgtoerdithmmu lwtilhaiycehr rfeedeudc-efos rtwhea rqdusaudpraetrivc ieserrdorle satrannidnagrdn e[t3w5]o. rk
traiInne tdheb yMaLPb,a ncko- pvarolupeasg aarteio fneda lbgaocrki tthom eawrlhieirc hlaryeedrsu caensdt thhee qduimadernastiiocnesr roofr thstea MndLaPr dar[e3 5d]e.sIcnritbheedM asL P,
nosvizaelu oefs ianrpeufte ldaybearc ×k stiozee aorfl iheirdldaeyne rlasyaenrd × tshizeed oimf oeuntspiuotn lsaoyfert h[4e1M,42L]P. Ianr ethdise sstcuridbye,d thaes isnipzeuto lfaiynepr ut
laywera×s tshiez espoefchtirdald denatlaa yoef rH×, Lswiz,e Porfro, uantpdu Stllna yleearv[e4s1., 4T2h]e. IRnBtFh iiss satlusdo ya, tfhuellyin cpountnleacyteerdw feaesdt-hfoerswpaecrdtr al
datnaeoufraHl ,nLewtw,Porrkr, wanitdhS alnn lienapvuets l.aTyheer, RaB hFidisdaenls olaayefur,l layncdo annn eocutetpdufte elady-efor r[w35a]r. dTnheeu vraarlianbeltews oorfk twhei th
aniinnppuutt alanyde ro,uatphuidt dlaeynelrasy were,raen tdhea nsaomuetp aust tlhayoseer [u3s5e]d. Tinh ethvea rMiaLbPle smoefththoed.i nTphue tmanadin oduitfpfeuretnlacyese rs
webreettwheeesna mtheesaes twthoo sneeuursaeld nientwthoerkMs aLrPe mtheatt hiond R.BTFh, ethme aaisnsodciiaffteerse bnectewsebeent wtheee ninpthuets aentdw oountpeuutr al
layers are not weighted and the transfer purposes on the hidden layer nodes are radially symmetric
networksarethatinRBF,theassociatesbetweentheinputandoutputlayersarenotweightedand
[43]. The capability of MLP and RBF for every classification model was determined by a hold-out
thetransferpurposesonthehiddenlayernodesareradiallysymmetric[43]. ThecapabilityofMLP
cross-validation method. Cross validation consecutively classified all variables, but the first one to
andRBFforeveryclassificationmodelwasdeterminedbyahold-outcross-validationmethod. Cross
develop a classification function and then classify the variable was left out [44]. The full dataset was
validationconsecutivelyclassifiedallvariables,butthefirstonetodevelopaclassificationfunction
randomly split into three datasets by partitioning the active dataset into training 60%, testing 20%,
andthenclassifythevariablewasleftout[44]. Thefulldatasetwasrandomlysplitintothreedatasets
and holdout 20% samples.
bypartitioningtheactivedatasetintotraining60%,testing20%,andholdout20%samples.
3. Results
3. Results
Figure 3a–c show the spectral signature of the early and late stage disease development of Lw,
Figure3a–cshowthespectralsignatureoftheearlyandlatestagediseasedevelopmentofLw,Prr,
Prr, and Sln as well as H, respectively. There are apparent reflectance differences in the red-edge
andSlnaswellasH,respectively. Thereareapparentreflectancedifferencesinthered-edgeandNIR
and NIR regions for early and late stage development for all treatments. Higher differences are
regionsforearlyandlatestagedevelopmentforalltreatments. Higherdifferencesarefoundbetween
found between the late stage of each treatment and the control, suggesting the potential to
thelatestageofeachtreatmentandthecontrol,suggestingthepotentialtodiscriminatediseaseor
discriminate disease or damage at that level of infection. Spectral signatures at the early stage are
damageatthatlevelofinfection. Spectralsignaturesattheearlystageareverysimilartothoseofthe
very similar to those of the control and doing more difficult to discriminate at that stage of
condterovlealonpdmdeonitn. gInm thoere ladtief fisctauglte,t oexdtiesrcnraiml siynmatpetaotmths ahtasdt adgeevoelfodpeevde, lloepamveesn wt.eIrne twheiltleadte osrt adgees,icecxatteerdn al
symanpdto wmesrhe aad dduellv eglroapye-gdr,eleena voers bwroewrenw cioltloedr worhdilee sliecacvaetesd ina nthdew eaerrleya sdtauglel gwrearye- gsrtielel ngroerebnr aonwdn jucostl or
whbileeglienanvinegs itno tlohseee taurrlgyidstiatyg.e werestillgreenandjustbeginningtoloseturgidity.

(a)
Figure3.Cont.
Agriculture 2016, 6, 56 7 of 14
Agriculture2016,6,56 6of13
Figure 3. Cont.

(b)

(c)
Figure 3. Spectral signature for laurel wilt disease (a); Phytophthora root rot (b); and salinity
Figure3.Spectralsignatureforlaurelwiltdisease(a);Phytophthorarootrot(b);andsalinitydamage
damage (c) in early and advanced stages vs. control.
(c)inearlyandadvancedstagesvs.control.
3.1. Early Stage
3.1. EarlyStage
In Figure 3c, the early stage of the Sln treatment shows higher reflectance values in the near
In Figure 3c, the early stage of the Sln treatment shows higher reflectance values in the near
infrared domain than for the late stage, which is typical for green vegetation. The same trend
infrareddomainthanforthelatestage,whichistypicalforgreenvegetation. Thesametrendoccurred
occurred for Lw and Prr diseases (Figure 3a, b). Figure 3b shows a similar spectral signature for H
forLwandPrrdiseases(Figure3a,b). Figure3bshowsasimilarspectralsignatureforHandPrrat
and Prr at the early stage, since after 14 days very few early symptoms of Prr appeared. Figure 3
theearlystage,sinceafter14daysveryfewearlysymptomsofPrrappeared. Figure3alsoshowsthe
also shows the reflectance of different bands for control, Lw, and Prr where it does not show a
reflectanceofdifferentbandsforcontrol,Lw,andPrrwhereitdoesnotshowasignificantdifference
significant difference for spectral reflectance in the blue bands. On the other hand, Lw disease
for spectral reflectance in the blue bands. On the other hand, Lw disease always exhibited lower
always exhibited lower reflectance in the NIR domain (750–950 nm) than H, Prr, and Sln.
reflectanceintheNIRdomain(750–950nm)thanH,Prr,andSln.
Table 1 shows the classification results obtained from MLP and RBF for 10 nm and 40 nm
Table 1 shows the classification results obtained from MLP and RBF for 10 nm and 40 nm
bandwidth data, as well as the set of suitable wavelengths for that damaged and control leaf
bandwidth data, as well as the set of suitable wavelengths for that damaged and control leaf
discrimination. The results obtained with MLP were better than those achieved with RBF where
discrimination. The results obtained with MLP were better than those achieved with RBF where
correct classification percentages ranged from 96% to 99% in all stages and comparisons (Table 1).
correct classification percentages ranged from 96% to 99% in all stages and comparisons (Table 1).
RBF obtained the lowest classification rate (65%) for 10 nm at the early stage between H, Lw, and
RBFobtainedthelowestclassificationrate(65%)for10nmattheearlystagebetweenH,Lw,andSal.
Sal. Table 2 lists the percentage of correct classifications for every class in the MLP model at both
Agriculture2016,6,56 7of13
Table2liststhepercentageofcorrectclassificationsforeveryclassintheMLPmodelatboththeearly
andlatestagesforbothstudiedbandwidths,10and40nm. Inmostofthecases,theobtainedperclass
accuracyisaround100%,withthelowestvalueof89%obtainedforSlnclasswhentheearlystage
datasetwasusedintheMLPanalysisforthe40nmbandwidth. TheseresultsprovethatMLPallows
fortheclassificationthatdisease,sincetheprobabilitythataclassifiedpixelactuallyrepresentsthat
categoryinrealityishighenough. TheconfusionmatrixanalysistodeterminetheaccuracyoftheMLP
methodbycomparingthepercentageofclassifiedpixelsofH,LwandPrratearlystage-10nmwiththe
verifiedgroundtruthclass,subsequentlyindicatingthecorrectassessment,isshownintheTable3.
Table1. Hyperspectralclassificationofhealthy(H),laurelwilt(Lw)disease,Phytophthorarootrot
(Prr),andsalinitydamage(Sln)inavocadoandbestwavebandsselectionusingmultilayerperceptron
(MLP)andradialbasisfunction(RBF)classificationfordifferentgrowthstages.
StageandParameterSetting ImportanceVariable(nm) OverallClassification(%)
Earlystage10nm
MLP:H,Lw,Sln 717,750,739,526,952,772 96
MLP:H,Lw,Prr 750,739,952,728,717,772 96
RBF:H,Lw,Sln 615,627,649,660,671,681 65
RBF:H,Lw,Prr 794,806,783,817,705,693 71
Earlystage40nm
MLP:H,Lw,Sln 738,780,944,615,697,656 97
MLP:H,Lw,Prr 491,780,697,944,615,862 97
RBF:H,Lw,Sln 944,697,903,615,738,862 73
RBF:H,Lw,Prr 409,421,433,445,457,469 60
Latestage10nm
MLP:H,Lw,Sln 885,874,952,852,841,761 99
MLP:H,Lw,Prr 491,780,697,944,615,862 99
RBF:H,Lw,Sal 761,874,885,728,952,739 77
RBF:H,Lw,Prr 409,421,433,445,457,469 97
Latestage40nm
MLP:H,Lw,Sln 862,780,738,944,410,903 98
MLP:H,Lw,Prr 862,738,410,944,697,903 97
RBF:H,Lw,Sln 738,944,903,410,862,697 77
RBF:H,Lw,Prr 410,532,450,574,862,697 93
Thevaluesgiveninboldrepresenttheoverallaccuracybestperformance.
Table2. Percentageofcorrectclassificationofearlyandlatestageofhealthy(H),laurelwilt(Lw),
Phytophthorarootrot(Prr)andsalinitydamage(Sln)using10and40nmbandwidthintheinthe
multilayerperceptron(MLP)model.
PerClassAccuracy(%) Overall
DataSet Classes
H Lw Prr Sln Classification(%)
H,Lw,Prr 99 91 96 - 96
Earlystage10nm
H,Lw,Sln 100 91 - 97 96
H,Lw,Prr 95 98 98 - 97
Earlystage40nm
H,Lw,Sln 100 100 - 89 97
H,Lw,Prr 99 99 99 - 99
LateStage10nm
H,Lw,Sln 100 100 - 98 99
H,Lw,Prr 98 96 98 - 97
Latestage40nm
H,Lw,Sln 99 97 - 98 98
Agriculture2016,6,56 8of13
Table3.Classificationmatrixobtainedfromthemultilayerperceptron(MLP)neuralnetwork(for2010)
using10bandwidthdataandhealthy(H),laurelwilt(Lw),Phytophthorarootrot(Prr)earlystage.
PredictedSpectra
ObservedSpectra PerClassAccuracy
H Lw Sln
H 30 3 0 99.1%
Lw 0 39 0 90.9%
Sln 0 2 44 100%
Overallclassification 95.8%
Table4showstheclassificationresultsforSTEPDISCanalysisandthebandsthatwerechosenon
thebasisoftheirorderofentryintotheSTEPDISCprocedureselectiontodiscriminatebetweenH,Lw,
Sln,andPrrleavesatearlyandlatestages. Themostfrequentlyselectedwavelengthswerefoundin
thered-edgerange(717–750nm)forthe10nmbandwidthandinthered-edgeandbluerangesfor
40nm. Ineachofthecasesstudied,asmallWilks’slambda(0.0–0.2)wasobtainedindicatingthehigh
discriminatorypowerofeverysetofselectedwavebands.
Table4. Hyperspectralclassificationoflaurelwilt(Lw), Healthy(H),Phytophthorarootrot(Prr),
andsalinitydamage(Sln)andthebestwavebandsselectionusingstepwisediscriminant(STEPDISC)
analysisfordifferentgrowthstages.
Stage-Waveband Wilks’s Cross
BestBandsSelected
InputData Lambda Validation
Earlystage-10nm
H,Lw,Prr 806,761,548,638,885,941,537 0.2 82%
H,Lw,Sln 908,745,852,504,445,638,604 0.2 86%
Earlystage-40nm
H,Lw,Prr 618,781,822,904,657,575,410,740,451,945,493 0.2 83%
H,Lw,Sln 595,904,781,410,534,740,698,657,616,493 0.2 82%
Latestage-10nm
H,Lw,Prr 817,829,761,409,941,548,560,421 0.0 98%
H,Lw,Sln 560,761,806,750,885,829,409,421,504,593,515 0.05 92%
Latestage-40nm
H,Lw,Prr 817,829,761,409,941,548,560,421 0.0 98%
H,Lw,Sln 560,761,806,750,885,829,409,421,504,593,515 0.05 92%
3.2. LateStage
ThedevelopmentstagesofLwdiseasevariedinthevisibleandNIRranges. Thelatestageshad
highreflectanceinthe1300–2500nmrange. Thelatedevelopmentalstageshowedalowerreflectance
intheNIRrangeof700–950nm. ReflectancecurvesforplantsaffectedbySlnvariedfromstagetostage
dependingonsalinityconcentration. Figure3ashowsLw’sspectralreflectanceinthevisiblerange
from550to700nmincreasedduringthelatestage. Lw’slatestageshowedasignificantdifference
inspectralreflectancespectraforhealthyplants. ThePrrlatestagedisplayedlowerreflectancethan
thatofthehealthyplants(Figure3b). Thespectralsignaturecurveshowedvariationinreflectance
values, especially for Prr late stages in the NIR domain. The overall classification percentage was
92%–99%forallMLP.STEPDISCwasthesecond-bestmethodwithRBFbeingtheworstofthethree
classificationmethods.
3.3. CombinationofEarlyandLateStages
Table5showstheMLPandRBFclassificationresultsforthecombinationoftheearlyandlate
growth stages for the 10 and 40 nm bandwidths. RBF produced a low value and a high value of
Agriculture2016,6,56 9of13
classificationforthecombinationofH,Lw, andPrrfor40and10nm(65%and71%, respectively).
MLP resulted in a high classification value for the combination of H, Lw, and Prr at 10 nm, while
the combinations of H, Lw, and Sal at 40 nm and H, Lw, and Prr at 40 nm had the lowest value.
NIR(700–900nm)wasthemostfrequentrange, especiallythered-edge, withtheexceptionofthe
combinationofH,Lw, andPrrat10nmunderMLPforwhichtheblueandgreenbandswerethe
mostcommon(445,515,504,433nm). StepwiseclassificationvaluesareshowninTable6. Thelowest
classificationinTable6isthatofthecombinationofH,Lw,andSalat40nmhavingacrossvalidation
value of 80%. The average of all categories at 10 nm had a higher value than the average of the
categoriesat40nm. Thewavebandsbetween700–800nmwerethemostcommonforallcategories.
Thecorrectclassificationpercentagewas92%–98%forallMLP.STEPDISCwasthesecondbestmethod,
withRBFbeingtheworstofthethreeclassificationmethods.
Table5. ReflectanceclassificationusingMLPandRBFmethodsforlaurelwilt(Lw), Healthy(H),
Phytophthorarootrot(Prr),andsalinitydamage(Sal)inearlyandlatestagecombined.
Stageand Neural Overall
ImportanceWavebands(nm)
ParameterSettings Network Classification
H,Lw,Sal,10nm MLP 750,717,433,705,772,693,604 94%
H,Lw,Sal,10nm RBF 705,537,548,409,717,560,526,941,952,433,615 67%
445,515,504,433,638,705,750,817,421,739,952,
H,Lw,Prr,10nm MLP 98%
493,615,761,604
H,Lw,Prr,10nm RBF 671,681,660,649,638,693,627 71%
H,Lw,Sal,40nm MLP 738,780,944,862,903,450 93%
944,903,862,738,656,532,450,820,491,615,574,
H,Lw,Sal,40nm RBF 70%
410,780,697
H,Lw,Prr,40nm MLP 944,821,656,862 93%
H,Lw,Prr,40nm RBF 944,697,903,532,738,656,574,615,862,780,821,410 65%
Table6.ClassificationreflectanceresultusingSTEPDISCforH,Lw,salinity,andPrrinearlyandlate
stagecombined.
Wilk’s Cross
InputData BestBandsSelected(nm)
Lambda Validation
H,Lw,Sln:10nm 772,794,504,548,705,638,693,681,604,717,615,445 0.0 87%
H,Lw,Prr:10nm 750,885,829,952,548,649,693,638,705,537,515,681 0.06 89%
H,Lw,Sln:40nm 780,820,491,532,944,410,574,697,450,656 0.05 80%
H,Lw,Prr:40nm 780,862,944,532,574,738,697,656,410,450,615 0.03 81%
4. Discussion
In the early stage for Lw and all other stress factors, the spectral reflectance showed varying
reflectancesignatures,forexample,inthevisiblerangethespectralsignatureincreasedintheblue
andredrangewhileinthegreenrange,itshowedeitheranincreaseoradecreaseofreflectancevalue.
Althoughwewereabletoobservethechangingofthecolorsoftheleavesformanyofthecategories,
notallhadthesamelevelofchange. Thechallengewasfortreesthatwereintheveryearlystage
“asymptomatic”. At this stage, it was not easy to differentiate between healthy and infected trees.
The chlorophyll and pigment in leaves strongly absorb blue and red light in order to conduct the
photosynthesisthatisnecessaryforplantmetabolism,meaningthatanychangeinconcentrationofthe
chlorophyllwillleadtochangesinspectralreflectancedependingonthepigmentratio[45,46]. NIR
domainreflectancecanprovidegoodindicatorsofcellstructuredamagetoleavescausedbydisease,
drought,andnutrientdeficiency. Anyoverallchangeintheleafwillimpactthespectralreflectance
Agriculture2016,6,56 10of13
which therefore makes it a good indicator to use in hyperspectral detection methods [25,47]. The
spectralsignatureforhealthyplantshadhighabsorptioninredandblue,withagreenpeakat550nm.
Eachdiseasedevelopmentstagehaduniquespectralreflectancesignatures,indicatingarelationship
betweenspectralsignatureandtheappearanceofdiseasesymptoms[27,28].
IntheNIRdomain,thereflectanceincreasedordecreaseddependingonthediseasesymptom
stage,thusleafcelldamageforLw,Sln,andPrrshoweddifferentreflectancecurvesthanforthehealthy
plant. Watercontentisanotherfactorthatmighthaveaneffectonreflectance. Watercontentcould
affecttheabsorbanceorreflectanceoflightdependingonthewaterquantityintheleaf[48–51]. The
ambrosiabeetleblockstheflowofwaterinthexylemtosupportitsgrowth,andthereforewaterflow
orwatercontentisaveryimportantindicatorforthisdisease. Allofthesefactors(e.g.,chlorophyll
concentration,watercontent,wallcelldamage,andseverityofdisease)couldaffectspectralsignature
andshoweddifferentcurvesignatures. Eachaffectingfactorhasadifferentcurvesignatureforall
categories[24–26].
ThemainpurposeofthisstudywastoselectspectralbandsintheVIS-NIRrangethatcanbeused
todifferentiatebetweenhealthy,laurelwiltdiseased,salinitydamaged,andPhytophthorarootrot
infestedtrees. Severalwavebandswereselectedforclassifyingtreeswithdifferentstagesandtypesof
diseaseandstressfactors. Therearemanystudiesonbandselectiontoobtainthebestbandsinthe
VIS-NIRrangetoreducethehugenumberofnarrowbands,minimizeredundancy,andtoobtaina
statisticalmodelwithoutalongprocess(i.e.,smallernumberofspectralbandsaremoreconvenient
andmoreefficientinhyperspectralanalysis)[48,52–55]. TheMLPclassificationmethodresultedin
classificationvaluesforallofthesetreatments. Thisisrelatedtotheprogressivenatureofdisease
development,whichmakesiteasiertodifferentiatethelatestagethantheearlystage. Bydetecting
diseaseattheearlystage,particularlyatasymptomaticstage,growerscansprayorremovetheaffected
treeandmanagethediseasemorecosteffectivelybeforethediseasespreadstotherestofthegrove[36].
Usingsmallernumberofspectralbandsitispossibletouseinexpensiveremotesensingtoolssuchas
smalldronestodetecttheLWinfectedtrees[36]. Theresultsofthisstudyagreewiththeresultsof
otherstudiesusingclassificationmethods[35],sotheclassificationaccuracywerehigherinlatestage
ofinfectioncomparetotheotherstageofthedisease. Thepurposeofcombiningtheearlystageand
latestagedatatoobtainwavebandscouldbeusedintheearlyandlatestageatthesametimebecause
notalltreeleavesshowalotofsymptomsatthesametime. Diseaseistransmitteddependingon
severityofdiseaseanddiseasedevelopmentstage[24,37].Environmentalfactorsandthepopulationof
thevectorsalsoaffectthedynamicdiseasetransmissioninplants[14,56,57]. Sometimesthesymptoms
wouldbeobviousinthecrownanddonotappearonthebottomsideofthetree. Fastdecision-making
andtime-effectivesprayingatspecificlocationswillleadtolesspollutionoftheenvironment.Accurate
bandselectionandclassificationareimportanttoencouragethegrowertousethesemethodsfortimely
andappropriatedecision-making. Thisisthefirstinstanceinwhichthespectralsignaturesofleaves
fromavocadotreeshavebeenusedtodeterminespectralreflectanceandtorecommendappropriate
bandstodetectLwdiseaseandotherfactors.
5. Conclusions
The results of this study showed that it is possible to classify Lw-, Phytophthora root rot-,
and salinity-damaged leaves of avocado trees from healthy leaves using visible and near-infrared
(400–950nm)spectralbandsforthedatacollectedunderlaboratoryconditions. Theclassificationof
differentdiseasesandstressfactorsstudiedinthisworkcanbedoneatboththeearlyandthelate
stageofdiseaseinfection;however,developingaclassificationtechniquethatworksforbothearly
andlatestageofdifferentdiseasesandstressfactorswillresultinlowerdetectionaccuracy. Since
forpracticalpurposes,itismoredesirabletodetectthediseaseattheearlystageofdiseaseorstress
infection,itisbettertouseaclassificationmodelthatisdevelopedforearlystage. Comparingthe
classification results for 10-nm and 40-nm bandwidths showed similar results, indicating that the
narrowerbandwidthdoesnotproducebetterresultsandthatthe40nmwidebandscanbeasgoodas