optimising configuration of a hyperspectral imager for on-line … · 2020. 2. 17. · 1 1...
TRANSCRIPT
1
Optimisingconfigurationofahyperspectralimagerforon-linefield1
measurementofwheatcanopy2
RebeccaLWhettona,TobyWWainea,andAbdulMMouazena,b*3
aCranfieldSoilandAgriFoodInstitute,CranfieldUniversity,BedfordshireMK430AL,UK.4
bDepartmentofSoilManagement,GhentUniversity,Coupure653,9000Gent,Belgium.5
E-mailofcorrespondingauthor:[email protected]
Abstract7
There is a lack of information on optimal measurement configuration of hyperspectral8
imagers for on-line measurement of a wheat canopy, this paper aims at identifying this9
configurationusingapassivesensor(400-750nm).Theindividualandinteractioneffectsof10
cameraheightandangle,sensorintegrationtimeandlightsourcedistanceandheightonthe11
spectra’s signal-to-noise ratio (SNR)wereevaluatedunder laboratory scanningconditions,12
fromwhichanoptimalconfigurationwasdefinedandtestedunderon-linefieldmeasurement13
conditions.Theinfluencesofsoiltotalnitrogen(TN)andmoisturecontent(MC)measured14
with an on-line visible and near infrared (vis-NIR) spectroscopy sensor on SNRwere also15
studied. Analysis of variance and principal component analysis (PCA) were applied to16
understand the effects of the laboratory considered factors and to identify the most17
influencingcomponentsonSNR.18
Results showed that integration time and camera height and angle are highly influential19
factorsaffectingSNR.Among integration timesof10,20and50ms, thehighestSNRwas20
obtainedwith1.2m,1.2mand10°valuesoflightheight,lightdistanceandcameraangle,21
respectively. The optimum integration time for on-line field measurement was 50 ms,22
2
obtainedatanoptimalcameraheightof0.3m.On-linemeasuredsoilTNandMCwerefound23
tohavesignificanteffectsontheSNRwithKappavaluesof0.56and0.75, respectively. In24
conclusion, an optimal configuration for a tractor mounted hyperspectral imager was25
establishedforthebestqualityofon-linespectracollectedforwheatcanopy.26
27
Keywords28
Hyperspectralimager,signal-to-noiseratio,wheatcanopy,principalcomponentanalysis,29
soilproperties.30
31
1 Introduction32
Advancedmethodsforearlydiseasedetectionincropsisvitalforimprovingtheefficacyof33
treatment, reducing infectionandminimising the losses toyieldandquality.Traditionally,34
diseasedetectioniscarriedoutmanually,whichiscostly,timeconsumingandrequiresspecial35
expertise (Schmale & Bergstrom, 2003; Bock et al., 2010). Developments in agricultural36
technologyhaveledtodemandsforanon-destructive,automatedapproachforcropdisease37
detection that should be ideally rapid, disease specific, and sensitive to early symptoms38
(Lópezet al., 2003).Optical sensingmethods arenon-destructive, allowing repeateddata39
acquisitionthroughoutthegrowingseasonwithoutinhibitingcropgrowth.Spectroscopyand40
imagingtechniqueshavebeenusedindiseaseandstressmonitoring(Hahn,2009).However,41
their in-situ application although established in other industries (e.g. health services,42
pharmacologyandfoodsafety)isstillratherlimited.BothLenketal.(2007)andSankaranet43
al. (2010) focused on implementing the technology in the field, as a mobile (on-line)44
3
application for mapping crop disease. Yuan et al. (2016) have used high spatial satellite45
imagery inthedetectionofpowderymildew.Remotespectralsensingfor identificationof46
weedsinwheatfieldshasbeentestedbymeansofgroundcollecteddata(Gómez-Caseroet47
al.,2009).Herrmannetal.(2013)haveappliedproximalhyperspectralimageryinthefieldfor48
weeddetection(e.g.,bothbroadleafandgrassweeds),reporting85%accuracy.Okamoto&49
Lee (2009) collected in situ hyperspectral images for the detection of green citrus fruits,50
reporting promising results for identification of citrus fruits from background objects. In51
contrast,non-mobile (off-line)and laboratorymethods fordiseaseclassificationandplant52
growingconditionshavebeenstudiedanddemonstrated(Roggoetal.,2003;Wuetal.,2008).53
Hahn (2009) claims that spectroscopic and imaging techniques could be integrated with54
agricultural vehicles, providing non-invasive and reliable systems for the monitoring and55
mappingofcropdiseases,withfurtherpotentialforearlydiseasedetection.Moshouetal.56
(2005) have shown that hyperspectral imaging for the recognition of in-situ disease can57
provideidentificationwithahighdegreeofaccuracy.Dependingonthemethodofanalysis58
anddatafusion,anerrorbetween1-16.5%wasreported.59
Spectral reflectance in vegetation canopies is dependent on several factors including the60
illuminationangle, thecanopyarchitectureandtheradiativepropertiesof theplants.The61
reflectanceofcropcanopiesisnon-lambertianscattering,varyingwiththesunposition,view62
positionsandmeteorologicalconditionsincludingcloudcover(Pinter&Jackson,1985;Asner,63
1998). Plant species, maturity, phenology, level of foliage and nutrient status are plant64
properties affecting reflectance (Asner, 1998; Coops et al., 2003; Gnyp et al., 2014).65
Geometrical arrangement of objects can affect the spectral reflectance such as leaf66
orientation,whichcannotbecontrolledduringon-linemeasurements (Asner,1998;Coops67
etal.,2003).Thiscreatesproblemsassociatedwithreducedreflectionfromlightscattering.68
4
Shadowsatsmallscalecanbereducedbyadditionallightsourcesandthatopposinglighting69
canhelpreduceshadows(Barbedoetal.,2015).Obertietal.(2014)arguedthattheangle70
betweenthecanopyandcameraintherangebetween0°to60°affectsthesensitivityofa71
mountedon-line(mobile)sensorduetolightbackscattering,suggestingthepotentialofan72
obliquecameraangle,toreducetheimpactonsignal-to-noiseratio(SNR)variation.73
A tractormounted hyperspectral imager allows for on-line field crop canopy sensing and74
mapping,however,anoptimalconfigurationofthecamera,lightsourceandintegrationtime75
needstobeestablishedforoptimalqualityofimageryandspectratobecollected.Spectral76
quality ispredominantlyaffectedbysensor integrationtime,cameraorientation,andlight77
heightandanglefromtheobject(leaforcanopy).Integrationtimeistheperiodoverwhich78
thedetectorcollectsphotonsoflight.Thegreatertheintegrationtimeandlightintensity,the79
morereflectedlightisexpectedtobecapturedbythedetector,providingahigherSNRand80
pronunciationofthespectralpeaks.Thoughwhenrelyingonsunlighttheintensitycanbe81
variable. When applying a spectral technique to a forward moving platform (on-line82
measurement) longer integration times result in an average spectrumover a larger area,83
reducingthesensitivity.Furthermore,thegreaterthedistancebetweenthecameraandits84
subject, the larger the area observed and captured by a single pixel, reducing spatial85
resolution.Therefore,optimisingthemeasurementconfigurationisessentialbeforeon-line86
fieldmeasurementscanbesuccessfullycarriedout.Furthermore,backgroundsoilinfluences87
canopyspectra,andeffortshavebeenmadetoremovethisinfluence(Huete,1988).Based88
onremotesensingdataofthesurfacesoil,Demetriades-Shahetal.(1990)suggestedusinga89
secondorderderivativetoremovedeviationscausedbythesoilbackground.However,none90
ofthesestudieshaveinvestigatedtheinfluencesofon-linemeasured(atadepthof15-2091
5
cm)soilproperties[e.g.,moisturecontent(MC)andtotalnitrogen(TN)]onthequalityofcrop92
canopyspectra.93
Thispaperevaluates,under laboratoryconditions,theindividualandinteractioneffectsof94
cameraheightandangle,integrationtimeandlightdistanceandheightonthespectralSNR95
of a wheat plant canopy captured with a hyperspectral line imager. Furthermore, the96
influenceofon-linemeasuredsoilMCandTNonSNRofplantspectracollectedon-lineinthe97
field is also assessed. This was essential to inform optimal configuration and operational98
conditionsforon-linefieldmeasurementofcropcanopyanddiseases.99
100
Table1.Factorsincludedinconfiguringhyperspectralimager(multipleconfigurations101considered)102
103
2 Materialsandmethods104
2.1Hyperspectralconfigurationinthelaboratory105
WinterwheatTriticumSativum(Solsticevariety)wasgrownoutdoorsin600x400mmtrays106
(depthof120mm)with100seedsevenlysownandspacedin5parallellines.Afterseeding107
thetrayswerepredominantlyrainfed,toreduceinputofexcesssaltsfromtreatedtapwater.108
Camera angle,
deg �
Light height,
m
Light distance,
m
Camera height,
m
Integration
time, ms
0 0.90 0.60 0.15 10
5 1.2 0.90 0.30 20
10 - 1.2 0.45 50
- - - 0.60 1000
6
Apushbroomhyperspectralimager(spectrograph)(HSspectralcameramodelfromGilden109
PhotonicsLtd.,UK)wasusedtocapturehigh-resolutionlineimageswitharesolutionof1,608110
pixelsover1second,usingadiodearraydetector.Itisa12bitBaslerpiA1600-35gmcamera,111
with Schneider-Kreuznach XNP1.4/23 lens and has a pixel pitch of 7.4 μm112
interpolated/averagedto0.6nmreadingswithaspectralrangeof400-750nm.Thereflected113
light from the target travels through the lens, past an entrance slit through a series of114
inspector optics in the spectrograph and then split by the prism dispersing element into115
differentwavelengths. This sensorwas chosen for its potential for being applied to crop116
canopymeasurements,andwasoflowpricecomparedtocomparablesensors,commercially117
availableinthemarket.118
Thedatacapturedisintheformofalinearray,witheachpixelcontainingaspectrumand119
onedetectorperpixelacrosstheswath.Inordertocompileafullimage,everylineacrossa120
target must be captured (Gilden Photonics Ltd, Glasgow, UK). When configured on a121
consistentmovingplatform,theimagersweepsacrossanareatobuildupanimage.Dueto122
practical restraintsofapplyingaconsistentmovingplatformthespectraSENSv3.3 (Gilden123
Photonics Ltd, Glasgow, UK) software was adapted to record a single line array, which124
requiredanadditionalRGBphototakenbya5megapixelcamerawitha3.85mmf/2.8lens125
atthesametimeofimagecapture,sothatthescannedareacouldbecomprehended.Two126
laserpointerswereaddedateachsideofthehyperspectralimagertoindicatetheareaofthe127
canopytobescanned.Thelaserpointerswereshutoffwhenthespectralimagewascaptured128
to remove any interference. Before data analysis, the collected scans were corrected by129
meansofadarkandwhitereference,whichwerecollectedjustbeforespectralcapture,and130
at 10 minute intervals until scanning was completed. The white reference used was a131
7
commerciallyavailableSpectralonTeflonwhitecalibrationpanelwith99.9%whitereflectance132
value.133
Factorialanalysiswasundertakentounderstandandquantifytheinfluenceofconfiguration134
parametersonSNR.Thewheatwasatgrowthstage47accordingtoZadoksscale(Zadokset135
al.,1974)(whilsttheheadwasbooting)atthetimeofscanning.Thestudiedconfiguration136
parametersareshowninTable1.Thesameareawasscannedintriplicateforthedifferent137
combinationofconfigurations.Thelaboratory(simulated-field)measurementconfiguration138
is shown in Figure1. For the indoorenvironment, two500wattdiffusedbroad spectrum139
halogen lamps were positioned at either end of the crop sample tray. The additional140
illuminationusedinthecurrentworkwasshownbyexperiencetoreducetheinfluenceof141
shadowwithinthecomplexandnon-homogenouscanopystructure.Imagerydatawasthen142
capturedatdifferentcameraandlightheights,lightdistances,cameraanglesandintegration143
times(measuredinmilliseconds(ms)),asillustratedinTable1.Lightanglewaskeptconstant144
at45°,whichisdebatedastheoptimalangletoprovidethestrongestresponse(Huadong,145
2001).Additionalopposinglightingwasusedtoreduceshadows(Barbedoetal.,2015).Four146
integrationtimesof10,20,50and1000mswereadoptedasthesecoverthemostpractical147
ranges.The1000msintegrationtimeillustratesthehighestpotentialtime,duringwhichthe148
systemwillabsorbthereflectedlighthence;thisisexpectedtogivethesmoothestspectra.149
Having determined a suitable configuration in the laboratory the next experiments were150
designedtoapply theconfigurations toa fieldenvironment,andassess the impactof the151
environmental factors;e.g., soilmoistureand totalnitrogen,onSNR.Fieldmeasurements152
wereconductedina9hafieldatDuckEndfarm,Wilstead,Bedfordshire,UK(52°05'46.3"N153
0°26'41.4"W),withanaverageannualrainfallof598mm.Thefarmhasacroprotationof154
8
barley, wheat and oil seed rape. Wheat was cultivated during the experiment in 2013155
croppingseason.Thedominantsoiltypeinthefieldisaclayloam,buthasasandfractiondue156
tounderlyinggraveldeposits.157
158
Figure1.Schematicillustrationofthelaboratory(simulatedfield)configurationsusedandthe159variablesimplementedtoobtainthehyperspectraldatawiththehighestsignal-to-noiseratio160(SNR).Thehyperspectralimagerisapassivesensor,buthasbeenappliedwithanexternal161halogenlightsource.TheLaserpointersallowtopreciselypositionthehyperspectralimager162overthetarget.163
2.2On-linesoilandcropmeasurementsinthefield164
Theon-linefieldmeasurementsincludedcropspectraandsoilMCandTN.Thereasonwhy165
MCandTNweretheselectedsoilproperties,isthattheformeraffectsthesoilphysicaland166
mechanicalconditions,influencingthesoildynamicbehaviourbelowthetractortyresduring167
theon-linemeasurement,whereasbothmaywellbelinkedwithcropgrowth(assumingthat168
TNisdirectlylinkedtomineralnitrogen).Itisworthmentioningthatmineralnitrogen(e.g.169
nitrateandammonia)cannotbemeasuredwithvisible(400-780nm)andnearinfrared(780-170
2200nm)(vis-NIR)spectroscopy(Kuangetal.,2012).Furthermore,theliteratureconfirmed171
9
thepotentialofthevis-NIRtomeasureMCandTN,whichisattributedtothedirectspectral172
responseofthesepropertiesintheNIRrange(Stenbergetal.,2010;Kuangetal.,2012).The173
SNRvaluesfromthecanopieshyperspectraldatawascomparedagainsttheon-linemeasured174
soil properties at the same location via Kappa statistics and visual comparisons. Thiswas175
essentialtoevaluatewhetherornottheoptimalmeasurementconfigurationestablishedin176
thelaboratoryisapplicableinthefield,andwhethermodificationsshouldbeconsidered.177
Anon-linevis-NIRsoilsensordevelopedbyMouazen(2006)wasusedinthisstudytomeasure178
soilMCandTN,withtheobjectiveofmappingthespatialvariabilityofthesetwoselectedsoil179
properties. The system consists of a subsoiler, opening a smooth trench at 15 cm depth180
(Mouazen et al., 2005). The sensor was mounted on a three-point linkage of a tractor181
travellingataspeedof3kmh-1andcollectingspectralsoildataat10mparallelintervals.In182
order tomeasure soil spectraanAgroSpecmobile, fibre type, vis–NIR spectrophotometer183
(Tec5TechnologyforSpectroscopy,Oberursel,Germany),withameasurementrangeof305–184
2200nmandalightsourceof20Wtungstenhalogenlampwereused(Kuang&Mouazen,185
2013).Adifferentialglobalpositioningsystem(DGPS)(EZ-Guide250,Trimble,California,USA)186
recordedthepositionoftheon-linespectrawithsub-meteraccuracy.Thecollectionofsoil187
spectraandDGPSreadingstookplaceat1secsamplingresolutionusingAgroSpecsoftware188
(Tec5TechnologyforSpectroscopy,Oberursel,Germany).189
The same hyperspectral imager (Gilden Photonics Ltd, Glasgow, UK) as that used in the190
laboratorytooptimisemeasurementconfigurationwasusedforon-linemeasurementthe191
wheat canopy in the field. The following hyperspectral measurement configuration was192
considered: an integration time of 50ms, a camera height of 0.3m and light height and193
distanceof1.2mandacameraangleof10°.Thehyperspectral imagerwasmountedona194
10
tractorboom(Fig.2)travelingatapproximately4kmh-1.Thedirectionandangleoftheimager195
waskeptconsistent,andadaywithuniformlyovercastweather(completecloudcover)was196
selected,whichhelpedpreventissuesofmovingshadowsfromlateralsunmovementonthe197
data(Westetal.,2003).Nevertheless,ahandheldLUXmeter(RS180–7133,RSComponents198
&Controls,India)wasutilizedtocheckthesunlightandreadingsrangedbetween1950and199
2000,indicatingnosignificantdifference.Thehyperspectralcamerawasmountedtotheside200
ofthetractor.Itcapturesimagesof1608pixelsperline,overaone-secondinterval,whichis201
subsequentlyloggedandgeo-locatedusingaDGPS.Thecollectedscanswerecorrectedby202
meansofadarkandawhitereference(spectralon99%whitereflectancepanel).Thelatter203
wasusedbeforespectralcapture,andatamaximumof30minuteintervalsuntilscanning204
wascompleted.205
206
Figure 2. Illustrates the on-line field hyperspectral measurement using hyperspectral207measurementconfiguration.208
2.3Dataanalyses209
2.1.1 Spectrographspectraldataprocessingandevaluation210
Quality of the wheat canopy spectra (measured both in the laboratory and field) was211
evaluatedbyanalysingSNR.TheSNRisdefinedhereasaratioofthesignalstrengthtothatof212
11
unwanted interference. A strong signal devoid of interference is the desired outcome in213
measurements. If the data is too noisy it can hide key features of the spectrum, and data pre-214
processingsuchassmoothingcanresultinthembeingremoved(Dasu&Johnson,2003).Anoisy215
spectrumcanresultinpoorcalibrationmodels,duetonoisebeingconsideredasafeature.There216
aremanyapplicationsforestimatingtheSNRfromsourcessuchaselectrical,chemical,and217
spectral.DifferentmethodsareoftenappliedtoestimatetheSNRvalue,dependingonthe218
datainput.Curran&Dungan(1989)usedabrighthomogenoussurfacetoestimatetheSNR219
andproducedamethodtermedthegeostatisticalmethodforremovalofperiodicnoisein220
images. Van derMeer (2000) used amethod outlined by Lee (1990) for remote spectral221
sensing, to asses SNR of Landsat Thematic imagery. Analysis of image SNR has also been222
conductedthroughproductionofhistogramsofanimage(Ramamurthyetal.,2004).Asimilar223
approachwasoutlinedearlierbysmith(1999),wherespectrawerecollectedfromgrayscale224
images.Smith(1999)exampledaSNRrangebetween0.5and2.0,statingthatthereisonly225
anissueiftheSNRvaluedropsbelow1.0.Wehaveusedacropcanopyinthisstudyinstead226
ofawhiteorgreyreferencepaneltocalculateSNR,sincetheintensionwastousetheoptimal227
configurationforon-linemeasurementofcropcanopyinthefield,wherevariationsincanopy228
architectureandleaforientationareforeseen.Similartothecurrentwork,Daumardetal.229
(2010)reliedoncropcanopyspectratomaximiseSNR,althoughdetailsofthecalculationof230
SNRwerenotprovided.However,theyconsideredcentralpixelsonly intheircalculations,231
whilstwehaveconsideredallpixelsinalineimageafterremovingnon-cropcanopyspectra.232
Thefollowingassumptionsweremadeinthecurrentstudytojustifytheselectionofacrop233
canopyforachievinganoptimalmeasurementsetup:234
1-Non-cropcontaminatedspectraincludingsoilandsoil-plant,etc.canbeexcludedfromthe235
analysis.Thiswasdonebycalculatingnormaliseddifferentialvegetationindex(NDVI)ofall236
12
pixels.SpectraofpixelswithNDVIvaluessmallerthan0.3wereremovedfromthecalculation237
ofSNR.ThismethodisusedbyBravoetal.(2004)andRouseetal.(1974).238
2-Spectraofremainingpixelscontainbothnoiseandactualcanopysignal,theintensityof239
whichdependsonthepixelpositionwithinacomplexcanopystructureofthecrop.Thiswill240
allowsmimickingthemeasurementofactualcanopy.241
3-Variationinpositionofpixelsforaseriesofscanscanbeminimisedbyfixingtheposition242
of crop trays, so that line images are collected from the same target area for different243
measurementconfigurations.Inthiscase,whilstuniformintensitycannotbeachievedacross244
thespectruminonescan,eachpixelhasalmostthesametargetobjectthroughoutallthe245
scans.246
ThecalculationoftheSNRwasdoneinthisstudyfollowingasimilarapproachadoptedby247
Ramamurthyetal.(2014)anddescribedearlierbySmith(1999).Asthedatacollectedissingle248
line 2-dimensional captures, the data was assessed as spectra rather than images. This249
methodofSNRwasselectedasitcouldbeusedasabasisofcomparisonbetweenindividual250
spectraldatacaptures.WecalculatedtheSNRofindividualwavelength(SNRw)asfollows:251
252
SNRw=Mw/SDw (1)253
254
Where:Mwisthemeanreflectancevalueofindividualwavelengththroughallthepixels,and255
SDwisthestandarddeviationofindividualwavelengthofallpixels(Fig.3).Itisworthnoting256
thatallpixelswereconsideredimportantinthecurrentworktocalculateSNR.Analternative257
13
approachtothismethodwouldhavebeentoselectcentralpixelsonlytocalculateSNR,a258
methodappliedbyDaumardetal.(2010).259
Themeanspectralsignaldescribeswhatisbeingmeasured,whereasthestandarddeviation260
representsnoiseandotherinterferenceforeachpixel(Smith,1999).Mwvaluesfordifferent261
wavelengthswerecalculatedonremaining967wavebands,afterremovingthespectralrange262
outsideofthe400to750nmrange,sincetheywerefoundtobenoisy.OnceMwandSDw263
arecalculatedforeachindividualwavelength,theSNRforaspectrum(SNRs)wascalculated264
asfollows(seeFig.3):265
266
SNRs=Ms/SDs (2)267
268
Where:Ms ismean reflectanceofallwavelengths ina spectrum (a scan)andSDs ismean269
standarddeviationofallwavelengthsinaspectrum.270
The SNRswas used in this study to evaluate the strength of scans, hence, the quality of271
spectralsignal.ItisworthnotingthatSDisnotimportantinitself,butonlyincomparisonto272
themean.Whileconvenientandstraightforward,thedeviationisastatisticdoesn'tfitwell273
with the physics of how signals operate (Smith, 1999). Therefore, a strong SNR with274
pronouncedabsorptionpeaksprovidesincreasedrecognitionfortheassociationofspectral275
signaturestoasubject.276
A principal component analysis (PCA) was also undertaken on laboratory measured277
hyperspectraldataonlyusingStatistica software (StatSoft inc.,OklahomaUSA) to identify278
14
parameterswiththegreatestimpactonSNR.PCAisastatisticalmethod,whichanalysesthe279
distribution of data inmultidimensional space (principal components) or similaritymaps,280
where similarities (within groups of variable) and differences between groups can be281
evaluated(Dytham,2011).Inthisinstance,SNR,integrationtime,lightheightanddistance,282
andcameraangleandheightwereusedasinputvariablesforthePCAanalysis.Additionally,283
a two-way analysis of variance (ANOVA) was carried out with RStudio software (RStudio284
Boston,MA)toestimatesignificantinfluencesofindividualvariablesandinteractionbetween285
variablesonSNR(Webster,2007;Dytham,2011).286
287
2.1.2 On-linesoilsensorcalibration288
LaboratorymeasurementsofMCandTNwerecarriedoutusingstandardreferencemethod.289
SoilMCwasmeasuredwith oven drying of samples at 105ºC for 24 h, whereas TNwas290
measuredwithaTrusSpecCNSspectrometer(LECOCorporation,St.Joseph,MI,USA),using291
theDumascombustionmethod(Dumas,1826ascitedbyBuckee,1994).Theon-linecollected292
soilspectraweresubjectedtopre-processingbeforemodelling.Pre-processingincludednoise293
cutbyremovingwavelengthssmallerthan400nmandlargerthan1900nm.Noisecutwas294
followedsuccessivelybymaximumnormalisation,firstderivativeandsmoothing.Partialleast295
squaresregression(PLSR)analysiswithleave-one-outfullcross-validationwascarriedoutto296
establishcorrelationsbetweensoilspectraandlaboratorymeasuredMCandTN.Spectrapre-297
processingandPLSRanalysiswerecarriedoutusingUnscrambler7.8software(CamoInc.;298
Oslo,Norway).299
300
15
2.4Mapping301
Maps for on-line vis-NIR predicted MC and TN and on-line spectrograph measured crop302
canopyofwheatweredevelopedusedArcGIS10(ESRI,California,USA)software.Krigingwas303
usedtodevelopmaps,assumingthatthedistanceordirectionbetweensamplepointsreflects304
aspatialcorrelationthatcanbeusedtoexplainspatialvariations.Theadvancedparameters305
option inArcGIS10software(ESRI,California,USA)allowedcontrolofthesemi-variogram306
usedforkriging,selectingsphericalasthebestfit.Thesemi-variogramvalueswerecalculated307
inRStudio(RStudio,Boston,MA).308
Thesimilarityassessmentbetweenmapscanbeperformedbyvisualinspectionandstatistical309
tests(Tekinetal.,2013).Thesimplestwayofcomparingbetweenmapsisbyvisualinspection,310
toconcludeonsimilaritiesthatmayexistornot.However,thisisinsufficient,asquantitative311
estimation of similarity is a more robust approach to adopt. To compare statistical312
relationship of pairs ofmaps, Kappa statistics (Cohen, 1960) analyseswere performed to313
calculateKappavalue(κ),usingSPSS(StatisticalPackagefortheSocialSciences,IBM,Armonk,314
New York, USA). However, before running Kappa statistics data was subjected to raster315
analyses to have the same 5 m by 5 m grid size for all maps, after which the data was316
normalised.317
16
318
Figure3.Schematicillustrationoutlininghowmean,standarddeviationandsignal-to-noise319ratio for wavelength (Mw, SDw, SNRw, respectively) and a spectrum (Ms, SDs, SNRs,320respectively)werecalculated.321
3 ResultsandDiscussion322
3.1Spectralqualityinthelaboratory323
Typical crop canopy spectra can be observed in Figure 4, which shows clear, noisy and324
saturatedspectra.Theclearandsaturatedspectracanbeobservedtobemorepronounced,325
whereasaweakabsorptioninthespectralsignatureandinterferenceinthenoisyspectrum326
leads to reduced quality, low SNR, masking detail in the signature and causing a loss of327
importantspectralinformationthroughtheentirespectralrangestudied.Theclearspectrum328
isthebestquality,andthetargettobeobtained.Thenoisyspectrumiscausedbythelow329
integration times, and greater distance of the Halogen light source. A strong SNR with330
pronounced absorption peaks would allow for a greater success in analysis of crop331
assessments and disease presence. Although, pre-processing of spectral data includes332
techniquessuchassmoothing, iftheprocessofcleaningthedatais intensiveduetonoisy333
spectraitcanleadtothelossofimportantspectralfeatures,andthusimpactonthesuccess334
ofanalysis(Dasu&Johnson,2003).335
17
Saturation predominantly occurs within the central pixels associated with the highest336
reflectance,causingdata inthepeaksoftheelectromagneticspectrumtobelost(starting337
around650nm).LeafreflectanceintheNIRrangeisaffectedbythestructureoftheplant338
leaves(Gatesetal.,1965),andcanberelatedtotheleafwaxcoating(Cameron1970).Inthe339
case of spectral saturation the data becomes unusable. In the remaining parts of the340
spectrum,however,particularlythevisiblerange(400–700nm),thereisalowerreflectance341
duetoalargerabsorptionofthelight,whichisattributedtothephotosyntheticpigmentsof342
theplantsleaves,(Gatesetal.,1965).343
When analysing the entire spectral signature saturation causes a reduction in sensitivity.344
Therefore,sensorconfigurationsleadingtosaturatedspectraldatawereremovedfromthe345
analysisofSNR.Allsaturateddatawasobtainedwithintegrationtimeof1000ms,thisisin346
line with the literature, where larger integration times can cause saturation of the data347
(Dell’Endice,2008).348
349
Figure4.Examplesofsmooth(thegrey linewith1.6signal-to-noiseratio(SNR),noisy(the350blacklinewith1.2SNR),andsaturated(thedottedlinewith1.8SNR)spectraofwheatcanopy.351
18
Thesaturatedspectrumflattensoffat680-750nmandimportantinformationislostonthe352peakaround700nmthatisotherwiseillustratedbytheclearspectra.353
3.2Hyperspectralimagesconfigurationparameters;laboratory354
TheintegrationtimeconsiderablyaffectedtheSNR,expressedasaverageandSDvalues(Eq.355
1),asshowninTable2.Thelongertheintegrationtimethegreatertheopportunityformore356
energy to be captured by the spectrograph. The 1000ms configuration was the highest357
potential integration time and was trialled (under field simulations) to see the potential358
highestSNR,whilstthemanufacturerrecommendedaround20msforlaboratoryuse.The359
resultsfromthe1000msintegrationtimeprovidedahighpercentageofsaturatedresults,360
thenon-saturatedrecordingswereofthesameSNRvaluestothatofthe50msintegration361
times, so integration times higher than 50mswere not trialled further.With varying the362
integrationtime,parametershaveconstantvaluesapartfromthecameraheight.Theoptimal363
cameraheightdecreaseswiththeintegrationtime,thecloserthecameratotheobjectthe364
shortertheintegrationtimerequired.Thismaymeanthatthemostinfluencingparameters365
ontheSNRaretheintegrationtimeandcameraheight.Otherstudiesdemonstratedthatthe366
movement(adjustments,bounceorvibration)betweenthe imagerandthesubjectduring367
integrationtimecancauseacompiledimagetobewarpedornoisy(Zhongetal.,2011).For368
on-linemeasurementstheeasiestvariabletocontrolisintegrationtime,asangleandheight369
can alter slightly as ground is uneven, crop stands vary and unavoidable movement in370
mountings. It becomes necessary to have the optimal configurations set initially but to371
understand therewouldbe slightdeviations.ANADIR cameraanglewas selected,due to372
reportsfromObertietal.(2014)andVanBeeketal.(2013).373
ANOVA showed that all individual variables and interactions of variables have significant374
influencesonSNRat the (0.005)95%confidence level,except for the interactionsof light375
19
distanceandintegrationtime(0.001),andlightdistance,cameraheightandintegrationtime376
(0.01). Therefore, the null hypothesise ‘that variables and subsequent interactions of377
variableswillhavenosignificanteffectonSNR’canberejected.378
Figure5illustratesthesimilaritymapsofprincipalcomponent(PC)1and2(a),andPC1and3379
(b)accountingfor47.62%and42.26%ofvariance,respectively.ExaminingtheplotofPC1vs380
PC2onecanobservethattheintegrationtime,SNRandcameraheightaregatheredinone381
group, which explains these to be closely related. Integration time shows the strongest382
influenceandcorrelationwithSNR,whereas,cameraheightdemonstratesthesecondclosest383
corresponding variable on SNR with the latter having a weaker influence (Fig. 5a). Light384
distance,lightheightandcameraangleseemtohaveonlyminorinfluencesonSNR,asthey385
make a separate group associated with PC2 with minor variance associated with PC1.386
Disregardingothervariables, itbecomesclearthata longer integrationtimeandasmaller387
cameraheightresultinahigherSNR.388
AlthoughthelightdistanceandheightarestronglyassociatedinthePCAsimilaritymaps(Fig.389
5a&b),theyhaveanegligibleinfluenceontheSNR.Thisdoesnotmeanthattheabsenceof390
lightwouldhavenoeffect.Aslongastherewasamplediffusedlight,inorderforthedetector391
tocollectphotons,lightvariablesappearedtohavelittleimpact.Althoughtheinfluenceof392
cameraheightwasthesecondlargestinthePC1-PC2similaritymap,inthePC1-PC3similarity393
map,thecameraheighthassmallinfluenceonSNRsimilartothelightvariables.Furthermore,394
the close correlation between camera height and integration time shown in the PC1-PC2395
similaritymapisnotobservedinthePC1-PC3similaritymap.Inthelattersimilaritymap(Fig.396
5b),thecameraanglehasthesecondlargestinfluenceonSNRaftertheintegrationtime.This397
means that both camera height and angle have strong influences on the SNR. Under398
20
laboratorymeasurementconditions,largercameraangleswerereportedtobebeneficialfor399
powderymildewrecognition in leavesofgrapevinesbyObertietal. (2014).Smallerview400
anglesarediscussedbyPiseketal.(2009)totheoreticallybebetterduetoobservingalarger401
area.Rautiainenetal.(2008)concludedgreaternadiranglesaremoresuitableforviewing402
thetopcanopyandhadaverylimitedviewoftheunderstory.Thisissupportedbythefinding403
that80%ofthecropyieldiscalculatedfromthehealthofthetop3leaves(HGCA,2008).The404
correlationbetweengreenleafretentionandyieldhasbeenobservedinanumberoftrials405
(Reynolds et al., 2009; Ali et al., 2010; Hunt & Poole, 2010) and can be observed with406
referencetotrialworkinbarleyconductedbytheauthorsin2009,whereevery1%reduction407
ingreenleafareaonflag-1atGS80correlatedtoa20kg/halossinyield.Thisistruewhen408
applyingacameraangle.Butwhenassessinglightconditions,anoff-nadiranglecancreate409
more lightingvariability,duetoshadyandsunlitareasof thecrop.VanBeeketal. (2013)410
foundthatforsmalleroff-nadirviewingangles(<20°)ofthesensor,thesunorientationwas411
foundnottobeimportant.This,alongwithissuesofshadowfromtheinfieldmountings,is412
whysmallanglesfortheconfigurationswereselected.413
21
414
Figure5.Principalcomponentanalysissimilaritymapsdevelopedforprincipalcomponents1415and2(a)andforprincipalcomponents1and3(b).Inputvariablesarecameraandlightheight,416lightdistance,cameraangle,integrationtimeandsignal-to-noiseratio(SNR)417
418
22
TomakeadecisiononanoptimalconfigurationthatwouldresultinthehighestSNR,wasnot419
a straightforward process. However, when SNR values are arranged according to the420
integration time (Table 2), a clear trend is observed. For example, SNR increased with421
integrationtime,wherethehighestvaluesofSNRwereobtainedwithanintegrationtimeof422
1000ms, however, theywere onlymarginally higher than the SNR values at 50ms. Since423
several variable combinations resulted in similar SNR values, it is perhaps premature to424
suggesttheoptimalconfigurationasthesinglehighest,soconfigurationswithSNRvariability425
of less than 5% from the highest reading (of each integration time)were considered for426
furtherevaluation.ThefurtheranalysisconfirmsthatthehighestSNRoccurswiththesame427
configurations of 1.2m, 1.2m, and 10° of light height, light distance and camera angle,428
respectively.Asstatedearlier,thecameraheightisnegativelycorrelatedwiththeintegration429
time(Table2).Apossibletheoryforreoccurrenceoftheseconfigurationscouldbethatthey430
allowforthegreatestamountofreflectedlighttobecapturedbythedetector.Amongthe431
integration time steps of 10, 20, 50 and 1000 ms selected, a practical range for on-line432
measurementisbetween10msand50ms,arangewhichissuggestedbythemanufacturers.433
Assumingthatthebest integrationtimeforpracticalon-line(mobile)measurement inthe434
field is 50ms, the optimal configuration parameters of 1.2m, 1.2m, 0.3m and 10° are435
recommendedforlightheight,lightdistance,cameraheight,andcameraangle,respectively.436
TheaverageSNRfor this integrationtime is1.669 (seen in table2),which in referenceto437
Smith (1999)webelieve is a strong signal for a crop canopy.Theseoptimal configuration438
parameterswereadoptedfortheon-linemeasurementofwheatcanopymeasurementinthis439
study.440
441
442
23
Table2.Averageandstandarddeviation(SD)valuesofthehighestsignal-to-noise-ratio443(SNR)obtainedwithdifferentintegrationtime,cameraandlightsourcesettingswhen444scanningawheatcanopy.Theoreticalforwarddistancetravelled(andcapturedtoasingle445dataline)ifappliedonamovingplatformatfieldscale.446
447
3.3Hyperspectralimager;on-linemeasurement448
Duringtheon-linemeasurement,itwasnoticedthattherewasanunavoidablebounceinthe449
boomobservedtobeat±0.2moftheoriginalheightofthemounting(setat0.3mabovethe450
crop canopy). On inspection of the data the camera height from the object affected the451
uniformityoflightintensitymeasuredacrossthepixels,(particularlyatthebeginningandend452
of the captured line). Therefore, for calibrating the on-line hyperspectral scans, it is453
recommendedtoovercomethisfluctuationbyremovingthefirstandlast320pixelsfromthe454
spectral data. This is specific to this hyperspectral imager but is an interesting factor for455
consideration.456
Comparingthelaboratorywithon-linefieldmeasuredcanopyspectra,onecanobservethat457
thelaboratoryreflectancescanstobemarginallyhigherandmoreconsistentthantheon-line458
scans(Fig.6).Initially,theon-linespectrashowedvariationincanopyspectrathroughoutthe459
field,whichappearedtobeinresponsetodifferentcropspatialconditions.Forexample,the460
fieldscan1referstocanopyofwaterstressedwheatplants,whereasfieldscan2refersto461
Average,
SNR SD
Integration
time,
ms
Light
height,
m
Light
distance,
m
Camera
height,
m
Camera
angle,
deg
Distance
travelled,
m
1.688 0.102 1000 1.2 1.2 0.15 10 4
1.669 0.160 50 1.2 1.2 0.30 10 0.2
1.471 0.103 20 1.2 1.2 0.45 10 0.08
1.386 0.078 10 1.2 1.2 0.45 10 0.04
24
healthierwheatplants.Cropstartstoshowyellowcolourasasymptomofwaterstress,which462
leadstoreducelightabsorptionandincreasedreflectanceasshowninfieldscan1(Fig.6).463
More detailed analysis of crop and soil properties needs to be assessed to understand464
differencesinqualityofcanopyspectracollectedinthefield.465
466
Figure6.Comparisonofcanopyspectraofwheatcropobtainedinthelaboratory(thedashed467line)andon-lineinthefield(thedottedandgreylines).Cropwasofthesamecultureandat468comparable growth stages. Laboratory scanswere collected under the suggested optimal469configurations.Theon-linefieldscansare:1)fieldscan1(thegreyline)isofamorewater470stressedplantand2)fieldscan2(thedottedline)isofalesswaterstressedplant.471
472
25
473
Figure7.Mapsofon-linemeasuredsoilmoisturecontent(MC)(a)totalnitrogen(TN)(b),and474theaveragesignal-to-noiseratio(SNR)perscan(c).475
476
3.4 Influence of soil properties on signal-to-noise ratio during the on-line477
measurement478
Theon-linemeasuredsoilMCandTNmapsofthefieldprovidedavisualexplanation(Fig.7)479
foradropinabsorbancewithincertainareasofthefield.ComparingtheMCmapwiththe480
SNRmap,onecandrawageneralconclusionthatareasoflowMCcorrespondwithareasof481
26
highSNR,andviceversa(areasofhighMCbeingoflowSNR).Thiscanbeexplainedbythe482
factthatsoildeformationunderthetractortyreinwetsoilsislargerthanthatindriersoil483
conditions.Alargersoildeformationshouldresultinalargerfluctuationincameraheightand484
angle,comparedtothebaselinesetup,whichmayleadtoreducingtheSNR.Workcarriedout485
bySöhne(1958)showedthatanincreaseintheMCofasoilandtheincreaseinpayloadona486
tyrebothincreasethedepthofsoildeformation.487
Bothon-linemeasuredMCandTNwerefoundtohavesignificanteffectsontheSNRofthe488
wheatcanopyspectraat95%confidence.ThekappavaluesbetweentheSNRmaponone489
handandTNandMCmapsontheotherhandconfirmspatialsimilarityordifferencethefield.490
Landis& Koch (1977) classified Kappa values into different categories: 0–0.20, 0.21–0.40,491
0.41–0.60, 0.61–0.80, and 0.81–1, which indicate slight, fair, moderate, substantial, and492
almostperfectagreement,respectively.ResultsshowthattheKappavaluebetweenTNand493
SNRisrathersmall(kappa=0.56),indicatingmoderatesimilaritybetweenthesetwomaps.494
SinceTNisthemainsourceofmineralnitrogenessentialforcropgrowthanddevelopment,495
itcanbeinfluentialonthequalityofcropcanopyandSNR.ThekappavaluebetweenMCand496
SNR was much higher (kappa = 0.75) than that between TN and SNR map, confirming497
substantialsimilaritybetweenthetwomaps.Thissupportstheearliersuggestionaboutthe498
influence of MC on soil deformation that changes the initial (optimal) hyperspectral499
measurement configuration obtained in the laboratory and implemented for on-line500
measurementinthefield.Inordertoreducethenegativeinfluenceofsoildeformationon501
SNR due to high soil MC during on-line scanning, proper spectra pre-processing is502
recommendede.g.,normalisation,derivation,andmultiplicativescattercorrections.Further503
researchisrequiredtoconfirmthisassumption,aswhenmeasuringon-line,deviationsfrom504
27
the optimum configuration (due to tractor vibration, bounce in the boom and soil505
deformation)wereunavoidable.506
507
4 Conclusions508
This study was undertaken to determine an optimal measurement configuration of a509
hyperspectrallineimager(400-750nm)byevaluatingtheindividualandinteractioneffectsof510
thesystemsconfigurations,onthehyperspectralcalculatedSNRformeasurementofawheat511
canopy under laboratory scanning conditions. Optimal configurationwas determined and512
implementedforon-linefieldmeasurement.Theinfluenceofon-linemeasuredsoilmoisture513
content(MC)andtotalnitrogen(TN)onSNRwasevaluated.Resultsallowedthefollowing514
conclusionstobedrawn:515
1-Theintegrationtimefollowedbythecameraheightandcameraangleappearedtohave516
thelargestinfluenceoftheSNR.Alongintegrationtime(>50ms)wasofanegligibleinfluence517
andonlyslightlyincreasedtheSNR,butresultinspectralsaturation,henceshouldbeavoided.518
3-ThePCAsimilaritymapshowedthatthelightheightanddistancehaveastrongcorrelation519
witheachotherbutaminimalinfluenceonSNR.520
4-Bothon-linemeasuredMCandTNwerefoundtohavesignificanteffectsontheSNRofthe521
wheatcanopyspectraat95%confidence.Theon-linesoilmeasurementsrevealedstronger522
spatialsimilaritybetweenthehyperspectralSNRandMCmaps(kappavalue=0.75),which523
wasattributedtosoildeformationbelowthetractortyre.524
5-Thevariablereflectedlightintensitycapturedbythedifferentpixelsacrossthelineimagery525
isaninterestingfactortotakeintoaccount,duetotheimpactofvaryingcameraheightduring526
28
the on-linemeasurement.Whilst the solution suggested here is appropriate, it is camera527
specific.528
FurtherworkisplannedtoovercomevariationinSNRassociatedwithcameraheightchanges529
(vibration,bounceintheboom,andsoildeformationduringtheon-linemeasurement)bythe530
implementationofaproper spectrapre-processing. It isalsoplanned to implement these531
hyperspectral measurement configurations for on-line measurements of crop canopy for532
detectionofcrophealthanddiseasepresence.533
534
Acknowledgement535
Weacknowledge the funding received forFarmFUSEproject fromthe ICT-AGRIunder the536
EuropeanCommission’sERA-NETschemeunderthe7thFrameworkProgramme,andtheUK537
DepartmentofEnvironment,FoodandRuralAffairs(contractno:IF0208).538
539
References540
AliM.A.,HussainM.,KhanM.I.,AliZ.,ZulkiffalM.,AnwarJ.,SabirW.,ZeeshanM.(2010).541
Source-sink relationship between photosynthetic organs and grain yield attributes during542
grainfillingstageinspringwheat(Triticumaestivum).InternationalJournalofAgricultureand543
Biology,12:pp.509–515.544
Asner,G.P.(1998).Biophysicalandbiochemicalsourcesofvariabilityincanopyreflectance.545
RemotesensingofEnvironment,64(3),234-253.546
29
Barbedo, J.G., Tibola,C. S.,& Fernandes, J.M. (2015).Detecting Fusariumheadblight in547
wheatkernelsusinghyperspectralimaging.BiosystemsEngineering,131,65-76.548
Bock,C.H.,Graham,J.H.,Gottwald,T.R.,Cook,A.Z.,&Parker,P.E. (2010).Windspeed549
effectsonthequantityofXanthomonascitrisubsp.citridisperseddownwindfromcanopies550
ofgrapefruittreesinfectedwithcitruscanker.PlantDisease,94(6),725-736.551
Bravo,C.,Moshou,D.,Orberti,R.,West,J.,McCartney,A.,Bodria,L.andRamon,H.(2004).552
Foliar disease detection in the field using optical sensor fusion.Agricultural Engineering553
International:CIGRJournalofScientificResearchandDevelopment.ManuscriptFP04008.Vol.6,554
14p.555
Buckee, G. K. (1994). Determinationoftotal nitrogeninbarley,maltandbeerbyKjeldahl556
proceduresandtheDumascombustionmethod—Collaborativetrial.JournaloftheInstitute557
ofBrewing1994,100(2)pp.57–64.558
Cameron,R.J.(1970).LightintensityandthegrowthofEucalyptusseedlings:II.Theeffectof559
cuticular waxes on light absorption in leaves of Eucalyptus species. Australian Journal of560
Botany,18:pp.275–284.561
Cohen, J., 1960. Kappa: Coefficient of concordance.Educational Psychologist,562
Measurement,20,p.37563
Coops,N.C.,Smith,M.L.,Martin,M.E.,&Ollinger,S.V.(2003).Predictionofeucalyptfoliage564
nitrogencontentfromsatellite-derivedhyperspectraldata.IEEETransactionsonGeoscience565
andRemoteSensing,41(6),1338-1346.566
30
Curran,P.J.andDungan,J.L.,1989.Estimationofsignal-to-noise:anewprocedureappliedto567
AVIRISdata.IEEETransactionsonGeoscienceandRemoteSensing,27(5):pp.620-628.568
Dasu, T.,& Johnson, T. (2003).Exploratorydatamininganddata cleaning(Vol. 479). John569
Wiley&Sons.570
Daumard,F.,Champagne,S.,Fournier,A.,Goulas,Y.,Ounis,A.,Hanocq, J.F.andMoya, I.,571
2010. A field platform for continuous measurement of canopy fluorescence.IEEE572
TransactionsonGeoscienceandRemoteSensing,48(9),pp.3358-3368.573
Dell’Endice, F. (2008). Improving the Performance of Hyperspectral Pushbroom Imaging574
Spectrometers for Specific Science Applications. InISPRS 2008: Proceedings of the XXI575
Congress: Silk Road for Information from Imagery: The International Society for576
PhotogrammetryandRemoteSensing(pp.215-220).577
Demetriades-Shah,T.H.,Steven,M.D.andClark,J.A.,1990.Highresolutionderivativespectra578
inremotesensing.RemoteSensingofEnvironment,33(1):pp.55-64.579
Dytham,C.(2011).Choosingandusingstatistics:abiologist'sguide.JohnWiley&Sons.580
Gates, D.M., Keegan, H.J., Schleter, J.C. andWeidner, V.R. (1965). Spectral properties of581
plants.Appliedoptics,4(1),pp.11-20.582
Gnyp,M.L.,Miao,Y., Yuan,F.,Ustin, S.L., Yu,K., Yao,Y.,Huang,S.andBareth,G. (2014).583
Hyperspectralcanopysensingofpaddyriceabovegroundbiomassatdifferentgrowthstages.584
FieldCropsResearch,155:pp.42-55.585
Gómez-Casero,M.T.,Castillejo-González, I.,García-Ferrer,A.,Peña-Barragán, J.M., Jurado-586
Expósito,M.,García-Torres,L.andLópez-Granados,F.(2010).Spectraldiscriminationofwild587
31
oatandcanarygrassinwheatfieldsforlessherbicideapplication.AgronomyforSustainable588
Development,30(3),pp.689-699.589
Hahn,F.,(2009).Actualpathogendetection:Sensorsandalgorithms—areview.Algorithms2,590
301–338.591
Herrmann, I., Shapira, U., Kinast, S., Karnieli, A. and Bonfil, D.J., 2013. Ground-level592
hyperspectral imageryfordetectingweeds inwheatfields.Precisionagriculture,14(6):pp.593
637-659.594
HomeGrownCerealAuthority(HGCA)(2008).Thespringwheatdiseasemanagementguide.595
Warwickshire:HomeGrownCerealsAuthority(HGCAPublication).596
Huadong,G. (2001).ApplicationsofRadarRemoteSensing inChina.NewYork:Taylorand597
Francisinc.56.598
Huete,A.R.(1988).Asoilvegetationadjustedindex(SAVI).RemoteSensingofEnvironment,599
25:pp.295–309.600
HuntJ.,&PooleN.(2010).Simulatingleafareadurationtopredictyieldresponsetofoliar601
fungicide in wheat and barley. In: Dove H, Culvenor RA, editors. Food security from.602
sustainableagriculture.Proceedingsof15thAgronomyConference;2010Nov;Lincoln,New603
Zealand.604
Kuang,B.,Mahmood,H.S.,Quraishi,M.Z.,Hoogmoed,W.B.,Mouazen,A.M.&vanHenten,605
E.J. (2012). 4 Sensing Soil Properties in the Laboratory, In Situ, and On-Line: A606
Review.AdvancesinAgronomy,114(1),pp.155-223.607
32
Kuang,B.&Mouazen,A.M.,(2013).Non-biasedpredictionofsoilorganiccarbonandtotal608
nitrogen with vis–NIR spectroscopy, as affected by soil moisture content and609
texture.Biosystemsengineering,114(3),pp.249-258.610
Landis,J.R.,&Koch,G.G.(1977).Themeasurementofobserveragreementforcategorical611
data.Biometrics,33,159-174.612
Lee,J.B.,Woodhyatt,S.,andBerman,M.(1990).Enhancementofhighspectralresolution613
remote-sensingdatabyanoise-adjustedprincipalcomponentstransform.IEEETransactions614
onGeoscienceandRemoteSensing,28:pp.295–304615
Lenk,S.,Chaerle,L.,Pfündel,E.E.,Langsdorf,G.,Hagenbeek,D.,Lichtenthaler,H.K.,VanDer616
Straeten,D.,&Buschmann,C.,(2007).Multispectralfluorescenceandreflectanceimagingat617
theleaflevelanditspossibleapplications.JournalofExperimentalBotany58(4),807–814.618
López,M.M.,Bertolini,E.,Olmos,A.,Caruso,P.,Gorris,M.T.,Llop,P.,Penyalver,R.,&Cambra,619
M., (2003). Innovative tools for detection of plant pathogenic viruses and bacteria.620
InternationalMicrobiology6,233–243.621
Moshou,D.,Bravo,C.,Oberti,R.,West, J.,Bodria,L.,McCartney,A.,&Ramon,H., (2005).622
Plant disease detection based on data fusion of hyper-spectral and multi-spectral623
fluorescenceimagingusingKohonenmaps.Real-TimeImaging11(2),75–83.624
Mouazen, A.M. (2006). Soil SurveyDevice. International publication published under the625
patent cooperation treaty (PCT). World Intellectual Property Organization, International626
Bureau. International Publication Number: WO2006/015463; PCT/BE2005/000129; IPC:627
G01N21/00;G01N21/00.628
33
Mouazen,A.M.;Anthonis, J.;&Ramon,H., (2005).Anautomaticdepthcontrolsystemfor629
online measurement of spatial variation in soil compaction, Part 4: improvement of630
compactionmapsbyusingaproportionalintegrativederivativedepthcontroller.Biosystems631
Engineering,90(4),409-418.632
Oberti,R.,Marchi,M.,Tirelli,P.,Calcante,A.,Iriti,M.,&Borghese,A.N.(2014).Automatic633
detection of powderymildew on grapevine leaves by image analysis: Optimal view-angle634
rangetoincreasethesensitivity.ComputersandElectronicsinAgriculture,104,1-8.635
Okamoto, H. & Lee, W.S. (2009). Green citrus detection using hyperspectral636
imaging.ComputersandElectronicsinAgriculture,66(2):pp.201-208.637
Pinter, JR, P.J. and Jackson, R.D. (1985). Sun-angle and canopy-architecture effects on the638
spectralreflectanceofsixwheatcultivars.INT.J.REMOTESENSING,6(12):pp.1813-1825.639
Pisek, J., Chen, J.M.,Miller, J. R., Freemantle, J. R., Peltoniemi, J. I., & Simic, A., (2009).640
Mapping forest background reflectance in a boreal region using multi-angle Compact641
AirborneSpectrographic Imager (CASI)data. IEEETransactionsonGeoscienceandRemote642
Sensing,47.doi:10.1109/TGRS.2009.2024756.643
Ramamurthy,R.,Canning,C. F., Scheetz, J.P.,&Farman,A.G. (2004). Impactofambient644
lightingintensityanddurationonthesignal-to-noiseratioof imagesfromphotostimulable645
phosphor plates processed usingDenOptix® and ScanX® systems. Dentomaxillofac Radiol,646
33(5),307-11647
Rautiainen,M.,Lang,M.,Mõttus,M.,Kuusk,A.,Nilson,T.,Kuusk,J.&Lükk,T.,2008.Multi-648
angular reflectance properties of a hemiboreal forest: An analysis using CHRIS PROBA649
data.RemoteSensingofEnvironment,112(5),pp.2627-2642650
34
Reynolds,M.,Foules,M.J.,Slafer,G.A.,Berry,P.,Parry,M.A.,Snape,J.W.,Angus,W.J.(2009).651
Raisingwheatpotential.JournalofExperimentalBotany.60:pp.1899–1918.652
Roggo,Y.,Duponchel,L.,Huvenne,J.P.(2003).Comparisonofsupervisedpatternrecognition653
methodswithMcNemar’sstatisticaltest:applicationtoqualitativeanalysisofsugarbeetby654
near-infraredspectroscopy.AnalyticaChimicaActa,477(2):pp.187–200.655
Rouse, J.W. Jr., R.H. Haas, D.W. Deering, J.A. Schell and J.C. Harlan. 1974.Monitoring the Vernal656
Advancement andRetrogradation (GreenWave Effect) ofNatural Vegetation.NASA/GSFC Type III657
FinalReport,Greenbelt,MD.,371.658
Sankaran,S.,Mishra,A.,Ehsani,R.,&Davis,C.(2010).Areviewofadvancedtechniquesfor659
detectingplantdiseases.ComputersandElectronicsinAgriculture,72(1),1-13.660
Schmale,D.G.,III,&Bergstrom,G.C.(2003).Fusariumheadblightinwheat.ThePlantHealth661
Instructor.http://dx.doi.org/10.1094/PHI-I-2003-0612-01.662
Smith,S.W.(1999).TheScientistandEngineer'sGuidetoDigitalSignalProcessing.2nded.663
SanDiego:CaliforniaTechnicalPublishing,(Chapter2pp13-16).664
Sohne,W.(1958).Fundamentalsofpressuredistributionandsoilcompactionundertractor665
tires.AgriculturalEngineering,39,290666
Tekin,Y.,Kuang,B.,&Mouazen,A.M.(2013).Potentialofon-linevisibleandnearinfrared667
spectroscopy for measurement of pH for deriving variable rate lime668
recommendations.Sensors,13(8),10177-10190.).669
VanBeek,J.,Tits,L.,Somers,B.,&Coppin,P.(2013).Stemwaterpotentialmonitoringinpear670
orchardsthroughWorldView-2multispectralimagery.RemoteSensing,5(12),6647-6666.671
35
Van derMeer, F. and De Jong, S.M., 2000. Improving the results of spectral unmixing of672
Landsat Thematic Mapper imagery by enhancing the orthogonality of end-673
members.InternationalJournalofRemoteSensing,21(15):pp.2781-2797.674
Webster, R. (2007). Analysis of variance, inference, multiple comparisons and sampling675
effectsinsoilresearch.EuropeanJournalofSoilScience,58(1),pp.74-82.676
West, J.S., Bravo, C., Oberti, R., Lemaire, D.,Moshou, D. andMcCartney, H.A., 2003. The677
potentialofopticalcanopymeasurementfortargetedcontroloffieldcropdiseases.Annual678
reviewofPhytopathology,41(1),pp.593-614.679
Wu,D.,Feng,L.,Zhang,C.,&He,Y.,(2008).EarlydetectionofBotrytiscinereaoneggplant680
leavesbasedonvisibleandnear-infraredspectroscopy.TransactionsoftheASABE51(3),pp.681
1133–1139.682
Yuan,L.,Pu,R.,Zhang,J.,Wang,J.andYang,H.(2016).Usinghighspatialresolutionsatellite683
imagery formappingpowderymildewat a regional scale.PrecisionAgriculture,17(3): pp.684
332-348.685
Zadoks,J.C.;T.T.Chang,C.F.Konzak(1974)."Adecimalcodeforthegrowthstagesofcereals.".686
WeedResearch14(6):415–421.doi:10.1111/j.1365-3180.1974.tb01084.x.687
Zhong,C.,Ding,Y.,&Fu,J.(2011).ImageForwardMotionCompensationMethodforSome688
Aerial Reconnaissance Camera Based on Neural Network Predictive Control.Advanced689
ResearchonComputerScienceandInformationEngineering.152(1),pp.237-244.690