• Aucun résultat trouvé

Textural approaches for vineyard detection and characterization using very high spatial resolution remote sensing data

N/A
N/A
Protected

Academic year: 2021

Partager "Textural approaches for vineyard detection and characterization using very high spatial resolution remote sensing data"

Copied!
13
0
0

Texte intégral

(1)

HAL Id: hal-01196900

https://hal.archives-ouvertes.fr/hal-01196900

Submitted on 15 May 2020

HAL is a multi-disciplinary open access

archive for the deposit and dissemination of

sci-entific research documents, whether they are

pub-lished or not. The documents may come from

teaching and research institutions in France or

abroad, or from public or private research centers.

L’archive ouverte pluridisciplinaire HAL, est

destinée au dépôt et à la diffusion de documents

scientifiques de niveau recherche, publiés ou non,

émanant des établissements d’enseignement et de

recherche français ou étrangers, des laboratoires

publics ou privés.

characterization using very high spatial resolution

remote sensing data

Carole Delenne, Sylvie Durrieu, Gilles Rabatel, Michel Deshayes,

Jean-Stéphane Bailly, Camille Lelong, Pierre Couteron

To cite this version:

Carole Delenne, Sylvie Durrieu, Gilles Rabatel, Michel Deshayes, Jean-Stéphane Bailly, et al..

Textu-ral approaches for vineyard detection and characterization using very high spatial resolution remote

sensing data. International Journal of Remote Sensing, Taylor & Francis, 2008, 29 (4), pp.1153-1167.

�10.1080/01431160701311259�. �hal-01196900�

(2)

Textural approaches for vineyard detection and characterization using very high

spatial resolution remote-sensing data

C.DELENNE

,S. DURRIEU

, G.RABATEL

,M.DESHAYES

, J.S.BAILLY

, C.LELONG

, P.

COUTERON

§

UMRTETIS,CEMAGREF -CIRAD-ENGREF,34093Montpellier,France

UMRITAPCEMAGREF-ENSA-CIRAD,34196Montpellier,France

§

FrenchInstituteofPondicherry,PB33,605001Pondicherry,India (Received00Month200x;Innalform 00Month200x)

Vine-plotmappingandmonitoringarecrucialissuesinlandmanagement,particularlyforareas wherevineyardsaredominantlikein someFrench regions.Inthiscontext, theavailabilityof anautomatictoolforvineyard detection andcharacterizationwouldbevery useful.Theobjectiveofthestudyistocomparetwodierentapproachestomeetthisneed.Therstoneusesdirectionalvariationsof thecontrastfeaturecomputedfromHaralick'scooccurrencematricesandthesecondoneisbasedonalocalFourierTransform.Foreach pixel,a`vineindex'iscomputedonaslidingwindow.Tofosterlarge-scaleapplications,testandvalidationwerecarriedoutonstandard veryhighspatialresolutionremote-sensingdata.70.8%and86% ofthe271plotsofthe studyareawerecorrectly classiedusingthe cooccurrenceand the frequencymethod respectively.Moreover, the latterenabled an accuratedetermination (less than3% error) of interrowwidthandroworientation.

Keywords:Texture;Imageanalysis;Cooccurrence;FourierTransform;Vineyard

1 Introduction

Thanks to the increased availability of remote sensing data and of more powerful computers, automatic

analysismethodscan be developedto buildorupdategeographicaldatabasesforlandmanagement.

Accu-ratedigitalmappingofvineyardsforwine-growingregionssuchasLanguedoc-Roussillon(France)couldbe

extremely usefulfor many reasons. For example,these mapscan beintegrated within Geographical

Infor-mationSystems(GIS)whichcanbeusedbywinegrowercooperativestoimprovethemonitoring ofquality

compliance in areas registered in the list of Controlled Origin Denomination. The management of

pollu-tion, erosion and ood risks is another eld that can take advantage of these maps. Indeed, these risks,

depending on culture and soil surface condition, are worsened by mechanization and intensive cropping

practices (Wassenaar etal., 2005;Vincini etal., 2004).

User demand usually concerns 1) locating vine plots and 2) identifying some characteristics that can

be connected to cropping practices or cropquality(interrow width, orientation of rows, presence ofgrass

between rows...).

Mostvineyard relatedstudies usingremotesensingdatameet thesecondrequirement bydetecting vine

rows(Bobilletetal.,2003)forexample,orbycharacterizingtrainingmode(Wassenaaretal.,2002)orfoliar

density(Hall et al., 2003) for previously delimitedplots. Those dealing with vineyard plots identication

anddelineation often usemulti-spectralinformation onover-metricspatial resolution images, provided by

satellites Landsat, Ikonos or airborne sensors (Rodriguez et al., 2006; Johnson et al., 2001; Gong et al.,

2003). However, the increasing availability of Very High Spatial Resolution (VHSR) images oers a lot

of new potential applications: the object shape and spatial structure are becoming more distinguishable,

providing greater discrimination and characterization opportunities. Indeed, according to the

Shannon-Nyquist theorem

1

,periodic patterns resulting from the spatial arrangement of vine plants (often in lines

or grid),become perceptible withaspatial resolution thatisat leasttwiceassmallasthepattern period.

1

(3)

In the study area, like in many wine-growing regions, the minimum distance between two vine rows, is

about 1.5 m; consequently, image spatial resolution should be lower than 0.75 m. However, as they deal

withspatialstructuresor shapes,these newapplications alsorequire newimage processing approaches.

Inarecentstudy(Warneretal.,2005),aclassicationalgorithmbasedonananalysisofautocorrelograms

was developed and tested using Ikonos panchromatic imagery of Granger (Washington). This method,

althoughprovidinggoodresultsintheapplicationpresented,couldhardlybegeneralizedinolderEuropean

wine-growing regions where the heterogeneity among researched patternsis high. Because of theperiodic

organization of vineyards, frequency analysis appears as a suitable approach for vine detection. Wavelet

analysispresented in(Ranchin et al., 2001)is applied to 25 cm resolution images for vine/non-vine pixel

classication. Using a plot basis validation, 78 % of plots were accurately classied; but this approach

is complex and needs signicant user intervention. A Fourier Transform based analysis should be more

straightforward and quite as eective since this tool is perfectly suited for oriented and periodic texture

detection. Its eciency has been demonstrated to characterize and monitor natural periodic vegetation

(CouteronandLejeune, 2001;Couteron,2002). Wassenaar(Wassenaaretal.,2002)successfullyuseditfor

vine/non-vineclassication and characterization ofpreviously delimitedplotson 25cmresolution images.

Onasampleof46`extremelyvariedeldpatterns',vine/nonvineclassicationwascorrectforalltheplots

andonlyveerrorswereencounteredconcerningtrainingmodeclassicationofthe41vineplots.Moreover,

this method gave a very precise (less than 1 % error) estimation of interrow width and row orientation.

Prat(Prat, 2002)employed asimilar methodto identify vineplots inanimage.Thisone wasrstdivided

into small square windows (of 12.5 m side) on which ve indices were deduced from Fourier spectrum

andimageradiometry.Then, amultidimensional supervisedclassicationusingmaximumlikelihoodledto

correctclassication of 81%of vine pixels.

OtherverypopularapproachesfortexturalanalysisarebasedonHaralick'sresearches,accordingtowhom

`the texture informationin an image iscontained inthe overall or average spatial relationship which the

gray tones in the image have to one another' (Haralick et al., 1973). He then introduced the `gray-level

spatial dependency' (cooccurrence) matrices, which had remained unused for many years as they were

tootime-consuming. Withtheamazingincreaseof computerpower, cooccurrencebecameone of themost

popular characterization toolsbecauseitisbasedonsecondorderstatistics, wellsuited forthedescription

of textural properties, which the human eye is most sensitive to. A lot of studies have demonstrated its

relevancefor texturalanalysis(Chenet al.,1979)andits usefulnessfor manyapplications: urbanplanning

(Moralesetal.,2003),medicine(Smuteketal.,2003),scienticpolice(Vermaetal.,2002),textileindustry

(Abdulhady etal., 2002)... and even remote-sensingfor agro-forestry (Arvis et al.,2004).

The general objective of this work was to develop an automatic method for vineyard detection and

characterization usingveryhighspatialresolutionremote-sensingdataandwithoutanyaprioriknowledge

oftheparcel plan. Indeed,this latterisnot available inmostEuropean wine-growing regionsand,when a

georeferencedcadastreisavailable,itgenerallydoesnotcorrespondtoagriculturalplotsactuallyobservable

in the eld. To foster large-scale applications, image used was a `standard' orthophotography in natural

colour, with a 50 cm spatial resolution, similar to data available on the whole French territory. In this

paper, the relevanceof cooccurrencebased analysisisevaluated incomparison witha frequencyapproach

to meet the need for vine plot detection. Moreover, characterizations of row orientation and/or interrow

width,deduced fromthese approaches, arecompared.

2 Study area

The study area is part of theLa Peyne watershed(110 km

2

) and is located in theLanguedoc-Roussillon

region - France (Figure 1). This zone is representative of the French Mediterranean coastal plain with

respectto geology,agriculturalpracticesandvineyard management(Wassenaaretal.,2002). Two subsets,

of 2 km

2

and 1 km

2

, have been selected from this area near Roujan municipality (43

30'N, 3

18'E).

Despitea general decrease, vine cultivationis still predominant and coversabout 70% ofthe271 plots of

thestudy area.

The diversity of agricultural practices in the study area leads to a great heterogeneity among vine

(4)

Paris

Montpellier

Roujan

Figure1. LocalizationofthestudyareainFranceandEurope.

(a)

(b)

Figure2. Vinetrainingmodes:a)Goblet(gridpattern),b)Trellis(rowpattern).

spectral information for vineyard detection. However, on VHRS remote sensing data, two main patterns

can be observed accordingto trainingmode (gure 2):

Grid pattern: about a quarter of the vineyard considered in this study is trained as `goblet'. This old methodofvinetraininginvolvesnowiresorothersystemofsupport:vinestocksareplantedaccordingto

a gridpattern,oftensquare,withapproximately 1.5m

×

1.5mspacing inthestudyareabut sometime up to 3mspacing indryregions.

Linepattern:mostoftherecent vineyardsaretrainedusinghorizontalwirestowhichthefruitingshoots are tied. Spacing separating two wires is higher than spacing between vine stocks guided by thesame

wire (often 1 m

×

2.5 m spacing in the study area), which leads to row patterns. More adapted to mechanization, thistraining mode namedtrellis or wire-training, ismainly used.

These patterns can be observed on each spectral band and are less dependant on the previously cited

heterogeneities. Then, withvineyard detectioninaim,methods should be more robust whendealing with

(5)

0

25

50

100

Meters

´

Figure3. Zoomonthestudyarea.Alltheplotsofthestudyareahavebeensurveyedandintegratedwiththeircharacteristicsina geodatabase.

3 Data acquisition

DataacquisitionwasmadeduringtherstweekofJuly 2005,whenfoliardevelopmentwassuchthatboth

vineandsoilwerevisibleonaerialphotographs.AdigitalcamerawasusedaboardanUltraLightMotorized

(U.L.M.) to acquire photographs in natural colours (Red, Green and Blue). Images were geometrically

correctedandgeoreferencedusingArcGis

r

(ESRI),mosaickedusingERDAS

r

Imagine(LeicaGeosystem)

andresampled toa50 cmresolution.Theresultingimages have characteristics similartothose ofthe

BD-Ortho

r

coverage oftheFrench geographicinstitute(IGN),which iswidelyusedand coversalmostallthe

French territory.

For result validation, ground-truth information was collected at the same time as image acquisition.

Each ofthe 271 vine and non-vine plots of the site hasbeen digitized ina GISdatabase (gure 3) which

alsocontains information concerninglanduseand aseries of characteristics for vine plots:training mode,

interrow width, orientation, rough estimates of vine height and width, soil surface condition... Row

orientation and interrow width were obtained by precise on-screen measurements: row orientation was

measured with a 1

precision and interrow width wascalculated by dividing thewidth of the whole plot

bythenumber ofinterrows.

4 Textural analysismethods

Both methods compared inthis paper were implemented to calculate textural characteristics on the

sur-roundingof each pixelusing asliding window.

4.1 Cooccurrence analysis: use of Haralick's contrast feature

Therstmethodpresentedinthisstudyhasbeendevelopedfromcooccurrencematricesdenedin(Haralick

et al., 1973). Element

p

i,j

of each matrix

P

d

x

,d

y

contains the number of transitions from grey level

i

to

j

between twopixels of image

I

,distant from

d

x

pixels incolumn and

d

y

inline (equation1):

(6)

(1,0)

(1,1)

(1,-1)

(0,-1)

(-1,-1)

(-1,0)

(-1,1) (0,1)

x

45°

90°

135°

d=1

26.5°

d=2.24

d=1.41

~60°

d=21.9

d: distance in pixels

y

Figure4. Cooccurrencecalculationdistancesvsorientations.Left:nearestneighbours.Right:moreorientationsimplieslonger calculationdistances(greaterthaninterrowwidths).

p

i,j

= P

d

x

,d

y

(i, j) = #



((x, y), (x

0

, y

0

)) / x

0

= x + d

x

, y

0

= y + d

y

;

(I

x,y

= i & I

x

0

,y

0

= j) or (I

x,y

= j & I

x

0

,y

0

= i)



(1)

where # denotes the number of elements inthe set and

I

x,y

is thegrey level of pixel of coordinate (x,y). Eachcooccurrencematrix isthensymmetricand for a

N

g

grey levelimage, itssize is equalto

N

g

× N

g

.

Depending on the spatial resolution used (50 cm) and row spacing encountered (from 1.4 mto 2.5 m)

analysismust be done on transitions between one pixeland its direct neighbours inorder to characterize

soil-vinetransition:

|d

x

|

,

|d

y

| ∈ {0, 1}

.Onlyfourdirectionsarethenexplored:

θ

=0

,45

,90

or135

.With

image coordinates increasing from upper leftto lowerright corner:

d

0

= (d

x

, d

y

) = (0, −1)

,

d

45

= (1, −1)

,

d

90

= (1, 0)

and

d

135

= (1, 1)

.Searchformoredirectionswouldimplylongercalculationdistances,unsuited to interrow widths(gure 4).

Fromcooccurrencematrices, Haralickdened 14textural characterization features, someofwhichbeing

correlated. As preliminary comparative analysis (unpublished), they have been computed on a sliding

windowapplied ona synthetic image imitating3 vine plots(with roworientedat 0

,45

and90

)aswell

asanon-vineplot,modeledbyarandomtexture.Somefeatures(e.g.correlationorangularsecondmoment)

could be used to highlight vine plots, but their histogram have a highdispersion, which would hamper a

good pixel classication in vine/non-vine. That is not the case for contrast feature (equation 2), which

appeared to be well suited for vineyard detection. The higher the local variations inthe sliding window,

the higher the contrast, strongly depending on orientations of both vine row and feature calculation.

Consequently,contrast is highwhen calculated in a direction thatis perpendicularto vine rows and very

lowwhen calculatedinrow direction.

f

2

(P

d

x

,d

y

) =

N

g

−1

X

n=0

n

2

X

|i−j|=n

p

i,j

(2)

Wethenproposea `vineindex'basedonthisproperty,whichcan beusedto distinguishrowpatternsfrom

othernon-orientedhighcontrastedpatterns(e.g.checkerboard-like). Indeed,vineyardwillbecharacterized

byahighdierenceofcontrastcalculatedintwoperpendiculardirections.Foreachpixel,signeddierences

between thefour pairs of perpendicular directionsare compared. The highest dierence is thevine index

andthetwo directionsassociated provide a classof row orientation to thefocalpixel. Theoretically,when

calculatedonvineyardswithroworientation

θ ∈ [23, 68]

,contrast shouldbehighfordirection

d

135

andlow

for

d

45

,sothatthese vineyardsshouldbeclassiedinclass

C

45

;likewise,class

C

90

correspondstovinerow orientations in

θ ∈ [69, 112]

,

C

135

to

θ ∈ [113, 158]

,and

C

180

to

θ ∈ [158, 180]

or

θ ∈ [1, 23]

. Contrast is all themore interesting asitcan becomputed directlyonimage withoutpreviouscalculation ofcooccurrence

matrices; this considerably reduces calculation time. Figure 5 describes the classication method applied

(7)

Contrast

V=max D

q

Vine index V

Orientation q

1

2

3

Original image:

0

255

synthetic image with

3 oriented and 1 non

oriented texture

d

135

d

45

-d

135

d

45

d

135

-d

45

d

90

d

0

d

90

-d

0

d

0

-d

90

D

0

D

90

D

45

D

135

C

180

C

135

C

90

C

45

Figure5. VineyarddetectionusingHaralick'scontrast.1)Contrastcalculationonslidingwindowinorientations:0

,45

,90

and 135

.Thebrighterthepixel,thehigherthecontrast.2)Dierenceofcontrastbetweentwoorthogonaldirections.3)Themaximum valueofthedierencesgivesthe`vineindex'andanestimationofroworientationwithaclassicationinfourclasses.

4.2 Frequency analysis:use of local Fast FourierTransform

Thesecondmethoddevelopedinthisstudy isbasedontheworksof(Wassenaaretal.,2002)whousedthe

Fourier Transform to characterize already delimited plots.Here, we test thesame kind ofapproach when

theonly availabledatais theaerialimage (the maingoal beingvineyard detection).

Fourier theory statesthat almost anysignal, including images, can be expressedas a sum of sinusoidal

waves oscillating at dierent frequencies. The discrete Fourier transform (FT) of an image is computed

usingtheFastFourierTransform(FFT)algorithm.Takingthemodulusofthecomplex-valuedFFTresults

yields the FT amplitude (or spectrum), which can be represented in the frequency domain as an image

of the same size as the initial image,

I

. In theconventional representation, this image is symmetricwith respecttoitscentre,whichcontainstheaverageof

I

i.e.theamplitude ofthenullfrequency

F

0

.Eachpixel corresponds to a particular spatial frequency increasing the further it is from centre. Its value codes the

amplitudeofFourier spectrum,whichdependsonthe frequencypresenceinimage

I

.Theamplitudeofthe discrete Fourier Transform of

I

isdened byequation3:

a(u, v) =

1

N

x

N

y

N

x

−1

X

x=0

N

y

−1

X

y=0

I

x,y

exp



−j2π

 ux

N

x

+

vy

N

y



(3)

where

(N

x

, N

y

)

is the size (column, line) of both images,

x = 0 . . . N

x

− 1

,

y = 0 . . . N

y

− 1

are spatial indexes(inimage

I

)and

u = 0 . . . N

x

− 1

,

v = 0 . . . N

y

− 1

arefrequencyindexes(intheFourierspectrum). The method consists inapplying the FFT algorithm on a sliding window. When this window contains

vineyard arranged in rows, two peaks will be present on the Fourier image, and will be symmetric with

respectto thecentre; for the gridpattern ofa gobletvine, fourpeaks willbepresent at90

(see gure6).

TheFFT algorithm assumes thatthe data isperiodical, i.e. theimage repeats from end to end innitely.

Therefore,FFT calculation on a nite window maylead to aliasing artefacts (Gibbs'phenomenon) when

pixel values at the edges of the window do not match. To avoid these artefacts, which could introduce

additional peaks,pixel values arerst ofall multipliedbya Hanningwindow (byVon Hann)which shape

ishalfacycleofacosine wave andisnullattheedges(see gure7).Threecharacteristics canbededuced

fromthepeakvalueand position:

The distance

r

of the peaks from theimage centre correspondsto thepattern frequency inthewindow and,consequently,isconnectedtothevineyard interrowwidth,whichisequaltothesize

N

ofthesliding windowdivided by

r

.Peaks can thenbesought inan annular ring, correspondingto potential vineyard interrowwidthsto avoidconfusion withother periodicpatterns(e.g.orchards, characterized a bylarger

interrow).

The angle

θ

,between horizontal lineand one peak,determines thewave direction inapolarcoordinate

(8)

Figure6. ExamplesofFourierTransform.Theleft-handsideshows3images,whicharethesamesizeastheslidingwindow,extracted fromagobletvine(up),atrellisvine(middle)andanon-vine(down);thecentreshowstheirrespectiveFourierTransformandthe

right-handsideshowsthefrequenciesremainingafterthresholding(thesamethresholdisusedforthethreeimages).

FFT

Vine Index V

Orientation q

Interrow N/r

r

q

Fourier spectrum of

the sliding window

synthetic image with

3 oriented and 1 non

oriented texture

Original image:

1

2

Peak V

Hanning window

3

1m

3m

180°

0

255

Figure7. VineyarddetectionusingFourierTransform(FT).1)UseofaHanningslidingwindowofsizeN

×

N.2)FTcalculationon thewindow.3)SearchforthemaximumFTamplitudeamongthepotentialvineyardfrequencies;foreachpixel,thisvalueissavedas

`vineindex'anditspositiongivesroworientationandinterrowwidth.

system, whichis equal torow directionina geographical coordinate system(90

oset).

Peak amplitude is the `vine index': the higher theamplitude, thehigher theprobability of thewindow beingina vineyard.

5 Implementationof the texturalanalysismethods

Bothmethods wereimplemented inClanguage andapplied on thestudy area.

Asensitivityanalysistothewindowsize hasbeencarriedout sinceaccuracyofdetectionand

characteri-zationdependsonthenumberofpixelsinthewindow.Ononehand,thiswindowmustbe largeenoughto

takeintoaccounttherepetitionofroworgridpatterns,soalargewindowprovidesmorepreciseinformation

whenlocatedinsideaplot. Ontheother handitdecreasesclassication resultsnearplots boundariesasit

cancontainseveralpatternsatthe sametime,andofcourse,increasesthecalculationtimes. Eightwindow

sizes have been tested from

11 × 11

to

39 × 39

pixels. For the frequency based method, results become acceptablefora

27 × 27

windowsize.(13%ofbadlyclassiedpixels)andthelowestrateofmisclassication (12.2%)isreachedfor

31 × 31

.Extendingwindowsize upto

39 × 39

pixelsdoesnot improveresults(12.4%

(9)

(a)

(b)

(c)

Figure8. Resultsonasubsetofthestudyarea.a)originalimage,b)vine-indexissuedfromthecontrastmethod,c)vine-indexissued fromthefrequencymethod.Themanualsegmentationissupperposedingreyoneachimage.

ofmisclassiedpixels)whiledoublingcomputationaltime.Consequently,thebesttrade-ofor thewindow

sizeisabout

31 × 31

pixels,which cancontainfromveto tenvinerows inthestudyarea.Through visual assessment ofthe dierent window sizes, thislatter alsoappears tobe thebest forthecontrast approach.

Themethods have been testedoneachofthe threechannelsoftheimage.Vineindex,produced byboth

methods, is an indicator of the probability for a pixel to belong to a vineyard. For vineyard detection,

a threshold has been dened to separate two classes: `vine' and `non vine'. The pixels whose vine index

is lower than the threshold are classied as `non vine', the others as `vine'. Threshold determination is

often empirical; here, it was chosen to minimize global classication error for a representative sample of

thedatabase plots.Therefore,omissionerror (vinedetected asnonvine)is chosen lowerbut almostequal

to commissionerror (non vine detected asvine).Some testshave shownthat asamplecontaining10 %of

theplots was large enough to determine a threshold valuethat isvery close to the one obtained using all

theplots. The sample must be representative enough of the study area, particularly in termsof landuse

andvineyard trainingmode.

6 Validationmethod

Validation is performed on a plot basis using all the 271 vine and non-vine digitized plots of the study

area. A simple classication rule is employed: a plot is classied as `vine' if at least 75 % of its pixels

are `vine', as `non vine' if at least 75 % of its pixels are `non vine' and not classied otherwise. Then,

vine plot characteristics (orientation class obtained by Haralick's contrast and orientation and interrow

widthgiven byFourier Transform) arechosen to be the majority value amongthe pixels of the plot. For

validation,resultsof plotclassication and characterization arecompared to theinformation contained in

theground-truthdatabase.

7 Results

BestresultswereobtainedwiththeRedchannel,probablybecauseitprovidesthehighestcontrast between

vine and soil surface, even when covered by grass. Therefore, we only present results derived while using

this channel. Figure 8shows vine-indexof bothmethods on asubset of thestudy area.

7.1 Classicationresults

Table1 gives confusionmatrices which enablethe estimation ofplot classication qualitybycomparingit

togroundtruth data.Consideringboth vineand non-vineplots,70.8%ofthe271 plotsarewell-classied

(10)

and81nonvineplotsusingRedchannel:aplotiswell-classiedifatleast75%ofitspixelsarewell-classied, badlyclassiediflessthan25%ofitspixelsarewell-classied,notclassiedotherwise.

Contrastmethod Frequencymethod

vine nonvine notclassied Total vine nonvine notclassied Total

Vine 72% 7% 21% 100% 89% 5% 6% 100%

Nonvine 1% 68% 31% 100% 2% 79% 19% 100%

Becauseof the thresholdchosen, both methods lead to a worse classication for non vine than for vine

plots.Themaincauseofnondetectioniswhenthevineistooyoungi.e.lessthanthreeyearsold.Vegetation

isthusnot suciently developed for therows tobevisible onaerialphotographs; consequently,these vine

plotshavegoodclassication ratesofonly 26%and47%for contrastandfrequencymethods respectively

(seegure 9bfor an example).

Globally,results provided bycontrast methodarepoorerthan thoseprovided byfrequency method.For

nonvineplots,oneexplanation maybethatcontrastdoesnottakeintoaccounttheperiodicityofpatterns:

aroad,forexample,canleadto adierenceofcontrastintwoorthogonaldirectionsthatisashighasthat

ofavine but doesnothave apeakoffrequency correspondingto vine interrow width.Figure 9ashows an

exampleofnonvinedetectedasvinebycontrastmethodbutnotbyfrequencymethod.Likewise,confusion

could theoretically appear between vine and orchards. However, since the sliding window size is adapted

tovineyardsinterrowwidths, vineindex willbelower onorchardsbecausetheir interrow widthsaremuch

larger.

For vine plots, results must be analyzed according to training mode. Indeed, goblet vines benet from

a good classication rate of only 49 %using contrast method against 89.7 % for `adult' trellis vines (all

gobletvinesareadult). Likewise,frequency method leads toa goodclassication rateof 88.6%for goblet

vinesand 95.7%for `adult'trellis vines.

For both methods, thepoorer results obtained for goblet vines mainly have two originswhich lead to a

lowvisibilityof soilbetween rows.Firstly,gobletvinesarenot stressed bywires andcan growfreelyinall

directions;secondly,interrowwidthsofgobletvinesaregenerallysmallerthan thoseoftrellisvines(onthe

study area, 67 % of the goblet vines have an interrow width lower than 160 cm against only 2 % of the

trellisvines).

Concerning goblet vine classication, the big dierence (39.6 %) between methods is due to the fact

that goblet vines are often planted on a square grid so that contrast is identical in both perpendicular

directions,whichhamperdetectionbythecontrast method(seegure9cforexample).Infact,most goblet

vines properlyidentied by the contrast method are pruned along one direction, which leads to a higher

contrast intheperpendicular direction.

7.2 Results of vine plot characterization

Estimationof orientation andinterrowwidthobtained byboth methods arenowexamined.

With the contrast method, 78 % of the plots correctly classied as vine have been allocated with the

correctorientation class(amongthe fourused). Asevoked previously,deningmore than four orientation

classescannotbeconsideredwiththisresolutionbecauseitwouldimply,forcontrastcalculation,adistance

toolargeinfront ofinterrowwidth(gure4).Indeed,ifa30

classissought,thedistanceinpixelsneeded

to compute contrast feature will be at least

(d

x

, d

y

) = (5, 3)

i.e. anEuclidian distance ofabout 2.9m(for 30.96

and 10.98mfor 30.07

),largerthan most ofinterrowwidth.

Figure10a showscharacterization results for well classied vine plots, using frequency method.Fourier

Transformleadstomoreaccurateresultsforvineroworientation.Indeed,betweenon-screenmeasurements

and method estimation, an average absolute dierence of 3.5

was found, which is lessthan a 2 % error.

Moreover, errordistribution is almostcentered (gure10b).

Interrow widthcalculation is also very precise with an average absolute dierence of 6.2 cm, i.e. about

3 % error (see gure 10d for error distribution). The four outliers shown in gure 10c, concern two vine

(11)

(a)

(b)

(c)

(d)

Figure9. Examplesofplots.a)Noncultivatedplotrecognized asvinebycontrastmethod:anorientedpatternisvisiblebutwithno particularfrequency.b)Veryyoungtrellisvine,badlyclassiedbybothmethods:rowsarehardlyvisible.c)Gobletvineclassiedas

nonvinebycontrastmethodandwell-classied(with98%ofpixels)byfrequencymethod.d)Trellisvine,wellclassiedbyboth methods.

y = 0,9769x

R

2

= 0,9961

0

20

40

60

80

100

120

140

160

180

0

20

40

60

80

100

120

140

160

180

Row orientation measured (°)

R

o

w

o

ri

e

n

ta

ti

o

n

c

o

m

p

u

te

d

)

Orientation

Bisector

Regression line

0 0 0 0

3

2

5

6

4

12

14

16

20

16

13

16

13

5 5

7

6

4

2

0

0

5

10

15

20

25

-11 -9

-7

-5

-3

-1

1

3

5

7

9

11

Errors (°)

N

u

m

b

e

r

o

f

p

lo

ts

1

0

1

5

13

22

51

33

29

4

4

2

0

0

0

10

20

30

40

50

60

-30

-20

-10

0

10

20

30

Errors (cm)

N

u

m

b

e

r

o

f

p

lo

ts

y = 1,0035x

R

2

= 0,9457

100

150

200

250

300

100

150

200

250

300

Interrow measured (cm)

In

te

rr

o

w

c

o

m

p

u

te

d

(c

m

)

Interrow

outliers

Bisector

Regression line

b)

a)

c)

d)

Figure10. Orientationandinterrowcharacterizationusingfrequencymethod.Right:comparisonofretrievedcharacteristicsandplot measurements;regressionandbisectorlinesarealmostconfounded.Left:errordistributions.

interrow width, and two vine plots ploughed between rows, for which interrow width determined is half

on-screenmeasurement. Characterization resultshighly depend on thesize ofthecalculation window (see

below),whichiswhy(Wassenaaretal.,2002),whoappliedtheFFTalgorithm ontheentireplot,obtained

1%errorsfor both orientation and interrowwidth.

8 Conclusionand discussion

Twomethodswerecomparedforvineyarddetectionandcharacterizationfromaerialphotographpresenting

`standard' characteristics. The rst one was based on Haralick's cooccurrence analysis, which had been

successfullytested onmanyapplications but notyetfor vineyard detection. Thesecondone wasbased on

Fourier analysis, awell-triedapproach for periodic andorientedpattern recognition.

The originality of the proposed cooccurrence approach lies in the comparison of the contrast feature

(12)

than withthefrequencymethod mainlyfor two kindsof plots:

non-vine: contrast is sometimes higher in one direction than for its perpendicular, due to ploughing or roads for example, and this leads to a classication as `vine' of 32 % of non-vine plots. However,

theresulting patternshaveno particular frequencyand aregloballywell-classiedbyfrequencymethod

(21%of commission error).

Goblet vines: with this training mode, when there is no privileged direction of pruning, contrast isthe same inboth perpendicular direction,which leads to a `non-vine' classication of 51 % ofgoblet vines.

Onthecontrary,since their overall patternis periodic, theseplots arewell-classied withthefrequency

method(only21.4 %oferror).

The poorer results of goblet detection using both methods are strongly linked to the relation between

thepattern period (interrow width) and the image resolution: thelimit of theShannon-Nyquist theorem

is reached. However, this highlights the fact that a coarser resolution could be usedin many other

wine-growing regions, especially dry ones such as Castilla-la-Mancha in Spain, where interrow widths are up

to three meters. Theseapproachescould also be applied to orchardswith other resolutionsas longasthe

periodicpattern isvisible.

Incomparison withthecontrast method,the frequencyapproachnot only permitsa better

`vine'/`non-vine'classication (86%against70 %),but alsoa verypreciseestimation ofroworientation andinterrow

width (2 % and 3 % errors respectively) whereas only four classes of orientation could be dened and

distinguishedwith anaccuracy of78 %usingthe contrast feature.

In addition to the utility of characterization as such, orientation and interrow width estimation could

be usedto increaseplot classication qualityin prospectfor asegmentation stage.Ontheone hand,they

could be used to better separatedetected plots; indeed, some plots,which are spatially very close, would

be grouped within a same polygon unless they have dierent row orientation or interrow width. On the

other hand, these characteristics could help the discrimination of badly classied non-vine plots; indeed,

vine classied plotswithno particular orientation or interrow widthcouldbereclassied asnon-vine.

Inprospect,to meet the second userrequirement, characteristics of roworientation and interrow width

couldalso beused for anautomatic detection of eachvine row. Thiswould enable theevaluationof more

vineyard characteristicssuchasmissingvinetrees or soilsurface condition(e.g.presenceof grass between

rows). Moreover, vine index could also be used on vine plots as an indicator of vine quality, since its

intensitydependson the pattern contrast.

Acknowledgements

This work is part of two projects: Bacchus European Project (http://www.bacchus-project.com) and

Mobhydic project ofthe French national program of research inhydrology (PNRH).The authors thank

C.Debain (CEMAGREF)for his contribution concerningthe cooccurrenceapproach.

We wish to pay tribute to the late J.M. Robbez-Masson who died in a mountain accident. We are

deeplygratefulfor hiskindness, availabilityand for the very constructive discussions weshared withhim

concerningthis work atmeetings and on eldtrips.

References

Abdulhady, M., Abbas, H. Nassar, S., 2002, Fabric fault classication using neural trees. Proceedings

of the IEEE International Conference on Systems,Man and Cybernetics, 6,114117.

Alonso,F.,Algorri, M.E.Flores-Mangas,F.,2004,Compositeindexfor thequantitativeevaluation

of image segmentation results. Proceedings of the 26th Annual International Conference of the IEE

EMBS, SanFrancisco, USA.

Arvis, V., Debain,C., Berducat, M. Benassi, A.,2004, Generalization ofthe cooccurrencematrix for

colour images: application to colour textureclassication. ImageAnalysis andStereology, 23, 6372.

(13)

resolution remote sensing images of vine elds, European Conference on Precision Agriculture, Berlin,

8187.

Chen,P.C.Pavlidis,T.,1979,Segmentationbytextureusingacooccurrencematrixandasplit-and-merge

algorithm, ComputerGraphes andImage Processing,10, 172182.

Couteron, P.,2002,Quantifyingchangeinpatterned semi-aridvegetationbyFourier analysisofdigitised

airphotographs, International Journal of RemoteSensing,23, 34073425.

Couteron, P. Lejeune, O., 2001, Periodic spotted patterns in semiarid vegetation explained by a

propagation-inhibition model,Journal of Ecology,89, 616628.

Gong,P.Mahler, S.A.,Biging, G.SNewburn,D.A.,2003,Vineyard identicationinanoakwoodland

landscapewithairbornedigitalcameraimagery,InternationalJournalof RemoteSensing,24(6),1303

1315.

Hall, A., Louis, J., Lamb, D., 2003, Characterising and mapping vineyard canopy using

high-spatial-resolution aerial multispectralimages, Computersand Geosciences, 29,813822.

Haralick, R.M.,Shanmugam, K.,Dinstein,I.,1973,TexturalFeatures forImageClassication, IEEE

Transaction on Systems,Man, and Cybernetics,3 (6), 610621.

Johnson,L.F.,Roczen,D.,Youkhana,S.,2001,VineyardcanopydensitymappingwithIkonossatellite

imagery,Presented at the Third International Conference on Geospatial Information in Agriculture and

Forestry,Denver(Colorado), 5-7November2001.

Morales, D.I., Moctezuma, M., Parmiggiani, F., 2003, Urban and Non Urban Area Classication

by Texture Characteristicsand Data Fusion, InternationalGeoscience andRemote Sensing Symposium

(IGARSS),6, 35043506.

Prat, S., 2002, Caractérisation de végétations régulières par télédétection à très haute résolution

spa-tiale: Application à l'analyse texturale de parcelles viticoles, DESSStatistique et Traitement du Signal,

UniversitéBlaise Pascal,Clermont-Ferrand.

Ranchin,T.,Naert,B.,Albuisson,M.,Boyer,G.,Astrand,P.,2001,AnAutomaticMethodforVine

DetectioninAirborneImageryUsingWaveletTransformandMultiresolutionAnalysis,Photogrammetric

Engineering and RemoteSensing,67(1), 9198.

Rodriguez, J. R., Miranda, D., Alvarez, C. J., 2006, Application of Satillite Images to Locate and

Inventory VineyardsintheDesignationofOrigin"Bierzo"inSpain, America SocietyofAgricultural and

Biological Engineers, 49(1), 277290.

Smutek,D.,’ára,R.,Sucharda,P.,Tjahjadi,T.,’vec,M.,2003,Imagetextureanalysisofsonograms

inchronic inammationsof thyroidgland, Ultrasound inMedicine andBiology, 29(11), 15311543.

Verma, M.S,Pratt,L., Ganesh, C.,Medina, C.,2002, Hair-MAP: Aprototype automatedsystemfor

forensichair comparison andanalysis, ForensicScience International, 129 (3),168186.

Vincini, M., Frazzi, E., Assessment of erosion-related vineyards features inTuscany by object-oriented

classication of high resolution images Proceedings of Agro Environ 2004,20-24, October 2004, Udine

(Italy).

Warner, T.A., Steinmaus, K., 2005, Spatial classication of orchards and vineyards with high spatial

resolution panchromatic imagery,Photogrammetric Engineering andRemote Sensing,71(2), 179187.

Wassenaar, T., Robbez-Masson, J.-M., Andrieux, P., Baret, F., 2002, Vineyard identication and

description of spatial crop structure by per-eld frequency analysis, International Journal of Remote

Sensing,23(17), 33113325.

Wassenaar, T., Andrieux, P., Baret, F., Robbez-Masson, J.-M., 2005, Soil surface inltration

ca-pacityclassication basedon the bi-directional reectancedistribution function sampled by aerial

pho-tographs. Thecases ofvineyardsina Mediterranean area, Catena,62,94110.

Figure

Figure 1. Localization of the study area in France and Europe.
Figure 3. Zoom on the study area. All the plots of the study area have been surveyed and integrated with their characteristics in a
Figure 4. Cooccurrence calculation distances vs orientations. Left: nearest neighbours
Figure 5. Vineyard detection using Haralick's contrast. 1) Contrast calculation on sliding window in orientations: 0
+5

Références

Documents relatifs

1) Building the search space: In order to implement these 2 algorithms, a representation of the search space under the form of a meta-graph has been built. Each node of this new

Maize and sunflower biomass estimation in southwest France using high spatial and tempo- ral resolution remote sensing data... In this context, we investigated the perspective

ANDROMEDA: Applications on various experimental data 6.1.4 Results from ANDROMEDA applied to the GJ504 SPHERE-IRDIS images In order to compare the results obtained with the

™ Results : The 500-m landscape closure index was positively correlated with higher serological incidence (p < 0.05).. Contact: *soti@teledetection.fr *vs.nev@ntropic.fr ) ) ) ) )

Then, landscape variables were defined for each pond in the area based on bibliographic knowledge of the vector ecology: a landscape closure index calculated within a 100, 500 and

Despite the lack of temporal model, the natural evolution of landscape and the evolution in- duced by the sensors, many valuable techniques exist that perform change detection from

L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des

Thresholds of the detection methods used in the study, namely spectral distance similarity (SDS), spectral correlation similarity (SCS), spectral similarity value (SSV), spectral