research article a novel approach to developing a

Research ArticleA Novel Approach to Developing a SupervisedSpatial Decision Support System for Image ClassificationA Study of Paddy Rice Investigation

Shih-Hsun Chang

Department of Information Networking and System Administration Ling Tung University Taichung 408 Taiwan

Correspondence should be addressed to Shih-Hsun Chang sschangteamailltuedutw

Received 27 May 2014 Accepted 21 July 2014 Published 20 August 2014

Academic Editor Teen-Hang Meen

Copyright copy 2014 Shih-Hsun Chang This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

Paddy rice area estimation via remote sensing techniques has been well established in recent years Texture information andvegetation indicators are widely used to improve the classification accuracy of satellite images Accordingly this study employstexture information and vegetation indicators as ancillary information for classifying paddy rice through remote sensing imagesIn the first stage the images are attained using a remote sensing technique and ancillary information is employed to increase theaccuracy of classification In the second stage we decide to construct an efficient supervised classifier which is used to evaluatethe ancillary information In the third stage linear discriminant analysis (LDA) is introduced LDA is a well-known method forclassifying images to various categories Also the particle swarm optimization (PSO) algorithm is employed to optimize the LDAclassification outcomes and increase classification performance In the fourth stage we discuss the strategy of selecting differentwindow sizes and analyze particle numbers and iteration numbers with corresponding accuracy Accordingly a rational strategyfor the combination of ancillary information is introduced Afterwards the PSO algorithm improves the accuracy rate from 8226to 8931 The improved accuracy results in a much lower salt-and-pepper effect in the thematic map

1 Introduction

Paddy rice is the major food crop in Taiwan The main con-tributions of this crop in Taiwan include regional ecofriendlyenvironment flood control and improvement of air qualityFood shortage has become a serious issue for many coun-tries The estimation of crop area is important because thisinformation is related to national food policy Thereforedeveloping a fast and accurate method for estimating croparea is desirable With the progress of spatial data surveytechniques in the geosciences massive data or informationcan be easily collected and monitoredThus the collection ofinfluencing variables of investigated target category becomescomplicated Advancements in applying spatial data technol-ogy have led to the effective approaches in the measurementof given categories from predictive models to the actualor practical remote sensing data For instance searching atarget category in image classificationmust rely on governingancillary attributes with specified rules

On the other hand paddy cultivation draws the attentionof governments around the world to problems caused byfood shortages The evaluation of paddy cultivation area maybecome a crucial problem in the near future Rice is one ofthe major crops cultivated in Taiwan and the Agricultureand Food Agency of Taiwan government has dedicated sub-stantial efforts towards the estimation of cultivated areas andcorresponding harvests One of the best possible solutions isto use satellite image data to precisely handle themanagementof paddy rice area [1] Moreover numerous studies haveendeavored to construct target GIS maps by means of remotesensing image classification [2 3]

Geographic Information System (GIS) is an extensivelyused tool for processing spatial data and displaying theresults GIS can be used to handle a variety of datasetsprovides remedial measures and aids in decision analysisGupta and Joshi [4] used GIS in assessing landslide hazardzones Recently several GIS-based approaches to assessing

Hindawi Publishing CorporationMathematical Problems in EngineeringVolume 2014 Article ID 804548 11 pageshttpdxdoiorg1011552014804548

2 Mathematical Problems in Engineering

landslides have been reported [5 6] One significant advan-tage of GIS over traditional field examination and mappingmethods is the ability to process various layers of data andcomprehensively display the results of spatial assessments Inour study GIS is used as a data-processing tool for assessingland cover categories and displaying the results

Data mining is one of the fastest growing fields in thecomputer industry Specifically classification is a task ofgrouping data with multiple attributes into relevant cate-gories Data mining has become the most valuable processfor learning the implicit knowledge among datasets Theobjective of classification herein is to classify the categoriesof paddy ricenon-paddy rice in consideration of multipleattributes of ancillary information and site conditions

Ancillary information is one of the most popular materi-als for assisting scientists to categorize green plants thus itis also suitable for our paddy rice study This information issummarized as the following

(1) Vegetation Indicator This is a simple numerical indi-cator that is widely used to analyze remote sensingmeasurements typically but not necessarily from aspace platform and to assess whether the target beingobserved contains live green vegetation The NDVIABI MSAVI and RVI indicators are applied in thisstudyThus the goal of collecting these attributes is toextract the most influential attributes which lead tothe best discernibility

(2) Geostatistical Indices One of the key factors in geo-statistical modeling is the semivariogram a functiondescribing the spatial dependence of the spatial vari-able The semivariogram has been widely used inremote sensing to determine spatial structures [7 8]In general a semivariogram is employed as a tool tomodel the spatially varying phenomenon of naturalland covers We incorporate two semivariograms(1) direct semivariogram and (2) semimadogramA direct semivariogram model constructed fromknown physical properties is commonly used tomeasure general texture material in remote sensingThe semimadogram is a texture measure commonlydefined as half of the average absolute differencebetween pairs of points separated by a given vectorPlease refer to the work of Chica-Olmo and Abarca-Hernandez [9] for details In addition this studyemploys gray-level cooccurrence matrices (GLCM)proposed byHaralick et al [10] GLCM is defined overan image as the distribution of cooccurring values of agiven offset In this study four different GLCM valuesare employed as additional texture information

In general the use of a supervised classifier shouldconsider two crucial points (a) proper training samplesand (b) an effective learning process As a matter of factengineers and scientists encounter obstacles to attain theaforementioned samples and rules Thus an effective knowl-edge classifier can lead to an interesting solution for imageclassification This is the goal of this study When applyingparticle swarm optimization (PSO) on LDA three questionsarise

(1) Iteration Number How many iteration numbers arerequired to obtain acceptable accuracy

(2) Number of Particles In PSO the number of particlesneeds to be initialized before starting calculationHence different numbers of particles should be testedto approach the best performance

(3) Attribute Extraction Different combinations of vari-ables are used based upon PSO in which the classi-fication error rate must be examined in each epochAccordingly the error matrices are calculated withregard to a combination of variables The fitnessnumber of each variable should be computed througheach epoch

The uncertainty of an image classification problem maybe produced by deficiencies in the description of variouscategories and feature spaces [11 12] To resolve this prob-lem extensive studies have carried out the augmentation ofancillary information to improve classification accuracy Toimprove the quality of classification results many scientistshave used supervised classifiers (MLH ANN and FuzzyClassifier) to tackle image-processing problems [9 10 13ndash18] Specifically some justification of the use of a variogramand GLCM as texture measures for the optimization ofLDA approaches would be useful as they play a key part inthe procedures outlined [19] Our solution is to develop anenhanced supervised classifier in our rice decision supportsystem

The study is divided into four parts The first partdiscusses the development of vegetation indicators and geo-statistical indices for the study area In the second part thetraditional LDA method is introduced The third part brieflyintroduces the use of PSO to reduce the dimensions of theattributes in LDA (so called PSOLDA)The fourth part showsthe results of a parallel analysis through the (a) LDA methodand (b) PSOLDA method

2 Materials and Methods

21 Study Area and Materials An area of paddy rice (locatedin Taichung Taiwan) is selected as a case study to demon-strate the plan of this research The study area is locatedat Tanzi County Taichung Taiwan The study region hascomplex categories of paddy rice grass bare land buildingsasphalt roads water bodies and others as shown in Figure 1Figure 2 shows a Quickbird image with 2096 times 2096 pixelsrepresenting an area of 215 ha Figure 3 expresses the distri-bution of samples It includes training samples and learningsamples A 7 by 7 moving window is used to calculate thetexture values of the subsamples of the image data The sizeof the moving window and the spectral bands selected tocalculate the texture measures GLCM and variogram aredetermined by statistical methods More detailed informa-tion and results are presented in Section 3 The goal of ourstudy is to develop an effective supervised classifier thatemploys the above spatial variables in the decision supportsystem

Mathematical Problems in Engineering 323∘09984000998400998400N

24∘09984000998400998400N

25∘09984000998400998400N

23∘09984000998400998400N

24∘09984000998400998400N

25∘09984000998400998400N

120∘09984000998400998400E 121

∘09984000998400998400E 122

∘09984000998400998400E

120∘09984000998400998400E 121

∘09984000998400998400E 122

∘09984000998400998400E

70000 14000 28000 42000 56000(km)

N

W

S

E

Figure 1 The study area (Tanzi County Taichung City Taiwan)

Figure 2 Quickbird RS image of the study area

22 Material Preprocessing and Study Steps The inputs forour decision support classifier include five major steps (1)executing image fusion (2) employing ancillary information(3) selection of windows size (4) selecting proper trainingand testing datasets and (5) developing a PSO + LDA modelfor comparison These steps are described as follows

Step 1 (image fusionmdashcombine spectrum image and panchro-matic image) The Quickbird image resolutions of the spec-tral bands are 288mThedrawback of this resolution is that itcannot provide any adequate information for distinguishing

Figure 3 Sample distribution on the studied map (blue circle ispaddy rice and red x is non-paddy rice)

vegetation categories such as grass and paddy rice To attaina higher resolution of image data on the previous studymaterial we combine the image data with some ancillaryinformation In this study we integrate a multispectral image(with a resolution of 288m) with a higher spatial resolutionpanchromatic image (with a resolution of 069m) fromQuickbird by using ERDAS image software with the use ofthe PCA (principal component analysis) method

Step 2 (ancillary informationmdashreinforce better classificationperformance) In addition to spectral information a seriesof vegetation indices are included in the building of ourclassifier Furthermore to improve the classification accuracyof land covers with close spectral measures such as grassand paddy rice the spatial structures measures are includedIn this study they are GLCM contrast GLCM homogeneityGLCM energy GLCM entropy a direct semivariogram and asemimadogram Please refer to Table 2 for all the conditionalattributes used in this study

Step 3 (selection of window size) We propose an approachwhich resolves the problem of varying window size selectionfor a wide class of classifiers Window size is consideredas a variable estimation and testing a series of differentwindow sizes can lead to a better understanding of windowsize selection The texture measures were calculated fordifferent window sizes land covers and spectral bands Allsamples which include various land covers were used inthe calculation to attain the mean texture values and thenthey were depicted in figures for the sake of comparisonWe present a number of results which demonstrate how thewindow size rules were selected in our study cases

Step 4 (preparing training and testing datasets) The trainingdataset consists of 455 sample points which are comprisedof 135 paddy rice samples and 320 non-paddy rice samplesPlease refer to Table 1 for the distribution of samples withvarious land cover types These data are input into ourenhanced decision support system in the training processFollowing this process all of the image data are classifiedinto two categories (paddy and non-paddy rice) The 119870-fold


Table 1 Number of samples distribution over various land covers

Land cover

Paddy rice Levee Grass Woods Dryfarmland Road Building Shadow

Number of samples 135 50 46 45 40 42 56 41

Table 2 All conditional attributes used in this study

Numbering1 2 3 4 5 6 7 8 9

Attribute R G B IR NDVI CFMI BR SQBR VINumbering

10 11 12 13 14 15 16 17 18

Attribute SAVI MSAVI ABI GLCMcontrast

GLCMenergy

GLCMhomogeneity

GLCMentropy

Directsemivariogram Semimadogram

cross-validation method was applied We used 119896 = 5 in ourstudy which means 80 of the sample dataset was randomlyselected for training and the remaining 20 was used forvalidation The value of each cell on the error matrix (Tables3 4 5 6 7 8 and 9) was obtained by averaging the 20 timesof the aforementioned 119896-fold cross-validation calculation

Step 5 (develop a PSO + LDA computer program) In thepresent study the weight coefficients of the LDA equationwere obtained through the training dataset by using theMATLAB code we developed This equation serves as aclassification rule This rule can be used to determine theclass of land cover of each pixel in the RS image The PSOalgorithm was incorporated into the LDA code to optimizethe classification outcome by selecting different attributecombinations

23 Research Method

231 Particle SwarmOptimization Particle swarmoptimiza-tion is a group intelligence optimization method proposedby Kennedy and Eberhart in 1995 [20] This method hasbeen successfully applied in many areas It is inspired bybird flocking behaviors in which a temporary destination isdetermined by the cognition and global direction of the entiregroup In PSO a population of particles is created and eachparticle is assigned with an initial position and velocity Eachparticle moves to a new position in each calculation iterationwith regard to the value of fitness function The particlemovement is based on individual best fitness and the grouprsquosbest fitness Assume in a D-dimensional space that there are119899 particles described by 119883 = (119883

1 1198832 119883

119899) where 119883

119894=

(1199091198941

1199091198942

119909119894119863

)119879 denotes the position of the 119894th particleTheposition of the particles is the potential solution in questionThe velocity of the 119894th particle is 119881

119894= (1198811198941

1198811198942

119881119894119863

)119879 Thebest individual position and best global position with regardto optimizing fitness are 119875

119894= (1198751198941

1198751198942

119875119894119863

)119879 and 119875119892=

(1198751198921

1198751198922

119875119892119863

)119879 respectively The velocity and position of

each particle are updated in each iteration with the followingequations

119881119896+1

119894119889

= 120596119881119896

119894119889

+ 11988811199031(119875119896

119894119889

minus 119883119896

119894119889

) + 11988821199032(119875119896

119892119889

minus 119883119896

119894119889

) (1)

119883119896+1

119894119889

= 119883119896

119894119889

+ 119881119896+1

119894119889

(2)

where 119889 = 1 2 119863 119894 = 1 2 119899 119896 is the current iterationstep 120596 is the inertial weight 119888

1represents the cognition

learning factor 1198882denotes the social learning factor and 119903

1

and 1199032are random numbers

The basic steps of the PSO algorithm can be described asfollows

Step 1 create a number of particles assigned withinitial positions and velocitiesStep 2 calculate the fitness of each particleStep 3 calculate the velocity of each particle using (1)Step 4 update the position of each particle using (2)Step 5 stop the iteration process if termination crite-rion is met otherwise return to Step 2 and continuethe process

In this study the PSO algorithm is used to accomplishfeature selection The fitness function is the function thatreturns classification accuracy through the LDA algorithmThe fitness function 119869 is defined as the summation ofEuclidean distance between the data points to its associatedgroup center Consider

119869 =

119872

sum119894=1

sum119883isin120596119894

10038171003817100381710038171003817119883 minus 119883120596119897

10038171003817100381710038171003817

2

(3)

where119872 is the number of classes 120596119894is a specific class 119883 is

the vector of data points and119883120596119897 is the center of classThe feature of each training sample acts as a position

variable 119909119894119889

and its value is normalized and is bound to be[0 1] The result of the particle position after PSO process isexamined and those features with 119909

119894119889

lt 05 are discardedDetailed illustrated examples can be found in work of Lin andChen [21]

Mathematical Problems in Engineering 5

Table 3 Error matrix for all land cover types and 4 spectral bands (data is obtained by averaging 20 trials of calculation)

All land cover typesRGB + IR

Kappa = 05950Classification result

Paddy rice Non-paddy rice Producer accuracy

Ground truth classPaddy rice 2275 425 08426

Non-paddy rice 1249 5151 08048

User accuracy 06456 09238 Overall accuracy08160 std = 00038

Table 4 Error matrix for all land cover types and 4 spectral bands + NDVI (data is obtained by averaging 20 trials of calculation)

All land cover typesRGB + IR + NDVI






Table 5 Error matrix for all land cover types and 4 spectral bands + NDVI + texture (data is obtained by averaging 20 trials of calculation)

All land cover typesRGB + IR + NDVI + texture






Table 6 Non-PSO versus PSO error matrix for all land cover types

All land cover typesNon-PSO Kappa = 05982

PSO Kappa = 07510Classification result (non-PSOPSO)


Ground truth class

Paddy rice 21622348 538352 0800708696Non-paddy rice 1076621 53245779 0831909030

User accuracy 0667707908 0908209426

Overall accuracy08226

std = 00380overallaccuracy

08931 std = 00048

Table 7 Non-PSO versus PSO error matrix for paddy rice versus grass

Land cover paddy rice versus grassNon-PSO Kappa = 04528


Paddy rice Grass Producer accuracy

Ground truth class

Paddy rice 23082620 39208 0854809704Grass 366255 554665 0602207228

User accuracy 0863109113 0585608926Overall accuracy07906 std =

00598overall accuracy09075 std = 00073


Table 8 Non-PSO versus PSO error matrix for paddy rice versus levee

Land cover paddy rice versus leveeNon-PSO Kappa = 08502


Paddy rice Levee Producer accuracy

Ground truth class

Paddy rice 26112611 089089 0967009670Levee 127080 873920 0873009200



Table 9 Non-PSO versus PSO error matrix for paddy rice versus woods

Land cover paddy rice versus woodsNon-PSO Kappa = 07716


Paddy rice Woods Producer accuracy

Ground truth class

Paddy rice 25672567 089089 0950709507Woods 171171 729729 0810008100

User accuracy 0937509375 0845708457


std=00443overallaccuracy

09156 std = 00443

232 Linear Discriminant Analysis (LDA) Linear discrimi-nant analysis (LDA) is a popular statistical method used forclassification In general it is composed of linear discrim-inant equations which are obtained by definite and simpleprocedures Due to the contribution of recent advancesin satellite photography technology high resolution imagesare now well accepted for analysis However uncertaintyinformationmay exist in such images leading to the decreaseof classification accuracy It is thus expected that with theefforts of attribute reduction and the data preprocessing ofraw data the classification accuracy of satellite images can beprofoundly improved

3 Results and Discussions

31 Selecting Window Sizes and Spectral Bands When Calcu-lating Texture Information In this study GLCM and semi-variogram texture information are a part of the conditionattributes To attain these data it is required to determinethe size of the moving window and which spectral bandshould be used for calculation Four GLCM attributesincluding contrast homogeneity energy and entropy arecalculated for different window sizes at each sample pixelTheir mean values are obtained by averaging the samples oftheir corresponding land covers The GLCM versus windowsize distributions (spectral band R and IR) are shown inFigures 4 and 5 respectively The 119909-axis is the window sizeand the 119910-axis is the value of GLCM texture informationUnder different curves (land covers) the larger the separationbetween curves the better the discernibility rate It is seenfrom the figures that the IR band has better discernibility than

does the R band Paddy rice and grass are types of land coverthat are close in spectral distributions and thus are difficultto classify For the paddy rice field and grass the IR bandobviously depicts higher discernibility The distributions forthe B and G bands were also obtained However they donot present better discernibility than the IR band does andare thus not shown in the figures Accordingly the IR bandis selected for calculating GLCM texture information in ourstudy Similar distributions for semivariogram texture casesare shown in Figures 6 and 7 In this case it is depicted that theR band depicts better discernibility Therefore the R band isselected for calculating semivariogram texture information inour study As seen in Figures 4ndash7 although the larger windowsize cases tend to have better discernibility they may includemore uncertainties and probably enclose pixels of other landcovers Accordingly we decide to select window size 7 in ourstudy In summary the IR band is used for calculatingGLCMthe R band is used for calculating semivariogram and thewindow size is 7

32 Effects of Ancillary Attributes To study the effective-ness of ancillary attributes NDVI and texture information(GLCM and semivariogram) and 3 different conditionalattribute combinations are used to obtain the classificationoutcomes Table 3 presents the error matrix with only 4 spec-tral bands as conditional attributes Table 4 adds NDVI as anadditional conditional attribute Table 5 further adds textureinformation (4 GLCM data and 2 semivariogram data) asadditional attributes All land cover types are included in theabove calculation ComparingTables 3 4 and 5 it is seen thatwith the inclusion of ancillary attributes the classification


3 4 5 6 7 8 9 10 11

005

01

015

02

025

Window size

Mea

n G

LCM

cont

rast

GLCM contrast distribution for various land covers (band R)

(a)

3 4 5 6 7 8 9 10 11

088

09

092

094

096

098

Window size

Mea

n G

LCM

hom

ogen

eity

GLCM homogeneity distribution for various land covers (band R)

(b)

Window size3 4 5 6 7 8 9 10 11

065

07

075

08

085

09

Mea

n G

LCM

ener

gy

GLCM energy distribution for various land covers (band R)

Paddy riceLevees

GrassWoods

(c)

Window size3 4 5 6 7 8 9 10 11

0

02

04

06

08

1

12

14

Mea

n G

LCM

entro

py

GLCM entropy distribution for various land covers (band R)

Paddy riceLevees

GrassWoods

(d)

Figure 4 GLCM versus window size for different land covers (band R)

Table 10 Dropped attributes after PSO process

Land cover All land covertypes Paddy rice versus grass Paddy rice versus levee Paddy rice versus

woods

Droppedattributes IR VI ABI

R G NDVI CMFI SQBRSAVI MSAVI ABI GLCM

energy GLCMhomogeneity directsemivariogram

BR VI MSAVI GLCMcontrast GLCM energy None

outcome improves The overall accuracy rates are increasedfrom 8160 to 8343 with NDVI included and to 8501with texture information included

33 PSO Attribute Reduction with LDA

331 Determining the Number of Particles and MaximalEpochs in PSOLDA This study incorporates PSO with LDAas an optimization tool to find the best combination ofconditional attributes This kind of approach is generallyreferred to as an attribute extraction process in data miningPSO is an iterative calculation process in which it is necessaryto set up initial conditions The inertial weight 120596 is set to 10and the cognition learning factor and social learning factor1198881and 119888

2 are both set to 08 However two other initial

conditions maximal epoch number and number of particlesmust also be determined Figure 8 depicts the iterationevolutions for various maximal epoch numbers (the particlenumber is fixed at 30) Figure 8 shows that a higher maximalepoch does not always lead to a lower classification errorrate Figure 9 depicts the iteration evolutions for differentnumbers of particles Similarly a higher number of particlesset up in initial condition do not always lead to a lowerclassification error rate The classification error rate alwaysvaries at different discrete values This is due to the fact thatthe nature of the problem in study is an optimal combina-tion of attributes which is inherently discrete Consideringthe balance between classifier optimization and additionalrequired computing time themaximal epoch is set to 400 andthe number of particles is set to 40 in the rest of the study


3 4 5 6 7 8 9 10 11

0

005

01

015

02

025

03

035

Mea

n G

LCM

cont

rast

Window size

GLCM contrast distribution for various land covers (band IR)

(a)

3 4 5 6 7 8 9 10 11

088

09

092

094

096

098

1

Window size

Mea

n G

LCM

hom

ogen

eity

GLCM homogeneity distribution for various land covers (band IR)

(b)

Window size3 4 5 6 7 8 9 10 11

065

07

075

08

085

09

095

1

Mea

n G

LCM

ener

gy

GLCM energy distribution for various land covers (band IR)

Paddy riceLevees

GrassWoods

(c)

3 4 5 6 7 8 9 10 11

0

02

04

06

08

1

12

14

16

18

Window size

Mea

n G

LCM

entro

py

GLCM entropy distribution for various land covers (band IR)

Paddy riceLevees

GrassWoods

(d)

Figure 5 GLCM versus window size for different land covers (band IR)

332 Classification Outcome with and without PSO AttributeReduction PSO is used as a dimension reduction techniqueon LDA equations in this study To compare the results ofthe PSOLDA classifier with those obtained with the LDAclassifier only 4 different initial conditions are studiedincluding all land cover types paddy rice field versus grasspaddy rice field versus levee and paddy rice versus woodsThe number of conditional attributes listed in Table 2 beforeapplying PSO is 18 and is reduced to a smaller number 15after applying PSO Tables 6ndash9 present the error matricesof non-PSO and PSO cases with different land covers Thebold face numbers are obtained through PSO It is clear fromTables 6ndash8 that with attribute reduction incorporating PSOthe classification outcomes are improved However as seenin Table 9 in the paddy rice versus woods case no attributesare eliminated after incorporating PSOTheoverall land covercase shows an 857 accuracy improvement and the caseof paddy rice versus grass reveals 1478 improvement Theeliminated attributes are summarized in Table 10 It is notedthat some of the attributes used in this study are correlatedtherefore the dropped attributes could still be influentialin classifying land covers PSO here serves to optimize theclassification outcome by employing different combinationsof attributes For various correlated attributes such as NDVI

versus R and IR it is possible that either one of them couldbe extracted or eliminated under the PSO process

34 Thematic Map Comparison of Non-PSO versus PSO Athematic map is useful for visually examining the perfor-mance of the developed classifier and estimating the areaof paddy rice field The classification outcome discussed inprevious section presents better results by using PSOLDAas compared with using LDA alone To further examinethe benefit of incorporating PSOLDA two thematic mapsFigures 10 and 11 are generated for non-PSO and PSO for thesake of comparisonThese two figures are created by using theclassifiers (LDA versus PSOLDA) through the same trainingdatasets Salt-and-pepper effect can be easily observed inFigure 10 Much lower salt-and-pepper effect is depicted inthe PSO case as shown in Figure 11 This is because the PSOcase has a higher overall classification accuracy which resultsin fewer misclassified points on the thematic map

4 Conclusion

Rice is a crop of global importance Thus remote sensingtechniques have been applied for evaluating its productionThis study combines PSO with LDA as an optimization tool


3 4 5 6 7 8 9 10 110

1

2

3

4

5

6

7

8

9

Window size

Mea

n di

rect

sem

ivar

iogr

am

Direct semivariogram distribution for various land covers (band R)

Paddy riceLevees

GrassWoods

(a)

3 4 5 6 7 8 9 10 11

04

05

06

07

08

09

1

11

12

13

Window size

Mea

n ab

solu

te se

miv

ario

gram

Absolute semivariogram distributionfor various land covers (band R)

Paddy riceLevees

GrassWoods

(b)

Figure 6 Semivariogram versus window size for different land covers (band R)

3 4 5 6 7 8 9 10 110

2

4

6

8

10

12

14

16

18

Window size

Mea

n di

rect

sem

ivar

iogr

am

Direct semivariogram distributionfor various land covers (band IR)

Paddy riceLevees

GrassWoods

(a)

3 4 5 6 7 8 9 10 1104

06

08

1

12

14

16

18

2

22

24

Window size

Mea

n ab

solu

te se

miv

ario

gram

Absolute semivariogram distributionfor various land covers (band IR)

Paddy riceLevees

GrassWoods

(b)

Figure 7 Semivariogram versus window size for different land covers (band IR)


0 100 200 300 400 500 600 700 800

008

0085

009

0095

01

0105

011

Number of epoch

Clas

sifica

tion

erro

r rat

e

Effect of maximal epoch in PSOLDA iteration evolution

Max epoch = 50

Max epoch = 50

Max epoch = 100

Max epoch = 100

Max epoch = 200

Max epoch = 200

Max epoch = 300

Max epoch = 300

Max epoch = 400

Max epoch = 400

Max epoch = 800

Max epoch = 800

Figure 8 Effect of maximal epoch in PSOLDA iteration evolution

0 50 100 150 200 250 300 350 400

0075

008

0085

009

0095

01

0105

011

0115

Number of epoch

Clas

sifica

tion

erro

r rat

e

Effect of particle number in PSOLDA iteration evolution

Particle number = 20








Figure 9 Effect of particle number in PSOLDA iteration evolution

for finding the best combination of conditional attributesFive conclusions are made

(1) This study proposes a method PSOLDA to improveclassification accuracy in remote-sensing image clas-sification tasks

(2) This study presents a process to select window sizesand spectral bands for calculating texture variables In

Figure 10 Paddy rice thematic map using merely LDA

Figure 11 Paddy rice thematic map applying PSO on LDA

this study for GLCM texture the IR band has betterdiscernibility than does the R band however forsemivariogram the R band has better discernibilitythan the IR band does Larger window size cases tendto have better discernibility but may include moreuncertainties and probably enclose pixels of otherland covers

(3) Incorporating ancillary attributes such as vegetationindices and texture measures helps to improve classi-fication accuracy

(4) By incorporating PSO into LDA the number ofattributes is reduced and classification accuracy isimproved The proposed method leads to accuracyimprovements of 857 and 1478 in the overallland cover case and paddy rice versus grass caserespectively

(5) Applying PSOLDA greatly reduces the salt-and-pepper effect in the thematic map when comparedwith merely applying LDA

Conflict of Interests

The author declares that there is no conflict of interestsregarding the publication of this paper


References

[1] Y Shao X Fan H Liu et al ldquoRice monitoring and productionestimation using multitemporal RADARSATrdquo Remote Sensingof Environment vol 76 no 3 pp 310ndash325 2001

[2] A Cheriyadat and L M Bruce ldquoWhy principle componentanalysis is not an appropriate feature extraction method forhyperspectralrdquo in Proceedings of the IEEE Conference Geo-sciences and Remote Sensing (IGARSS rsquo03) pp 3420ndash3422 2003

[3] Y Tian P Guo and M R Lyu ldquoComparative studies onfeature extraction methods for multispectral remote sensingimage classificationrdquo in Proceedings of the IEEE InternationalConference on Systems Man and Cybernetics Society vol 2 pp1275ndash1279 October 2005

[4] R P Gupta and B C Joshi ldquoLandslide hazard zoning usingthe GIS approachmdasha case study from the Ramganga catchmentHimalayasrdquo Engineering Geology vol 28 no 1-2 pp 119ndash1311990

[5] B Pradhan ldquoA comparative study on the predictive ability of thedecision tree support vector machine and neuro-fuzzy modelsin landslide susceptibility mapping using GISrdquo Computers andGeosciences vol 51 pp 350ndash365 2013

[6] W Huabin L Gangjun XWeiya andW Gonghui ldquoGIS-basedlandslide hazard assessment an overviewrdquo Progress in PhysicalGeography vol 29 no 4 pp 548ndash567 2005

[7] P J Curran ldquoThe semivariogram in remote sensing an intro-ductionrdquoRemote Sensing of Environment vol 24 no 3 pp 493ndash507 1988

[8] P M Atkinson and P Lewis ldquoGeostatistical classification forremote sensing an introductionrdquo Computers and Geosciencesvol 26 no 4 pp 361ndash371 2000

[9] M Chica-Olmo and F Abarca-Hernandez ldquoComputing geo-statistical image texture for remotely sensed data classificationrdquoComputers and Geosciences vol 26 no 4 pp 373ndash383 2000

[10] R M Haralick K Shaunmmugam and I Dinstein ldquoTexturalfeatures for image classificationrdquo IEEE Transaction SystemMan and Cybernetics vol 67 pp 786ndash804 1973

[11] Z Sun D Huang Y Cheung J Liu and G Huang ldquoUsingFCMC FVS and PCA techniques for feature extraction ofmultispectral imagesrdquo IEEE Geoscience and Remote SensingLetters vol 2 no 2 pp 108ndash112 2005

[12] C Wang M Menenti and Z Li ldquoModified Principal Compo-nent Analysis (MPCA) for Feature Selection of HyperspectralImageryrdquo in Proceeding of the IEEE Conference Geoscience andRemote Sensing Symposium (IGARSS 03) vol 6 pp 3781ndash3783July 2003

[13] P J Sellers ldquoCanopy reflectance photosynthesis and transpira-tionrdquo International Journal of Remote Sensing vol 6 no 8 pp1335ndash1372 1985

[14] S Yu S D Backer and P Scheunders ldquoGenetic feature selectioncombined with composite fuzzy nearest neighbor classifiersfor high-dimensional remote sensing datardquo IEEE InternationalConference on Systems Man and Cybernetics vol 3 no 3 pp1912ndash1916 2000

[15] NKosaka SMiyazaki andU Inoue ldquoVegetable green coverageestimation from an airborne hyperspectral imagerdquo in Proceed-ings of the IEEE International Geoscience and Remote SensingSymposium (IGARSS rsquo02) pp 1959ndash1961 June 2002

[16] T Y Chou T C Lei S Wan and L S Yang ldquoSpatial knowledgedatabases as applied to the detection of changes in urban landuserdquo International Journal of Remote Sensing vol 26 no 14 pp3047ndash3068 2005

[17] H Fang S LiangM PMcClaran et al ldquoBiophysical characteri-zation andmanagement effects on semiarid rangeland observedfrom landsat ETM+ datardquo IEEE Transactions on Geoscience andRemote Sensing vol 43 no 1 pp 125ndash133 2005

[18] S Wan T C Lei and T Y Chou ldquoAn enhanced supervisedspatial decision support system of image classification con-sideration on the ancillary information of paddy rice areardquoInternational Journal of Geographical Information Science vol24 no 4 pp 623ndash642 2010

[19] C D Lloyd S Berberoglu P J Curran and P M Atkinson ldquoAcomparison of texture measures for the per-field classificationof Mediterranean land coverrdquo International Journal of RemoteSensing vol 25 no 19 pp 3943ndash3965 2004

[20] J Kennedy and R Eberhart ldquoParticle swarm optimizationrdquoin Proceedings of the IEEE International Conference on NeuralNetworks vol 4 pp 1942ndash1948 IEEE Service Center PerthAustralia 1995

[21] S Lin and S Chen ldquoPSOLDA a particle swarm optimizationapproach for enhancing classification accuracy rate of lineardiscriminant analysisrdquo Applied Soft Computing Journal vol 9no 3 pp 1008ndash1015 2009

Submit your manuscripts athttpwwwhindawicom

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

MathematicsJournal of


Mathematical Problems in Engineering

Hindawi Publishing Corporationhttpwwwhindawicom

Differential EquationsInternational Journal of

Volume 2014

Applied MathematicsJournal of


Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal of


Mathematical PhysicsAdvances in

Complex AnalysisJournal of


OptimizationJournal of


CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of


Operations ResearchAdvances in

Journal of


Function Spaces

Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of Mathematics and Mathematical Sciences


The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014


Algebra

Discrete Dynamics in Nature and Society



Decision SciencesAdvances in

Discrete MathematicsJournal of


Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Stochastic AnalysisInternational Journal of


landslides have been reported [5 6] One significant advan-tage of GIS over traditional field examination and mappingmethods is the ability to process various layers of data andcomprehensively display the results of spatial assessments Inour study GIS is used as a data-processing tool for assessingland cover categories and displaying the results

Data mining is one of the fastest growing fields in thecomputer industry Specifically classification is a task ofgrouping data with multiple attributes into relevant cate-gories Data mining has become the most valuable processfor learning the implicit knowledge among datasets Theobjective of classification herein is to classify the categoriesof paddy ricenon-paddy rice in consideration of multipleattributes of ancillary information and site conditions

Ancillary information is one of the most popular materi-als for assisting scientists to categorize green plants thus itis also suitable for our paddy rice study This information issummarized as the following

(1) Vegetation Indicator This is a simple numerical indi-cator that is widely used to analyze remote sensingmeasurements typically but not necessarily from aspace platform and to assess whether the target beingobserved contains live green vegetation The NDVIABI MSAVI and RVI indicators are applied in thisstudyThus the goal of collecting these attributes is toextract the most influential attributes which lead tothe best discernibility

(2) Geostatistical Indices One of the key factors in geo-statistical modeling is the semivariogram a functiondescribing the spatial dependence of the spatial vari-able The semivariogram has been widely used inremote sensing to determine spatial structures [7 8]In general a semivariogram is employed as a tool tomodel the spatially varying phenomenon of naturalland covers We incorporate two semivariograms(1) direct semivariogram and (2) semimadogramA direct semivariogram model constructed fromknown physical properties is commonly used tomeasure general texture material in remote sensingThe semimadogram is a texture measure commonlydefined as half of the average absolute differencebetween pairs of points separated by a given vectorPlease refer to the work of Chica-Olmo and Abarca-Hernandez [9] for details In addition this studyemploys gray-level cooccurrence matrices (GLCM)proposed byHaralick et al [10] GLCM is defined overan image as the distribution of cooccurring values of agiven offset In this study four different GLCM valuesare employed as additional texture information

In general the use of a supervised classifier shouldconsider two crucial points (a) proper training samplesand (b) an effective learning process As a matter of factengineers and scientists encounter obstacles to attain theaforementioned samples and rules Thus an effective knowl-edge classifier can lead to an interesting solution for imageclassification This is the goal of this study When applyingparticle swarm optimization (PSO) on LDA three questionsarise

(1) Iteration Number How many iteration numbers arerequired to obtain acceptable accuracy

(2) Number of Particles In PSO the number of particlesneeds to be initialized before starting calculationHence different numbers of particles should be testedto approach the best performance

(3) Attribute Extraction Different combinations of vari-ables are used based upon PSO in which the classi-fication error rate must be examined in each epochAccordingly the error matrices are calculated withregard to a combination of variables The fitnessnumber of each variable should be computed througheach epoch

The uncertainty of an image classification problem maybe produced by deficiencies in the description of variouscategories and feature spaces [11 12] To resolve this prob-lem extensive studies have carried out the augmentation ofancillary information to improve classification accuracy Toimprove the quality of classification results many scientistshave used supervised classifiers (MLH ANN and FuzzyClassifier) to tackle image-processing problems [9 10 13ndash18] Specifically some justification of the use of a variogramand GLCM as texture measures for the optimization ofLDA approaches would be useful as they play a key part inthe procedures outlined [19] Our solution is to develop anenhanced supervised classifier in our rice decision supportsystem

The study is divided into four parts The first partdiscusses the development of vegetation indicators and geo-statistical indices for the study area In the second part thetraditional LDA method is introduced The third part brieflyintroduces the use of PSO to reduce the dimensions of theattributes in LDA (so called PSOLDA)The fourth part showsthe results of a parallel analysis through the (a) LDA methodand (b) PSOLDA method

2 Materials and Methods

21 Study Area and Materials An area of paddy rice (locatedin Taichung Taiwan) is selected as a case study to demon-strate the plan of this research The study area is locatedat Tanzi County Taichung Taiwan The study region hascomplex categories of paddy rice grass bare land buildingsasphalt roads water bodies and others as shown in Figure 1Figure 2 shows a Quickbird image with 2096 times 2096 pixelsrepresenting an area of 215 ha Figure 3 expresses the distri-bution of samples It includes training samples and learningsamples A 7 by 7 moving window is used to calculate thetexture values of the subsamples of the image data The sizeof the moving window and the spectral bands selected tocalculate the texture measures GLCM and variogram aredetermined by statistical methods More detailed informa-tion and results are presented in Section 3 The goal of ourstudy is to develop an effective supervised classifier thatemploys the above spatial variables in the decision supportsystem


24∘09984000998400998400N

25∘09984000998400998400N

23∘09984000998400998400N

24∘09984000998400998400N

25∘09984000998400998400N

120∘09984000998400998400E 121

∘09984000998400998400E 122

∘09984000998400998400E

120∘09984000998400998400E 121

∘09984000998400998400E 122

∘09984000998400998400E

70000 14000 28000 42000 56000(km)

N

W

S

E












Land cover




Numbering1 2 3 4 5 6 7 8 9


10 11 12 13 14 15 16 17 18


GLCMenergy

GLCMhomogeneity

GLCMentropy




23 Research Method


1 1198832 119883

119899) where 119883

119894=

(1199091198941

1199091198942

119909119894119863


119894= (1198811198941

1198811198942

119881119894119863


119894= (1198751198941

1198751198942

119875119894119863

)119879 and 119875119892=

(1198751198921

1198751198922

119875119892119863



119881119896+1

119894119889

= 120596119881119896

119894119889

+ 11988811199031(119875119896

119894119889

minus 119883119896

119894119889

) + 11988821199032(119875119896

119892119889

minus 119883119896

119894119889

) (1)

119883119896+1

119894119889

= 119883119896

119894119889

+ 119881119896+1

119894119889

(2)




1





119869 =

119872

sum119894=1

sum119883isin120596119894

10038171003817100381710038171003817119883 minus 119883120596119897

10038171003817100381710038171003817

2

(3)



variable 119909119894119889


119894119889




























Ground truth class


User accuracy 0667707908 0908209426



08931 std = 00048





Ground truth class

Paddy rice 23082620 39208 0854809704Grass 366255 554665 0602207228








Ground truth class

Paddy rice 26112611 089089 0967009670Levee 127080 873920 0873009200







Ground truth class

Paddy rice 25672567 089089 0950709507Woods 171171 729729 0810008100

User accuracy 0937509375 0845708457



09156 std = 00443







3 4 5 6 7 8 9 10 11

005

01

015

02

025

Window size

Mea

n G

LCM

cont

rast


(a)

3 4 5 6 7 8 9 10 11

088

09

092

094

096

098

Window size

Mea

n G

LCM

hom

ogen

eity


(b)

Window size3 4 5 6 7 8 9 10 11

065

07

075

08

085

09

Mea

n G

LCM

ener

gy


Paddy riceLevees

GrassWoods

(c)

Window size3 4 5 6 7 8 9 10 11

0

02

04

06

08

1

12

14

Mea

n G

LCM

entro

py


Paddy riceLevees

GrassWoods

(d)




woods











3 4 5 6 7 8 9 10 11

0

005

01

015

02

025

03

035

Mea

n G

LCM

cont

rast

Window size


(a)

3 4 5 6 7 8 9 10 11

088

09

092

094

096

098

1

Window size

Mea

n G

LCM

hom

ogen

eity


(b)

Window size3 4 5 6 7 8 9 10 11

065

07

075

08

085

09

095

1

Mea

n G

LCM

ener

gy


Paddy riceLevees

GrassWoods

(c)

3 4 5 6 7 8 9 10 11

0

02

04

06

08

1

12

14

16

18

Window size

Mea

n G

LCM

entro

py


Paddy riceLevees

GrassWoods

(d)





4 Conclusion



3 4 5 6 7 8 9 10 110

1

2

3

4

5

6

7

8

9

Window size

Mea

n di

rect

sem

ivar

iogr

am


Paddy riceLevees

GrassWoods

(a)

3 4 5 6 7 8 9 10 11

04

05

06

07

08

09

1

11

12

13

Window size

Mea

n ab

solu

te se

miv

ario

gram


Paddy riceLevees

GrassWoods

(b)


3 4 5 6 7 8 9 10 110

2

4

6

8

10

12

14

16

18

Window size

Mea

n di

rect

sem

ivar

iogr

am


Paddy riceLevees

GrassWoods

(a)

3 4 5 6 7 8 9 10 1104

06

08

1

12

14

16

18

2

22

24

Window size

Mea

n ab

solu

te se

miv

ario

gram


Paddy riceLevees

GrassWoods

(b)



0 100 200 300 400 500 600 700 800

008

0085

009

0095

01

0105

011

Number of epoch

Clas

sifica

tion

erro

r rat

e


Max epoch = 50

Max epoch = 50

Max epoch = 100

Max epoch = 100

Max epoch = 200

Max epoch = 200

Max epoch = 300

Max epoch = 300

Max epoch = 400

Max epoch = 400

Max epoch = 800

Max epoch = 800


0 50 100 150 200 250 300 350 400

0075

008

0085

009

0095

01

0105

011

0115

Number of epoch

Clas

sifica

tion

erro

r rat

e























References





























Volume 2014




Journal of











Journal of


Function Spaces






Algebra










24∘09984000998400998400N

25∘09984000998400998400N

23∘09984000998400998400N

24∘09984000998400998400N

25∘09984000998400998400N

120∘09984000998400998400E 121

∘09984000998400998400E 122

∘09984000998400998400E

120∘09984000998400998400E 121

∘09984000998400998400E 122

∘09984000998400998400E

70000 14000 28000 42000 56000(km)

N

W

S

E












Land cover




Numbering1 2 3 4 5 6 7 8 9


10 11 12 13 14 15 16 17 18


GLCMenergy

GLCMhomogeneity

GLCMentropy




23 Research Method


1 1198832 119883

119899) where 119883

119894=

(1199091198941

1199091198942

119909119894119863


119894= (1198811198941

1198811198942

119881119894119863


119894= (1198751198941

1198751198942

119875119894119863

)119879 and 119875119892=

(1198751198921

1198751198922

119875119892119863



119881119896+1

119894119889

= 120596119881119896

119894119889

+ 11988811199031(119875119896

119894119889

minus 119883119896

119894119889

) + 11988821199032(119875119896

119892119889

minus 119883119896

119894119889

) (1)

119883119896+1

119894119889

= 119883119896

119894119889

+ 119881119896+1

119894119889

(2)




1





119869 =

119872

sum119894=1

sum119883isin120596119894

10038171003817100381710038171003817119883 minus 119883120596119897

10038171003817100381710038171003817

2

(3)



variable 119909119894119889


119894119889




























Ground truth class


User accuracy 0667707908 0908209426



08931 std = 00048





Ground truth class

Paddy rice 23082620 39208 0854809704Grass 366255 554665 0602207228








Ground truth class

Paddy rice 26112611 089089 0967009670Levee 127080 873920 0873009200







Ground truth class

Paddy rice 25672567 089089 0950709507Woods 171171 729729 0810008100

User accuracy 0937509375 0845708457



09156 std = 00443







3 4 5 6 7 8 9 10 11

005

01

015

02

025

Window size

Mea

n G

LCM

cont

rast


(a)

3 4 5 6 7 8 9 10 11

088

09

092

094

096

098

Window size

Mea

n G

LCM

hom

ogen

eity


(b)

Window size3 4 5 6 7 8 9 10 11

065

07

075

08

085

09

Mea

n G

LCM

ener

gy


Paddy riceLevees

GrassWoods

(c)

Window size3 4 5 6 7 8 9 10 11

0

02

04

06

08

1

12

14

Mea

n G

LCM

entro

py


Paddy riceLevees

GrassWoods

(d)




woods











3 4 5 6 7 8 9 10 11

0

005

01

015

02

025

03

035

Mea

n G

LCM

cont

rast

Window size


(a)

3 4 5 6 7 8 9 10 11

088

09

092

094

096

098

1

Window size

Mea

n G

LCM

hom

ogen

eity


(b)

Window size3 4 5 6 7 8 9 10 11

065

07

075

08

085

09

095

1

Mea

n G

LCM

ener

gy


Paddy riceLevees

GrassWoods

(c)

3 4 5 6 7 8 9 10 11

0

02

04

06

08

1

12

14

16

18

Window size

Mea

n G

LCM

entro

py


Paddy riceLevees

GrassWoods

(d)





4 Conclusion



3 4 5 6 7 8 9 10 110

1

2

3

4

5

6

7

8

9

Window size

Mea

n di

rect

sem

ivar

iogr

am


Paddy riceLevees

GrassWoods

(a)

3 4 5 6 7 8 9 10 11

04

05

06

07

08

09

1

11

12

13

Window size

Mea

n ab

solu

te se

miv

ario

gram


Paddy riceLevees

GrassWoods

(b)


3 4 5 6 7 8 9 10 110

2

4

6

8

10

12

14

16

18

Window size

Mea

n di

rect

sem

ivar

iogr

am


Paddy riceLevees

GrassWoods

(a)

3 4 5 6 7 8 9 10 1104

06

08

1

12

14

16

18

2

22

24

Window size

Mea

n ab

solu

te se

miv

ario

gram


Paddy riceLevees

GrassWoods

(b)



0 100 200 300 400 500 600 700 800

008

0085

009

0095

01

0105

011

Number of epoch

Clas

sifica

tion

erro

r rat

e


Max epoch = 50

Max epoch = 50

Max epoch = 100

Max epoch = 100

Max epoch = 200

Max epoch = 200

Max epoch = 300

Max epoch = 300

Max epoch = 400

Max epoch = 400

Max epoch = 800

Max epoch = 800


0 50 100 150 200 250 300 350 400

0075

008

0085

009

0095

01

0105

011

0115

Number of epoch

Clas

sifica

tion

erro

r rat

e























References





























Volume 2014




Journal of











Journal of


Function Spaces






Algebra











Land cover




Numbering1 2 3 4 5 6 7 8 9


10 11 12 13 14 15 16 17 18


GLCMenergy

GLCMhomogeneity

GLCMentropy




23 Research Method


1 1198832 119883

119899) where 119883

119894=

(1199091198941

1199091198942

119909119894119863


119894= (1198811198941

1198811198942

119881119894119863


119894= (1198751198941

1198751198942

119875119894119863

)119879 and 119875119892=

(1198751198921

1198751198922

119875119892119863



119881119896+1

119894119889

= 120596119881119896

119894119889

+ 11988811199031(119875119896

119894119889

minus 119883119896

119894119889

) + 11988821199032(119875119896

119892119889

minus 119883119896

119894119889

) (1)

119883119896+1

119894119889

= 119883119896

119894119889

+ 119881119896+1

119894119889

(2)




1





119869 =

119872

sum119894=1

sum119883isin120596119894

10038171003817100381710038171003817119883 minus 119883120596119897

10038171003817100381710038171003817

2

(3)



variable 119909119894119889


119894119889




























Ground truth class


User accuracy 0667707908 0908209426



08931 std = 00048





Ground truth class

Paddy rice 23082620 39208 0854809704Grass 366255 554665 0602207228








Ground truth class

Paddy rice 26112611 089089 0967009670Levee 127080 873920 0873009200







Ground truth class

Paddy rice 25672567 089089 0950709507Woods 171171 729729 0810008100

User accuracy 0937509375 0845708457



09156 std = 00443







3 4 5 6 7 8 9 10 11

005

01

015

02

025

Window size

Mea

n G

LCM

cont

rast


(a)

3 4 5 6 7 8 9 10 11

088

09

092

094

096

098

Window size

Mea

n G

LCM

hom

ogen

eity


(b)

Window size3 4 5 6 7 8 9 10 11

065

07

075

08

085

09

Mea

n G

LCM

ener

gy


Paddy riceLevees

GrassWoods

(c)

Window size3 4 5 6 7 8 9 10 11

0

02

04

06

08

1

12

14

Mea

n G

LCM

entro

py


Paddy riceLevees

GrassWoods

(d)




woods











3 4 5 6 7 8 9 10 11

0

005

01

015

02

025

03

035

Mea

n G

LCM

cont

rast

Window size


(a)

3 4 5 6 7 8 9 10 11

088

09

092

094

096

098

1

Window size

Mea

n G

LCM

hom

ogen

eity


(b)

Window size3 4 5 6 7 8 9 10 11

065

07

075

08

085

09

095

1

Mea

n G

LCM

ener

gy


Paddy riceLevees

GrassWoods

(c)

3 4 5 6 7 8 9 10 11

0

02

04

06

08

1

12

14

16

18

Window size

Mea

n G

LCM

entro

py


Paddy riceLevees

GrassWoods

(d)





4 Conclusion



3 4 5 6 7 8 9 10 110

1

2

3

4

5

6

7

8

9

Window size

Mea

n di

rect

sem

ivar

iogr

am


Paddy riceLevees

GrassWoods

(a)

3 4 5 6 7 8 9 10 11

04

05

06

07

08

09

1

11

12

13

Window size

Mea

n ab

solu

te se

miv

ario

gram


Paddy riceLevees

GrassWoods

(b)


3 4 5 6 7 8 9 10 110

2

4

6

8

10

12

14

16

18

Window size

Mea

n di

rect

sem

ivar

iogr

am


Paddy riceLevees

GrassWoods

(a)

3 4 5 6 7 8 9 10 1104

06

08

1

12

14

16

18

2

22

24

Window size

Mea

n ab

solu

te se

miv

ario

gram


Paddy riceLevees

GrassWoods

(b)



0 100 200 300 400 500 600 700 800

008

0085

009

0095

01

0105

011

Number of epoch

Clas

sifica

tion

erro

r rat

e


Max epoch = 50

Max epoch = 50

Max epoch = 100

Max epoch = 100

Max epoch = 200

Max epoch = 200

Max epoch = 300

Max epoch = 300

Max epoch = 400

Max epoch = 400

Max epoch = 800

Max epoch = 800


0 50 100 150 200 250 300 350 400

0075

008

0085

009

0095

01

0105

011

0115

Number of epoch

Clas

sifica

tion

erro

r rat

e























References





























Volume 2014




Journal of











Journal of


Function Spaces






Algebra



































Ground truth class


User accuracy 0667707908 0908209426



08931 std = 00048





Ground truth class

Paddy rice 23082620 39208 0854809704Grass 366255 554665 0602207228








Ground truth class

Paddy rice 26112611 089089 0967009670Levee 127080 873920 0873009200







Ground truth class

Paddy rice 25672567 089089 0950709507Woods 171171 729729 0810008100

User accuracy 0937509375 0845708457



09156 std = 00443







3 4 5 6 7 8 9 10 11

005

01

015

02

025

Window size

Mea

n G

LCM

cont

rast


(a)

3 4 5 6 7 8 9 10 11

088

09

092

094

096

098

Window size

Mea

n G

LCM

hom

ogen

eity


(b)

Window size3 4 5 6 7 8 9 10 11

065

07

075

08

085

09

Mea

n G

LCM

ener

gy


Paddy riceLevees

GrassWoods

(c)

Window size3 4 5 6 7 8 9 10 11

0

02

04

06

08

1

12

14

Mea

n G

LCM

entro

py


Paddy riceLevees

GrassWoods

(d)




woods











3 4 5 6 7 8 9 10 11

0

005

01

015

02

025

03

035

Mea

n G

LCM

cont

rast

Window size


(a)

3 4 5 6 7 8 9 10 11

088

09

092

094

096

098

1

Window size

Mea

n G

LCM

hom

ogen

eity


(b)

Window size3 4 5 6 7 8 9 10 11

065

07

075

08

085

09

095

1

Mea

n G

LCM

ener

gy


Paddy riceLevees

GrassWoods

(c)

3 4 5 6 7 8 9 10 11

0

02

04

06

08

1

12

14

16

18

Window size

Mea

n G

LCM

entro

py


Paddy riceLevees

GrassWoods

(d)





4 Conclusion



3 4 5 6 7 8 9 10 110

1

2

3

4

5

6

7

8

9

Window size

Mea

n di

rect

sem

ivar

iogr

am


Paddy riceLevees

GrassWoods

(a)

3 4 5 6 7 8 9 10 11

04

05

06

07

08

09

1

11

12

13

Window size

Mea

n ab

solu

te se

miv

ario

gram


Paddy riceLevees

GrassWoods

(b)


3 4 5 6 7 8 9 10 110

2

4

6

8

10

12

14

16

18

Window size

Mea

n di

rect

sem

ivar

iogr

am


Paddy riceLevees

GrassWoods

(a)

3 4 5 6 7 8 9 10 1104

06

08

1

12

14

16

18

2

22

24

Window size

Mea

n ab

solu

te se

miv

ario

gram


Paddy riceLevees

GrassWoods

(b)



0 100 200 300 400 500 600 700 800

008

0085

009

0095

01

0105

011

Number of epoch

Clas

sifica

tion

erro

r rat

e


Max epoch = 50

Max epoch = 50

Max epoch = 100

Max epoch = 100

Max epoch = 200

Max epoch = 200

Max epoch = 300

Max epoch = 300

Max epoch = 400

Max epoch = 400

Max epoch = 800

Max epoch = 800


0 50 100 150 200 250 300 350 400

0075

008

0085

009

0095

01

0105

011

0115

Number of epoch

Clas

sifica

tion

erro

r rat

e























References





























Volume 2014




Journal of











Journal of


Function Spaces






Algebra














Ground truth class

Paddy rice 26112611 089089 0967009670Levee 127080 873920 0873009200







Ground truth class

Paddy rice 25672567 089089 0950709507Woods 171171 729729 0810008100

User accuracy 0937509375 0845708457



09156 std = 00443







3 4 5 6 7 8 9 10 11

005

01

015

02

025

Window size

Mea

n G

LCM

cont

rast


(a)

3 4 5 6 7 8 9 10 11

088

09

092

094

096

098

Window size

Mea

n G

LCM

hom

ogen

eity


(b)

Window size3 4 5 6 7 8 9 10 11

065

07

075

08

085

09

Mea

n G

LCM

ener

gy


Paddy riceLevees

GrassWoods

(c)

Window size3 4 5 6 7 8 9 10 11

0

02

04

06

08

1

12

14

Mea

n G

LCM

entro

py


Paddy riceLevees

GrassWoods

(d)




woods











3 4 5 6 7 8 9 10 11

0

005

01

015

02

025

03

035

Mea

n G

LCM

cont

rast

Window size


(a)

3 4 5 6 7 8 9 10 11

088

09

092

094

096

098

1

Window size

Mea

n G

LCM

hom

ogen

eity


(b)

Window size3 4 5 6 7 8 9 10 11

065

07

075

08

085

09

095

1

Mea

n G

LCM

ener

gy


Paddy riceLevees

GrassWoods

(c)

3 4 5 6 7 8 9 10 11

0

02

04

06

08

1

12

14

16

18

Window size

Mea

n G

LCM

entro

py


Paddy riceLevees

GrassWoods

(d)





4 Conclusion



3 4 5 6 7 8 9 10 110

1

2

3

4

5

6

7

8

9

Window size

Mea

n di

rect

sem

ivar

iogr

am


Paddy riceLevees

GrassWoods

(a)

3 4 5 6 7 8 9 10 11

04

05

06

07

08

09

1

11

12

13

Window size

Mea

n ab

solu

te se

miv

ario

gram


Paddy riceLevees

GrassWoods

(b)


3 4 5 6 7 8 9 10 110

2

4

6

8

10

12

14

16

18

Window size

Mea

n di

rect

sem

ivar

iogr

am


Paddy riceLevees

GrassWoods

(a)

3 4 5 6 7 8 9 10 1104

06

08

1

12

14

16

18

2

22

24

Window size

Mea

n ab

solu

te se

miv

ario

gram


Paddy riceLevees

GrassWoods

(b)



0 100 200 300 400 500 600 700 800

008

0085

009

0095

01

0105

011

Number of epoch

Clas

sifica

tion

erro

r rat

e


Max epoch = 50

Max epoch = 50

Max epoch = 100

Max epoch = 100

Max epoch = 200

Max epoch = 200

Max epoch = 300

Max epoch = 300

Max epoch = 400

Max epoch = 400

Max epoch = 800

Max epoch = 800


0 50 100 150 200 250 300 350 400

0075

008

0085

009

0095

01

0105

011

0115

Number of epoch

Clas

sifica

tion

erro

r rat

e























References





























Volume 2014




Journal of











Journal of


Function Spaces






Algebra










3 4 5 6 7 8 9 10 11

005

01

015

02

025

Window size

Mea

n G

LCM

cont

rast


(a)

3 4 5 6 7 8 9 10 11

088

09

092

094

096

098

Window size

Mea

n G

LCM

hom

ogen

eity


(b)

Window size3 4 5 6 7 8 9 10 11

065

07

075

08

085

09

Mea

n G

LCM

ener

gy


Paddy riceLevees

GrassWoods

(c)

Window size3 4 5 6 7 8 9 10 11

0

02

04

06

08

1

12

14

Mea

n G

LCM

entro

py


Paddy riceLevees

GrassWoods

(d)




woods











3 4 5 6 7 8 9 10 11

0

005

01

015

02

025

03

035

Mea

n G

LCM

cont

rast

Window size


(a)

3 4 5 6 7 8 9 10 11

088

09

092

094

096

098

1

Window size

Mea

n G

LCM

hom

ogen

eity


(b)

Window size3 4 5 6 7 8 9 10 11

065

07

075

08

085

09

095

1

Mea

n G

LCM

ener

gy


Paddy riceLevees

GrassWoods

(c)

3 4 5 6 7 8 9 10 11

0

02

04

06

08

1

12

14

16

18

Window size

Mea

n G

LCM

entro

py


Paddy riceLevees

GrassWoods

(d)





4 Conclusion



3 4 5 6 7 8 9 10 110

1

2

3

4

5

6

7

8

9

Window size

Mea

n di

rect

sem

ivar

iogr

am


Paddy riceLevees

GrassWoods

(a)

3 4 5 6 7 8 9 10 11

04

05

06

07

08

09

1

11

12

13

Window size

Mea

n ab

solu

te se

miv

ario

gram


Paddy riceLevees

GrassWoods

(b)


3 4 5 6 7 8 9 10 110

2

4

6

8

10

12

14

16

18

Window size

Mea

n di

rect

sem

ivar

iogr

am


Paddy riceLevees

GrassWoods

(a)

3 4 5 6 7 8 9 10 1104

06

08

1

12

14

16

18

2

22

24

Window size

Mea

n ab

solu

te se

miv

ario

gram


Paddy riceLevees

GrassWoods

(b)



0 100 200 300 400 500 600 700 800

008

0085

009

0095

01

0105

011

Number of epoch

Clas

sifica

tion

erro

r rat

e


Max epoch = 50

Max epoch = 50

Max epoch = 100

Max epoch = 100

Max epoch = 200

Max epoch = 200

Max epoch = 300

Max epoch = 300

Max epoch = 400

Max epoch = 400

Max epoch = 800

Max epoch = 800


0 50 100 150 200 250 300 350 400

0075

008

0085

009

0095

01

0105

011

0115

Number of epoch

Clas

sifica

tion

erro

r rat

e























References





























Volume 2014




Journal of











Journal of


Function Spaces






Algebra










3 4 5 6 7 8 9 10 11

0

005

01

015

02

025

03

035

Mea

n G

LCM

cont

rast

Window size


(a)

3 4 5 6 7 8 9 10 11

088

09

092

094

096

098

1

Window size

Mea

n G

LCM

hom

ogen

eity


(b)

Window size3 4 5 6 7 8 9 10 11

065

07

075

08

085

09

095

1

Mea

n G

LCM

ener

gy


Paddy riceLevees

GrassWoods

(c)

3 4 5 6 7 8 9 10 11

0

02

04

06

08

1

12

14

16

18

Window size

Mea

n G

LCM

entro

py


Paddy riceLevees

GrassWoods

(d)





4 Conclusion



3 4 5 6 7 8 9 10 110

1

2

3

4

5

6

7

8

9

Window size

Mea

n di

rect

sem

ivar

iogr

am


Paddy riceLevees

GrassWoods

(a)

3 4 5 6 7 8 9 10 11

04

05

06

07

08

09

1

11

12

13

Window size

Mea

n ab

solu

te se

miv

ario

gram


Paddy riceLevees

GrassWoods

(b)


3 4 5 6 7 8 9 10 110

2

4

6

8

10

12

14

16

18

Window size

Mea

n di

rect

sem

ivar

iogr

am


Paddy riceLevees

GrassWoods

(a)

3 4 5 6 7 8 9 10 1104

06

08

1

12

14

16

18

2

22

24

Window size

Mea

n ab

solu

te se

miv

ario

gram


Paddy riceLevees

GrassWoods

(b)



0 100 200 300 400 500 600 700 800

008

0085

009

0095

01

0105

011

Number of epoch

Clas

sifica

tion

erro

r rat

e


Max epoch = 50

Max epoch = 50

Max epoch = 100

Max epoch = 100

Max epoch = 200

Max epoch = 200

Max epoch = 300

Max epoch = 300

Max epoch = 400

Max epoch = 400

Max epoch = 800

Max epoch = 800


0 50 100 150 200 250 300 350 400

0075

008

0085

009

0095

01

0105

011

0115

Number of epoch

Clas

sifica

tion

erro

r rat

e























References





























Volume 2014




Journal of











Journal of


Function Spaces






Algebra










3 4 5 6 7 8 9 10 110

1

2

3

4

5

6

7

8

9

Window size

Mea

n di

rect

sem

ivar

iogr

am


Paddy riceLevees

GrassWoods

(a)

3 4 5 6 7 8 9 10 11

04

05

06

07

08

09

1

11

12

13

Window size

Mea

n ab

solu

te se

miv

ario

gram


Paddy riceLevees

GrassWoods

(b)


3 4 5 6 7 8 9 10 110

2

4

6

8

10

12

14

16

18

Window size

Mea

n di

rect

sem

ivar

iogr

am


Paddy riceLevees

GrassWoods

(a)

3 4 5 6 7 8 9 10 1104

06

08

1

12

14

16

18

2

22

24

Window size

Mea

n ab

solu

te se

miv

ario

gram


Paddy riceLevees

GrassWoods

(b)



0 100 200 300 400 500 600 700 800

008

0085

009

0095

01

0105

011

Number of epoch

Clas

sifica

tion

erro

r rat

e


Max epoch = 50

Max epoch = 50

Max epoch = 100

Max epoch = 100

Max epoch = 200

Max epoch = 200

Max epoch = 300

Max epoch = 300

Max epoch = 400

Max epoch = 400

Max epoch = 800

Max epoch = 800


0 50 100 150 200 250 300 350 400

0075

008

0085

009

0095

01

0105

011

0115

Number of epoch

Clas

sifica

tion

erro

r rat

e























References





























Volume 2014




Journal of











Journal of


Function Spaces






Algebra










0 100 200 300 400 500 600 700 800

008

0085

009

0095

01

0105

011

Number of epoch

Clas

sifica

tion

erro

r rat

e


Max epoch = 50

Max epoch = 50

Max epoch = 100

Max epoch = 100

Max epoch = 200

Max epoch = 200

Max epoch = 300

Max epoch = 300

Max epoch = 400

Max epoch = 400

Max epoch = 800

Max epoch = 800


0 50 100 150 200 250 300 350 400

0075

008

0085

009

0095

01

0105

011

0115

Number of epoch

Clas

sifica

tion

erro

r rat

e























References





























Volume 2014




Journal of











Journal of


Function Spaces






Algebra










References





























Volume 2014




Journal of











Journal of


Function Spaces






Algebra
















Volume 2014




Journal of











Journal of


Function Spaces






Algebra









research article a novel approach to developing a

Documents