research article classification of etm+ remote sensing image based...

Hindawi Publishing CorporationMathematical Problems in EngineeringVolume 2013 Article ID 719756 8 pageshttpdxdoiorg1011552013719756

Research ArticleClassification of ETM+ Remote Sensing Image Based onHybrid Algorithm of Genetic Algorithm and Back PropagationNeural Network

Haisheng Song12 Ruisong Xu1 Yueliang Ma1 and Gaofei Li1

1 Guangzhou Institute of Geochemistry Chinese Academy of Sciences 510640 Guangzhou China2University of Chinese Academy of Sciences 100049 Beijing China

Correspondence should be addressed to Haisheng Song songhsustcedu

Received 10 April 2013 Revised 26 May 2013 Accepted 26 May 2013

Academic Editor Erik Cuevas

Copyright copy 2013 Haisheng Song et alThis 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

The back propagation neural network (BPNN) algorithm can be used as a supervised classification in the processing of remotesensing image classification But its defects are obvious falling into the local minimum value easily slow convergence speed andbeing difficult to determine intermediate hidden layer nodes Genetic algorithm (GA) has the advantages of global optimizationand being not easy to fall into local minimum value but it has the disadvantage of poor local searching capability This paper usesGA to generate the initial structure of BPNNThen the stable efficient and fast BP classification network is gotten throughmakingfine adjustments on the improved BP algorithm Finally we use the hybrid algorithm to execute classification on remote sensingimage and compare it with the improved BP algorithm and traditionalmaximum likelihood classification (MLC) algorithm Resultsof experiments show that the hybrid algorithm outperforms improved BP algorithm and MLC algorithm

1 Introduction

Satellite remote sensing data is widely used in resourceexploration military reconnaissance environmental disastermonitoring land use crop yield assessment urban planningand many other areas [1] Classification of remote sensingimage has been a focus and a difficulty in the field of remotesensing and a critical step of transformation from remotesensing technology to practical application [2] The mostcommonly used classification processing method of remotesensing image is statistical pattern recognitionmethod whichis based on the spectral characteristics of the image data [3]But sometimes the classification effect of this method is notvery satisfactory It often makes people see the confusingresults because the method is only based on the statisticalcharacteristics of gray-scale data of each band [4]

Jia [5] used BP algorithm to classify three-band RGBcolor images for automatic visual inspection of seed maizeand compared it with minimum distance and maximumlikelihood classifications Xiong et al [6] combined BPNNand ground truth database and attained good classification

effect Mousavi Rad et al [7] used BPNN classifier to classifyrice varieties In order to classify they extracted twenty-twofeatures from sixty color and texture features Experimentsshowed that classification result was satisfactory Yang et al[8] and Liu et al [9] applied improved BPNN algorithm toclassify TM image and achieved good classification result

Using BPNN algorithm for classification of remote sens-ing image can eliminate fuzziness and uncertainty to someextent But simple BPNN algorithm used in remote sensingimage classification has great limitations mainly in fiveways [10 11] Firstly BP algorithm learning process has thepossibility of falling into local minimum easily so it cannotguarantee that the networkrsquos learning process always tends tobe in globally stable state Secondly BP algorithm has somedefects such as slow convergence velocity and requiring alarge amount of training samples Thirdly the determinationof learning rate of network depends on onersquos experienceFurthermore the choice method of learning rate can directlyinfluence the stability and efficiency of the learning processFourthly determination of the number of nodes in the hiddenlayer is based not on some theory but on onersquos experience

2 Mathematical Problems in Engineering

Fifthly there exists overfitting phenomenon in the process oflearning Classification errors generate as thus [12ndash14]

On the contrary GA has advantages of strong ability ofglobal optimization and being not easy to fall into local min-ima But it has poor ability of local optimization Combinedwith BP algorithmrsquos strong ability of partial optimizationHybrid algorithm is constructed Cao and Jin [15] fusedLandsat ETM+ image and ERS-2 SAR image with princi-pal component analysis (PCA) and constructed a hybridalgorithm of BP-ANNGA Then they applied the hybridalgorithm to classify urban terrain surfaces in Pudong NewArea of Shanghai China They got the satisfactory resultespecially in building and road classification But in hishybrid algorithm traditional BPNN with one hidden layerand twelve nodes of hidden layer was adopted It has weakpoints of slow adjustment of learning rate being fit for urbanarea with less ground features

In this paper we constructed a hybrid algorithm ofGABPNN with GA and improved BPNN instead of tradi-tional BPNN Initial structure of BPNN is generated by usingGA Itmakes the best of both algorithms By using this hybridalgorithm we got the excellent classification results on ETM+remote sensing image

2 Data Source

The data in this study comes from American Landsat-7satellite EnhancedThematicMapper plus (ETM+) imageTheimage acquisition time is January 10 2003 The image hasa spatial resolution of 30 meters and the image coverageof the area is Shunde District Foshan City GuangdongProvince China In order to fulfill research needs 500 times

500-pixel subimages were tailored from the original imageas a study area Because the sixth band of ETM+ images isthermal infrared band and it has a resolution of 60 metersmeanwhile the eighth band is panchromatic wave band andit has a resolution of 15 meters the two bands have beenremoved [16 17] According to the information entropy ofthe remaining six bands and the correlation between eachoriginal image band this paper chooses the third fourth andfifth bands to composite pseudocolor image because of theirrich information and small correlation [18] Figure 1 showsthe pseudo-color composition image with band 5 4 and 3

3 Methodology

By means of visual interpretation and geographical knowl-edge of land use in Shunde the study area can be classified asseven kinds of ground objects (Table 1) [2] that is farmland(C1) woodland (C2) grass (C3) water (C4) residential area(C5) traffic land (C6) and other construction lands (C7)

A total of 1400 samples were gotten from the land useclassification map of the same period Each category has 200samples Among these 200 samples 80 were used for trainingand 120 were used to validate the classification accuracy Inorder to carry out a comparative study three classificationmethods were implemented in this paper The first methodis the traditional maximum likelihood classification (MLC)

Table 1 Land cover types in classification

Classname Land-cover class Description

C1 Farmland Crop fields and vegetable fields

C2 Woodland Deciduous or evergreen forest landorchards and tree groves

C3 Grass Lawn grassland or rangelandC4 Water Reservoir pond lake river and wetland

C5 Residential area Rural and urban residencescommercial land

C6 Traffic land Expressway road railway urban roadcountry road

C7Otherconstructionlands

Industrial and other developed lands

Figure 1 Pseudo-color composition image in Shunde district ofGuangdong China

The method is realized by software of the Environment forVisualizing Images (ENVI) and the version of ENVI is 48The second classification method is improved BP neuralnetwork algorithm (BPNN) A 6-9-7 three-tier networkarchitecture was adoptedThat is to say the input layer has sixnodes namely band numbers of the remote sensing imageThe hidden layer number is one and the node number of thislayer is 9 by experience The output layer node number is thesame as classification category number namely 7The classifi-cationmethod is realized by the software ofMatlab 2011bThethird classification method is GABPNN hybrid algorithmGA is used here for getting initial structure of BPNN andclassification is realized by modified BPNNThe input nodesof the BP network are band numbers namely 6 The outputnodes are the category numbers namely 7 Layers numberof hidden layers and nodes number of each hidden layerthresholds and connecting weights of each node in hiddenor output layer are unknownThese unknown parameters are

Mathematical Problems in Engineering 3

determined through hybrid algorithmThe hybrid algorithmis realized by the software of Microsoft Visual C++ 2008

4 Algorithm Principle

41 Improved BP Algorithm The dynamic adaptive adjust-ment for learning rate is adopted in this paper to improveefficiency of common BP algorithm Learning or training ofBPNN is process of modification on weight vector of networkrepeatedly according to least mean square error principlewhere the squared error is between the desired output andthe actual output of the network The learning process willnot stop until the output value of the network is close to thedesired output value In other words the total error is lessthan a given number or the iteration program ceases at thegiven maximum cycle times For a training sample 119901 thesquare error 119864119901 is defined as follows

119864119901 =

1

2

sum

119896

(119905119901119896 minus 119900119901119896)

2

(1)

where 119905119901119896 represents the desired output and 119900119901119896 represents theactual output and the total mean square error of the wholeBPNN is expressed as

119864 =

1

119899

sum 119864119899 (2)

where 119899 is the total number of training samples Adjustmentfor weight vector of the output layer and the hidden layer iscarried out by

119908 (119896 + 1) = 119908 (119896) + 120578 (119896) 119863 (119896) + 120572 (119908 (119896) minus 119908 (119896 minus 1)) (3)

where 119908(119896) and 119908(119896 minus 1) represent weight vectors of 119896

moment and before respectively Momentum factor 120572 isused to prevent the learning process from falling into a localminimum Dynamic adjustment formula for learning rate 120578

is as follows

120578 (119896) = 4120582120578 (119896 minus 1) (4)

120578(119896) is learning rate for network of 119896moment In formula (4)120582 refers to adjustment coefficient where

120582 = sign (119863 (119896) 119863 (119896 minus 1)) (5)

119863(119896) is the negative gradient of 119896 moment which is definedas

119863 (119896) = (minus

120597119864 (119908)

120597119908

)

10038161003816100381610038161003816100381610038161003816119908=119908(119896)

(6)

So there are two situations of 120582 value If two consecutiveiterations have the same gradient direction it means thatthe decline speed is too slow then 120582 equals to +1 Thelearning rate increased to four timesWhile if two consecutiveiterations have contrary gradient direction it means thatthe decline speed is too fast then 120582 equals to minus1 Thelearning rate decreased to 25 percent Such adjustment of thelearning rate not only can avoid falling into local minima and

misconvergence but also can help to shorten the learning ortraining time

Sigmoid function is chosen as activation function ofnodes for hidden layer and output layer Consider

119891 (119909) =

1

1 + exp (minus119883)

(7)

where 119883 is the sum of all input value multiplied by thecorresponding weight plus the threshold value of the nodeUsing sigmoid type excitation function as incentive functioncan ensure that the function is nonlinear and continuouslydifferentiable

However net input value of the node 119895 for hidden oroutput layer is

net119895 = sum

119895

119908119895119894119909119894 (8)

Therefore the output of the node 119895 is

119900119895 = 119891 (net119894 + 119887119895) = 119891 (sum

119895

119908119895119894119909119895 + 119887119895) (9)

Thus the threshold 119887119895 of node 119895 can be regarded as aspecial weight vector and adjustment of the threshold canalso use (3)

42 Genetic Algorithm Genetic algorithm is used to obtainthe initial structure of the improved BP network

421 Encoding In this study floating-point encoding isadopted Individual gene encoded string is composed of allinput weights thresholds learning rate momentum factorlayers number of hidden layers and nodes number of eachhidden layer Initial population of individuals is generatedrandomly Population size of 40 is more appropriate in thisstudy because a larger size increases the training time and theglobal optimal solution cannot be gotten by a smaller size

422 Fitness Function According to principle of single valuecontinuity nonnegative value maximization rationalityconsistency compatibility and small amount of calculationfor the fitness function [10] and because the network traininggoal is minimum mean square error (MSE) the minimumvalue problem can be converted into the maximum valueproblem so the fitness function of individual can beexpressed as the following

fitness (119909) =

1

1 + 119864 (119909)

(10)

where 119864(119909) represents the total MSE of the output accordingto BP network structure of decoding of the individual 119909Hence the fitness of an individual is determined by the totalMSE The bigger the error is the smaller the fitness is

43 Genetic Operators

(1) Selection The best individual is saved preferentially andthe remaining individuals are selected stochastically Namely


the individual with maximum fitness value is preservedfirstly and then the remaining individuals are chosen atrandomThe selectionmethod provides a zero-deviation andminimum individual extension Individuals of small fitnessvalue may be selected under the premise of reservation of thebest individual so that population can keep diverse

(2) Crossover For this asexual reproduction operator owingto floating-point encoding for individual heterogeneousarithmetical crossover is adopted in this study Two parentindividuals 119883

119905

119860and 119883

119905

119861are selected from the population by

using the stochastic selection operation firstly and then twonew offspring individuals are created by crossover operationas

119883119905+1

119860= 120574119883119905

119861+ (1 minus 120574) 119883

119905

119860

119883119905+1

119861= 120574119883119905

119860+ (1 minus 120574) 119883

119905

119861

(11)

where parameter 120574 isin (0 1) is not a constant but a variabledetermined by the current evolutional generation Value of 120574

can be expressed as

120574 =

1

1 + 119892

(12)

where 119892 is number of current evolutional generation Arith-metical crossover probability 119901119888 plays a dominant role inGA from the beginning to the end It directly affects theconvergence of the algorithm The greater 119901119888 value is thefaster new individuals generate But if it is too large then thegenetic model will be destroyed easily Therefore it generallyranges between 04 and 099 In this study 119901119888 takes a value of066

(3) Mutation Uniformmutationmethod is used in this studyFirstly the parent individual 119883(1199091 1199092 119909119899) is chosen byselection operator beforemutation operationThen a randomsingle point component 119909119894 119894 isin [1 119899] from the coding ofthe parent individual 119883 is selected randomly Accordingly119909119894 is replaced by 119909

1015840

119894 1199091015840

119894= 119909119894(1 minus 120574) where 120574 is the same of

(12) Mutation probability119901119898 recommended generally rangesbetween 00001 and 01 In this study 119901119898 is given value of0005

The property of combining these crossover and mutationoperators can make a uniform search in the initial space inearly generations and very locally at a later stage favoringlocal tuning It also greatly reduces the risk of prematureconvergence

44 Hybrid Algorithm for Classification Before implement-ing the hybrid algorithm remote sensing data must be pre-processed Preprocessing of remote sensing imagery datarefers to the normalized processing of remote sensing imagedata The normalized transform uses the following formula[19 20]

1199091015840

119894=

(119909119894 minus 119872min)

(119872max minus 119872min)

(13)

where 119872min represents the minimum value of a band ofETM+ image Correspondingly 119872max is the maximum valueof the identical band Accordingly all data after normalizingare distributed in the region of 0 and 1 This reduces thesolving difficulty to a certain extent

All weights thresholds and other parameters of BPNNare initialized at random firstly They are encoded as realnumbers Initial population of GA is generated

Training samples are input and fitness of GA is calcu-latedThe individual is assessed by program If the individualgets specified precision then program exits GA and entersinto modified BPNN or else program executes selectioncrossover andmutation operator to generate new species GAwill not exit until program reaches the maximum number ofevolution generation or the individual gets specified preci-sion

Inmodified BPNN the best individual is decoded by pro-gram and then initial structure of BPNN is gotten Weightsthresholds of all nodes layers number of hidden layers andnode number of hidden layers are obtained Then programexecutes training Trainingwill not exit until program reachesthe maximum number of iterations or overall error is lessthan a specified minimum value Finish of training meansthat the best stable effective BPNN is constructed

Finally program executes classification with trainedBPNN and outputs the classified results Flow chart ofGABPNN hybrid algorithm can be seen in Figure 2

5 Classification Results

Figure 3(a) illustrates the classification result by using con-ventional MLC algorithmThis algorithm is implemented byENVI 48 In this software training samples are defined asregion of interest (ROI)

Figure 3(b) indicates the classification result by usingimproved BPNNalgorithm In this classificationmethod a 6-9-7 three-tier neural network architecture was adopted Theinput layer has six nodes namely band numbers of the imageThere is only one hidden layer and the node number of thislayer is 9 by experience The output layer node number isthe same as classification category number namely 7 Theanticipated outputs for the 7 surface classes C1 C2 C3 C4C5 C6 and C7 are defined as (1 0 0 0 0 0 0) (0 1 0 0 0 00) (0 0 1 0 0 0 0) (0 0 0 1 0 0 0) (0 0 0 0 1 0 0) (0 0 0 00 1 0) and (0 0 0 0 0 0 1) respectively The error betweenthe real output and the desired output are evaluated If theBPANN output is beyond the allowed error then the errorfeedback is propagating in the opposite direction to adjust theweights until the satisfied results are acquiredWe first choosethe training data for this modified BPANN from the ETM+image at six bands and 80 samples for each class Then themaximum number of iterations is set to 15000 and the initiallearning rate is set to 005 Momentum factor is set to 05Theminimum MSE is set to 001 After 14520 times of trainingan excellent BPNN is constructed The training performanceconvergence curve is shown in Figure 4 Input ETM+ remotesensing image data and the classification result are gotten asshown in Figure 3(b)


Preprocessing of remote sensing image data

Initialization for all weights thresholds and otherparameters of BPNN at random real coding

generation for initial population of GA

Inputting training samplescalculating the fitness

assessing the individual

(1) Reach the maximum number of evolution generation(2) Reach the specified precision of the best individual

Decoding generating initial BPNN structureobtaining weights thresholds of all nodes and

number of hidden layers and nodes

Training with modified BP algorithm

(1) Reach the maximum number of iterations(2) Overall error is less than a specified minimum error

Classification with trained BPNN

Output classification results

end

Selectioncrossovermutation

generation ofnew species

No

No

Yes

Yes

Figure 2 Flow chart of GABPNN hybrid algorithm

Figure 3(c) shows the classification result by usingGABPNNhybrid algorithm First we usedGA to generate theinitial structure of BPNN the population size is set to 40 thecrossover probability is set to 066 themutation probability isset to 0005 the maximum evolution generation is set to 500The GA program exited after the 458th generation and thenthemost suitable 4-layer structure BP network is obtained Inthis BP network there are two hidden layers the nodes num-ber of hidden layer near the input layer is 15 and the nodesnumber of hidden layers near the output layer is 18Thereforea 6-15-18-7 structure of 4-layer BPNN is constructed Thenwe utilized the previous improved BP algorithm for gettingthe exact solution of weights and thresholds The maximumnumber of iterations is also set to 15000 The initial learning

rate is set to 005Themomentum rate factor is set to 05Theminimum MSE is set to 001 After 11200 times of trainingan efficient stable BPNN is constructedThe hybrid algorithmtraining performance convergence curve is shown in Figure 5Using this fine-tuned BPNN we obtained the classified resultas shown in Figure 3(c)

6 Classification Assessment

In order to further verify the correctness of these three typesof remote sensing image classification algorithm we car-ried out the quantitative comparison for their classificationaccuracy [15] The method is conventional that is confusion


(a)

(b)

WaterFarmlandWoodlandGrass

Residential areaTraffic landOther construction lands

(c)

Figure 3 Classified images derived from (a) MLC (b) BPNN and(c) hybrid algorithm of GABPNN

0 2000 4000 6000 8000 10000 12000 140000

002

004

006

008

01

012

014

Epochs

MSE

Figure 4 Training performance of modified BPNN

Table 2 Comparison of production and user accuracies as apercentage

Accuracy C1 C2 C3 C4 C5 C6 C7Productionaccuracy

MLC 708 825 733 883 917 625 742BPNN 750 833 775 917 917 667 750GABPNN 792 917 875 958 925 750 817

Useraccuracy

MLC 697 750 772 922 719 781 824BPNN 726 781 816 948 729 816 826GABPNN 841 775 833 991 847 891 883

Table 3 Comparison of the overall accuracy

MLC BPNN GABPNNOverall accuracy 7762 8012 8619Kappa coefficient 0739 0768 0838

matrix production and user accuracy overall accuracy andkappa coefficient [21]

Every 120 validation samples for each type are selectedfrom existing LULCmapping or field investigation with glob-al positioning system (GPS) receiver The production anduser accuracy overall accuracy and kappa coefficient are thencomputed [22] They are shown in Tables 2 and 3

From the previous classification images and accuracytables we find that farmland woodland and grass are proneto misclassification or leakage points because of the phe-nomena of ldquosame object with different spectrardquo and ldquosamespectrum with different objectsrdquo whereas water area hasrelatively high accuracy in three classification methods Thereason is that the spectral curves of water and others aredifferent obviously while farmlandwoodland and grass havesimilar or same spectrum Residential area traffic land and


0 50 100 150 200 250 300 350 400 450 5000

01

02

03

04

05

06

07

Generation

MSE

(a) GA

0 2000 4000 6000 8000 100000

002

004

006

008

01

012

014

Epochs

MSE

(b) modified BPNN after GA

Figure 5 Hybrid algorithm training performance

other construction land are easy to be misclassified becausethese classes contain large amounts of cement and otherbuildingmaterials which caused their spectra similar or alike

As shown in Table 2 the user accuracy of farmland inMLC is 697 and it is 726 and 841 in modified BPNNand GABPNN respectively The user accuracy of GABPNNis 144 higher than that of MLC and it is nearly 12 higherthan the modified BPNN algorithm the same for other clas-sification types

The overall accuracy and kappa coefficient are shown inTable 3 From this table we can find that the overall accuracyis from 7762 of MLC to 8012 of BPNN and to 8612 ofGABPNN hybrid algorithm It increased significantly as wellas kappa coefficient

7 Conclusion and Discussion

This paper developed a hybrid algorithm of GABPNN tooptimize the networkrsquos structure and make fast convergenceof BPANN The algorithm is applied on ETM+ image inShunde District Guangdong China

(1) The training speed of modified BPNN after GA isspeeded up and the classification accuracy for all surfaces isimproved

(2) Applying the GA algorithm to initialize the structureof the BPANN can take its advantage of optimization andovercome the shortcomings of the BPANNrsquos slow conver-gence and falling into the local minima easily

The probability distribution density of the training data ofeach typewith a normal distribution characteristic is requiredin the conventional MLC but the actual chosen sample datamay be deviated from the normal distribution thereforeclassification precision of MLC is low relatively

But on the contrary there are no strict requirementsno strict limitations and no need of normal distributioncharacter on training data in BPNN In addition BPNN hasthe ability of complex nonlinear mapping Hence it is more

suitable for the classification processing of massive high-dimensional remote sensing image data

On the other hand the ordinary BPNN classifier hasdefects of falling into local minima easily convergence diffi-cultly and being time consuming while the GABPNN devel-oped in this study not only avoids shortcomings of BPNNbut also improves the accuracy and efficiency of classificationExperience has proved that GABPNN hybrid algorithm isstable and effective Results of experiments in this papershow that the hybrid algorithm outperforms improved BPalgorithm and MLC algorithm It combines the advantagesof both and it can always get the global optimal solution It isan excellent competent promising algorithm of classificationprocessing for moderate resolution multispectral remotesensing data and it has a strong practical value

Acknowledgments

This research is supported by the National Natural ScienceFoundation of China (Grant no 41072247) and NaturalScience Foundation of Guangdong Province China (Grantno 9151064004000004)

References

[1] J B Campbell Introduction To Remote Sensing Taylor amp Fran-cis London UK 2000

[2] Y M Luo M H Liao J Yan C Y Zhang and S L ShangldquoDevelopment and demonstration of an artificial immunealgorithm for mangrove mapping using landsat TMrdquo IEEEGeoscience and Remote Sensing Letters vol 10 no 4 pp 751ndash755 2013

[3] R O Duda P E Hart and D G Stork Pattern ClassificationJohn Wiley amp Sons New York NY USA 2nd edition 2001

[4] J A Richards and X P Jia Remote Sensing Digital Image Anal-ysis An Introduction Springer Berlin Germany 3rd edition1999


[5] J Jia ldquoPattern classification of RGB colour images using a BPneural network classifierrdquo Proceedings of the SPIE-The Inter-national Society for Optical Engineering vol 1989 pp 248ndash2561993

[6] Z Y Xiong K Chen and C D Gu ldquoAn algorithm of imageclassification based on BP neural networkrdquo in Proceedings ofthe IEEE International Conference on Computer Science andAutomation Engineering (CSAE rsquo12) pp 523ndash526 2012

[7] S J Mousavi Rad F A Tab and K Mollazade ldquoClassificationof rice varieties using optimal color and texture features and BPneural networksrdquo inProceedings of the 7th IranianConference onMachineVision and Image Processing (MVIP rsquo11) p 5 November2011

[8] B Yang Z Liu YXing andC Luo ldquoRemote sensing image clas-sification based on improved BPneural networkrdquo inProceedingsof the International Symposium on Image and Data Fusion(ISIDF rsquo11) August 2011

[9] Q L Liu B L Jiao and L Liu ldquoStudy On remote sensing imageclassification method based on improved BP neural networkmodelrdquo Electronics Optics amp Control vol 16 no 8 pp 65ndash672009

[10] F Yu Y S Zhao and H T Li ldquoSoil moisture retrieval basedon GA-BP neural networks algorithmrdquo Journal of Infrared andMillimeter Waves vol 31 no 3 pp 283ndash288 2012

[11] D-Y Ge X-F Yao and W-J Xiang ldquoApplication of BPneural network and genetic algorithm in testing of micro-drillrsquosrounded cornerrdquo Sensor Letters vol 9 no 5 pp 1943ndash1947 2011

[12] C F Luo C Y Wang and Y H Liu ldquoClassification of land usein Xinjiangwith BP classifier usingMODIS imageryrdquoArid LandGeography vol 28 no 2 pp 258ndash262 2005

[13] S Yin N Shu and X Liu ldquoClassification of remote sensingimage based on adaptive GA and improved BP algorithmrdquoGeo-matics and Information Science ofWuhan University vol 32 no3 pp 201ndash204 2007

[14] J F Mas and J J Flores ldquoThe application of artificial neural net-works to the analysis of remotely sensed datardquo InternationalJournal of Remote Sensing vol 29 no 3 pp 617ndash663 2008

[15] G Cao and Y Q Jin ldquoA hybrid algorithm of the BP-ANNGAfor classification of urban terrain surfaces with fused data ofLandsat ETM+andERS-2 SARrdquo International Journal of RemoteSensing vol 28 no 2 pp 293ndash305 2007

[16] B Guo S R Gunn R I Damper and J D B Nelson ldquoBandselection for hyperspectral image classification using mutualinformationrdquo IEEE Geoscience and Remote Sensing Letters vol3 no 4 pp 522ndash526 2006

[17] L Zhang Y Zhong B Huang J Gong and P Li ldquoDimen-sionality reduction based on clonal selection for hyperspectralimageryrdquo IEEE Transactions on Geoscience and Remote Sensingvol 45 no 12 pp 4172ndash4186 2007

[18] F A Kruse A B Lefkoff J W Boardman et al ldquoThe spectralimage processing system (SIPS)-interactive visualization andanalysis of imaging spectrometer datardquo Remote Sensing ofEnvironment vol 44 no 2-3 pp 145ndash163 1993

[19] F Ratle G Camps-Valls and JWeston ldquoSemisupervised neuralnetworks for efficient hyperspectral image classificationrdquo IEEETransactions on Geoscience and Remote Sensing vol 48 no 5pp 2271ndash2282 2010

[20] GACarpenterMNGjaja SGopal andC EWoodcock ldquoArtneural networks for remote sensing vegetation classificationfrom landsat tm and terrain Datardquo IEEE Transactions onGeoscience and Remote Sensing vol 35 no 2 pp 308ndash325 1997

[21] H Li Y Wang Y Li and X Wang ldquoPixel-unmixing moderate-resolution remote sensing imagery using pairwise couplingsupport vector machines a case studyrdquo IEEE Transactions onGeoscience and Remote Sensing vol 49 no 11 pp 4298ndash43062011

[22] X Liu X Li L Liu J He and B Ai ldquoAn innovative method toclassify remote-sensing images using ant colony optimizationrdquoIEEE Transactions on Geoscience and Remote Sensing vol 46no 12 pp 4198ndash4208 2008

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


Fifthly there exists overfitting phenomenon in the process oflearning Classification errors generate as thus [12ndash14]

On the contrary GA has advantages of strong ability ofglobal optimization and being not easy to fall into local min-ima But it has poor ability of local optimization Combinedwith BP algorithmrsquos strong ability of partial optimizationHybrid algorithm is constructed Cao and Jin [15] fusedLandsat ETM+ image and ERS-2 SAR image with princi-pal component analysis (PCA) and constructed a hybridalgorithm of BP-ANNGA Then they applied the hybridalgorithm to classify urban terrain surfaces in Pudong NewArea of Shanghai China They got the satisfactory resultespecially in building and road classification But in hishybrid algorithm traditional BPNN with one hidden layerand twelve nodes of hidden layer was adopted It has weakpoints of slow adjustment of learning rate being fit for urbanarea with less ground features

In this paper we constructed a hybrid algorithm ofGABPNN with GA and improved BPNN instead of tradi-tional BPNN Initial structure of BPNN is generated by usingGA Itmakes the best of both algorithms By using this hybridalgorithm we got the excellent classification results on ETM+remote sensing image

2 Data Source

The data in this study comes from American Landsat-7satellite EnhancedThematicMapper plus (ETM+) imageTheimage acquisition time is January 10 2003 The image hasa spatial resolution of 30 meters and the image coverageof the area is Shunde District Foshan City GuangdongProvince China In order to fulfill research needs 500 times

500-pixel subimages were tailored from the original imageas a study area Because the sixth band of ETM+ images isthermal infrared band and it has a resolution of 60 metersmeanwhile the eighth band is panchromatic wave band andit has a resolution of 15 meters the two bands have beenremoved [16 17] According to the information entropy ofthe remaining six bands and the correlation between eachoriginal image band this paper chooses the third fourth andfifth bands to composite pseudocolor image because of theirrich information and small correlation [18] Figure 1 showsthe pseudo-color composition image with band 5 4 and 3

3 Methodology

By means of visual interpretation and geographical knowl-edge of land use in Shunde the study area can be classified asseven kinds of ground objects (Table 1) [2] that is farmland(C1) woodland (C2) grass (C3) water (C4) residential area(C5) traffic land (C6) and other construction lands (C7)

A total of 1400 samples were gotten from the land useclassification map of the same period Each category has 200samples Among these 200 samples 80 were used for trainingand 120 were used to validate the classification accuracy Inorder to carry out a comparative study three classificationmethods were implemented in this paper The first methodis the traditional maximum likelihood classification (MLC)

Table 1 Land cover types in classification

Classname Land-cover class Description

C1 Farmland Crop fields and vegetable fields

C2 Woodland Deciduous or evergreen forest landorchards and tree groves

C3 Grass Lawn grassland or rangelandC4 Water Reservoir pond lake river and wetland

C5 Residential area Rural and urban residencescommercial land

C6 Traffic land Expressway road railway urban roadcountry road

C7Otherconstructionlands

Industrial and other developed lands

Figure 1 Pseudo-color composition image in Shunde district ofGuangdong China

The method is realized by software of the Environment forVisualizing Images (ENVI) and the version of ENVI is 48The second classification method is improved BP neuralnetwork algorithm (BPNN) A 6-9-7 three-tier networkarchitecture was adoptedThat is to say the input layer has sixnodes namely band numbers of the remote sensing imageThe hidden layer number is one and the node number of thislayer is 9 by experience The output layer node number is thesame as classification category number namely 7The classifi-cationmethod is realized by the software ofMatlab 2011bThethird classification method is GABPNN hybrid algorithmGA is used here for getting initial structure of BPNN andclassification is realized by modified BPNNThe input nodesof the BP network are band numbers namely 6 The outputnodes are the category numbers namely 7 Layers numberof hidden layers and nodes number of each hidden layerthresholds and connecting weights of each node in hiddenor output layer are unknownThese unknown parameters are





119864119901 =

1

2

sum

119896

(119905119901119896 minus 119900119901119896)

2

(1)


119864 =

1

119899

sum 119864119899 (2)


119908 (119896 + 1) = 119908 (119896) + 120578 (119896) 119863 (119896) + 120572 (119908 (119896) minus 119908 (119896 minus 1)) (3)



is as follows

120578 (119896) = 4120582120578 (119896 minus 1) (4)


120582 = sign (119863 (119896) 119863 (119896 minus 1)) (5)


119863 (119896) = (minus

120597119864 (119908)

120597119908

)

10038161003816100381610038161003816100381610038161003816119908=119908(119896)

(6)




119891 (119909) =

1

1 + exp (minus119883)

(7)



net119895 = sum

119895

119908119895119894119909119894 (8)


119900119895 = 119891 (net119894 + 119887119895) = 119891 (sum

119895

119908119895119894119909119895 + 119887119895) (9)





fitness (119909) =

1

1 + 119864 (119909)

(10)







119905

119860and 119883

119905



119883119905+1

119860= 120574119883119905

119861+ (1 minus 120574) 119883

119905

119860

119883119905+1

119861= 120574119883119905

119860+ (1 minus 120574) 119883

119905

119861

(11)


can be expressed as

120574 =

1

1 + 119892

(12)



1015840

119894 1199091015840





1199091015840

119894=

(119909119894 minus 119872min)


(13)






















end



No

No

Yes

Yes







(a)

(b)



(c)


0 2000 4000 6000 8000 10000 12000 140000

002

004

006

008

01

012

014

Epochs

MSE




MLC 708 825 733 883 917 625 742BPNN 750 833 775 917 917 667 750GABPNN 792 917 875 958 925 750 817

Useraccuracy

MLC 697 750 772 922 719 781 824BPNN 726 781 816 948 729 816 826GABPNN 841 775 833 991 847 891 883







0 50 100 150 200 250 300 350 400 450 5000

01

02

03

04

05

06

07

Generation

MSE

(a) GA

0 2000 4000 6000 8000 100000

002

004

006

008

01

012

014

Epochs

MSE














Acknowledgments


References































Volume 2014




Journal of











Journal of


Function Spaces






Algebra













119864119901 =

1

2

sum

119896

(119905119901119896 minus 119900119901119896)

2

(1)


119864 =

1

119899

sum 119864119899 (2)


119908 (119896 + 1) = 119908 (119896) + 120578 (119896) 119863 (119896) + 120572 (119908 (119896) minus 119908 (119896 minus 1)) (3)



is as follows

120578 (119896) = 4120582120578 (119896 minus 1) (4)


120582 = sign (119863 (119896) 119863 (119896 minus 1)) (5)


119863 (119896) = (minus

120597119864 (119908)

120597119908

)

10038161003816100381610038161003816100381610038161003816119908=119908(119896)

(6)




119891 (119909) =

1

1 + exp (minus119883)

(7)



net119895 = sum

119895

119908119895119894119909119894 (8)


119900119895 = 119891 (net119894 + 119887119895) = 119891 (sum

119895

119908119895119894119909119895 + 119887119895) (9)





fitness (119909) =

1

1 + 119864 (119909)

(10)







119905

119860and 119883

119905



119883119905+1

119860= 120574119883119905

119861+ (1 minus 120574) 119883

119905

119860

119883119905+1

119861= 120574119883119905

119860+ (1 minus 120574) 119883

119905

119861

(11)


can be expressed as

120574 =

1

1 + 119892

(12)



1015840

119894 1199091015840





1199091015840

119894=

(119909119894 minus 119872min)


(13)






















end



No

No

Yes

Yes







(a)

(b)



(c)


0 2000 4000 6000 8000 10000 12000 140000

002

004

006

008

01

012

014

Epochs

MSE




MLC 708 825 733 883 917 625 742BPNN 750 833 775 917 917 667 750GABPNN 792 917 875 958 925 750 817

Useraccuracy

MLC 697 750 772 922 719 781 824BPNN 726 781 816 948 729 816 826GABPNN 841 775 833 991 847 891 883







0 50 100 150 200 250 300 350 400 450 5000

01

02

03

04

05

06

07

Generation

MSE

(a) GA

0 2000 4000 6000 8000 100000

002

004

006

008

01

012

014

Epochs

MSE














Acknowledgments


References































Volume 2014




Journal of











Journal of


Function Spaces






Algebra












119905

119860and 119883

119905



119883119905+1

119860= 120574119883119905

119861+ (1 minus 120574) 119883

119905

119860

119883119905+1

119861= 120574119883119905

119860+ (1 minus 120574) 119883

119905

119861

(11)


can be expressed as

120574 =

1

1 + 119892

(12)



1015840

119894 1199091015840





1199091015840

119894=

(119909119894 minus 119872min)


(13)






















end



No

No

Yes

Yes







(a)

(b)



(c)


0 2000 4000 6000 8000 10000 12000 140000

002

004

006

008

01

012

014

Epochs

MSE




MLC 708 825 733 883 917 625 742BPNN 750 833 775 917 917 667 750GABPNN 792 917 875 958 925 750 817

Useraccuracy

MLC 697 750 772 922 719 781 824BPNN 726 781 816 948 729 816 826GABPNN 841 775 833 991 847 891 883







0 50 100 150 200 250 300 350 400 450 5000

01

02

03

04

05

06

07

Generation

MSE

(a) GA

0 2000 4000 6000 8000 100000

002

004

006

008

01

012

014

Epochs

MSE














Acknowledgments


References































Volume 2014




Journal of











Journal of


Function Spaces






Algebra






















end



No

No

Yes

Yes







(a)

(b)



(c)


0 2000 4000 6000 8000 10000 12000 140000

002

004

006

008

01

012

014

Epochs

MSE




MLC 708 825 733 883 917 625 742BPNN 750 833 775 917 917 667 750GABPNN 792 917 875 958 925 750 817

Useraccuracy

MLC 697 750 772 922 719 781 824BPNN 726 781 816 948 729 816 826GABPNN 841 775 833 991 847 891 883







0 50 100 150 200 250 300 350 400 450 5000

01

02

03

04

05

06

07

Generation

MSE

(a) GA

0 2000 4000 6000 8000 100000

002

004

006

008

01

012

014

Epochs

MSE














Acknowledgments


References































Volume 2014




Journal of











Journal of


Function Spaces






Algebra










(a)

(b)



(c)


0 2000 4000 6000 8000 10000 12000 140000

002

004

006

008

01

012

014

Epochs

MSE




MLC 708 825 733 883 917 625 742BPNN 750 833 775 917 917 667 750GABPNN 792 917 875 958 925 750 817

Useraccuracy

MLC 697 750 772 922 719 781 824BPNN 726 781 816 948 729 816 826GABPNN 841 775 833 991 847 891 883







0 50 100 150 200 250 300 350 400 450 5000

01

02

03

04

05

06

07

Generation

MSE

(a) GA

0 2000 4000 6000 8000 100000

002

004

006

008

01

012

014

Epochs

MSE














Acknowledgments


References































Volume 2014




Journal of











Journal of


Function Spaces






Algebra










0 50 100 150 200 250 300 350 400 450 5000

01

02

03

04

05

06

07

Generation

MSE

(a) GA

0 2000 4000 6000 8000 100000

002

004

006

008

01

012

014

Epochs

MSE














Acknowledgments


References































Volume 2014




Journal of











Journal of


Function Spaces






Algebra



































Volume 2014




Journal of











Journal of


Function Spaces






Algebra









research article classification of etm+ remote sensing image based...

Documents