An Unsupervised Algorithm for Change Detection in …downloads.hindawi.com/journals/js/2017/9702612.pdf · 2019-07-30 · Traditional unsupervised change-detection algorithms are

Research ArticleAn Unsupervised Algorithm for Change Detection inHyperspectral Remote Sensing Data Using Synthetically FusedImages and Derivative Spectral Profiles

Youkyung Han1 Anjin Chang2 Seokkeun Choi3 Honglyun Park3 and Jaewan Choi3

1School of Convergence amp Fusion System Engineering Kyungpook National University Sangju 37224 Republic of Korea2School of Engineering and Computing Sciences Texas AampM University-Corpus Christi 6300 Ocean DrCorpus Christi TX 78412 USA3School of Civil Engineering Chungbuk National University 1 Chungdae-ro Seowon-gu CheongjuChungbuk 28644 Republic of Korea

Correspondence should be addressed to Jaewan Choi jaewanchoichungbukackr

Received 25 April 2017 Accepted 9 July 2017 Published 10 August 2017

Academic Editor Hyung-Sup Jung

Copyright copy 2017 Youkyung Han 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

Multitemporal hyperspectral remote sensing data have the potential to detect altered areas on the earthrsquos surfaceHowever dissimilarradiometric and geometric properties between the multitemporal data due to the acquisition time or position of the sensorsshould be resolved to enable hyperspectral imagery for detecting changes in natural and human-impacted areas In addition datanoise in the hyperspectral imagery spectrum decreases the change-detection accuracy when general change-detection algorithmsare applied to hyperspectral images To address these problems we present an unsupervised change-detection algorithm basedon statistical analyses of spectral profiles the profiles are generated from a synthetic image fusion method for multitemporalhyperspectral images This method aims to minimize the noise between the spectra corresponding to the locations of identicalpositions by increasing the change-detection rate and decreasing the false-alarm rate without reducing the dimensionality ofthe original hyperspectral data Using a quantitative comparison of an actual dataset acquired by airborne hyperspectral sensorswe demonstrate that the proposed method provides superb change-detection results relative to the state-of-the-art unsupervisedchange-detection algorithms

1 Introduction

Hyperspectral imaging based on spaceborne or airborneimagery has strong potential for applications in remote sens-ing because each pixel of the hyperspectral data which arecomposed of a continuous spectral profile includes a detaileddescription of the spectral features of the image Thesedescriptions allow for an analysis of the specific differencesin the characteristics of the earthrsquos surface Several satellite-based hyperspectral sensors such as the Hyperion spectrom-eter of the National Aeronautics and Space Administration(NASA) and various airborne sensors including the Air-borne Imaging Spectrometer for Applications (AISA) Com-pact Airborne Spectrographic Imager (CASI) and HyMapare currently available In addition several hyperspectral

sensors are planned for future launches including the Hyper-spectral Imager SUIte (HISUI) Hyperspectral Precursorof the Application Mission (PRISMA) HYPererspectral-XIMagery (HYPXIM) and the Environmental Mapping andAnalysis Program (ENMAP) [1] Thus the development ofcore technologies for various applications of future hyper-spectral sensor systems is necessary

Change detection is one of the most commonly usedapplications in remote sensing In general change detection isdefined as the use of remotely sensed data for the same area atdifferent times to identify altered areas on the earthrsquos surfaceWhen studying human-induced or natural disasters suchas floods earthquakes landslides oil spills and industrialaccidents change-detection technologies based on remotelysensed data can be effectively used to detect and estimate the

HindawiJournal of SensorsVolume 2017 Article ID 9702612 14 pageshttpsdoiorg10115520179702612

2 Journal of Sensors

extent of the damaged area Several researchers have analyzedthe existing change-detection methods for applications inthe remote sensing data including image differencing imagerationing vegetation index regression data transformationchange vector postclassification comparison and GIS-basedmethods [2ndash5] According to the literature change-detectiontechniques comprise two types supervised change detec-tion and unsupervised change detection [5] In particularunsupervised change-detection techniques are preferred forremote sensing applications because the training data donot need to be manually collected even though super-vised methods can provide change aspects of specific land-cover classes and ldquofrom-tordquo information of change detection[3]

Traditional unsupervised change-detection algorithmsare focused on proposing change-detection indices andapplying automatic thresholding methods for generating abinary map of the changed area For example change-detection indices such as change vector analysis (CVA)image correlation and Relative Dimensionless Global Error(ERGAS) have been proposed [6 7] In addition variousthresholding techniques have been developed and appliedto multitemporal remote sensing data [8] Molina et al [9]integrated different change-detection indices to improve thestatistical properties of indices for single change detectionBruzzone and Prieto proposed modified image differencingand thresholding of adaptive parcels based on a homogeneousregion [10] Markov random fields (MRFs) expectation-maximization (EM) algorithms and neural networks havebeen used to determine optical thresholds for automaticchange detection [11 12] In addition Bruzzone and Bovolo[13] analyzed change-detection techniques for very-high-resolution (VHR) satellite imagery and developed a novelframework using the top-down approach considering theradiometric characteristics of multitemporal images

With the launches and high availability of hyperspectralsensors new change-detection methodologies have beendeveloped and existing methods have been modified tomake them appropriate for hyperspectral imagery Eismannet al [14] proposed an algorithm based on linear predictorsto detect subtle targets against a complex background andKim [15] modified matched filtering using target signalexclusion Song et al [16] proposed unsupervised change-detection algorithms using spectral unmixing and iterativeerror analysis (IEA) Hao et al [17] applied hyperspectral datato detect changes in urban forest resources in natural disasterzones Image transformation techniques such asmultivariatealteration detection (MAD) and principal component anal-ysis (PCA) which can be applied to general multispectralimages have been extended to include hyperspectral data [1819] In particular iterative regularized multivariate alterationdetection (IR-MAD) is considered a state-of-the-art change-detection algorithm due to its stability and outstandingchange-detection results [20ndash22] However dimensionalityreduction through PCA or the minimum noise fraction(MNF) is required prior to the application of IR-MAD-based algorithms Meola et al [23] proposed a model-based change-detection approach that reduces false alarmscaused by shadow differences using calibrated hyperspectral

data and Liu et al [24] proposed a hierarchical change-detection algorithm using endmember detection and clustermerging in multitemporal hyperspectral images Junior etal [25] used CVA based on distance (Euclidean distanceand Mahalanobis distance) and similarity (spectral anglemapper SAM and spectral correlation mapper SCM) mea-surements and concluded that CVA using the Euclideandistance and SAM is superior to other measures Wu et al[26] developed a subspace-based algorithm using two typesof information from Hyperion and from the Chinese HJ-1A satellite proving that measurements based on the SAMare identical to the orthogonal subspace projection- (OSP-)based measurements Based on this concept the authorshave proposed local and adaptive measures based on OSPto minimize registration errors Other advanced algorithmsbased on target detection anomaly detection and changedetection have also been developed recently [27ndash29]

These related approaches have generally focused on thedevelopment of change-detection algorithms that increaseaccuracy by effectively using hyperspectral bands to maxi-mize the advantages of hyperspectral sensors In other wordsthese approaches have focused on using abundant spectralinformation However the dissimilarity of radiometric andgeometric properties between multitemporal data basedon the acquisition time or position of the sensors shouldbe resolved to enable hyperspectral imagery for detectingchanges in natural and human-impact areas Moreover fewconsiderations related to data noise have been included in thespectrum of hyperspectral imagery potentially decreasingthe change-detection accuracy when general change-detec-tion algorithms are applied to hyperspectral images withoutdimensionality reduction

Most studies have also considered hyperspectral datawithlow or medium spatial resolutions due to the technical limi-tations of generating hyperspectral sensors with high spatialresolutions However considering the future development ofhyperspectral sensors hyperspectral data with high spatialresolutions are necessaryThus approaches for change detec-tion should consider high spectral and spatial resolutionssimultaneously High-resolution images acquired at varioustimes demonstrate local geometric location differences andspectral variation by phenomenological or temporal differ-ences between dates of imagery even when the data are geo-referenced These geometric differences can cause poorchange-detection results based on false alarms of the changedarea caused by the differences [30] Information exclusivelyobtained from the data acquired at two times that is beforeand after the change occurs is insufficient for accurate changedetection

The main objective of this study is to develop a change-detection index suitable for hyperspectral imagery Weapplied unsupervised change detection using an image fusiontechnique and an ordered combination of similaritymeasuresto identify the changed regions To minimize the radiometricand geometric dissimilarities between multitemporal imageswith high spatial and spectral resolutions a temporal-data-based synthetic image fusion method is applied which gen-erates combined images at two data points at different timesA combination of the synthetically fused images is then used

Journal of Sensors 3

N

1208040200(m)

354100000000

353900000000

354000000000

354100000000

353900000000

354000000000

4049500000000

4049700000000

4049600000000

4049500000000

4049700000000

4049600000000

(a)

(b) (c)

Figure 1 Study site with targets for change detection (a) CASI image of study site generated with an RGB composite (red 657 nm green557 nm and blue 471 nm) (b) example of the area changed by camouflage nets (red circle of Figure 1(a)) and (c) example of the area changedby artificial turf (green circle of Figure 1(a))

to optimize the change-detection results We propose a newmeasure for change detection that is appropriate for hyper-spectral data without the loss of dimensionality To evaluatethe proposed method we install a test-bed acquired froman airborne hyperspectral sensor taken at different timesFinally the obtained experimental results are compared withthe performance of the state-of-the-art algorithms

2 Study Site and Dataset Description

The proposed change-detection algorithm was applied toairborne hyperspectral images collected by CASI The CASIimages for this experiment were collected over Gangnae-myeon Cheongwon-gun and Chungcheongbuk-do Koreaon 22 June 2013 (127∘ 221015840 E 36∘ 341015840 N)The data specificationsare described in Table 1

A total of 45 spectral bands with spectral wavelength of4134ndash10449 nm excluding noise bands were used and asubset of 500 times 500 image sizes was extracted for the exper-iments The constructed study site is shown in Figure 1(a)

Table 1 CASI data specifications

Sensor CASISpatial resolution 05mBands 48Wavelength 370ndash1044 nmSpectral resolution (nm FWHM) 36Flight height 750m

To artificially construct the changed areas within the studysite we installed two types of targets camouflages (45m times45m 15m times 15m and 05m times 05m) and an artificialturf (45m times 45m) that have colors similar to the vegetationand ground in the study site (Figures 1(b) and 1(c)) Inaddition we covered a car using a camouflage net as shownin the left side of Figure 1(b) The ginseng field (yellow circleof Figure 1(a)) includes an inclined canvas giving rise todifferent reflectance values depending on the position of theairborne hyperspectral sensor


(a) (b) (c)

ltChanged materials spectragt

Wavelength (nm)1000900800700600500400

0

01

02

03

04

05

06

07

08

camouflage_browncamouflage_brown_refercamouflage_vegecamouflage_vege_refertartan turftartan turf_refer

Refle

ctan

ce

(d)

Figure 2 Study area (a) reference data (b) target data (c) ground-truth data and (d) specification of changed area in the study area

Figures 2(a) and 2(b) show the reference and target datafor the experiment and Figures 2(c) and 2(d) show theground-truth data of the altered regions for the evaluation ofthe accuracy using image interpretation [31]

3 Methodology

Theworkflow of ourmethod is illustrated in Figure 3 Follow-ing geometric and radiometric preprocessing of hyperspectral

data acquired at different times with deliberate differ-ences synthetic image fusion is applied Then a spectralsimilarity is calculated using the fused images Finally thechanged area is detected by the proposed unsupervised simi-larity measure A detailed explanation of each step isprovided below

31 Preprocessing Prior to change detection geometric andradiometric correctionsmust be applied to themultitemporal


Hyperspectral data (time 2)target image

Preprocessing

Hyperspectral data (time 1)reference image

Generation of synthetic fused images

Spectral similarity comparison between synthetically fused images

Unsupervised detection of changed area

Figure 3 Workflow of the proposed method

images In our proposed method image-to-image registra-tion using a manually extracted ground control point (GCP)is used to obtain georeferenced images Registered hyper-spectral images are atmospherically corrected using theatmospheric and topographic correction (ATCOR) moduleand then radiometrically adjusted using an empirical line-calibration algorithm based on linear predictions [32] Fourtarps with different reflectance (35 23 35 and 53)are used as reference targets for the radiometric calibrationDuring the preprocessing step spectral information for thesame location may include identical characteristics eventhough partial noise data are presented

32 Generation of Synthetically Fused Images Syntheticallyfused images are generated based on the cross-sharpeningmethodology that was first developed forVHR satellite-basedmultitemporal data to minimize spatial dissimilarities whilepreserving spectral distortion [33] A general pansharpeningalgorithm for VHR images is applied to obtain a high-spatial-resolution multispectral image using high-spatial-resolution panchromatic and low-spatial-resolution multi-spectral images acquired at the same time and at the samesensor position The cross-sharpening algorithm is a modi-fied version of the algorithm of Chen et al [33 34] Cross-sharpened images are effective for minimizing change-detec-tion errors caused by geometric displacement and spectralvariation including the noise in multitemporal images [33]The cross-sharpening algorithm uses panchromatic andmul-tispectral datasets acquired at different times and at differentsensor positions However general airborne hyperspectraldatasets do not include high-spatial-resolution panchromaticdata Therefore to apply the cross-sharpening algorithm tothe hyperspectral dataset we generate a synthetic dataset withhigh-spatial-resolutionmultispectral images and low-spatial-resolution hyperspectral images and then employ a block-based fusion algorithm for sharpening the hyperspectral

images using the multispectral images Initially the hyper-spectral imagery obtained from CASI sensor is grouped inthree blocks 4134ndash5859 nm (corresponding to bands 1ndash13 ofthe hyperspectral images from CASI) 6003ndash6864 nm (cor-responding to bands 14ndash20 of the hyperspectral images fromCASI) and 7007ndash10449 nm (corresponding to bands 21ndash45of the hyperspectral images from CASI) Then we generatea synthetic dataset with high-spatial-resolution multispectralimages and low-spatial-resolution hyperspectral images cor-responding to each block [35 36] Here HSℎ119905

1

and HSℎ1199052

arethe original hyperspectral images with high spatial resolu-tions before the change (time 1199051) and after the change(time 1199052) respectively The workflow for the synthetic datasetgeneration is shown in Figure 4

Themultispectral data with high spatial resolution (MSℎ1199051

)at time 1199051 are obtained by spectrally degrading HSℎ119905

1

byaveraging all bands in each block of the hyperspectral image[37] Then hyperspectral data with low spatial resolution(HS119897119905

1

) are created by downscaling HSℎ1199051

via a Gaussian pointspread function (PSF) [38] In the downscaling process theoriginal hyperspectral data with 05m spatial resolution aredownscaled to 1m spatial resolution By applying this processto HSℎ119905

2

we obtain two synthetic multitemporal datasetsMSℎ119905

2

and HS1198971199052

After the synthetic multitemporal images aregenerated a cross-sharpening algorithm is appliedThe cross-sharpening method is defined as special image fusion that isaccomplished using hyperspectral and multispectral imagesobtained at the same or at different times using [35]

119891 (MSℎ1199051

HS1198971199051

) = 119865ℎ11990511199051

(1)

119891 (MSℎ1199052

HS1198971199052

) = 119865ℎ11990521199052

(2)

119891 (MSℎ1199051

HS1198971199052

) = 119865ℎ11990511199052

(3)

119891 (MSℎ1199052

HS1198971199051

) = 119865ℎ11990521199051

(4)

where 119891 is the specific fusion algorithm and 119865ℎ119898119899 is the syn-thetically fused image with high spatial resolution generatedusing themultispectral image at time119898 and the hyperspectralimage at time 119899 To generate the cross-sharpened hyperspec-tral image we employ a block-based fusion algorithm usinga multispectral band corresponding to the range of wave-lengths of hyperspectral bands [35 36 39] In contrast to typ-ical fusion processing with multispectral and panchromaticdata block-based fusion algorithm considers the multispec-tral bands as a panchromatic band and hyperspectral bandsas multispectral bands Each multispectral band is regardedas the panchromatic image and the corresponding hyper-spectral bands are considered the multispectral dataset Asmentioned above we divide the wavelength range into threeblocks Accordingly the multispectral image MSℎ119905

1

is com-posed of three bands MSℎ(1)119905

1

MSℎ(2)1199051

and MSℎ(3)1199051

and eachband of MSℎ119905

1

is regarded as a panchromatic image for pan-sharpening The hyperspectral image HS119897119905

1

is also partitionedas three hyperspectral images HS119897(1)119905

1

HS119897(2)1199051

and HS119897(3)1199051



Spectral downgrading


Spatially downgrading

High-spatial-resolution

multispectral data (time 1)



Low-spatial-resolution

hyperspectral data (time 1)



－３ℎt1

－３ℎt2

（３lt1

（３lt2

Figure 4 Workflow of the synthetic multitemporal dataset generation corresponding to each block of hyperspectral imagery

(1) (2) (3) (1) (2) (3)

Pansharpeningof block 1

Fℎ(1)t1t1


Fℎ(2)t1t1


Fℎ(3)t1t1

Layer-stacking of pansharpened images

Fℎt1t1

－３ℎt1

－３ℎt1

－３ℎt1

－３ℎt1

（３lt1

（３lt1

（３lt1

（３lt1

Block 1 Block 2 Block 3 Block 1 Block 2 Block 3

Figure 5 Workflow of the block-based pansharpening algorithm

corresponding to the wavelength of each band of MSℎ1199051

Thepansharpening algorithm is applied using generated pan-chromatic and partitioned hyperspectral images After apply-ing the block-based pansharpening approach fused multi-spectral bands of each block are integrated as 119865ℎ119905

11199051

Figure 5represents the workflow of block-based fusion using multi-spectral bands MSℎ119905

1

and hyperspectral bands HS1198971199051

In the case of the pansharpening algorithm we use the

Gram-Schmidt (GS) method which is a representative and

efficient pansharpening algorithm [40] In (1)ndash(4) the spatialresolution the number of bands and spectral wavelengthof the synthetically fused images are equivalent to those ofthe original hyperspectral image In addition fused imagescorresponding to the same multispectral images have similarspatial characteristics indicating that some of the geometricerrors in the change detection can be minimized using a pairof fused images with similar spatial information Figure 6shows an example in which the geometric error is minimizedin the synthetically fused images The spatial characteristics


(a) (b) (c)

Figure 6 Example of an image fusion result using synthetic hyperspectral images (a) 119865ℎ11990511199051 (b) 119865ℎ11990521199052 and (c) 119865ℎ11990511199052

of 119865ℎ11990511199051

in the tiled roof of the house are more similar to 119865ℎ11990511199052

than to 119865ℎ11990521199052

33 Spectral Similarity Measure for Detecting Change In thisstudy two representative normalized spectral measures areused First the spectral angle distance (SAD)measure is usedto calculate the spectral similarity between the spectra oftwo pixels in the originalmultitemporal hyperspectral images[41] We assume that 119909 = (1199091 1199092 119909119871) is a spectrum ofthe reference data and that 119910 = (1199101 1199102 119910119871) is a spectrumof the target data in a hyperspectral dataset with L bandsAccording to these spectral vectors SAD is calculated asfollows

SAD (119909 119910) = cosminus1 ( ⟨119909 119910⟩119909 10038171003817100381710038171199101003817100381710038171003817) (5)

where ⟨119909 119910⟩ represents the inner product between the twospectrum vectors and 119909 and 119910 represent the magnitudesof spectra 119909 and 119910 respectively The SAD range is [0-1]

Having obtained the SAD measure the Euclidean dis-tance (ED) is also used to measure the spectral dissimilaritybetween the pixels The ED a representative similarity dis-tance measure is defined as the square root of the sum of thesquared differences between the corresponding spectra TheED between two spectra is expressed as

ED (119909 119910) = [ 1119871 minus 1119871sum119894=1

(119909119894 minus 119910119894)2]12

(6)

The ED is mathematically simple conducive for rapid pro-cessing and sensitive to the absolute reflectance valuesIn addition the values are not normalized Therefore thevalues of ED between the two spectra are normalized by themagnitude of the spectral vector as follows

ED = ED minus EDminEDmax minus EDmin

(7)

where EDmax and EDmin are the maximum and minimumvalues of the calculated ED respectively According to (7) theED value can be normalized as [0-1]

34 Unsupervised Change Detection Using a Derivative Spec-tral Profile When change detection between 119865ℎ119905

11199051

and 119865ℎ11990521199052

of (1)-(2) is applied based on SAD and ED measures someunchanged areasmay be identified as changed areas given thedissimilarity of the geometric and radiometric characteristicsof the features However as noted in Section 32 the synthet-ically fused image 119865ℎ119905

11199052

has a tendency to adhere to the spatialproperties of the multispectral image at time 1 (MSℎ119905

1

) and toadhere to the spectral properties of the hyperspectral image attime 2 (HS119897119905

2

) Therefore if the fused image 119865ℎ11990511199052

is used withthe fused image at time 1 that is 119865ℎ119905

11199051

for change detectionthen the errors from the geometric difference between thetwo images are minimized while maintaining the spectraldifference between the two sets of data This characteristicis more significant in the derivative pixel spectra [42]Figure 7 illustrates the spectral profile (Figure 7(b)) andsecond-derivative spectral profile (Figure 7(c)) of the pixelthat is actually changed (Figure 7(a)) corresponding to acamouflage net that exists only at time 1 Comparing thesecond-derivative spectral profiles of fused images 119865ℎ119905

11199051

and119865ℎ119905

21199052

we find that 119865ℎ11990511199052

exhibits a high profile variation whichis beneficial for the identification of the changed region

Therefore the SAD and ED measures are integratedto generate a new similarity distance (integrated similaritydistance ISD) based on the combination of the original andderivative spectral profiles of the synthetically fused imagesThe ISD is defined as

ISD = SAD (119865ℎ11990511199051

119865ℎ11990521199052

)lowastmin (ED (11986510158401015840ℎ119905

11199051

11986510158401015840ℎ11990511199052

) ED (11986510158401015840ℎ11990521199052

11986510158401015840ℎ11990521199051

)) (8)

where 11986510158401015840 is the second-derivative spectral profile min is theminimum value and lowast is the convolution filter If a pixelhas a larger ISD value than a predefined threshold it isdetermined to be a changed pixel In the evaluation of theoriginal spectral profile SAD measurements between twohyperspectral datasets tend to exhibit different spectral andspatial characteristics However these results may include


Fℎt1t1

Fℎt2t2

Fℎt1t2

(a)

3530 40 452515105 200

Band number

0

50

100

150

200

250

300

350

400

450

Refle

ctan

ce

F11F22F12

(lowast1000)

(b)

Refle

ctan

ce

Band number454035302520151050

150

0

minus50

minus100

minus150

50

100

F11F22F12

(lowast1000)

(c)

Figure 7 Spectral property comparison of fused images for detecting changed areas (a) sample target of fused images 119865ℎ11990511199051 119865ℎ11990521199052 and 119865ℎ11990511199052 (b) spectral profile of pixel values and (c) spectral profile of the second-derivative pixel values (black line 119865ℎ11990511199051 red dot line 119865ℎ11990521199052 and bluedot line 119865ℎ11990511199052 )

false-positive regions of change detection The ED of thederivative spectral profile between the synthetically fusedimage pair demonstrates similar spatial characteristics thuswe calculated the ED through second-derivative spectralprofiles of the hyperspectral data In particular the overallinformation acquired from the synthetic image fusion pairs((119865ℎ119905

11199051

119865ℎ11990511199052

) and (119865ℎ11990521199052

119865ℎ11990521199051

)) is used for change detection tooptimize the accuracy of the change-detection results

4 Results and Discussion

41 Change-Detection Results To estimate the performanceof the proposed ISD measure various unsupervised state-of-the-art change-detection algorithms are applied The mag-nitude of the change vector (Euclidean distance of CVAor CVAED) and the spectral information divergence (SID)similarity measures are applied to detect the changed area

[24 41] Subspace-based change-detection (SCD) algorithmsthat is original SCD adaptive SCD (ASCD) and local SCD(LSCD) are also employed [26] ASCD and LSCD are appliedto estimate the effects of varying sensor positions or mis-registration between the multitemporal hyperspectral dataFinally IR-MAD is chosen as the change-detection algorithmfor quantitative estimations because it is frequently cited asa representative unsupervised change-detection algorithmIn the case of ASCD and LSCD a window size of 3 times3 is selected while considering the size of the camouflagenets at the test site To apply IR-MAD PCA is applied aspreprocessing for dimensionality reduction and the top fiveprincipal components including approximately 99 of theinformation are used for calculating IR-MAD

Each change-detection result map is calculated from twocorresponding hyperspectral datasets and then normalizedover a range of [0-1] Figure 8 presents the calculated results


(a) (b) (c) (d)

(e) (f) (g) (h)

(i) (j)

Figure 8 Comparison of the change-detection results (a) reference hyperspectral image (b) target hyperspectral image (c) ground-truthimage (d) result using the CVAED (e) result using the SID (f) result using the SCD (g) result using the ASCD (h) result using the LSCD(i) result using the IR-MAD and (j) result using the proposed algorithm

of the state-of-the-art and proposed methods these aredisplayed using a gamma correction The change-detectionresults obtained from the CVAED (Figure 8(d)) have highervalues in the mountains than the other change-detectionalgorithms The estimated results using SID (Figure 8(e))and SCD-based algorithms (Figures 8(f)ndash8(g)) show highlydissimilar values in the ginseng field in the lower part of thesite In the case of ASCD and LSCD some unchanged areasof edge or linear features were efficiently removed relative tothe result obtained by SCD However small changed areasthat have similar size with the sliding window represent lowervalues than other unchanged areas because these changedareas are blurred by the effect of local processing based onthe moving window of ASCD and LSCD In addition theproposed algorithm (Figure 8(j)) shows more distinguish-able and contrasting values over the camouflage nets andartificial turf relative to the results obtained using otherchange-detectionmethods including the IR-MAD algorithm(Figure 8(i)) even though the result obtained by the IR-MAD

shows better quality of change-detection output than that ofthe other existing algorithms

The change-detection results can be quantitatively eval-uated by estimating the receiver operating characteristic(ROC) curve andby evaluating the binary changemapusing athresholding technique The ROC curve provides a graphicalplot for the estimating the performance and selecting anoptimal model from the class distribution [26 43 44]The ROC curve is composed of the cumulative distributionfunction of the detection rate versus that of the false-alarmrate From the curve an appropriate threshold for separatingchanged and unchanged areas can be selected Using the ROCcurve the area under the curve (AUC) can be calculatedThe AUC describes the probability that the change-detectionalgorithm will rank a randomly chosen positive data pointhigher than a randomly chosen negative data point [44]Figure 9 shows the ROC curves corresponding to eachalgorithm showing that the proposed algorithm can detectchanged regions better than the other methods regardless


0

01

02

03

04

05

06

07

08

09

1

Det

ectio

n ra

te

01 02 03 04 05 06 07 08 09 10

False-alarm rate

Result by CVAEDResult by SIDResult by SCDResult by ASCD

Result by LSCDResult by IR-MADResult by proposed algorithm

Figure 9 ROC curves corresponding to each similarity measure forchange detection

Table 2 AUC values corresponding to each similarity measure forchange detection

Method AUC valueCVAED 07776SID 07112SCD 08628ASCD 07153LSCD 06385IR-MAD 09562Proposed algorithm 09776

of the threshold used for the unsupervised change-detectionprocess Poor results of ASCD and LSCD relative to the SCDresult are due to the blurring effect of the process based ona sliding window as shown in Figures 8(g) and 8(h) IR-MAD shows a higher value than the other existing algorithmsbut requires the use of a dimensionality reduction processMoreover the AUC by IR-MAD is lower than that obtainedusing the proposed algorithmThe efficiency of our algorithmcan be proven based on its high AUC value without the use ofdimensionality reduction techniques such as PCA and MNFas shown in Table 2 The ROC analysis indicates that theproposed method can optimize the change-detection resultswhile minimizing the false-alarm rate

An automatic thresholding technique is applied basedon the unsupervised change-detection results presented inFigure 8 The threshold determination is based on Rosinrsquosthreshold algorithm [45] Rosinrsquos threshold which is a rep-resentative unimodal threshold assumes that the histogramof the similarity distance image is a proportional unimodalfunction and that a nonchanged class of pixels (eg the back-ground) is considerably larger than the changed class of thepixels in the image Rosinrsquos threshold fits a straight line fromthe peak of the histogram Then the point of the maximumdeviation between the line and the histogram curve is selectedas the threshold Using the estimated threshold values we

extract the binary images corresponding to the changed areaAfter the thresholding a 3-by-3 median filter is applied toconsider the minimum size of the changed regions

Figure 10 depicts the change-detection results using theexisting and proposed algorithms with automatic threshold-ing for unsupervised change detection Figure 11 presents themagnified results near the targets for a visual assessment Asshown in Figures 10(a) and 11(c) ground-truth data are gener-ated including a total of 863 changed pixels Referring to thechange-detection results using CVAED the nearly changedregions are represented as nonchanged areas (Figure 10(b))The change-detection results based on the SID- and SCD-based algorithms are presented in Figures 10(c)ndash10(f) Inthese results the actual changed regions for example thecamouflage net and the artificial turf in the middle of thesite and the ginseng field in the lower part of the site aredetected as changed areas This result is attributed to theinclined canvas on the ginseng field which exhibits varyingradiometric and spatial properties based on the off-nadirangles of the given sceneThus many false alarms arise in thechange-detection results In addition the region correspond-ing to the brown camouflage net that has spectral propertiessimilar to the neighboring pixels is not detected as a changedregion In the results of the IR-MAD algorithm the artificialturf and certain building roofs (left side of Figure 10(a)) arenot detected as changed areas The result of the proposedalgorithm integrated by convoluting the similarity measuresbetween the synthetically fused images is presented in Fig-ure 10(h)The ginseng field is considered to be an unchangedarea whereas the camouflage net and artificial turf regionsare effectively detected as changed areas (Figure 11(j)) relativeto the IR-MAD result (Figure 11(i)) However in all cases thesmall targets of the camouflage net are not detected due tothe limited spatial resolution of the data With the exceptionof the small targets all of the changed regions are detectedusing the proposed algorithm and some regions that did notactually change are also detected These errors are attributedto the atmospheric correction and weather conditions Asshown in Figure 10(b) cloud shadows occurred on the rightside of the datasets Thus some change-detection errorsoccurred based on these areas To quantitatively evaluatethe change-detection results the detected changes overallaccuracy completeness correctness and false-alarm rate arecalculated based on the ground-truth data Evaluation factorsare estimated with the number of pixels and the results arepresented in Table 3 The change-detection result based onthe proposed algorithm shows higher completeness than theresults of the existing algorithms Therefore our algorithmdetected the highest portion of the actually changed areasThe CVAED IR-MAD and proposed algorithms showedbetter overall accuracy correctness and false-alarm ratesthan the SID- and SCD-based algorithms However the highcorrectness and low false-alarm rates of the CVAED and IR-MAD results indicate that these algorithms could not ade-quately detect the changed areas in contrast to the proposedalgorithmTherefore we conclude that our algorithm detectschanged areas efficiently and minimizes the overall error inthe change detectionwithout the use of dimensionality reduc-tion


(a) (b) (c) (d)

(e) (f) (g) (h)

Figure 10 Comparison of the change-detectionmaps (a) ground-truth image (b) result using the CVAED (c) result using the SID (d) resultusing the SCD (e) result using the ASCD (f) result using the LSCD (g) result using the IR-MAD and (h) result using the proposed algorithm

Table 3 Accuracy of the change-detection maps using the existing and proposed algorithms

Overall accuracy Completeness Correctness False-alarm rateCVAED 0995 0154 0109 0003SID 0905 0184 0005 0093SCD 0899 0538 0012 0100ASCD 0912 0328 0009 0087LSCD 0917 0152 0004 0081IR-MAD 0991 0461 0127 0007Proposed algorithm 0983 0768 0097 0017

5 Conclusions

In this study we proposed a novel algorithm appropriate fordetecting changes in hyperspectral data without the dimen-sionality reduction step that can lose specific information ofthe data Synthetically fused images that preserve spectraldifferences while minimizing the spatial differences betweenthe two images acquired at different times were used for thechange-detection procedure Based on the proposed simi-larity distance and the application of the second-derivativespectral profiles of pixels most of the changed regionsexcept for certain small targets were effectively detectedThe change-detection results using our algorithm showed thehighest AUC value relative to the results from the other state-of-the-art algorithms The high AUC value indicates that thechange-detection result can be optimized while minimizingthe false-alarm rate In the case of the evaluation from thebinary change-detection map the CVAED and IR-MAD

methods showed higher overall accuracies of the change-detection results than that of the proposed approach Theseresults arose due to the higher correctness and false-alarmrates of these methods The proposed method producedthe highest completeness values implying that the proposedalgorithm detected the largest portion of truly changed areasTherefore we conclude that our methodology can be usedto detect subtle changes in hyperspectral data acquired atdifferent times using the properties of the derivative spectralprofiles in pixels

However certain regions that were not changed betweenthe two data were incorrectly detected due to the spectraldissimilarity caused by atmospheric effects and cloud shad-ows Moreover small targets were also not detected due tothe limited spatial resolution of the acquired hyperspectraldataThis problemwill be addressed in our future work usingvarious hyperspectral data together with modifications of thesimilarity distance measure


(a) (b) (c) (d)

(e) (f) (g) (h)

(i) (j)

Figure 11 Comparison of the magnified change-detection maps (a) reference hyperspectral image (b) target hyperspectral image (c)ground-truth image (d) result using the CVAED (e) result using the SID (f) result using the SCD (g) result using the ASCD (h) resultusing the LSCD (i) result using the IR-MAD and (j) result using the proposed algorithm

Conflicts of Interest

The authors declare that they have no conflicts of interest

Acknowledgments

This work was supported by the Space Core TechnologyDevelopment Program through the National Research Foun-dation of Korea (NRF) andwas funded by theMinistry of Sci-ence ICT amp Future Planning (NRF-2012M1A3A3A02033469and NRF-2014M1A3A3A03034798)

References

[1] H Yamamoto R Nakamura and S Tsuchida ldquoRadiometriccalibration plan for the hyperspectral imager suite (HISUI)instrumentsrdquo in Proceedings of the Multispectral Hyperspectraland Ultraspectral Remote Sensing Technology Techniques andApplications IVConference 85270V Kyoto Japan October 2012

[2] R J Radke S Andra O Al-Kofahi and B Roysam ldquoImagechange detection algorithms a systematic surveyrdquo IEEE Trans-actions on Image Processing vol 14 no 3 pp 294ndash307 2005

[3] D Lu PMausel E Brondızio and EMoran ldquoChange detectiontechniquesrdquo International Journal of Remote Sensing vol 25 no12 pp 2365ndash2407 2004

[4] M Hussain D Chen A Cheng H Wei and D StanleyldquoChange detection from remotely sensed images from pixel-based to object-based approachesrdquo ISPRS Journal of Photogram-metry and Remote Sensing vol 80 pp 91ndash106 2013

[5] A Singh ldquoDigital change detection techniques using remotely-sensed datardquo International Journal of Remote Sensing vol 10 no6 pp 989ndash1003 1989

[6] D Renza E Martinez and A Arquero ldquoA new approach tochange detection in multispectral images by means of ERGASindexrdquo IEEE Geoscience and Remote Sensing Letters vol 10 no1 pp 76ndash80 2013

[7] J Im and J R Jensen ldquoA change detection model based onneighborhood correlation image analysis and decision tree


classificationrdquo Remote Sensing of Environment vol 99 no 3 pp326ndash340 2005

[8] S Patra S Ghosh and A Ghosh ldquoHistogram thresholdingfor unsupervised change detection of remote sensing imagesrdquoInternational Journal of Remote Sensing vol 32 no 21 pp 6071ndash6089 2011

[9] I Molina E Martinez A Arquero G Pajares and J SanchezldquoEvaluation of a change detection methodology by meansof binary thresholding algorithms and informational fusionprocessesrdquo Sensors vol 12 no 3 pp 3528ndash3561 2012

[10] L Bruzzone and D F Prieto ldquoAn adaptive parcel-based tech-nique for unsupervised change detectionrdquo International Journalof Remote Sensing vol 21 no 4 pp 817ndash822 2000

[11] L Bruzzone ldquoAutomatic analysis of the difference image forunsupervised change detectionrdquo IEEE Transactions on Geo-science and Remote Sensing vol 38 no 3 pp 1171ndash1182 2000

[12] G Pajares ldquoA hopfield neural network for image changedetectionrdquo IEEE Transactions on Neural Networks vol 17 no5 pp 1250ndash1264 2006

[13] L Bruzzone and F Bovolo ldquoA novel framework for the designof change-detection systems for very-high-resolution remotesensing imagesrdquo Proceedings of the IEEE vol 101 no 3 pp 609ndash630 2013

[14] M T Eismann J Meola and R C Hardie ldquoHyperspectralChangeDetection in the Presence ofDiurnal and Seasonal Vari-ationsrdquo IEEE Transactions on Geoscience and Remote Sensingvol 46 no 1 pp 237ndash249 2008

[15] K Kim ldquoStudy on Improving Hyperspectral Target Detectionby Target Signal Exclusion in Matched Filteringrdquo Korean Jour-nal of Remote Sensing vol 31 no 5 pp 433ndash440 2015

[16] A Song J Choi A Chang and Y Kim ldquoChange DetectionUsing Spectral Unmixing and IEA(Iterative Error Analysis) forHyperspectral Imagesrdquo Korean Journal of Remote Sensing vol31 no 5 pp 361ndash370 2015

[17] Z Hao H-J Song and B-C Yu ldquoApplication of hyper spectralremote sensing for urban forestrymonitoring in natural disasterzonesrdquo in Proceedings of the 2011 International Conference onComputer and Management CAMAN 2011 May 2011

[18] A A Nielsen ldquoThe regularized iteratively reweighted MADmethod for change detection in multi- and hyperspectral datardquoIEEE Transactions on Image Processing vol 16 no 2 pp 463ndash478 2007

[19] T Celik ldquoUnsupervised change detection in satellite imagesusing principal component analysis and 120581-means clusteringrdquoIEEE Geoscience and Remote Sensing Letters vol 6 no 4 pp772ndash776 2009

[20] M J Canty and A A Nielsen ldquoAutomatic radiometric normal-ization of multitemporal satellite imagery with the iterativelyre-weighted MAD transformationrdquo Remote Sensing of Environ-ment vol 112 no 3 pp 1025ndash1036 2008

[21] P R Marpu P Gamba and M J Canty ldquoImproving changedetection results of ir-mad by eliminating strong changesrdquo IEEEGeoscience andRemote Sensing Letters vol 8 no 4 pp 799ndash8032011

[22] M J Canty and A A Nielsen ldquoLinear and kernel methods formultivariate change detectionrdquoComputers and Geosciences vol38 no 1 pp 107ndash114 2012

[23] J Meola M T Eismann R L Moses and J N Ash ldquoDetect-ing changes in hyperspectral imagery using a model-basedapproachrdquo IEEE Transactions on Geoscience and Remote Sens-ing vol 49 no 7 pp 2647ndash2661 2011

[24] S Liu L Bruzzone F Bovolo and P Du ldquoHierarchical unsuper-vised change detection inmultitemporal hyperspectral imagesrdquoIEEE Transactions on Geoscience and Remote Sensing vol 53no 1 pp 244ndash260 2015

[25] O A C Junior R F Guimaraes A R Gillespie N C Silva andRA TGomes ldquoAnew approach to change vector analysis usingdistance and similarity measuresrdquo Remote Sensing vol 3 no 11pp 2473ndash2493 2011

[26] C Wu B Du and L Zhang ldquoA subspace-based change detec-tion method for hyperspectral imagesrdquo IEEE Journal of SelectedTopics in Applied Earth Observations and Remote Sensing vol 6no 2 pp 815ndash830 2013

[27] A Brook and E Ben-Dor ldquoAdvantages of the boresight effectin hyperspectral data analysisrdquo Remote Sensing vol 3 no 3 pp484ndash502 2011

[28] A Averbuch and M Zheludev ldquoTwo linear unmixing algo-rithms to recognize targets using supervised classification andorthogonal rotation in airborne hyperspectral imagesrdquo RemoteSensing vol 4 no 2 pp 532ndash560 2012

[29] Y Yuan Q Wang and G Zhu ldquoFast hyperspectral anomalydetection via high-order 2-d crossing filterrdquo IEEE Transactionson Geoscience and Remote Sensing vol 53 no 2 pp 620ndash6302015

[30] J R G Townshend C Gurney J McManus and C O JusticeldquoThe impact of misregistration on change detectionrdquo IEEETransactions on Geoscience and Remote Sensing vol 30 no 5pp 1054ndash1060 1992

[31] ldquoConstruction and data analysis of test-bed by hyperspectralimageryrdquo httpwwwhyperspectral-testbedcom

[32] A Brook and E B Dor ldquoSupervised vicarious calibration(SVC) of hyperspectral remote-sensing datardquo Remote Sensing ofEnvironment vol 115 no 6 pp 1543ndash1555 2011

[33] B Wang S Choi Y Byun S Lee and J Choi ldquoObject-basedchange detection of very high resolution satellite imagery usingthe cross-sharpening of multitemporal datardquo IEEE Geoscienceand Remote Sensing Letters vol 12 no 5 pp 1151ndash1155 2015

[34] Z Chen J Chen P Shi andM Tamura ldquoAn IHS-based changedetection approach for assessment of urban expansion impacton arable land loss in Chinardquo International Journal of RemoteSensing vol 24 no 6 pp 1353ndash1360 2003

[35] B Wang S-K Choi Y-K Han S-K Lee and J-W ChoildquoApplication of IR-MAD using synthetically fused images forchange detection in hyperspectral datardquo Remote Sensing Lettersvol 6 no 8 pp 578ndash586 2015

[36] Y Kim and J Choi ldquoEvaluation of block-based sharpeningalgorithms for fusion of Hyperion and ALI imageryrdquo Journal ofthe Korean Society of Surveying Geodesy Photogrammetry andCartography vol 33 no 1 pp 63ndash70 2015

[37] M T Eismann and R C Hardie ldquoHyperspectral resolutionenhancement using high-resolutionmultispectral imagery witharbitrary response functionsrdquo IEEE Transactions on Geoscienceand Remote Sensing vol 43 no 3 pp 455ndash465 2005

[38] N Yokoya T Yairi and A Iwasaki ldquoCoupled nonnegativematrix factorization unmixing for hyperspectral andmultispec-tral data fusionrdquo IEEE Transactions on Geoscience and RemoteSensing vol 50 no 2 pp 528ndash537 2012

[39] D Sylla A Minghelli-Roman P Blanc A Mangin and OHembise Fanton DrsquoAndon ldquoFusion of multispectral images byextension of the pan-sharpening ARSIS methodrdquo IEEE Journalof Selected Topics in Applied Earth Observations and RemoteSensing vol 7 no 5 pp 1781ndash1791 2014


[40] B Aiazzi S Baronti F Lotti and M Selva ldquoA comparisonbetween global and context-adaptive pansharpening of multi-spectral imagesrdquo IEEE Geoscience and Remote Sensing Lettersvol 6 no 2 pp 302ndash306 2009

[41] 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

[42] J Zhang B Rivard and A Sanchez-Azofeifa ldquoDerivative spec-tral unmixing of hyperspectral data applied to mixtures oflichen and rockrdquo IEEE Transactions on Geoscience and RemoteSensing vol 42 no 9 pp 1934ndash1940 2004

[43] L C Alatorre R Sanchez-Andres S Cirujano S Beguerıaand S Sanchez-Carrillo ldquoIdentification of mangrove areas byremote sensingThe ROC curve technique applied to the north-western Mexico coastal zone using Landsat imageryrdquo RemoteSensing vol 3 no 8 pp 1568ndash1583 2011

[44] T Fawcett ldquoAn introduction to ROC analysisrdquo Pattern Recogni-tion Letters vol 27 no 8 pp 861ndash874 2006

[45] P L Rosin and E Ioannidis ldquoEvaluation of global image thresh-olding for change detectionrdquoPattern Recognition Letters vol 24no 14 pp 2345ndash2356 2003

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Active and Passive Electronic Components

Control Scienceand Engineering

Journal of


International Journal of

RotatingMachinery


Hindawi Publishing Corporation httpwwwhindawicom

Journal of

Volume 201

Submit your manuscripts athttpswwwhindawicom

VLSI Design



Shock and Vibration


Civil EngineeringAdvances in

Acoustics and VibrationAdvances in



Electrical and Computer Engineering

Journal of

Advances inOptoElectronics


Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of


Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014


Chemical EngineeringInternational Journal of Antennas and

Propagation




Navigation and Observation



DistributedSensor Networks



extent of the damaged area Several researchers have analyzedthe existing change-detection methods for applications inthe remote sensing data including image differencing imagerationing vegetation index regression data transformationchange vector postclassification comparison and GIS-basedmethods [2ndash5] According to the literature change-detectiontechniques comprise two types supervised change detec-tion and unsupervised change detection [5] In particularunsupervised change-detection techniques are preferred forremote sensing applications because the training data donot need to be manually collected even though super-vised methods can provide change aspects of specific land-cover classes and ldquofrom-tordquo information of change detection[3]

Traditional unsupervised change-detection algorithmsare focused on proposing change-detection indices andapplying automatic thresholding methods for generating abinary map of the changed area For example change-detection indices such as change vector analysis (CVA)image correlation and Relative Dimensionless Global Error(ERGAS) have been proposed [6 7] In addition variousthresholding techniques have been developed and appliedto multitemporal remote sensing data [8] Molina et al [9]integrated different change-detection indices to improve thestatistical properties of indices for single change detectionBruzzone and Prieto proposed modified image differencingand thresholding of adaptive parcels based on a homogeneousregion [10] Markov random fields (MRFs) expectation-maximization (EM) algorithms and neural networks havebeen used to determine optical thresholds for automaticchange detection [11 12] In addition Bruzzone and Bovolo[13] analyzed change-detection techniques for very-high-resolution (VHR) satellite imagery and developed a novelframework using the top-down approach considering theradiometric characteristics of multitemporal images

With the launches and high availability of hyperspectralsensors new change-detection methodologies have beendeveloped and existing methods have been modified tomake them appropriate for hyperspectral imagery Eismannet al [14] proposed an algorithm based on linear predictorsto detect subtle targets against a complex background andKim [15] modified matched filtering using target signalexclusion Song et al [16] proposed unsupervised change-detection algorithms using spectral unmixing and iterativeerror analysis (IEA) Hao et al [17] applied hyperspectral datato detect changes in urban forest resources in natural disasterzones Image transformation techniques such asmultivariatealteration detection (MAD) and principal component anal-ysis (PCA) which can be applied to general multispectralimages have been extended to include hyperspectral data [1819] In particular iterative regularized multivariate alterationdetection (IR-MAD) is considered a state-of-the-art change-detection algorithm due to its stability and outstandingchange-detection results [20ndash22] However dimensionalityreduction through PCA or the minimum noise fraction(MNF) is required prior to the application of IR-MAD-based algorithms Meola et al [23] proposed a model-based change-detection approach that reduces false alarmscaused by shadow differences using calibrated hyperspectral

data and Liu et al [24] proposed a hierarchical change-detection algorithm using endmember detection and clustermerging in multitemporal hyperspectral images Junior etal [25] used CVA based on distance (Euclidean distanceand Mahalanobis distance) and similarity (spectral anglemapper SAM and spectral correlation mapper SCM) mea-surements and concluded that CVA using the Euclideandistance and SAM is superior to other measures Wu et al[26] developed a subspace-based algorithm using two typesof information from Hyperion and from the Chinese HJ-1A satellite proving that measurements based on the SAMare identical to the orthogonal subspace projection- (OSP-)based measurements Based on this concept the authorshave proposed local and adaptive measures based on OSPto minimize registration errors Other advanced algorithmsbased on target detection anomaly detection and changedetection have also been developed recently [27ndash29]

These related approaches have generally focused on thedevelopment of change-detection algorithms that increaseaccuracy by effectively using hyperspectral bands to maxi-mize the advantages of hyperspectral sensors In other wordsthese approaches have focused on using abundant spectralinformation However the dissimilarity of radiometric andgeometric properties between multitemporal data basedon the acquisition time or position of the sensors shouldbe resolved to enable hyperspectral imagery for detectingchanges in natural and human-impact areas Moreover fewconsiderations related to data noise have been included in thespectrum of hyperspectral imagery potentially decreasingthe change-detection accuracy when general change-detec-tion algorithms are applied to hyperspectral images withoutdimensionality reduction

Most studies have also considered hyperspectral datawithlow or medium spatial resolutions due to the technical limi-tations of generating hyperspectral sensors with high spatialresolutions However considering the future development ofhyperspectral sensors hyperspectral data with high spatialresolutions are necessaryThus approaches for change detec-tion should consider high spectral and spatial resolutionssimultaneously High-resolution images acquired at varioustimes demonstrate local geometric location differences andspectral variation by phenomenological or temporal differ-ences between dates of imagery even when the data are geo-referenced These geometric differences can cause poorchange-detection results based on false alarms of the changedarea caused by the differences [30] Information exclusivelyobtained from the data acquired at two times that is beforeand after the change occurs is insufficient for accurate changedetection

The main objective of this study is to develop a change-detection index suitable for hyperspectral imagery Weapplied unsupervised change detection using an image fusiontechnique and an ordered combination of similaritymeasuresto identify the changed regions To minimize the radiometricand geometric dissimilarities between multitemporal imageswith high spatial and spectral resolutions a temporal-data-based synthetic image fusion method is applied which gen-erates combined images at two data points at different timesA combination of the synthetically fused images is then used


N

1208040200(m)

354100000000

353900000000

354000000000

354100000000

353900000000

354000000000

4049500000000

4049700000000

4049600000000

4049500000000

4049700000000

4049600000000

(a)

(b) (c)










(a) (b) (c)


Wavelength (nm)1000900800700600500400

0

01

02

03

04

05

06

07

08


Refle

ctan

ce

(d)



3 Methodology






Preprocessing









1

and HSℎ1199052




1


1



2


2

and HS1198971199052


119891 (MSℎ1199051

HS1198971199051

) = 119865ℎ11990511199051

(1)

119891 (MSℎ1199052

HS1198971199052

) = 119865ℎ11990521199052

(2)

119891 (MSℎ1199051

HS1198971199052

) = 119865ℎ11990511199052

(3)

119891 (MSℎ1199052

HS1198971199051

) = 119865ℎ11990521199051

(4)


1


1

MSℎ(2)1199051

and MSℎ(3)1199051


1


1


1

HS119897(2)1199051

and HS119897(3)1199051














－３ℎt1

－３ℎt2

（３lt1

（３lt2


(1) (2) (3) (1) (2) (3)


Fℎ(1)t1t1


Fℎ(2)t1t1


Fℎ(3)t1t1


Fℎt1t1

－３ℎt1

－３ℎt1

－３ℎt1

－３ℎt1

（３lt1

（３lt1

（３lt1

（３lt1





11199051


1






(a) (b) (c)


of 119865ℎ11990511199051


than to 119865ℎ11990521199052


SAD (119909 119910) = cosminus1 ( ⟨119909 119910⟩119909 10038171003817100381710038171199101003817100381710038171003817) (5)



ED (119909 119910) = [ 1119871 minus 1119871sum119894=1

(119909119894 minus 119910119894)2]12

(6)



(7)



11199051

and 119865ℎ11990521199052


11199052


1


2



11199051


11199051

and119865ℎ119905

21199052

we find that 119865ℎ11990511199052



ISD = SAD (119865ℎ11990511199051

119865ℎ11990521199052

)lowastmin (ED (11986510158401015840ℎ119905

11199051

11986510158401015840ℎ11990511199052

) ED (11986510158401015840ℎ11990521199052

11986510158401015840ℎ11990521199051

)) (8)



Fℎt1t1

Fℎt2t2

Fℎt1t2

(a)

3530 40 452515105 200

Band number

0

50

100

150

200

250

300

350

400

450

Refle

ctan

ce

F11F22F12

(lowast1000)

(b)

Refle

ctan

ce

Band number454035302520151050

150

0

minus50

minus100

minus150

50

100

F11F22F12

(lowast1000)

(c)



11199051

119865ℎ11990511199052

) and (119865ℎ11990521199052

119865ℎ11990521199051







(a) (b) (c) (d)

(e) (f) (g) (h)

(i) (j)






0

01

02

03

04

05

06

07

08

09

1

Det

ectio

n ra

te

01 02 03 04 05 06 07 08 09 10

False-alarm rate











(a) (b) (c) (d)

(e) (f) (g) (h)




5 Conclusions





(a) (b) (c) (d)

(e) (f) (g) (h)

(i) (j)




Acknowledgments


References

















































RoboticsJournal of





Journal of



RotatingMachinery



Journal of

Volume 201


VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










N

1208040200(m)

354100000000

353900000000

354000000000

354100000000

353900000000

354000000000

4049500000000

4049700000000

4049600000000

4049500000000

4049700000000

4049600000000

(a)

(b) (c)










(a) (b) (c)


Wavelength (nm)1000900800700600500400

0

01

02

03

04

05

06

07

08


Refle

ctan

ce

(d)



3 Methodology






Preprocessing









1

and HSℎ1199052




1


1



2


2

and HS1198971199052


119891 (MSℎ1199051

HS1198971199051

) = 119865ℎ11990511199051

(1)

119891 (MSℎ1199052

HS1198971199052

) = 119865ℎ11990521199052

(2)

119891 (MSℎ1199051

HS1198971199052

) = 119865ℎ11990511199052

(3)

119891 (MSℎ1199052

HS1198971199051

) = 119865ℎ11990521199051

(4)


1


1

MSℎ(2)1199051

and MSℎ(3)1199051


1


1


1

HS119897(2)1199051

and HS119897(3)1199051














－３ℎt1

－３ℎt2

（３lt1

（３lt2


(1) (2) (3) (1) (2) (3)


Fℎ(1)t1t1


Fℎ(2)t1t1


Fℎ(3)t1t1


Fℎt1t1

－３ℎt1

－３ℎt1

－３ℎt1

－３ℎt1

（３lt1

（３lt1

（３lt1

（３lt1





11199051


1






(a) (b) (c)


of 119865ℎ11990511199051


than to 119865ℎ11990521199052


SAD (119909 119910) = cosminus1 ( ⟨119909 119910⟩119909 10038171003817100381710038171199101003817100381710038171003817) (5)



ED (119909 119910) = [ 1119871 minus 1119871sum119894=1

(119909119894 minus 119910119894)2]12

(6)



(7)



11199051

and 119865ℎ11990521199052


11199052


1


2



11199051


11199051

and119865ℎ119905

21199052

we find that 119865ℎ11990511199052



ISD = SAD (119865ℎ11990511199051

119865ℎ11990521199052

)lowastmin (ED (11986510158401015840ℎ119905

11199051

11986510158401015840ℎ11990511199052

) ED (11986510158401015840ℎ11990521199052

11986510158401015840ℎ11990521199051

)) (8)



Fℎt1t1

Fℎt2t2

Fℎt1t2

(a)

3530 40 452515105 200

Band number

0

50

100

150

200

250

300

350

400

450

Refle

ctan

ce

F11F22F12

(lowast1000)

(b)

Refle

ctan

ce

Band number454035302520151050

150

0

minus50

minus100

minus150

50

100

F11F22F12

(lowast1000)

(c)



11199051

119865ℎ11990511199052

) and (119865ℎ11990521199052

119865ℎ11990521199051







(a) (b) (c) (d)

(e) (f) (g) (h)

(i) (j)






0

01

02

03

04

05

06

07

08

09

1

Det

ectio

n ra

te

01 02 03 04 05 06 07 08 09 10

False-alarm rate











(a) (b) (c) (d)

(e) (f) (g) (h)




5 Conclusions





(a) (b) (c) (d)

(e) (f) (g) (h)

(i) (j)




Acknowledgments


References

















































RoboticsJournal of





Journal of



RotatingMachinery



Journal of

Volume 201


VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










(a) (b) (c)


Wavelength (nm)1000900800700600500400

0

01

02

03

04

05

06

07

08


Refle

ctan

ce

(d)



3 Methodology






Preprocessing









1

and HSℎ1199052




1


1



2


2

and HS1198971199052


119891 (MSℎ1199051

HS1198971199051

) = 119865ℎ11990511199051

(1)

119891 (MSℎ1199052

HS1198971199052

) = 119865ℎ11990521199052

(2)

119891 (MSℎ1199051

HS1198971199052

) = 119865ℎ11990511199052

(3)

119891 (MSℎ1199052

HS1198971199051

) = 119865ℎ11990521199051

(4)


1


1

MSℎ(2)1199051

and MSℎ(3)1199051


1


1


1

HS119897(2)1199051

and HS119897(3)1199051














－３ℎt1

－３ℎt2

（３lt1

（３lt2


(1) (2) (3) (1) (2) (3)


Fℎ(1)t1t1


Fℎ(2)t1t1


Fℎ(3)t1t1


Fℎt1t1

－３ℎt1

－３ℎt1

－３ℎt1

－３ℎt1

（３lt1

（３lt1

（３lt1

（３lt1





11199051


1






(a) (b) (c)


of 119865ℎ11990511199051


than to 119865ℎ11990521199052


SAD (119909 119910) = cosminus1 ( ⟨119909 119910⟩119909 10038171003817100381710038171199101003817100381710038171003817) (5)



ED (119909 119910) = [ 1119871 minus 1119871sum119894=1

(119909119894 minus 119910119894)2]12

(6)



(7)



11199051

and 119865ℎ11990521199052


11199052


1


2



11199051


11199051

and119865ℎ119905

21199052

we find that 119865ℎ11990511199052



ISD = SAD (119865ℎ11990511199051

119865ℎ11990521199052

)lowastmin (ED (11986510158401015840ℎ119905

11199051

11986510158401015840ℎ11990511199052

) ED (11986510158401015840ℎ11990521199052

11986510158401015840ℎ11990521199051

)) (8)



Fℎt1t1

Fℎt2t2

Fℎt1t2

(a)

3530 40 452515105 200

Band number

0

50

100

150

200

250

300

350

400

450

Refle

ctan

ce

F11F22F12

(lowast1000)

(b)

Refle

ctan

ce

Band number454035302520151050

150

0

minus50

minus100

minus150

50

100

F11F22F12

(lowast1000)

(c)



11199051

119865ℎ11990511199052

) and (119865ℎ11990521199052

119865ℎ11990521199051







(a) (b) (c) (d)

(e) (f) (g) (h)

(i) (j)






0

01

02

03

04

05

06

07

08

09

1

Det

ectio

n ra

te

01 02 03 04 05 06 07 08 09 10

False-alarm rate











(a) (b) (c) (d)

(e) (f) (g) (h)




5 Conclusions





(a) (b) (c) (d)

(e) (f) (g) (h)

(i) (j)




Acknowledgments


References

















































RoboticsJournal of





Journal of



RotatingMachinery



Journal of

Volume 201


VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation











Preprocessing









1

and HSℎ1199052




1


1



2


2

and HS1198971199052


119891 (MSℎ1199051

HS1198971199051

) = 119865ℎ11990511199051

(1)

119891 (MSℎ1199052

HS1198971199052

) = 119865ℎ11990521199052

(2)

119891 (MSℎ1199051

HS1198971199052

) = 119865ℎ11990511199052

(3)

119891 (MSℎ1199052

HS1198971199051

) = 119865ℎ11990521199051

(4)


1


1

MSℎ(2)1199051

and MSℎ(3)1199051


1


1


1

HS119897(2)1199051

and HS119897(3)1199051














－３ℎt1

－３ℎt2

（３lt1

（３lt2


(1) (2) (3) (1) (2) (3)


Fℎ(1)t1t1


Fℎ(2)t1t1


Fℎ(3)t1t1


Fℎt1t1

－３ℎt1

－３ℎt1

－３ℎt1

－３ℎt1

（３lt1

（３lt1

（３lt1

（３lt1





11199051


1






(a) (b) (c)


of 119865ℎ11990511199051


than to 119865ℎ11990521199052


SAD (119909 119910) = cosminus1 ( ⟨119909 119910⟩119909 10038171003817100381710038171199101003817100381710038171003817) (5)



ED (119909 119910) = [ 1119871 minus 1119871sum119894=1

(119909119894 minus 119910119894)2]12

(6)



(7)



11199051

and 119865ℎ11990521199052


11199052


1


2



11199051


11199051

and119865ℎ119905

21199052

we find that 119865ℎ11990511199052



ISD = SAD (119865ℎ11990511199051

119865ℎ11990521199052

)lowastmin (ED (11986510158401015840ℎ119905

11199051

11986510158401015840ℎ11990511199052

) ED (11986510158401015840ℎ11990521199052

11986510158401015840ℎ11990521199051

)) (8)



Fℎt1t1

Fℎt2t2

Fℎt1t2

(a)

3530 40 452515105 200

Band number

0

50

100

150

200

250

300

350

400

450

Refle

ctan

ce

F11F22F12

(lowast1000)

(b)

Refle

ctan

ce

Band number454035302520151050

150

0

minus50

minus100

minus150

50

100

F11F22F12

(lowast1000)

(c)



11199051

119865ℎ11990511199052

) and (119865ℎ11990521199052

119865ℎ11990521199051







(a) (b) (c) (d)

(e) (f) (g) (h)

(i) (j)






0

01

02

03

04

05

06

07

08

09

1

Det

ectio

n ra

te

01 02 03 04 05 06 07 08 09 10

False-alarm rate











(a) (b) (c) (d)

(e) (f) (g) (h)




5 Conclusions





(a) (b) (c) (d)

(e) (f) (g) (h)

(i) (j)




Acknowledgments


References

















































RoboticsJournal of





Journal of



RotatingMachinery



Journal of

Volume 201


VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation






















－３ℎt1

－３ℎt2

（３lt1

（３lt2


(1) (2) (3) (1) (2) (3)


Fℎ(1)t1t1


Fℎ(2)t1t1


Fℎ(3)t1t1


Fℎt1t1

－３ℎt1

－３ℎt1

－３ℎt1

－３ℎt1

（３lt1

（３lt1

（３lt1

（３lt1





11199051


1






(a) (b) (c)


of 119865ℎ11990511199051


than to 119865ℎ11990521199052


SAD (119909 119910) = cosminus1 ( ⟨119909 119910⟩119909 10038171003817100381710038171199101003817100381710038171003817) (5)



ED (119909 119910) = [ 1119871 minus 1119871sum119894=1

(119909119894 minus 119910119894)2]12

(6)



(7)



11199051

and 119865ℎ11990521199052


11199052


1


2



11199051


11199051

and119865ℎ119905

21199052

we find that 119865ℎ11990511199052



ISD = SAD (119865ℎ11990511199051

119865ℎ11990521199052

)lowastmin (ED (11986510158401015840ℎ119905

11199051

11986510158401015840ℎ11990511199052

) ED (11986510158401015840ℎ11990521199052

11986510158401015840ℎ11990521199051

)) (8)



Fℎt1t1

Fℎt2t2

Fℎt1t2

(a)

3530 40 452515105 200

Band number

0

50

100

150

200

250

300

350

400

450

Refle

ctan

ce

F11F22F12

(lowast1000)

(b)

Refle

ctan

ce

Band number454035302520151050

150

0

minus50

minus100

minus150

50

100

F11F22F12

(lowast1000)

(c)



11199051

119865ℎ11990511199052

) and (119865ℎ11990521199052

119865ℎ11990521199051







(a) (b) (c) (d)

(e) (f) (g) (h)

(i) (j)






0

01

02

03

04

05

06

07

08

09

1

Det

ectio

n ra

te

01 02 03 04 05 06 07 08 09 10

False-alarm rate











(a) (b) (c) (d)

(e) (f) (g) (h)




5 Conclusions





(a) (b) (c) (d)

(e) (f) (g) (h)

(i) (j)




Acknowledgments


References

















































RoboticsJournal of





Journal of



RotatingMachinery



Journal of

Volume 201


VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










(a) (b) (c)


of 119865ℎ11990511199051


than to 119865ℎ11990521199052


SAD (119909 119910) = cosminus1 ( ⟨119909 119910⟩119909 10038171003817100381710038171199101003817100381710038171003817) (5)



ED (119909 119910) = [ 1119871 minus 1119871sum119894=1

(119909119894 minus 119910119894)2]12

(6)



(7)



11199051

and 119865ℎ11990521199052


11199052


1


2



11199051


11199051

and119865ℎ119905

21199052

we find that 119865ℎ11990511199052



ISD = SAD (119865ℎ11990511199051

119865ℎ11990521199052

)lowastmin (ED (11986510158401015840ℎ119905

11199051

11986510158401015840ℎ11990511199052

) ED (11986510158401015840ℎ11990521199052

11986510158401015840ℎ11990521199051

)) (8)



Fℎt1t1

Fℎt2t2

Fℎt1t2

(a)

3530 40 452515105 200

Band number

0

50

100

150

200

250

300

350

400

450

Refle

ctan

ce

F11F22F12

(lowast1000)

(b)

Refle

ctan

ce

Band number454035302520151050

150

0

minus50

minus100

minus150

50

100

F11F22F12

(lowast1000)

(c)



11199051

119865ℎ11990511199052

) and (119865ℎ11990521199052

119865ℎ11990521199051







(a) (b) (c) (d)

(e) (f) (g) (h)

(i) (j)






0

01

02

03

04

05

06

07

08

09

1

Det

ectio

n ra

te

01 02 03 04 05 06 07 08 09 10

False-alarm rate











(a) (b) (c) (d)

(e) (f) (g) (h)




5 Conclusions





(a) (b) (c) (d)

(e) (f) (g) (h)

(i) (j)




Acknowledgments


References

















































RoboticsJournal of





Journal of



RotatingMachinery



Journal of

Volume 201


VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










Fℎt1t1

Fℎt2t2

Fℎt1t2

(a)

3530 40 452515105 200

Band number

0

50

100

150

200

250

300

350

400

450

Refle

ctan

ce

F11F22F12

(lowast1000)

(b)

Refle

ctan

ce

Band number454035302520151050

150

0

minus50

minus100

minus150

50

100

F11F22F12

(lowast1000)

(c)



11199051

119865ℎ11990511199052

) and (119865ℎ11990521199052

119865ℎ11990521199051







(a) (b) (c) (d)

(e) (f) (g) (h)

(i) (j)






0

01

02

03

04

05

06

07

08

09

1

Det

ectio

n ra

te

01 02 03 04 05 06 07 08 09 10

False-alarm rate











(a) (b) (c) (d)

(e) (f) (g) (h)




5 Conclusions





(a) (b) (c) (d)

(e) (f) (g) (h)

(i) (j)




Acknowledgments


References

















































RoboticsJournal of





Journal of



RotatingMachinery



Journal of

Volume 201


VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










(a) (b) (c) (d)

(e) (f) (g) (h)

(i) (j)






0

01

02

03

04

05

06

07

08

09

1

Det

ectio

n ra

te

01 02 03 04 05 06 07 08 09 10

False-alarm rate











(a) (b) (c) (d)

(e) (f) (g) (h)




5 Conclusions





(a) (b) (c) (d)

(e) (f) (g) (h)

(i) (j)




Acknowledgments


References

















































RoboticsJournal of





Journal of



RotatingMachinery



Journal of

Volume 201


VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










0

01

02

03

04

05

06

07

08

09

1

Det

ectio

n ra

te

01 02 03 04 05 06 07 08 09 10

False-alarm rate











(a) (b) (c) (d)

(e) (f) (g) (h)




5 Conclusions





(a) (b) (c) (d)

(e) (f) (g) (h)

(i) (j)




Acknowledgments


References

















































RoboticsJournal of





Journal of



RotatingMachinery



Journal of

Volume 201


VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










(a) (b) (c) (d)

(e) (f) (g) (h)




5 Conclusions





(a) (b) (c) (d)

(e) (f) (g) (h)

(i) (j)




Acknowledgments


References

















































RoboticsJournal of





Journal of



RotatingMachinery



Journal of

Volume 201


VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










(a) (b) (c) (d)

(e) (f) (g) (h)

(i) (j)




Acknowledgments


References

















































RoboticsJournal of





Journal of



RotatingMachinery



Journal of

Volume 201


VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation


















































RoboticsJournal of





Journal of



RotatingMachinery



Journal of

Volume 201


VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation
















RoboticsJournal of





Journal of



RotatingMachinery



Journal of

Volume 201


VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation









RoboticsJournal of





Journal of



RotatingMachinery



Journal of

Volume 201


VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation









Documents

An Unsupervised Algorithm for Change Detection in …downloads.hindawi.com/journals/js/2017/9702612.pdf · 2019-07-30 · Traditional unsupervised change-detection algorithms are