11.IJAEST Vol No 7 Issue No 1 Topographic Influence on Improved Change Vector Analysis Using MODIS...

8/6/2019 11.IJAEST Vol No 7 Issue No 1 Topographic Influence on Improved Change Vector Analysis Using MODIS Satellite Dat

1/8

Topographic Influence on Improved Change Vector

Analysis using MODIS Satellite Data of Western

Himalaya

Sartajvir Singh11M-Tech Student, E.C.E Deptt.

R.I.E.I.T, Railmajra,S.B.S. Nagar, Punjab, India

[email protected]

Prof. J.K. Sharma22Director, Engineering Deptt.

R.I.E.I.T, RailmajraS.B.S. Nagar, Punjab, [email protected]

Dr. V.D. Mishra33Scientist (E), R.S. Research Group

SASE, DRDOChandigarh, India

[email protected]

Abstract: -Change detection is a technique used to determinethe change between two or more time periods of a particular

area of study. Change detection is an important process of

monitoring and managing environment resources. There are

many methods available for change detection. In this paper,

we are dealing with improved change-vector analysis onwestern Himalayas (usually snow covered area) using MODIS

Satellite data. This method performed in two stages: in first

stage double- window flexible pace search (DFPS), which

aims at determining the threshold of change magnitude, and

in second stage, the direction cosines of change vectors for

determining change direction that combines single-date image

classification with a minimum-distance categorizing

technique. On other side, the topographic variability causes a

problem of differential illumination due to steep and varying

slopes in rugged Himalayan terrain. Therefore, topographic

corrections are essential for qualitative and quantitative

analysis of snow cover applications. Here we compared the

improved change vector analysis (ICVA) with and without

topographic correction on study area of lower and middleHimalaya, Himachal Pradesh, India. The experiment result

shown that ICVA with topographic correction gives more

accurate and hidden information as compare to ICVA

without topographic correction.

Keywords: - Improved change vector analysis, direction

cosines, topographic correction, MODIS.

I. INTRODUCTIONIn the past few years, there has been a growing interest in

the development of change detection techniques for the

analysis of multi-temporal remote sensing imagery. Thisinterest stems from the wide range of applications in whichchange detection methods can be used, like environmentalmonitoring, agricultural surveys, urban studies and snow covermonitoring. Change detection, by definition, requires imagesfrom two dates. Ideally, change detection procedures shouldinvolve data acquired by the same sensor, having the samespatial resolution, viewing geometry, spectral bands, samespatial location, radiometric resolution, and acquired at thesame time of day [1]. There are many techniques available formonitoring land cover changes, fall into two main categories:

The enhancement change detection techniques have theadvantage of generally being more accurate in identifyingareas of spectral change reported by[2]that includes: (1) imagedifferencing [3],(2) principal component analysis [4], (3)change vector analysis [5] and other is post classification

technique that involves the independent production andsubsequent comparison of spectral classifications for the samearea at two different time periods [6]. It is reported [7] thatimproved change vector analysis is valuable technique forchange detection analysis which includes a semiautomaticmethod, named double-window flexible pace search (DFPS),which aims at determining efficiently the threshold of changemagnitude, and a new method of determining change direction(change category) which combines a single imageclassification and a minimum-distance categorization basedupon the direction cosines of the change vector.

The topographic variability causes a problem of

differential illumination due to steep and varying slopes inrugged Himalayan terrain. Sun-facing illuminated slopes(south aspect) show more than expected spectral radiance orreflectance, whereas the effect is opposite in shaded relief area(north aspect) [8]. Therefore, topographic corrections areessential for qualitative and quantitative analysis of snowcover applications. There are many topographic correctionmethods available, fall into three main categories: (1)Empirical methods such as two stage normalization[9], (2)Lambertian methods such as C- correction, Cosine-T[10]etc.,(3) Non-Lambertian methods such as Minneartcorrection method [11], Slope match [12].It is reported[13]that slope match is most suitable technique for

topographic correction on Himalayan terrain.

II. STUDY AREAThe study area is a part of Lower and Middle Himalaya

and shown on MODIS image(Moderate Resolution ImagingSpectroradiometer) lies between latitude of 32.254 degree to32.999 degree North and longitude of 77.00 degree to 77.497degree East as shown in the Fig. 1. The lower part of the areais surrounded by forest and tree line exists up to 3100 m. Theupper part (Middle Himalaya) is devoid of forest. The average

Sartajvir Singh* et al. / (IJAEST) INTERNATIONAL JOURNAL OF ADVANCED ENGINEERING SCIENCES AND TECHNOLOGIESVol No. 7, Issue No. 1, 077 - 084

ISSN: 2230-7818 @ 2011 http://www.ijaest.iserp.org. All rights Reserved. Page 77
mailto:[email protected]:[email protected]:[email protected]:[email protected]


2/8

minimum temperature in winter is generally observed to be -12oC to -15oC in lower Himalaya (Pir-Panjal range) and -30oCto -35oC in Middle Himalaya (Greater Himalaya range). Pir-Panjal receives the highest snowfall (average 15-20 m) ascompared to Greater Himalayan range (12-15m) during thewinter period between October and May. The altitude in theentire study area varies from 1900 m to 6500 m with a meanvalue of 4700 m. The slope in the study area varies from 1-86

degree with mean value of 28 degree and aspect ranges from0-360 degree with mean values of 180 degree. Most of theslopes in the study regions are oriented to south aspect.

Figure 1 MODIS image of study area

III. SATELLITE DATASETSTwo almost cloud free satellite images of MODIS (Terra

platform) of 02nd November 2009(Pre image) and 21stNovember 2009(Post image) are used in the present work tostudy the influence of topography on improved spectralchange vector analysis. The salient specifications of MODISsensor are given in the Table 1.

IV. PRE-PROCESSINGA master scene of 56m spatial resolution of AWiFS

(Advance Wide field sensor) of study area is prepared afterrectification with high spatial resolution 23m of LISS-III(Linear Imaging self-Scanning) with 1:50,000 toposheet. Allsatellite images of MODIS were than geo-coded with AWiFSto the EVEREST datum by ERDAS/Imagine 9.1 (Leica

Geosystems GIS and Mapping LLC) software with sub pixelaccuracy. The preprocessed uncorrected image of 02nd

November 2009 and 21st November 2009 are shown in Fig. 2(a) and (b) respectively.

Figure 2 Uncorrected MODIS images(a) Pre image (02nd November 2009)(b) Post image (21st November 2009)

V. ESTIMATION OF REFLECTANCEWithout topographic consideration the atmospherically

corrected surface spectral reflectance under lambertianassumption for MODIS is computed using (1) as defined in[14, 15]:

( ) ( ) (1)TABLE 1SALIENT SPECIFICATIONS OF MODISSENSOR

Spectralbands

Spectralwavelength

(nm)

SpatialResolution

(m)

Quantization(bit)

Radiance Scale(mw/cm2/sr/m)

Radianceoffset

Solar Exoatmostphericspectral Irradiance

(mw/cm2/sr/m)

B1 620-670 250 12 0.0026144 0 160.327B2 841-876 250 12 0.0009926 0 98.70B3 459-479 250 12 0.0027612 0 209.071B4 545-565 250 12 0.0021087 0 186.4B5 1230-1250 250 12 0.0005568 0 47.6B6 1628-1652 250 12 0.0002572 0 23.8B7 2105-2155 250 12 0.0000787 0 8.7

(a) (b)




3/8

Where and are the transmittances of the atmospherein the view and illumination directions respectively calculatedfor MODIS bands using the equation proposed by authors[15], [16]. is the exo-atmospheric spectral irradiance (referTable1), is the solar zenith angle and calculated for each

pixel [17], d is the earth sun distance in astronomical unitsand calculated using the approach of [18], is thedownwelling diffused radiation and assumed zero according to[19]. is the path radiance and computed using [19].

VI. TOPOGRAPHIC CORRECTIONSTopographically corrected spectral reflectance using

MODIS imagery are estimated using slope matching methodgiven below [13]:

(2)Where is topographically corrected spectral

reflectance,

is spectral reflectance on the tilted surface,

and is maximum and minimum spectralreflectance and estimated from topographically uncorrectedreflectance image, is mean value ofillumination on the south aspect and is illumination andcalculated using equation proposed by [9]. is normalizationcoefficient for different satellite bands and estimated usingequation given in the literature [12]. The flow chart of detailedmethodology for topographic correction is given in the Fig. 3.Topographically corrected pre and post images are shown inFig. 4 (a) and (b) respectively.

Figure 3 Flow chart of pre-processing steps and Slope matching topographiccorrection technique.

VII. IMPROVED CHANGE VECTORANALYSISMalila [20] gave a general idea of change-vector analysis

(CVA). CVA is used widely to detect multispectral changedetection, and is one of the most effective pre-classificationchange detection techniques [21].

Figure 4 Topographic corrected MODIS images

(a) Pre image (02

nd

November 2009)(b) Post image (21st November2009)

Sohl [22] concluded that change vector analysis producedthe best result of all the techniques tested, due to itsgraphically rich content and its ability to detect urban andagricultural change with fairly good location information. Achange vector can be described by an angle of change (vectordirection) and a magnitude of change from date 1 to date 2[23].It is reported [7] that improved change vector analysis is avaluable technique for change detection which includes asemiautomatic method, named Double-Window Flexible PaceSearch (DFPS), which aims at determining efficiently thethreshold of change magnitude of multi temporal image and anew method of determining change direction (changecategory) which combines a single image classification and aminimum-distance categorization based upon the directioncosines of the change vector provided a way to find changetype discrimination. The improved change vector analysis isimplemented in this paper as change detection analysis. Thesequence of steps required to perform improved change vectoranalysis [7] is given in Fig. 5.

A. Change MagnitudeCVA is a multivariate technique, which accepts the

desired number of bands as input or spectral features fromeach scene pair [24].Change Vector is defined in (3).

G = H - G = (3)Where G includes all the change information between

the two dates, for a given pixel by G= andH = respectively and n is the number of bands.

MODIS Images

Geo-referencing

Atmospheric

Corrections

DEM

Slope, Aspect

IlluminationAngle

Estimation of

Reflectance

Estimation of

coefficients

Topographically Corrected Reflectance

(a) (b)




4/8

The magnitude of change vector is given according toEuclidian distance to produce magnitude of change.

Figure 5 Flow chart of Methodology for ImprovedChange Vector Analysis

The Change magnitude is given by

| G | = (4)Using (4), change magnitude was computed. A decision

on change is made based on whether the change magnitude

exceeds a specific threshold which is calculated as explainedin following sections.

B. Threshold search using Double-Window Flexible PaceSearch (DFPS)

DFPC mainly aims at determining the threshold ofchange magnitude. This method based upon selecting athreshold from a training sample as shown in Fig. 6 (a),(b), and (c), that contain all possible types of changes inour study area such as snow, soil, vegetation, shadow.As reported in the literature [7], training samples wereselected based upon: (1) it covers all types/ as much as

possible changes, (2) it must include only change typeinformation and (3) training samples should be encircled

by no-change pixels. A training sample has an innerboundary and outer boundary. Inner boundary is an areaof interest to find the change, outer boundary is used to

prevent the threshold from being too low.

Figure 6 Training sample area (Highlighted with white boundary)(a) Pre-image(b) Post image(c) Change magnitude

The threshold search range can be set as a differencebetween the minimum value x, and the maximum value y, ofchange in 1st search process, using formula:

= (5)Where is pace search and z is positive threshold value,

which can be set manually. Succession rate is used to find out

Multi-date without/with topographic Correction

Change Magnitude ImageImage Classification

Extraction of seed

points

Selection of training

Sample

Find Search range and

pace

Put values of thresholdin search process

Calculation of testing

Parameter

Determination ofoptimal Threshold

Condition

to exit

Change / No-change

Image

Calculation of Direction of change vector for everychange pixels

Change Type Discrimination

Accuracy Assessment Change

Comparison ICVA with and without Topographic

(a) (b)

(c)




5/8

best suitable threshold value for change. Succession rate ()can be calculated using following equation:

= % (6)Where, is number of change pixels detected inside an

inner training window,

is number of change pixels

detected inside outer training window incorrectly and I is totalnumber of pixels in inner training window. It should be notedthat outer window include one or two pixel only from every

side. When highest succession rate is achieved, the iteration isstopped, then a threshold value that have maximum successionrate is applied to an entire change magnitude image as shownin Table 2 (with topographic correction) and Table 3 (withouttopographic correction). The topographic uncorrectedchange/no-change image(white shows change part, blackshows no-change part) shown in Fig. 7(a) and the topographiccorrected change/no-change image(white shows change part,black shows no-change part) is shown in Fig. 7(b).

TABLE 2RESULTS OF DFPS WITH TOPOGRAPHIC

Range 90-50 Pace 10 Range 65-55 Pace 5 Range 64-56 Pace 2 Range 61-59 Pace 1 Range 60.5-59.5 Pace .5

Threshold Success % Threshold Success % Threshold Success % Threshold Success % Threshold Success %

90 2.50 65 55.80 64 58.44 61 58.44 60.5 58.4080 19.40 60 58.44 62 58.44 60 61.03 60.0 61.0370 41.55 55 57.15 60 61.03 59 58.44 59.5 59.7060 61.03 58 59.7450 51.90 56 57.16

TABLE 3RESULTS OF DFPS WITHOUT TOPOGRAPHIC

Range 100-50 Pace 10 Range 95-85 Pace 5 Range 94-86 Pace 2 Range 91-89 Pace 1 Range 90.5-89.5 Pace .5

Threshold Success % Threshold Success % Threshold Success % Threshold Success % Threshold Success %

100 32.5 95 40 94 40 91 52.5 90.5 52.590 55.0 90 55 92 47.5 90 55 90 55.080 42.5 85 45 90 55 89 52.5 89.5 52.570 42.5 88 5060 42.5 86 47.550 25.0

Figure 7 Change magnitude image(a) Without topographic correction(b) With topographic correction

C. Change type DiscriminationIt is reported in [7] that change type discrimination can be

obtained using a method which combines single imageclassification with minimum-distance categorizing based ondirection cosines of change vectors. The direction of a vectorcan be described by a series of cosine functions in a multi-dimensional space. This series is called direction cosines[25].Moreover, using the direction cosines instead of anglemeasurements can avoid the difficulty of baselineestablishment for angle measurement. First, Pre-image isclassified using supervised classification based on parametric

rule Maximum Likelihoods shown in Fig. 8. Pre imageclassification included four classes: (1) Snow, (2) Soil, (3)Vegetation and (4) Shadow. These classes are generated usingsignature file. Then the seed points were extracted fromdifferent class types. Seed points play an important role forchange type discrimination for all change pixels. Spectralchange vector between any two kinds of change type can becalculated based on classification and then Euclidean distanceof corresponding change is obtained by transplanted thesevalues in direction cosine. The direction of cosine is defined as[25].

(a) (b)

No-change Change




6/8

Cos = , Cos = Cos = (7)Where X (, ) is vector, n is number of bands

Figure 8 Pre image classifications(a) Without topographic correction(b) With topographic correction

Change magnitude (

is calculated using following

equation

(8)Change type is obtained by applying the minimum

distance rule because its effectiveness and its simplerequirement of only the estimation of the mean vector ofeach spectral class [26] based upon this unknown pixelis assigned to certain class. By superimposed theminimum distance classified image on pre classifiedimage, the classified post image has been obtained asshown in Fig. 9 (a) (without topographic correction) and

(b) (with topographic correction).

Figure 9 Minimum Distance Classified image(a) Without topographic correction(b) With topographic correction

D. Accuracy assessmentWith the experiment results, Influence of

topographic can be evaluated using accuracy assessmentof Change/ No-change error matrix. A kappa

coefficient of 0.8403 and overall accuracy of 92% wereachieved with topographic correction as shown in Table4. A kappa coefficient of 0.7136 and overall accuracy of86% were achieved without topographic corrections asshown in the Table 5. Further evaluation has been madefor accuracy assessment of From-To change. A kappacoefficient of 0.7779 and overall accuracy of 86% wereachieved with topographic correction as shown in Table6 and a kappa coefficient of 0.7572 and overall accuracyof 82% were achieved without topographic correctionsshown in Table 7.

TABLE 4ERRORMATRIX USING 50 SAMPLES FORCHANGE/NO-CHANGEDETECTION WITH TOPOGRAPHIC

Referenced Change

Change Pixels No-change Pixels Sum Commission Error

Classified change Change Pixels 23 3 26 11.50%No-change Pixels 1 23 24 4.16%Sum 24 26 50Commission Error 4.16% 11.5%Overall Accuracy= 92 %,Kappa Coefficient= 0.8403

(a) (b)

(a) (b)

Unclassified ShadowSnow SoilVegetation

Unclassified ShadowSnow SoilVegetation




7/8

TABLE 5ERRORMATRIX USING 50 SAMPLES FORCHANGE/NO-CHANGEDETECTION WITHOUT TOPOGRAPHIC

Referenced Change

Change Pixels No-change Pixels Sum Commission Error

Classified change Change Pixels 26 6 32 18.75%

No-change Pixels 1 17 18 5.55%Sum 27 23 50Commission Error 3.70% 26.08%Overall Accuracy= 86 %, Kappa Coefficient= 0.7136

TABLE 6ACCURACY ASSESSMENT USING 50 SAMPLES OF FROM-TOCHANGE DETECTION ICVA WITH TOPOGRAPHIC

22 /11/09 Shadow Snow Soil Vegetation Unclassified Sum

02/11/09 Shadow

Snow 2 25 1 28Soil 18 1 19Vegetation 2 2Unclassified 1 1

Sum 4 43 2 1 50Overall Accuracy =86%, Kappa Coefficient=0.7779

TABLE 7ACCURACY ASSESSMENT USING 50 SAMPLES OF FROM-TOCHANGE DETECTION ICVA WITHOUT TOPOGRAPHIC

22 /11/09 Shadow Snow Soil Vegetation Unclassified Sum

02/11/09 Shadow 1 2 3Snow 5 20 1 26Soil 3 15 18Vegetation

Unclassified 3 3Sum 9 37 1 3 50Overall Accuracy =82% , Kappa Coefficient= 0.7572

VIII. CONCLUSIONIn this paper, we have concluded that improved change

vector analysis is viable change detection technique to find thethreshold of change magnitude and change type for snowcover regions of Himalaya terrain. DFPS method isimplemented to estimate a threshold value for change/nochange area on satellite images with good succession rate. Thechange magnitude combines a single image classification anda minimum-distance categorization based upon the directioncosines of the change vector provided a way to find change

type discrimination. Topography plays a significant role inchange detection analysis. The inclusion of topographysignificantly improved the accuracy of change detection.Overall accuracy of 86% (Kappa coefficient 0.7779) has beenachieved with topographic model inclusion as compared to82% (Kappa coefficient 0.7572) for uncorrected images.

ACKNOWLEDGEMENT

The authors would like to thank Director, SnowAvalanche Study establishment, Department of DefenceResearch and Development Organization. We are alsothankful to Arun Chaudhary, Scientist, Snow Avalanche Study

Establishment for technical discussions.

REFERENCES

[1] Lillesand, Thomas and Ralph Kiefer, Remote Sensing and ImageInterpretation. 2000,New York: John Wiley and Sons.

[2] A. Singh, Digital Change Detection Techniques Using Remotely SensedData.International Journal of Remote Sensing 1989, 10: 989-1003.

[3] R.A. Weismiller, S.Y. Kristof, D.K. Scholz, P.E. Anuta, and S.A.Momin, 1977. Change detection in coastal zone environments,Photogrammetric Engineering and Remote Sensing, 43:15331539.

[4] T. Fung, and E. LeDrew, 1987. Application of Principle ComponentAnalysis to Change Detection. Photogrammetric Engineering andRemote Sensing 53: 1649-1658.

[5] R.D. Johnson, and E.S. Kasischke, 1998, Change vector analysis: Atechnique for the multispectral monitoring for land cover and condition,

International Journal of Remote Sensing , 19:411426.[6] J.F. Mas, 1999, Monitoring Land-Cover Changes: A Comparison of

Change Detection Techniques.International Journal of Remote Sensing20(1): 139152.

[7] Jin Chen, Peng Gong, Chunyang He, Ruiliang Pu, and Peijun Shi, 2003,Land-Use/Land-Cover Change Detection Using Improved Change-Vector Analysis 2003, Photogrammetric Engineering and RemoteSensing Vol. 69, No. 4, April 2003, pp. 369379.

[8] D. Riano, E. Chuvieco, J. Salas and I. Aguado 2003 Assessment ofdifferent topographic corrections in Landsat-TM data for mappingvegetation types;IEEE Trans. Geosci. Remote Sens. 41(5) 10561061.

[9] D. L. Civco, 1989 Topographic normalization of Landsat ThematicMapper digital imagery; Photogramm. Eng.Remote Sens. 55 13031309.

[10] P. M. Teillet, B. Guindon and D. G. Good enough 1982, on the slopeaspect correction of multispectral scanner data; Canada. J. Remote Sens.8 84106.

[11] M. Minneart, 1941, the reciprocity principle in lunar photometry; J.Astrophys. 93 403410.




8/8

[12] J. Nichol, L K Hang and W M Sing 2006 Empirical correction of lowsun angle images in steeply sloping terrain: a slope matching technique;Int. J. Remote Sens. 27(34) 629635.

[13] V. D. Mishra, J. K. Sharma, Singh K. K., Thakur N. K. and Kumar M.,Assessment of different topographic corrections in AWiFS satelliteimagery of Himalaya terrain. J. Earth Syst. Sci. 118, No. 1, February2009, pp. 1126.

[14] M. R. Pandya, R. P. Singh, K. R. Murali, P. N. Babu, A. S. Kiran kumarand V. K. Dadhwal, 2002 Bandpass solar exo-atmospheric irradianceand Rayleigh optical thickness of sensors on board Indian Remote

Sensing Satellites 1B, -1C, -1D, and P4; IEEE Trans. Geosci. RemoteSens. 40(3) 714717.

[15] C. Song, C. E. Woodcock, K. C. Seto, M. P. Lenney and A. S.Macomber 2001 Classification and change detection using Landsat TMdata: when and how to correct atmospheric effects; Remote Sens.Environ. 75230244.

[16] P.B. Rusell, J.M. Livingston, E.G. Dutton, R.F. Pueschel , J.A. Reagan,T.E. Defoor, M.A. Box, D. Allen, P. Pilewskie, B.M. Herman, S.A.Kinne, and D.J. Hoffman, 1993, Pinatubo and pre- Pinatuba and opticaldepth spectra: Mauna Loa measurement, Composition, inferred particledistribution, radiative effects, and relationship to lidar data, Journal ofGeophy. Res., V.98, pp.22969-22985.

[17] F. Kasten, 1962, Table of solar altitudes for geographical latitudes;CRREL Special Report, 57 U.S. Army Corps of Engineers, Hanover,New Hampshire.

[18] Van Der Meer F 1996 Spectral mixture modeling and spectralstratigraphy in carbonate lithofacies mapping; ISPRS J. Photogramm.Remote Sens. 51 150162.

[19] P. S. Jr. Chavez, 1989, Radiometric calibration of Landsat ThematicMapper multispectral images; Photogramm. Eng. Remote Sens. 5512851294.

[20] W.A. Malila , 1980. Change vector analysis: an approach for detectingforest changes with Landsat, Proceedings of the 6th Annual Symposiumon Machine Processing of Remotely Sensed Data, 03-06 June, PurdueUniversity, West Lafayette, Indiana, pp. 326335.

[21] S. Berberoglu and A. Akn, 2009, Assessing Different Remote SensingTechniques to Detect Land Use/Cover Changes in the EasternMediterranean.International Journal of Applied Earth Observation andGeoinformation. 11, 46 53.

[22] T.L. Sohl, 1999. Change Analysis in the United Arab Emirates: AnInvestigation of Techniques. Photogrammetric Engineering and RemoteSensing 65(4): 475 484.

[23] J.R. Jensen, 1996. Introductory Digital Image Processing: A RemoteSensing Perspective, Second Edition, Prentice Hall, Upper SaddleRiver, New Jersey, 316 p.

[24] Murat zyavuz1, Onur atr And Bayram Cemil Bilgili, 2011, AChange Vector Analysis Technique To Monitor Land-Use/Land-CoverIn The Yildiz Mountains, Fresenius Environmental Bulletin by PSPVolume 20 No 5., Turkey.

[25] B. Hoffmann, 1975. About Vectors, Dover Publications, Inc., NewYorkN.Y. 134 p.

[26] J.A. Richards and X. Jia 1999. Remote Sensing Digital Image Analysis,Third Edition, Springer-Verlag, Berlin, Heidelberg, Germany, 363 p.



Date post:	07-Apr-2018
Category:	Documents
Upload:	helpdesk9532
View:	219 times
Download:	0 times

11.IJAEST Vol No 7 Issue No 1 Topographic Influence on Improved Change Vector Analysis Using MODIS...

Documents