APPLICATION OF BINARY DIVISION ALGORITHM FOR IMAGE ANALYSIS AND CHANGE DETECTION TO IDENTIFY THE HOTSPOTS IN MODIS IMAGES Harsh Kumar G R * an Dharmenra Sngh (hargrec@tr.ernet.n, harmfec@tr.ernet.n) Department of Electroncs an Computer Engneerng, Inan Insttute of Technology Roorkee, Roorkee, Ina. Commsson VI, Workng Group IV/3 KEY WORDS: Bnary Dvson Algorthm, Change Detecton, Hotspot, MODIS mages, Multresoluton fusng ABSTRACT: In ths paper, an approach base on Bnary Dvson Algorthm, to etect the Hotspots usng Mult-resoluton fuson of ban 1, ban 2, ban 31 an ban 32 of Moerate Resoluton Imagng Spectroraometer (MODIS), NASA Satellte, for the Jhara (Ina) Regon wth the a of mage analyss an change etecton technque has been propose. INTRODUCTION: The management of mne fres n coal mnng regon s of much concern n Ina. Mne fres apart from economc aspects; gve rse to evastatng envronmental effects. Most of the fres n Inan coalfel, whch cause a local rse n the surface temperature, take place ue to spontaneous heatng of coal, whch epens on varous mnng, geologcal an coal factors. The man objectve of ths paper s to use operatonal satellte ata whch s freely avalable an etect the hotspots whch may help fre managers wth the spatal allocaton of coal fre preventon or fre extncton resources. The use of satellte remote sensng has mae t possble to etect coal fres an has recently been consere as an effectve an economc alternatve for fre montorng gven that a number of factors nvolve n fre susceptblty estmaton may be erve at least partally, from satellte ata. Satellte observatons can prove both the extent of fre scares an the expanson of ongong fres n tme an space. Jhara coalfel n Jharkhan s the rchest coal bearng area n Ina whch contans a large number of coal mne fres whch have been burnng for several ecaes. The Jhara Coalfel s locate n the Dhanba strct of the state of Bhar an s name after the man mnng area of Jhara. The Jhara Coalfel s confne between lattues 23 38' N an 23 50' N an longtues 86 07' E an 86 30' E. The maxmum extent of the coalfel s about 38 km from East to West an 19 km from North to South. Hotspots are efne as mage pxels whose brghtness temperatures excee a pre-efne threshol value. Threshol values of 316-320 K have been reporte n scentfc lterature. Dfferent temperature threshol wll result n fferent hotspot counts. The lower the threshol, the hgher the number of hotspots etecte. MODIS-seres satelltes has become the most wely use satellte ata set for regonal fre etecton an montorng because of ts avalablty, spatal resoluton, spectral characterstcs, an low costs. * Corresponng Author. Phone: 09837483542.
The Terra an Aqua MODIS nstrument proves hgh raometrc senstvty (12 bt) n 36 spectral bans rangng n wavelength from 0.4 µm to 14.4 µm. The responses are custom talore to the nvual nees of the user communty an prove exceptonally low outof-ban response. Two bans are mage at a nomnal resoluton of 250 m at nar, wth fve bans at 500 m, an the remanng 29 bans at 1 km. A ±55-egree scannng pattern at the EOS orbt of 705 km acheves a 2,330-km swath an proves global coverage every one to two ays. In Ths paper, the Terra MODIS, ban 1 wth resoluton of 250m an banwth 620-670 nm, ban 2 wth resoluton of 250m an banwth 842-876nm an ban 31, 32 wth resoluton of 1km an banwth 10780-11280nm an 11770-12270nm respectvely. Change etecton s the process of entfyng fferences n the state of a feature or phenomenon by observng t at fferent tmes. There are many remote sensor system an envronmental parameters that must be consere whenever performng change etecton. Falure to unerstan the mpact of the varous parameters on the change etecton process can lea to naccurate results. Ieally, the remotely sense ata use shoul be acqure by a remote sensor system that hols temporal, spatal, spectral, an raometrc resolutons constant. In ths paper the MODIS mages of May 2005 an Nov 2005 have been consere. Bnary Dvson Algorthm, whch s one of the effcent clusterng methos, has been apple on these mages for etectng hotspots n Jhara regon. Clusterng gves us sgnfcant nformaton about ata strbuton n multmensonal feature space. There have been a lot of clusterng methos reporte. They are roughly ve nto three groups: agglomeratve methos, parttonng methos, an agglomeratve-parttonng methos. The agglomeratve methos have hgh effcency but low accuracy. On the other han, the parttonng methos are accurate an effcent, but they requre huge memory, an much tme. The agglomeratve-parttonng methos, mprove the accuracy by usng teratve processng, whch makes the metho neffcent. In ths paper, a parttonng metho, bnary vson Algorthm, for mprovng accuracy an effcency has been propose. An ISODATA clusterng Algorthm for agglomeratve-parttonng methos also exsts, whose performance to Bnary Dvson Algorthm s not of that accuracy an effcency, hence Bnary Dvson Algorthm, have been use to etect the hotspots Prncples an Proceures In ths paper, the ban 1 an ban 2 of MODIS (MOD09A1), an ban 31 an ban 32 of MODIS (MYD11A2) of May 2005 an Nov 2005 have been consere. The MOD09A1 have a resoluton of 500 m an MYD11A2 have a resoluton of 1000m. The MYD11A2 resoluton was ncrease to 500m by the applcaton of blnear nterpolaton technque. The MOD09A1 proves an estmate of the surface spectral reflectance an MYD11A2 proves an estmate of the surface emssvty. So the MOD09A1 s normalze wth respect to the MYD11A2 ata. On these mages the Bnary vson Algorthm s apple. Bnary vson Algorthm In the bnary vson, t s essental to etermne the cluster to be ve, the subset of the feature space to be use for the vson, an the vson threshol n the subspace. A. Selecton of Dvson Cluster: Fgure 1. Bnary Dvson Algorthm * Corresponng Author. Phone: 09837483542.
The fgure 1 shows an example of the bnary vson Algorthm. In ths example, mage ata are ve nto clusters 1 an 2 at noe n 0, an those n cluster 2 nto clusters 3 an 4 at a noe n 2. Cluster 3 s selecte for the next vson among all termnal clusters 1, 3, An 4. The prncple s escrbe usng ths example. Clusterng n Bnary vson Algorthm s regare as a mnmzaton process of the total wthn-group sum of squares (WGSS) of the mage ata. In ths Algorthm, mage ata are clustere so that the maxmum reucton s obtane n the ntragroup sum of squares (IGSS) among chl clusters prouce by the vson. It s known that a mxe cluster c 0 of clusters c 1 an c 2 has the varance Var(c 0 ) = Var(c 1 ) + Var(c 2 ) + [mean (c 1 ) - mean (c 2 )] 2 (1) Usng the same manner, we obtan the relaton for the sum of squares n N-mensonal feature space as S Parent = S chll + S chl2 + S between (2) Where S Parent s an N x N matrx for the WGSS of a parent cluster, S chll an S chl2 are those for chl clusters prouce by the vson, an &, S between means the IGSS between the chl clusters. We select the vson cluster among all termnal clusters so that the WGSS effcently falls to the mnmum. We efne a matrx D for evaluaton of the reucton of the WGSS, as D= S Parent (S chll + S chl2 ) (3) The reucton D, however, epens on the sum of squares of the parent cluster. So we efne the reucton rate A as a normalze nex A=trace(S E -1/2 DS E -1/2 ) (4) Where S E s the total WGSS of all termnal noes clusters, an s S 1 + S 3 + S 4 n the example shown n Fgure 1. Thus we select the next cluster among all termnal clusters so that a maxmum reucton rate s obtane. Next, we wll escrbe the algorthm for selectng the optmal subspace to ve the cluster nto two subclusters an for etermnaton of the threshol n the subspace. B. Subsets of Feature Space an Bounary Search: Search areas n whch the optmal bounary wll be selecte become larger as the menson (the number of spectral bans) ncreases. In orer to acheve hgher effcency, we ecrease the menson by projectng mage ata onto a small number of the feature spaces. We aopt the frst two canoncal components p 1 an p 2 to form the two-mensonal subspace as canoncal correlaton analyss s useful n ata compresson an n nose reucton. We use the followng sx projecton functons efne on a two-mensonal subspace. (5) Let V T be the varance-covarance matrx of ata n a cluster, an V e be that of nose. We obtan the egenvector matrx B whch makes BV T B an BV e B agonal matrces. The frst two canoncal components p 1 an p 2 are erve from egenvectors relate to the larger two egenvalues. An element σ j 2 of the matrx V e s estmate from fferental values u 2 1 j σ = j kl kl 20{( s 2)( t 2) 1} u u (6) = + + + kl k 1, l k, l + 1 k + 1, l k, l 1 4 k, l (7) Where we assume the mage has s columns an t lnes (s x t = M), an k,l means spectral ensty of a pxel at the poston of the k th column an the l th lne n an mage of the th k l * Corresponng Author. Phone: 09837483542.
ban. Next s the bounary search n the subspaces. Data projecte onto the subspaces are reassgne to nteger values from 0 to 255 for the followng processng. We use the hstogram of these values to search a canate for the optmal bounary. We employ an nex Q, a kn of pxel ensty, n a cluster for bounary search as L Q = (8) ( V +1) Where L means the number of pxels n a cluster, V s varance of ata n the cluster, an constant 1 s ae to avo the vergence of Q. The nex changes from Q parent to Q chll + Q chl2 wth vson. As the vson ecreases S between n (3), the ensty ncreases wth vson as Q parent Q chll + Q chl2 (9) The equalty hols when ata n the parent cluster are unformly strbute. The canate for the optmal bounary s selecte among all possble bounares so that the sum of enstes of chl clusters has the maxmum, that s C. Proceures: The proceures of Bnary Dvson Algorthm are summarze below. As Algorthm has no stoppng rule yet, the algorthm s stoppe when the number of clusters s equal to a specfe number N c.(the no of clusters got by applyng the ISODATA algorthm on these ata) 1) Specfy the number of clusters to be obtane. 2) Apply proceure 3) for all termnal clusters, an select a cluster for the next vson. 3) Apply canoncal correlaton analyss to a termnal cluster an project the frst two canoncal components to the subspaces. Select the optmal bounary among all possble bounares n the spaces. 4) Dve the cluster at the bounary. 5) Repeat proceures 2)-4) untl the number of clusters s equal to N c 6) Calculate the mean vectors of all termnal clusters, an replace the ensty of pxels wth the mean vector. 7) Yel the resultant mage. The clustere mage, s segmente to solate hotspots an change etecton s apple to ths clustere mage an the MOD14A2(most confent fre) of May 2005 mage RESULTS Q1 2 1 + 2 x (X) + Q (X) = max{q (x) Q (x)} (10) Where, we suppose cluster 0 s ve nto clusters 1 an 2, x s a bounary number, an X the selecte bounary. The optmal bounary s selecte among these canates obtane from all the subspaces. We use a normalze ensty G for comparson as G 1 ( X ) + Q2 ( X ) Q = (=1~6) (11) Q 0 Fgure 2. MODIS Ban 1 May 2005 Image Where specfes the subspace. The bounary X j n the j th subspace s the optmal bounary for vson of a termnal cluster when G = max{ G } (12) j * Corresponng Author. Phone: 09837483542.
Fgure 3. MODIS Ban 2 May 2005 Image Fgure 6. BDA Apple mage Fgure 4. MODIS Ban 31 May 2005 Image Fgure 7. Segmente mage Fgure 5. MODIS Ban 32 May 2005 Image Fgure 8. Change Detecte mage * Corresponng Author. Phone: 09837483542.
CONCLUSIONS Wth the applcaton of Bnary Dvson Algorthm, an change etecton technques, t s foun that hotspots are clearly etecte n MODIS mages, whch shows the effcent way to analyze the mage n unsupervse way. The result an propose algorthm are very helpful to evelop a atabase of hotspots montorng system for satellte ata. REFERENCES [1] H.Hanazum, S.Chno, an S.Fujmura: A Bnary Dvson Algorthm for Clusterng Remotely Sense Multspectral Images, IEEE Trans., 1M44-3, pp.759-763 (1995) [2] H.Hanazum an S.Chno: bnary vson algorthm usng a lnear scrmnant functon for the cluster analyss of remotely sense multspectral mages, SPIE Proceengs-- Volume 2579, Image an Sgnal Processng for Remote Sensng II, November 1995, pp. 182-187 [3] D. P. ROY, L. GIGLIO, J. D. KENDALL an C. O. JUSTICE: Mult-temporal actve-fre base burn scar etecton algorthm, Int. Journal remote sensng, 1999, vol. 20, no. 5, 1031-1038 [4] C.O. Justce, L. Gglo, S. Korontza, J. Owens, J.T. Morsette, D. Roy, J. Desclotres, S. Alleaume, F. Pettcoln an Y. Kaufman: The MODIS fre proucts, Remote Sensng of Envronment 83 (2002) 244 262 * Corresponng Author. Phone: 09837483542.