DELINEATION OF DISEASED TEA PATCHES USING MXL AND TEXTURE BASED CLASSIFICATION

Transcription

1 DELIEATIO OF DISEASED TEA PATCHES USIG MXL AD TEXTURE BASED CLASSIFICATIO Rshra Dutta a, *, Alfred Sten a,.r. Patel b a Department of Earth Observaton Scence, Internatonal Insttute for Geonformaton Scence and Earth Observaton ITC, 75 AA Hengelostraat, Enschede, The etherlands - (dutta13191@tc.nl and sten@tc.nl) b Agrculture and Sols Dvson, Indan Insttute of Remote Sensng (RSA), Department of Space, Government of Inda, Dehradun-481, Uttarakhand, Inda - nrpatel@rs.gov.n KEY WORDS: Remote Sensng, Geographc Informaton System (GIS), GLCM, Texture, Healthy, Moderately Healthy, Dseased Tea patches ABSTRACT: Recently, a rapd declne n the qualty of Indan tea producton has been observed due to the old age of the plantatons, dsease and pests nfestatons and frequent applcaton of pestcdes and nsectcdes. Ths paper shows an applcaton of remote sensng and GIS technologes for montorng tea plantatons. We developed an approach for montorng and assessng tea bush health usng texture and tonal varatons from Landsat, Aster and LISS III mages. The Gray Level Co-occurrence Matrx (GLCM) categorzes the tea nto healthy, moderately healthy and dseased tea. We observed from the study that the GLCM could be used for delneatng the affected and non-affected tea patches at dfferent resolutons. 1. ITRODUCTIO Inda contnues to be the world's largest producer and consumer of tea. Due to ncreasng competton from countres lke Sr Lanka, Kenya, Chna, Bangladesh and Indonesa, and problems n mantanng the qualty, Indan tea yeld has been wtnessng a downward trend. Tea yeld statstcs show that tea plants start to produce tea leaves from approxmately ther thrd year onwards, mantan a steady ncreasng trend upto some or 3 years of age, wthn whch perod they reach a peak n producton durng several years and then declne. Remote sensng offers an effcent and relable means of collectng the requred nformaton for mappng tea type and acreage. It also provdes structural nformaton on varous aspects of vegetaton health, as the spectral reflectance of a tea feld vares wth respect to the phenology, stage type and crop health. Thus, tea plantatons could be well montored and measured usng multspectral sensors. Informaton from remotely sensed data can be used as nput nto a Geographc Informaton System (GIS) or a smlar croppng system. Ths can then be combned wth ancllary data to provde nformaton of ownershp and management practces. Assessment of tea health, as well as early detecton of crop nfestatons, s crtcal n ensurng good tea productvty. Stress assocated wth, for example, mosture defcences, nsects, fungal and weed nfestatons, must be detected early enough to provde an opportunty for the planters to mtgate. To montor ths process effectvely requres that remote sensng magery s provded on a frequent bass. Remote sensng has a number of attrbutes that lend themselves to montorng tea bush health. Detectng damage and montorng stress requres hgh-resoluton magery and multspectral magng capabltes. A crtcal factor n makng magery useful to planters s a quck turnaround tme from data acquston to dstrbuton of crop nformaton. To receve an mage that reflects crop condtons two weeks earler nether helps real tme management nor damage mtgaton. Images are also requred at specfc tmes durng the growng season, and on a frequent bass. The tea ndustry n Assam grew n terms of acreage, but n the course of tme t started facng many problems. The present crss n the tea ndustry started n 1999, when drought durng the early part of the season resulted nto large producton cuts. Year saw a margnal mprovement n producton but n the same tme a sharp drop n prces throughout the year. Producton n Assam n 1 was low as compared to the natonal average. Export dropped by 7 mllon kg and Assam could export only 18 mllon kg of tea. It has been hypotheszed that Assam s losng export due to ncreasng pest and dsease occurrences, whch are caused by mantanng an nferor producton mx, the old age of tea plants, poor qualty control at the processng level and economc factors lke organzaton of small holder farmers and assocated hgh costs of producton,. A recurrent problem of tea gardens s pest nfestaton. Many tea gardens are affected by Helopelts, Red Spder and Lopper attacks. Such nfestatons can possbly be tmely montored and assessed usng remote sensng mages. The man obectve of ths study s to delneate dsease and pests nfested areas n tea gardens. To do so, the paper uses texture and tonal varatons from satellte magery of tea growng areas and nvestgate whether texture based classfcaton could be utlsed for dsease and pests detecton n tea plantaton. 1.1 Revew of Lterature Tea s a mono-cultured crop. It stands n feld stuaton wth lttle nter culture operaton and no crop rotaton. Such condtons ultmately lead to degradaton of the sol envronment ether beng physcal, bologcal or chemcal n * Correspondng author. 1693

2 The Internatonal Archves of the Photogrammetry, Remote Sensng and Spatal Informaton Scences. Vol. XXXVII. Part B4. Beng 8 nature. The mono-culture of tea reportedly causes a condton of mproper sol functonng known as sol sckness (Barua, 1969). Realzng the mportance of tea growng from both a manageral and a commercal pont of vew, Trpathy et al predcted tea yeld usng remote sensng and GIS and other key parameters wth a GIS. Several attempts have been carred out to estmate LAI usng varous types of remote sensng data (Badwhar et al., 1986; Peterson et al., 1987; Turner et al., 1999). It has been prmarly based on the emprcal relatonshp between the feldmeasured LAI and sensor observed spectral responses (Curran et al. 199 and Peddle et al. 1999). The normalzed dfference vegetaton ndex (DVI) has been a popular ndex wth whch to estmate LAI across dverse ecosystems. Recent studes have shown that the DVI may not be very senstve to values of LAI n partcular at the forest ecosystem havng the close canopy condton that the LAI value s relatvely hgh (Chen and Chlar, 1996, Turner et al. 1999). Spectral characterstcs of tea plants are mportant for montorng tea plantatons by remote sensng. In Sr Lanka an emprcal model between DVI and LAI of tea canopy was developed, relatng the LAI of the tea canopy to the spectral reflectance of dfferent types of tea clones for dfferent prunng ages n ffty tea felds (Raapakse et. al., 1). Murthy et al. (1995) reported the valdty of crop yeld models wth satellte derved normalzed dfference vegetaton ndex (DVI) determned by the strength of assocaton between the two varables. The benefts from usng RS and GIS technology depend on the level of success of ts applcaton for solvng a concrete task. In general, these benefts can be dvded nto four categores such as scentfc, technologcal, methodologcal, and economc effcency (Badarch, 199)..1 Selecton of the regon. METHODS For ths study, the study area chosen was the Sontpur dstrct of Assam, Inda. Sontpur dstrct s spread over an area of 534 Sq. Kms. on north bank of Brahmaputra Rver. In terms of area Sontpur s the second largest dstrct of Assam after Karb Anglong dstrct. The Dstrct les between 6 3 and 7 1 lattude and 9 16 E and E longtude. Located between mghty Brahmaputra Rver and Hmalayan foothlls of Arunachal Pradesh, the dstrct s largely plan wth some hlls. Brahmaputra Rver forms the south boundary of the dstrct. A number of rvers whch orgnate n the Hmalayan foothlls flow southwards and ultmately fall n Brahmaputra Rver. Accordng to the 1 Census, Sontpur Dstrct had a populaton of 16, 77,874, wth a densty of 315 persons per Sq. km. In terms of populaton t ranks thrd n Assam after Kamrup and agaon dstrcts. The people represent a heterogeneous lot. Rather, they are a mosac of ethnc groups, an admxture of dverse types of people. Sontpur Dstrct falls n the Sub- Tropcal clmatc regon, and enoys Monsoon type of clmate. Summer tme s hot and humd, wth an average temperature of 9 C. The hghest temperature s recorded ust pror to the onset of Monsoon (around May-early June). Summer ran s heavy, and s prncpally caused from late June to early September by the mosture-laden South-West Monsoon, on strkng the Hmalayan foothlls of the north. Such ran s both a boon and a bane for the people. Autumns are dry, and warm. Wnters extend from the month of October to February, and are cold and generally dry, wth an average temperature of 16 C. Agrculture s the man occupaton of the people of Sontpur dstrct. Tea gardens are next most mportant feature of economy of the Sontpur dstrct. There are all together seventy three tea gardens n Sontpur dstrct. The area covered under these tea gardens s approxmately ha. Monabar near Bswanath Charal s Asa s largest tea garden. Monabar tea estate comprses an area of 196 ha wth annual tea producton of kg n whle Borgang tea estate s the second largest wth 118 ha and an annual producton of kg. Most tea gardens were prevously owned by European companes lke Mceal & Magor, George Wllamson Ltd., McLeod Russel, Brtsh Assam Tea garden Company$ and Empre Plantaton Lmted. In recent years, however, many Indan owned companes lke Tata Tea and Brooke Bond have taken over the ownershp. Apart from the bg companes, recently small tea gardens wth areas of 5 to 14 ha have emerged n the vcnty of the bg tea gardens.. Data..1 ASTER: An ASTER (15m) mage has been used for montorng the tea areas. The mage has been procured n June, 4. ASTER products have the followng characterstcs that are an mprovement over past remote sensng products: A wde spectral range and a hgh spectral resoluton, coverng the spectral range of.5 to mcrons wth 14 bands. 15m, 3m, and 9m spatal resolutons n the vsble and near nfrared spectral regon, the shortwave nfrared spectral regon, and the thermal nfrared spectral regon, respectvely. For band 3 ( mcrons), both the usual nadr-lookng telescope and a backward-lookng telescope are used to produce stereoscopc mages acqured n the same orbt... IRS-1C/1D: The LISS III mage used n ths study was taken on February 4. It has four bands (red, green, blue and near nfra red). The red, green and near nfra red bands are of 3.5 m resoluton. It has a swath of 17 km for bands, 3 and 4 and 134 km for band 5 and IR. Its repetvty s 5 days...3 LADSAT-7 ETM+ Image: The mage was taken on December 1. It ncludes spectral regons of VIR (bands 1-4 wth 3 m resoluton), SWIR (bands 5 and 7 wth 3 m resoluton and TIR (band 6, wth 6 m resoluton) and one panchromatc band 8 wth 15 m resolutons. The Enhanced Thematc Mapper Plus (ETM+) s the nstrument payload on the LADSAT-7 spacecraft. The ETM+ s a dervatve of the Thematc Mapper (TM) nstrument on LADSAT-4 and LADSAT-5 (Markham, 3)...4 Base Maps: Base maps used n hs study nclude the study area boundares, lattude and longtude of the area, road maps, contour maps and attrbute data collected from toposheets and other sources. Moreover ndvdual garden maps have been collected for each zone at cadastral level. Each garden has ts well mantaned garden maps wth well demarcated sectons and ther coordnates. These garden maps could be used for dentfyng the sectons undergong replantaton from ther secton numbers and also the sectons showng varatons due to 1694

3 The Internatonal Archves of the Photogrammetry, Remote Sensng and Spatal Informaton Scences. Vol. XXXVII. Part B4. Beng 8 sol condtons, envronmental factors and cultural practces..3 Methodology.3.1 Delneaton of dseased tea patches: Remote sensng and GIS technologes have been effcently used for montorng several annual crops lke rce, wheat, etc. Therefore, developng a smlar approach for montorng tea plantatons has become a challenge. The lack of prevous studes n montorng tea usng remote sensng provded the dea to develop an approach that can ad n montorng the growth of plantatons and help n takng effectve measures when the need arses. An attempt has been made to assess tea bush health usng texture and tonal varatons from remotely sensed mages. The Gray Level Co-occurrence Matrx (GLCM) technque was appled to categorze the tea patches nto healthy, moderately healthy and dseased tea. The dseased patches were delneated usng both texture and the classfed based mages. The study was carred out on the orth Bank of Assam, Inda. LADSAT, LISS III and ASTER remote sensng mages were consdered for ths study. The LADSAT mage was acqured on December- 1, the LISS III mage on nd February 4 and the ASTER mage on June- 4. Supervsed and unsupervsed classfcatons were carred out usng the maxmum lkelhood classfer on all the mages. The overall accuracy for all the classfed mages was then calculated. Texture analyss has been extensvely used to classfy the remotely sensed mages..3. Image classfcaton: The basc assumpton for mage classfcaton s a specfc part of the feature space correspondng to a partcular class. Classes have to be dstngushed n an mage and classfcaton needs to have dfferent spectral characterstcs. Ths can be done by comparng spectral reflectance curves. Image classfcaton gves results to certan levels of relablty. The prncple of mage classfcaton s that a pxel s assgned to a class based on ts feature vector by comparng t to predefned clusters n the feature space. Two basc approaches exst for classfcaton: pxel based and obect orented. For ths study a pxel based mage analyss was used. Unsupervsed Classfcaton: Unsupervsed classfcaton was carred out for the three mages. The spectral classes obtaned from the unsupervsed classfcaton are based solely on natural groupngs n the mage values. The spectral classes obtaned from all the three mages were not ntally known. So takng the reference values, the classfed data was compared and the spectral classes were dentfed. Supervsed Classfcaton: Here the mage analyst supervses the pxel categorzaton process by specfyng, to the computer algorthm, numercal descrptors of varous landcover types present n the mage. Tranng samples that descrbes the typcal spectral pattern of land cover classes are defned. Pxels n the mage are compared numercally to the tranng samples and are labeled to landcover classes that have smlar characterstcs. Accuracy Assessment: In accuracy assessment the man assumpton s that the reference data or feld data are correct. Classfcaton accuracy wll be determned by usng three complementary measures whch are based on error matrces or confuson matrx derved from ndependent feld data. The two methods used for accuracy assessment are: The Error Matrx: Error matrx s a square matrx wth the same number of nformaton classes. umbers n rows are the classfcaton result and numbers n columns are reference data. In the error matrx the elements n the dagonal are pxels that are correctly classfed for each nformaton class. The error matrx s the most effectve way to represent map accuracy. Accuracy measures lke the Overall accuracy (OA), the Producer s accuracy, the User s accuracy (UA) and the Kappa Statstcs are well descrbed elsewhere (Congalton), assessng the accuracy of the classfed mages In ths study, accuracy assessment of MLC based crop/landuse classes was acheved n the form of error matrx by comparng classfed output wth the ground truth nformaton of ndependent stes, collected usng GPS. A total of 375 sample stes/feld was geo-located wth GPS for comparson wth classfed landcover types: healthy tea patches, moderately healthy patches, dseased tea patches, scrubs, rver, rver bed, barren land and settlements. OA was defned as the percentage of total ndependent reference pxels that were correctly classfed by the MLC. PA was calculated by dvdng the number of pxels correctly classfed for each class by the total number of ndependent reference pxels for that class whle the user s accuracy was the number of correctly classfed pxels dvded by the total number of classfed pxels for that class..3.3 Texture Analyss: Texture analyss has been extensvely used to classfy the remotely sensed mages. Landuse classfcaton where homogeneous regons wth dfferent types of terrans need to be dentfed s an mportant applcaton. Haralck et.al uses gray level co-occurrence features to analyze the remotely sensed mages. They computed the gray level co-occurrence matrces for a dstance of one wth four drectons (, 45, 9 and 135 ). Identfyng the perceve qualtes of texture n an mage s an mportant frst step towards buldng mathematcal models for texture. The ntensty varatons n an mage whch characterze texture are generally due to some physcal varatons n the scene. Texture s usually characterzed by the two dmensonal varatons n the ntenstes present n the mage. Texture s a property of areas and the texture of a pont s undefned. Texture nvolves the spatal dstrbuton of gray levels. Texture n an mage can be perceved at dfferent scales or levels of resoluton. A regon s perceved to have texture when the number of prmtve obects n the regon s large. Image texture has a number of perceved qualtes whch plays an mportant role n descrbng texture. Grey Level Co-occurrence Matrx, GLCM (also called the Grey Tone Spatal Dependency Matrx): The GLCM s a tabulaton of how often dfferent combnatons of pxel brghtness values (grey levels) occur n an mage. GLCM texture consders the relaton between two pxels at a tme, called the reference and the neghbour pxel. The Grey Level Co-occurrence Matrx (GLCM) has been descrbed n the mage processng lterature by a number of names ncludng Spatal Grey Level Dependence (SGLD) etc. As the name suggests, the GLCM s constructed from the mage by estmatng the par wse statstcs of pxel ntensty. Each element () of the matrx represents an estmate of the probablty that two pxels wth a specfed separaton havng grey levels and. 1695

4 The Internatonal Archves of the Photogrammetry, Remote Sensng and Spatal Informaton Scences. Vol. XXXVII. Part B4. Beng 8 a) Mean: The GLCM mean s not smply the average of all the orgnal pxel values n the mage wndow. 1 1 μ = ( ) μ = ( ) P, = P, = The left hand equaton calculates the mean based on the reference pxels, μ. It s also possble to calculate the mean usng the neghbour pxels, μ as n the rght hand equaton. For the symmetrcal GLCM, where each pxel n the wndow s counted once as a reference and once as a neghbour, the two values are dentcal. The summaton s from to (-1), not from 1 to. Snce the frst cell n the upper left of the GLCM s numbered (,), then the value () of ths cell s the same as the value of the reference pxel (). Smlarly, the second cell down from the top has an value of 1, and a reference pxel value of 1. The P value s the probablty value from the GLCM,.e. how many tmes that reference value occurs n a specfc combnaton wth a neghbour pxel. It s not a measure of how many tmes the reference pxel occurs, perod, whch would be the "regular" mean for the orgnal wndow. Multplyng by P effectvely dvdes the entry by the sum of entres n the GLCM, whch s the number of combnatons n the orgnal wndow. Ths s the same as s done when calculatng a mean n the "usual" way. b) Varance: Varance n texture calculates the dsperson of GLCM values around ther mean,. As GLCM varance uses GLCM values t deals specfcally wth the combnatons of reference and neghbour pxel. In that sense t dffers from the varance of grey levels n the orgnal mage. Varance (v) calculated usng the gray levels or gves the same result, snce the GLCM s symmetrcal. d) Homogenety: Homogenety weghts values by the nverse of the Contrast weght, wth weghts decreasng exponentally away from the dagonal. Homogenty = = P, 1+ ( ) e) Dssmlarty: Instead of weghts ncreasng exponentally (, 1, 4, 9, etc.) as we move away from the dagonal, the dssmlarty weghts ncreases lnearly (, 1,, 3 etc.). Dssmlarty = = Ι Ι f) Entropy: The concept comes from thermodynamcs. It refers to the quantty of energy that s permanently lost to heat ("chaos") every tme a reacton or a physcal transformaton occurs. Entropy cannot be recovered to do useful work. Because of ths, the term s used n non techncal speech to mean rremedable chaos or dsorder. Also, as wth ASM, the equaton used to calculate physcal entropy s very smlar to the one used for the texture measure. Energy s, n ths context, the opposte of entropy. Energy can be used to do useful work. In that sense t represents orderlness. Ths s why "Energy" s used for the texture that measures order n the mage. E = 1 = P [ ] P, ln P v 1 = P = ( μ ) v = v v 1 = P = v = v ( μ ) c) Contrast: Contrast s also called sum of squares varance. Ths measures the amount of local varaton n the mage and s the opposte of homogenety. Contrast = 1 = P ( ) The term P * ln(p) s maxmzed where ts dervatve wth respect to. P s. By the product rule, ths dervatve s P * d(ln(p))/d(p) + d(p)/d(p) * ln(p) whch smplfes to 1 + ln(p) =, yeldng P = 1/e. Ths means that the maxmum of the term to be summed occurs when P s 1/e, whch s about.378. However by defnton the sum of P = 1. Wth ths constrant, the overall maxmum of the sum (.e. of ET) s.5. Ths maxmum s reached when all probabltes are equal. Conceptually ths makes sense because when all probabltes of D pars are equal, we have a random dstrbuton of D values, whch would yeld maxmum "chaos" or entropy. g) Angular Second Moment (ASM): Ths measures homogenety of the mage. It s specfed by the matrx of relatve frequences P() wth whch two neghbourng pxels occur on the mage, one wth grey value and the other wth grey value. When and are equal, the cell s on the dagonal and (-) =. These values represent pxels entrely smlar to ther neghbour, so they are gven a weght of. If and dffer by 1, there s a small smlarty, and the weght s 1. If and dffer by, contrast s ncreasng and the weght s 4. The weghts contnue to ncrease exponentally as (-) ncreases. 1 P, = ASM = for any partcular drecton The square root of the ASM s sometmes used as a texture measure, and s called Energy. 1696

5 The Internatonal Archves of the Photogrammetry, Remote Sensng and Spatal Informaton Scences. Vol. XXXVII. Part B4. Beng 8 Energy = ASM h) Correlaton: The Correlaton texture measures the lnear dependency of grey levels on those of neghbourng pxels. GLCM Correlaton can be calculated for successvely larger wndow szes. The wndow sze at whch the GLCM Correlaton value declnes suddenly may be taken as one defnton of the sze of defnable obects wthn an mage. GLCM Correlaton s qute a dfferent calculaton from the other texture measures descrbed above. As a result, t s ndependent of them (gves dfferent nformaton) and can often be used proftably n combnaton wth another texture measure. It also has a more ntutve meanng to the actual calculated values: s uncorrelated, 1 s perfectly correlated. 1 = Correlaton P, =, ( μ )( ) ( )( ) μ δ δ The GLCM varance s zero when the area s unform. In a normalzed GLCM, a unform mage area has a sngle entry of 1 n the dagonal correspondng to the row and column headed by the orgnal GL value n the pxels of the mage. The mean then becomes μ = =. When calculatng the GLCM varance, P = for every value except the sngle entry of 1. The formula collapses to = 1( - μ), but snce μ = ths becomes 1 (=) =. In ths study, we use the Grey Level Co-occurrence Matrx (GLCM) method for studyng the texture of the tea areas to enable us to dstngush between healthy and affected tea patches. The GLCM technque was appled to all the mages and the dfferent parameters were studed. After generatng the mages they are compared wth maps showng dsease patterns to see whether the affects observed on those could be observed on the texture mages..4 Results and Dscussons.4.1 Delneaton of dseased tea patches: Tea garden patches were masked from the landuse/landcover maps and the dsease patches were delneated. From the mages the gradual spread of the dseases at three dfferent dates could be well observed. The spread was observed between 1 and 4. The LADSAT mage showed healthy, moderately affected and affected patches; these, however, could not be observed n the LISS III and ASTER mage. Comparng the Landsat, LISS III and the ASTER mages at three dfferent dates, we notce that the affected patches vsble at the LADSAT mage have changed to moderately affected at the LISS III mage, moderately affected patches to affected patches and the dseased patches to healthy patches n LISS III. Smlarly, comparng ASTER mage wth LISS III we observed that the moderately affected patches n LISS III changed to healthy patches and healthy patches to dseased patches n ASTER. Ths shows how the dsease spread between February and June n 4. Moreover a stronger nfestaton was found durng June 4 as there were heavy rans accompaned by hgh temperature and hgh relatve humdty durng that perod. Fgure 3. Images showng the gradual spread of the dseases at dfferent tme perods Creatng a texture mage: The result of a texture calculaton s a sngle number representng the entre wndow. Ths number s put n the place of the centre pxel of the wndow, then the wndow s moved one pxel and the process s repeated of calculatng a new GLCM and a new texture measure. In ths way an entre mage s bult up of texture values. Each cell n a wndow must st over an occuped mage cell. Ths means that the centre pxel of the wndow cannot be an edge pxel of the mage. If a wndow has dmenson x, a strp (-1)/ pxels wde around the mage wll reman unoccuped. The usual way of handlng ths s to fll n these edge pxels wth the nearest texture calculaton. 1697

6 The Internatonal Archves of the Photogrammetry, Remote Sensng and Spatal Informaton Scences. Vol. XXXVII. Part B4. Beng 8 Further delneaton of dsease tea patches was tred usng the texture based classfcaton n all the three dataset. Texture analyss for LADSAT, LISS III and ASTER mages were carred out usng varous wndows lke 3 3, 5 5, 7 7, 9 9, 11 11, 13 13, 15 15, and and wth band wse generated texture mages. It was found that only the 3 3 wndow yelded good texture results. The other wndows gave hazy and blurred texture mages from whch the tea patches could not be dentfed and mage analyss could not be done. The band gvng the best texture result n a 3 3 wndow was then selected for analyss purposes. All three bands gave good texture results for the ASTER mage, allowng the three patches could to be clearly observed. Healthy, moderately healthy and affected tea patches could be dstngushed wthn the ASTER mage. For the LADSAT mage, Band 4 (IR) and 5 (IR) gave good texture results. From the mean results, the tea patches could be well dentfed. The other parameters showed large nterclass mxng. From the mean results the healthy, moderately healthy and dseased patches could be separated, but mxng stll occurred among other landuse classes. Fgure 6. ASTER Images showng the healthy, moderately healthy and affected tea patches obtaned from texture based analyss. ASTER, beng a hgh resoluton mage, provdes good texture results as compared to LADSAT and LISS III mages. LADSAT and LISS III mages showed large nterclass mxng as compared to ASTER mages. Fgure 4. LADSAT Image showng the healthy and affected patches obtaned from texture based analyss. Band 3 (IR) and Band 4 (MIR) gave good texture results for LISS III, wth the tea patches promnently dentfable from the mean results. The texture mages of LISS III were further analyzed to dentfy between the healthy, moderately healthy and dseased tea patches. Here moderately healthy and dseased tea could not be separated. Therefore, the analyss could be done only for healthy and affected patches (both moderately affected and dseased patches combned). From ths we conclude that, for ths type of study, the texture analyss should be carred out usng hgh resoluton mages lke LISS IV or IKOOS. The other parameters of texture were also tested but none of them could produce good results due to excessve nterclass mxng..4. A Qualtatve Evaluaton of the Percentage of Healthy, Moderately Healthy and Dseased Tea Areas: The tea patches were masked out from the landuse/landcover maps for further analyss to fnd the percentage of healthy, moderately healthy and dseased tea patches. The formula used for estmatng the percentage of affected and non-affected tea s gven below: Healthy Tea Patches (%) = Area of Healthy Tea Patch / Total Tea Area Moderately Affected Tea Patches (%) = Area of Moderately Healthy Tea / Total Tea Area Dseased Tea Patches (%) = Area of Dseased Tea / Total Tea area Fgure 5. LISS III Images showng the healthy and affected patches obtaned from texture based analyss. Table 1. Table showng the percentage of affected and nonaffected tea patches 1698

7 The Internatonal Archves of the Photogrammetry, Remote Sensng and Spatal Informaton Scences. Vol. XXXVII. Part B4. Beng 8 Usng the LADSAT, LISS III and ASTER landuse/landcover mages, the area for the ndvdual classes were computed (n ha). Among the classes, the area of healthy, moderately affected and dseased tea patches were taken nto consderaton from whch the percentage of healthy, moderately healthy and dseased tea patches. From the table, the varatons n tea plantatons at three dfferent dates were observed. The LADSAT mage of December, 1 showed 6.4% area under healthy tea, 3.6% area under moderately affected tea and 16.% area under dseased tea. For the LISS III mage of February, 4, t was found that 43.9% area under healthy tea, 36.8% area under moderately affected tea and 19.3% area under dseased tea. Smlarly for ASTER mage of June, 4, area under healthy tea was found to be 4.9%, moderately healthy tea was found to be 5.1% and dseased tea to be 5.1%. Clearly, the dsease was most promnent n June. From June to October, the weather remaned hot and humd accompaned by heavy monsoon showers. Ths s the peak perod for the outbreak of Red Spder Mte attack and the Helopelts attack. The tea bushes are hghly prone to these two dseases durng ths perod and once the dsease breaks out t spreads rapdly..4.3 Detecton of Affected Plantatons from Remotely Sensed Images: We were able to dentfy and detect the affected and non-affected tea patches from remotely sensed mages. The dsease started n the year 1999 but was not severe at the ntal stage. Durng the year and 1, there was heavy monsoon showers accompaned by hgh temperature and humdty. These condtons have favoured the spread of the dsease. Between 3 and 5 the spread was qute severe. When the LADSAT ETM+ (December, 1), LISS III (February, 4) and ASTER (June, 4) mages were compared, t was observed that there was a gradual spread of the dsease from one place to another. The spread was mnutely observed n all the mages. It was further verfed on the feld. All affected tea patches from 1999 to 5 were dentfed from the garden records and from the feld vst. Feld records showed that there was a gradual spread of the dsease n the gardens. All the spots were vsted and marked n the mages and compared. From ths t was qute clear that whatever has been vsually nterpreted from the mage was correct and the spread of the dsease could be well establshed. 3. COCLUSIOS AD DISCUSSIO We conclude that texture analyss s useful for studyng tea bush health at tea plantatons. We could dstngush healthy, moderately healthy or dseased bushes. The Gray Level Cooccurrence Matrx (GLCM) technques have been able to delneate affected from non-affected tea patches. A qualtatve assessment was conducted to assess the percentage area wth affected and that wth non-affected tea plants. The LADSAT Image showed 6% area under healthy, 4% under moderately healthy and 16.% under dseased tea patches. The LISS III Image showed 44% area under healthy, 37% under moderately healthy and 19.3% area under dseased tea patches. Smlarly, the ASTER Images showed 4.9% area under healthy, 5.1% under moderately healthy and 5.1% area under dseased tea patches. From ths study t has been observed that ASTER mage could delneate the pest and dseased nfested areas nto healthy, moderately healthy and dseased / pests tea areas wthout any nterclass mxng whch was well observed n the LADSAT and LISS III mages. A decson on dentfyng the best texture analyss technque could be udged only when applyng avalable texture technques to a smlar study, lke the current one, and then comparng ther results. In addton, such an analyss should be carred out n hgh resoluton mages lke an IKOOS mage or a LISS IV mage. Further research should be carred out wth dfferent sensors of varous spatal, spectral and temporal resolutons to do a comparatve assessment or analyss. Dfferent texture analyss technques should be used and the results should be compared wth other technques to fnd the best sutable technque. Knowledge and obect based classfcaton should be further explored so that the results can be compared wth the texture based classfcaton. REFERECES Anys, H., et.al (1995). "Evaluaton of Textural and Multpolarzaton RADAR Features for Crop Classfcaton: IEEE Transactons on Geoscence and Remote Sensng, 33(5)." Azhar, H. S. (1993). Vegetaton Studes Usng Remote Sensng Technques. Malayasa, Unverst Teknolog. Bhattacharyya,., Tudu, Bpan, Bandyopadhyay, Rab, Bhuyan, Manabendra & Mud Raankanta (4). "Aroma Characterzaton of Orthodox Black Tea Wth Electronc ose." IEEE: Bordolo P. K. B., B.C. (1994). A Study on The Landuse Pattern And Ground Water Potental Through Remote Sensng Technque. Jorhat, Tokla Expermental Staton, Tea Research Assocaton. Chellappa, R. a. C., S. (1985). Classfcaton of Textures usng Gaussan Markov Random Felds. IEEE Transactons on Acoustcs Speech and Sgnal Processng. Congalton, R. G. (1991). A Revew of Assessng The Accuracy of Classfcatons of Remote Sensng Data: Remote Sensng of The Envronment. Dadhwal, V. K. "Crop Growth and Productvty Montorng and Smulaton Usng Remote Sensng and GIS." Satellte Remote Sensng and GIS Applcatons n Agrcultural Meteorology: pp Dadhwall, V. K. a. R., S.S. (). "Crop Assessment Usng Remote Sensng-Part II: Crop Condton and Yeld Assessment." Indan Journal of Agrcultural Economcs, (1947)-4. Duarah, B. P., Goswam I.D., P.P.Deka & Goswam D.C. (1993). Remote Sensng Applcaton For Ground Water Prospectng In The Tea Growng Areas of The Tera Regon, West Bengal. Guwahat Assam, Inda, Assam Remote Sensng Applcaton Centre. Dutta, R., Hnes, E.L., Gardner, J.W., Kashwan, K.R. & Bhuyan, M (3). "Determnaton of Tea Qualty By Usng A eural etwork Based Electronc ose." IEEE: Ghosh, A. (3). "GIS Anchored Integrated Plantaton Management: Tea." 1699

8 The Internatonal Archves of the Photogrammetry, Remote Sensng and Spatal Informaton Scences. Vol. XXXVII. Part B4. Beng 8 Gonzalez, R. C. a. W., R.E. (1993). Dgtal Image Processng. Haddon, J. F., et.al (1993). "Cooccurance Matrces for Image Analyss." IEE Electroncs and Communcaton Engneerng Journal Vol.5, o. : Hall-Beyer, M. (October, ). GLCM Texture: A Tutoral. Haralck, R. M., Shanmugam, K. and Dnsten, I. (ovember, 1973). Textural Features for Image Classfcaton. IEEE Transactons on Systems Man and Cybernetcs. Jan, A. K., et.al (1998). Texture Analyss, World Scentfc Publshng Co. Janssen, L. a. v. d. W., F. (1994). "Accuracy of Satellte Derved Land cover Data, A Revew, Photogrametrc Engneerng and Remote Sensng,(6)4." Janssen, L. L. F. a. G., B.G.H. (1). Chapter 1, Dgtal Image Classfcaton. Prncples of Remote Sensng. Internatonal Insttute for Aerospace Survey and Earth Scence (ITC), Enschede, The etherlands Kashyap, R. L., et.al (198). "Texture Classfcaton Usng Features Derved From Random Feld Models: Pattern Recognton Letters." 1:43-5. Lee, K. S., Park, Y.I., Km, S.H., Park, J.S., Woo, C.S. & Jang, K.C. "Remote Sensng Estmaton of Forest LAI n Close Canopy Stuaton." Lllesand, T. M. R. W. K. (). Remote Sensng and Image Interpretaton. ew York, Chchester, John Wlly & Sons. Megown, R. A., Webster, Mke and Jacobs, Shayne "Usng Landsat TM Imagery to Estmate LAI n a Eucalyptus Plantaton." Oala, T., et.al (1999). "Gray Level Cooccurance Hstograms va Learnng Vector Quantzaton." The 11th Scandnavan Conference on Image Analyss, Kangerlussuaq, Greenland. Panda, R. K., Stephans, W and Matthews, R. (8 October, ). "Modellng The Influence of Irrgaton on The Potental Yeld of Tea (Camella snenss) In orth-east Inda." Patel,. R., Bhattacharee, B., Mohammad, A.J., Tanuprya, B & Saha, S.K. (In the Press). "Remote Sensng of Regonal Yeld Assessment of Wheat n Haryana, Inda." Internatonal Journal of Remote Sensng. Raapakse, R. M. S. S., Trpathy, tn, K. and Honda, Kyosh (1). "Modellng Tea (Camella(L)O.Kuntze) Yeld Usng Satellte Derved LAI, Landuse and Meteorologcal Data." Raapakse, R. M. S. S., Trpathy, tn, K. and Honda, Kyosh (). "Spectral Characterzaton and LAI Modellng For Tea (Camella snenss (L) Kuntze) Canopy." Internatonal Journal of Remote Sensng Vol. 3, o. 18, Sun, X. (16 October, 3). "Comparson of Pxel Based and Obect Orented Approaches to Landcover Classfcaton Usng Hgh Resoluton IKOOS Satellte Data." Tuceryan, M. a. J., A.K., Ed. (1993). Texture Analyss, In Handbook of Pattern Recognton and Computer Vson. Sngapore, C.H. Chen, L.F. Pau and P.S.P. Wang, World Scentfc Publshng Co. Varlygun, D., Wrght, R.K., Goetz, S.J. and Prnce, S.D. (1). "Advances In Landcover Classfcaton From Applcaton Research: A Case Study From The Md Atlantc RESAC." Proceedngs of The Amercan Socety of Photogrammetry and Remote Sensng. 17