Your SlideShare is downloading. ×
Mr2521772184
Mr2521772184
Mr2521772184
Mr2521772184
Mr2521772184
Mr2521772184
Mr2521772184
Mr2521772184
Upcoming SlideShare
Loading in...5
×

Thanks for flagging this SlideShare!

Oops! An error has occurred.

×
Saving this for later? Get the SlideShare app to save on your phone or tablet. Read anywhere, anytime – even offline.
Text the download link to your phone
Standard text messaging rates apply

Mr2521772184

126

Published on

IJERA (International journal of Engineering Research and Applications) is International online, ... peer reviewed journal. For more detail or submit your article, please visit www.ijera.com

IJERA (International journal of Engineering Research and Applications) is International online, ... peer reviewed journal. For more detail or submit your article, please visit www.ijera.com

Published in: Technology
0 Comments
0 Likes
Statistics
Notes
  • Be the first to comment

  • Be the first to like this

No Downloads
Views
Total Views
126
On Slideshare
0
From Embeds
0
Number of Embeds
0
Actions
Shares
0
Downloads
2
Comments
0
Likes
0
Embeds 0
No embeds

Report content
Flagged as inappropriate Flag as inappropriate
Flag as inappropriate

Select your reason for flagging this presentation as inappropriate.

Cancel
No notes for slide

Transcript

  • 1. Gitanjali S. Korgaonkar, Dr. R. R. Sedamkar / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue 5, September- October 2012, pp.2177-2184 Hyperspectral Image Classification Gitanjali S. Korgaonkar, Dr. R. R. SedamkarGitanjali S. Korgaonkar is currently pursuing masters degree in electronics & telecommunication engineering in TCET,Kandivali,Mumbai University,India,Dr. R.S. Sedamkar is received his Ph.D. degree in 201 and is currently Dean-Academics, Professor and Head of Computer Department at Thakur college of engineering and technology, Mumbai.AbstractThis paper proposes weighted decision fusion for 1.1The Imaging Spectrometerclassification on a sample hyperspectral image Hyperspectral images are produced by in-which has two characteristics 1) Spatial and 2) struments called imaging spectrometers. The devel-Spectral. Spatial is nothing but within a single opment of these complex sensors has involved theband and Spectral means narrow spectral bands convergence of two related but distinct technologies:over a continuous spectral range.The Weighted spectroscopy and the remote imaging of Earth andMajority voting rule is introduced which is likely planetary surfaces. Spectroscopy is the study ofto enhance the performance. In this pixels close light that is emitted by or reflected from materialsto the spectral centroid of a segment are assigned and its variation in energy with wavelength. As ap-higher weights for voting because they are the plied to the field of optical remote sensing, spectros-best representatives of the segment, and pixels copy deals with the spectrum of sunlight that is dif-deviated from the centroid are not representa- fusely reflected (scattered) by materials at thetives and will have lower weights. Experimental Earth.s surface.results for SAMSON (Spectroscopic Aerial Map-ping System with Onboard Navigation) data Instruments called spectrometers (or spec-show that proposed scheme achieve Overall Ac- troradiometers) are used to make ground-based orcuracy, Kappa Coefficient and Average Accuracy laboratory measurements of the light reflected frombetter than Majority voting rule. a test material. An optical dispersing element such as a grating or prism in the spectrometer splits thisIndex Terms— Hyperspectral image classifica- light into many narrow, adjacent wavelength bandstion, Supervised classification, Unsupervised Classi- and the energy in each band is measured by a sepa-fication, Weighted majority voting Rule. rate detector. By using hundreds or even thousands of detectors, spectrometers can make spectral mea-Introduction surements of bands as narrow as 0.01 micrometers over a wide wavelength range, typically at least 0.4 Hyperspectral images contain a wealth of to 2.4 micrometers (visible through middle infrareddata, but interpreting them requires an understand-ing of exactly what properties of ground materials wavelength ranges).we are trying to measure, and how they relate to themeasurements actually made by the hyperspectral 1.2Characteristics of Electromagnetic Radiation:sensor.[1] Hyperspectral imagery involves the sensing of electromagnetic radiation. The electromagnetic spectrum that is used in remote sensing includes the ultraviolet (UV), visible, infrared, and microwave portions of the spectrum. Figure 2 illustrates the wavelength regions of the electromagnetic spectrum. The spectral por- tions of near IR and short wave infrared (0.7-3.0 μm) are called the reflective infrared because meas- ured radiation in this spectral region is mostly com- posed of reflected sunlight. In contrast, the IR spec- trum from 5.0 to 13.0 μm is termed thermal infrared because measurements in this spectral region are generally recording energy radiated from scene ele- ments [2]Figure 1. The Hyperspectral Imagary 2177 | P a g e
  • 2. Gitanjali S. Korgaonkar, Dr. R. R. Sedamkar / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue 5, September- October 2012, pp.2177-2184 to the other two in specific wavelength regions. Therefore, by knowing a materials spectral signa- ture, it is possible for this material to be detected in pixels within an image. Libraries with the character- istic spectra of various materials of interest exist, and these spectral signatures can be compared with measured spectra in order to identify the features in an image. In the early work on spectral imagery, computational limits prevented full exploitation of such data. Computational power in the latter portion of the 1990‘s made routine use of spectral imageryFigure 2 Wavelength regions of the electromagnetic much more practical.spectrum.Based on these wavelength regions, remote sensingcan be classified into three main categories: visible Rand reflective infrared remote sensing, thermal re- amote sensing, and microwave remote sensing. In dvisible and reflective remote sensing, the radiation ithat is measured has a solar origin, radiated with a apeak wavelength of 0.5 μm. In thermal remote sens- ning, the measured radiation comes from the ob- cserved objects. Materials with normal temperatures e(~300K) emit radiation with a peak wavelength of10.0 μm. Finally, in microwave remote sensing, ob-servations are generally due to reflection of energyradiated by the observing platform (i.e. radar). Inthis paper, hyperspectral images from the visible andreflective infrared spectrum are used. The majority Wavelength(nm)of currently available sensors with material identifi-cation (ID) capability utilize this portion of spec- Figure 3. Hyperspectral pixel spectra.trum. Typically the source of energy in hyper- Identifying groups of pixels that have simi-spectral imagery is the sun. The incident energy lar spectral characteristics and determining the vari-from the sun that is not absorbed or scattered in the ous features or land cover classes represented byatmosphere interacts with the earths surface, where these groups is an important part of image analysis.it is absorbed, transmitted, or reflected. Additionally, This form of analysis is known as classification.the electromagnetic radiation has specific properties Visual classification relies on the analysts ability tothat are predictable in its interaction with materialsand transmission through the atmosphere. Therefore, use visual elements (tone, contrast, shape, etc) toby measuring the electromagnetic radiation in nar- classify an image. Digital image classification is based on the spectral information used to create therow wavelength bands, the resulting spectral signa- image and classifies each individual pixel based ontures can be used, in principle, to uniquely charac- its spectral characteristics. The result of a classifica-terize and identify any given material. tion is that all pixels in an image are assigned to All materials have unique spectral charac- particular classes or themes (e.g. water, coniferousteristics because they absorb, reflect, and emit radia- forest, deciduous forest, corn, wheat, etc.), resultingtion in a unique way. For example, in the visibleportion of the spectrum, a leaf appears green be- in a classified image that is essentially a thematiccause it absorbs in the blue and red regions of the map of the original image. The theme of the classifi-spectrum and reflects in the green region. These cation is selectable, thus a classification can be per-variations in absorption, reflection, and emission are formed to observe land use patterns, geology, vege-due to the material composition. Differences in tation types, or rainfall. In classifying an image we must distin-spectral response due to absorption, transmission, guish between spectral classes and informationand reflection cause materials to have a unique spec- classes. Spectral classes are groups of pixels thattral signature. Figure 3 illustrates the spectral signa- have nearly uniform spectral characteristics. Infor-tures of three pixels, respectively dominated by mation classes are various themes or groups we areseawater, vegetation, and concrete. Comparing the attempting to identify in an image. Informationspectra between seawater and vegetation, it is ob-served that they reflect similarly in the visible wave- classes may include such classes as deciduous and coniferous forests, various crop types, or inland bod-lengths but differently in the infrared portion. Also,concrete has a different spectral signature compared ies of water. The objective of image classification is to match the spectral classes in the data to the in- 2178 | P a g e
  • 3. Gitanjali S. Korgaonkar, Dr. R. R. Sedamkar / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue 5, September- October 2012, pp.2177-2184formation classes of interest. 2.1.1.Support Vector Machine Different conditions like the sun illumina- One of the best algorithm used for super-tion, snow cover, other atmospheric conditions, crop vised classification is Support Vector Ma-growth in a particular season etc. conditions effect chine(SVM). A support vector machine (SVM) is athe classification process. concept in statistics and computer science for a set of related supervised learning methods that analyze2. TYPES OF CLASSIFICATION data and recognize patterns, used for classification Image classification is perhaps the most and regression analysis. The standard SVM takes aimportant part of digital image analysis. It is very set of input data and predicts, for each given input,nice to have a "pretty picture" or an image, showing which of two possible classes forms the input, mak-a magnitude of colors illustrating various features of ing the SVM a non-probabilisticbinarylinear classi-the underlying terrain, but it is quite useless unless fier. Given a set of training examples, each markedto know what the colors mean. Image classification as belonging to one of two categories, an SVMor the ‗partition of the feature space‘ can be done in training algorithm builds a model that assigns newtwo ways: Supervised and Unsupervised classifica- examples into one category or the other. An SVMtion. In the supervised classification, the operator model is a representation of the examples as pointsdefines his classes of interest by selecting the train- in space, mapped so that the examples of the sepa-ing samples in an image on the basis of his/her rate categories are divided by a clear gap that is asknowledge of the area. In the unsupervised classifi- wide as possible. New examples are then mappedcation, the image is partitioned into homogenous into that same space and predicted to belong to aspectral clusters using some clustering algorithm. category based on which side of the gap they fall on.These spectral clusters are accomplished on the ba- SVMs characterize classes using a geometrical crite-sis of some spectral similarities. In the classification rion rather than statistical criteria. They seek a sepa-application, we need to preserve the features that are rating hyperplane maximizing the distance with theuseful for discrimination among classes. closest training samples for two classes. This ap- proach allows SVMs to have a very high capability2.1 Supervised Classification of generalization and, as a consequence, only re- With supervised classification, we identify quire a few training samples. For non linearly sepa-examples of the Information classes (i.e., land cover rable data, SVMs use the kernel trick to map thetype) of interest in the image. These are called data onto a higher dimensional space where they are"training sites". The image processing software sys- linearly separable [4].tem is then used to develop a statistical characteriza- Here, we consider multiclass SVMs without anytion of the reflectance for each information class. feature reduction of the original hyperspectral data.This stage is often called "signature analysis" and The standardmay involve developing a characterization as simple Gaussian radial basis kernel with L2-normas the mean or the rage of reflectance on each bands, distance is used. This algorithm was proved to pro-or as complex as detailed analyses of the mean, vide interesting classification accuracy, even in thevariances and covariance over all bands. case of limited training set [5]. Note that other ker-Once a statistical characterization has been achieved nels could be considered, such as the spectral anglefor each information class, the image is then classi- mapper which basically computes the angle betweenfied by examining the reflectance for each pixel and two vectors in the vector space [5][6][7]. In the fol-making a decision about which of the signatures it lowing, we present the used classifier and start byresembles most[3]. briefly recalling the general mathematical formula- tion of SVMs. Starting from the linearly separable case, optimal hyperplanes are introduced, then the classification problem is modified to handle non- linearly separable data and a brief description of multiclass strategies is given. Finally, kernel meth- ods are presented. Linear SVM For a two-class problem in a n-dimensional space n, we assume that l training samples, xi n, are available with their corresponding labels yi = ±1, S = {(xi, yi) | i 2 [1, l]}. The SVM method consists of finding the hyperplane that maximizesFigure 4. Supervised Classification the margin i.e., the distance to the closest training data points in both classes. Noting w 2179 | P a g e
  • 4. Gitanjali S. Korgaonkar, Dr. R. R. Sedamkar / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue 5, September- October 2012, pp.2177-2184 n as the vector normal to the hyperplane and b find the hyperplane of maximal margin in the new feature space. The separating hyperplane becomes a as the bias, the hyperplane Hp is defined as linear function in the transformed feature space but (1) a nonlinear function in the original input space. LetWhere <W, X> is the inner product between w and x be a vector in the n-dimensional input space andx. If x Hp then f(x) = <W, X> + b is the distance j(·) be a nonlinear mapping function from the inputof x to Hp. The sign of f corresponds to decision space to the high-dimensional feature space. Thefunction y = sgn (f(x)). The optimal Parameters (w, hyperplane representing the decision boundary inb) are found by solving the feature space is defined as follows. +c ] (2) w·( x)−b = 0 (5)Subject to where w denotes a weight vector that can map the training data in the high dimensional fea- ((W,Xi) + b ) ≥ 1-ξi, ξi ≥ 0 i (3) ture space to the output space, and b is the bias. Us- ing the j(·) function, the weight becomes The solution vector is a linear combinationof some samples of the training set, whose _i is non-zero, called Support Vectors. The hyperplane deci- w  iyi (xi) (6)sion function can thus be written as: l The decision function ofyu  sgn( yii  Xu , Xi   b) i 1 (4)The maximum margin linear classifier is the linearclassifier with the, um, maximum margin. This is (7)the simplest kind of SVM (Called an LSVM) becomes (8) Note that the feature mapping functions in the opti- mization problem and also in the classifying func- tion always appear as dot products, e.g., j(xi) ·j(xj). j(xi) · j(xj) is the inner product between pairs of vectors in the transformed feature space. Computing the inner product in the transformed feature space seems to be quite complex and suffer from the course of dimensionality problem. To avoid this problem, the kernel trick is used. The kernel trick replaces the inner product in the feature space with a kernel function K in the original input space as fol- lows. K(u,v) =j(u) ·j(v) (9) The Mercer‘s theorem proves that a kernel function K is valid, if and only if, the following conditions are satisfied, for any function y(x).Figure 5 Maximum Margin Of SVMThere is one non separable vector in each class (10)Non linear support vector machine Where 2dx If the training data is not linearly separable, The Mercer‘s theorem ensures that the ker-there is no straight hyperplane that can separate the nel function can be always expressed as the innerclasses. In order to learn a nonlinear function in that product between pairs of input vectors in some high-case, linear SVMs must be extended to nonlinear dimensional space, thus the inner product can beSVMs for the classification of nonlinearly separable calculated using the kernel function only with inputdata. The process of finding classification functions vectors in the original space without transformingusing nonlinear SVMs consists of two steps. First, the input vectors into the high dimensionalthe input vectors are transformed into high- feature vectors.dimensional feature vectors where the training data The classification function becomes:can be linearly separated. Then, SVMs are used to 2180 | P a g e
  • 5. Gitanjali S. Korgaonkar, Dr. R. R. Sedamkar / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue 5, September- October 2012, pp.2177-2184 term GIS database maintenance. The reason is that (11) there are now systems that use clustering procedures that are extremely fast and require little in the nature Since K(·) is computed in the input space, of operational parameters. Thus it is becoming pos-no feature transformation will be actually done or no sible to train GIS analysis with only a general fa-j(·) will be computed, and thus the weight vector w miliarity with remote sensing to undertake classifi-= åaiyij(x) will not be computed either in nonlinear cations that meet typical map accuracy standards.SVMs. With suitable ground truth accuracy assessment pro-The followings are popularly used kernel functions. cedures, this tool canprovide a remarkably rapid• Polynomial: K(a,b) = (a ·b+1)d means of producing quality land cover data on a• Radial Basis Function (RBF): K(a,b) = exp(−g continuing basis[3].||a−b||2)• Sigmoid: K(a,b) = tanh(ka ·b+c)[8]. Figure 7 Unsupervised ClassificationFigure 6 Non Linear SVM ‗s Separation 2.2.1 K-Mean Algorithm The kernel function transforms the data K-means is one of the simplest unsuper-into a higher dimensional space to make it possible vised learning algorithms that solve the well knownto perform the separation. clustering problem. The procedure follows a simple and easy way to classify a given data set through a2.2 Unsupervised Classification certain number of clusters (assume k clusters) fixed Unsupervised classification is a method a priori. The main idea is to define k centroids, onewhich examines a large number of unknown pixels for each cluster. These centroids shoud be placed inand divides into a number of classed based on natu- a cunning way because of different location causesral groupings present in the image values. unlike different result. So, the better choice is to place themsupervised classification, unsupervised classification as much as possible far away from each other. Thedoes not require analyst-specified training data. The next step is to take each point belonging to a givenbasic premise is that values within a given cover data set and associate it to the nearest centroid.type should be close together in the measurement When no point is pending, the first step is completedspace (i.e. have similar gray levels), whereas data in and an early group age is done. At this point wedifferent classes should be comparatively well sepa- need to re-calculate k new centroids as barycentersrated (i.e. have very different gray levels). of the clusters resulting from the previous step. After The classes that result from unsupervised we have these k new centroids, a new binding has toclassification are spectral classed which based on be done between the same data set points and thenatural groupings of the image values, the identity nearest new centroid. A loop has been generated. Asof the spectral class will not be initially known, a result of this loop we may notice that the k cen-must compare classified data to some from of refer- troids change their location step by step until noence data (such as larger scale imagery, maps, or site more changes are done. In other words centroids dovisits) to determine the identity and informational not move any more.[9]values of the spectral classes. Thus, in the super- Finally, this algorithm aims at minimizingvised approach, to define useful information catego- an objective function, in this case a squared errorries and then examine their spectral separability; in function. The objective function:the unsupervised approach the computer determinesspectrally separable class, and then define their in-formation value. Unsupervised classification is becoming (12)increasingly popular in agencies involved in long 2181 | P a g e
  • 6. Gitanjali S. Korgaonkar, Dr. R. R. Sedamkar / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue 5, September- October 2012, pp.2177-2184 where Xi(j) - Cj∥ 2 is a chosen distance Decision fusion has been widely used tomeasure between a data point Xi(j) and the cluster improve the overall classification accuracy Mostcentre Cj, is an indicator of the distance of the n data decision fusion approaches mainly focus on super-points from their respective cluster centres. vised classifiers as base learner, i.e., all classifiersThe algorithm is composed of the following steps: need training, so the classification results can only1) Place K points into the space represented by the be as good as training data. To avoid the possibleobjects that are being clustered. These points repre- negative influence from the limited quality of train-sent initial group centroids. ing data, we are motivated to propose amethod that2) Assign each object to the group that has the clos- can combine supervised and unsupervised classifi-est centroid. ers.3) When all objects have been assigned, recalculate A supervised classifier may provide betterthe positions of the K centroids. classification than an unsupervised classifier. How-4) Repeat Steps 2 and 3 until the centroids no longer ever, in addition to training data limitation, a super-move. This produces a separation of the objects into vised classifier may result in over-classification forgroups from which the metric to be minimized can some homogeneous areas. Although it may be lessbe calculated. powerful, an unsupervised classifier can generallyAlthough it can be proved that the procedure will well classify those spectrally homogeneous areas.always terminate, the k-means algorithm does not Thus, fusing supervised and unsupervised classifica-necessarily find the most optimal configuration, tion may yield better performance since the impactcorresponding to the global objective function from trivial spectral variations may be alleviated andminimum. The algorithm is also significantly sensi- the subtle difference between spectrally similar pix-tive to the initial randomly selected cluster centres. els may not be exaggerated. Although individualThe k-means algorithm can be run multiple times to classifiers are pixel-based, the final fused classifica-reduce this effect. tion has a similar result to an objectbased classifier.K-means is a simple algorithm that has been adaptedto many problem domains. As we are going to see, it 4. VOTING RULEis a good candidate for extension to work with fuzzy In this paper, we used a weighted majorityfeature vectors. voting (WMV) rule to further improve classification accuracy. The final fusion output is dependent on3. DECISION FUSION not only the classification accuracy of the super- A decision fusion approach is developed to vised classifier and the chosen unsupervised classi-combine the results from supervised and unsuper- fier but also the fusion rule. The MV rule is the mostvised classifiers. The final output frequently used. It is simple and requires no train- takes advantage of the power of a support- ing. The MV rule may have some limitations. Forvector machine- based supervised classification in instance, it only counts the number of pixels in eachclass separation and the capability of an unsuper- class in a segment, but does not consider pixelvised classifier, such as K-means clustering, in re- ―quality‖ since it treats all the pixels equally; as aducing trivial spectral variation impact in homoge- result, the spectral variations from pixel to pixel isneous regions [10][11][12]. completely ignored in a relatively homogeneous segment. Intuitively, pixels close to the spectral cen- troid of a segment should be assigned higher weights for voting because they are the best repre- sentatives of the segment, and pixels deviated from the centroid are not representatives and should have lower weights. The resulting WMV rule can allevi- ate the impact from outliers, whose spectral signa- tures are quite different from the rest of pixels al- though they are grouped into the same segment. To gauge such deviation, the Mahalanobis distance can be used to measure the distance between a pixel and the centroid. 4.1 Accuracy Assessment: Accuracy assessment is an important step in the classification process. The goal is to quantita- tively determine how effectively pixels were grouped into the correct feature classes in the areaFigure 8 Decision Fusion Approach under investigation. The land cover types derived from digital image interpretation and analysis re- quires validation with data obtained from ground 2182 | P a g e
  • 7. Gitanjali S. Korgaonkar, Dr. R. R. Sedamkar / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue 5, September- October 2012, pp.2177-2184verification. For the accuracy assessment, a coeffi-cient of agreement between classified image dataand ground reference data were calculated usingKappa. The accuracy was determined by kappa sta-tistics in order to test whether any difference existsin the interpretation work. Briefly, Kappa statistic considers a measureof overall accuracy of image classification and indi-vidual category accuracy as a means of actualagreement between classification and observation.The value of Kappa lies between 0 and 1, where 0represents agreement due to chance only. Mean-while 1 represents complete agreement between thetwo data sets. Negative values can occur but theyare spurious. It is usually expressed as a percentage(%). [13] claimed that the Kappa statistic has beenshown to be a statistically more sophisticated meas- Figure 9 Hyperspectral Imageure of classifier agreement and thus gives betterinterclass discrimination than overall accuracy. 5.1 Classification output:5. RESULT AND ANALYSIS The software implementation of the algo-rithm is written in a Matlab environment using Mat-lab7.10 software. This hyperspectral dataset wasgenerated by the SAMSON sensor The SAMSONcomponents are integrated with the flight manage-ment system to enable real-time management of datacollection and flight orientation. K- Means SVM The HSI sensor incorporates improvedspectral alignment, radiometric calibrations and at-mospheric correction procedures. The charge cou-pled device (CCD) used in the SAMSON sensor is athinned, backside-illuminated CCD (produced bySarnoff Imaging). The CCD manufacturing processgreatly increases the quantum efficiency from about5% to 60% at 400 nm, and from about 40% to 85%at 700 nm.with a band width of 3.2nm. The data wascollected by the Florida Environmental Research Weighted MajorityInstitute as part of the GOES-R sponsored experi- Voting Majority Votingment. The instrument flown during the collect is the Figure 10.Classified Output for band 75SAMSON, a push-broom,visible to near IR, hyper-spectral sensor. This sensor was designed and devel- Bands OA= AA= OA= AA=oped by FERI. The following paper describes the 90.5886 90.585 99.0039 99basic design of the sensor: Kohler et al. (2006)[14]. Majority Voting Weighted MajorityThis dataset utilized the new radiometric calibration Kappa Coefficient Votingtechnique specifically designed to characterize and Kappa Coefficientcorrect the stray light found within the sensor. The 1 91.4158 99.9080following paper describes the approach: Kohler etal. (2004). It has samples = 952, 10 90.8216 99.2586lines = 952,bands = 156. 25 91.0987 99.5614 45 90.7436 99.1733 75 88.8225 97.0738 50 90.3567 98.7505 2183 | P a g e
  • 8. Gitanjali S. Korgaonkar, Dr. R. R. Sedamkar / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue 5, September- October 2012, pp.2177-2184100 89.0256 97.2958 multiclass classification of hyperspectral- remote sensing data,‖ in Proc. IEEE Int.125 91.4878 99.9867 Conf. on Acoustics, Speech, and Signal Processing, Toulouse, France,2006. [6] M. L. G. Mercier, ―Support vector ma-145 89.0309 97.3015 chines for hyperspectral image classifica- tion with spectral-based kernels,‖in Geo-150 89.0639 97.3376 science and Remote Sensing Symposium, vol. 1. IGARSS ‘03. Proceedings, July156 89.1105 97.3886 2003. [7] N. Keshava, ―Distance metrics and band selection in hyperspectral processing withTable 1. Comparision of different Bands application to material iden-tification and spectral librairies,‖ IEEE Trans. on Geo- science and Remote Sensing, vol. 42, pp. 1552–1565, July2004. [8] http://people.sabanciuniv.edu/berrin/cs512/l ectures/11-svmtutorial2.pdf by Hwanjo Yu andSungchul Kim[9]http://home.dei.polimi.it/matteucc/C lustering/tutorial_html/kmeans.html [10] H. Yang, B. Ma, and Q. Du, ―Decision fu- sion for supervised and unsupervised hyper spectral image classification,‖ in Proc. IEEE Geosci. Remote Sens. Symp., Jul. 2009, pp. 948–951. [11] M.Petrakos,J.A. Beniktsson, and I. Kanel- lopoulos,‖The effect of Classifier agree-Figure 11.Kappa Coefficient of hyperspectral image ment on the accuracy of the combined clas-at different wavelengths sifier in decision level fusion,‖IEEE Trans.Geosci. Remote Sens., Vol6. CONLUSION 39,no.11,pp.2539-2546, Nov.2001. Decision Fusion is the most widely used [12] A.Cheriyadat,L.M.Bruce and A. Mathur, ―methods for Classification of hyperspectral image. Decision Level Fusion with best-bases forIn this paper, we used a weighted decision fusion hyperspectral Classification,‖ in Proc.approach. The final output take advantage of the IEEE Workshop Adv. Tech. Anal. Remotelypower of supervised classification in class separa- Sensed Data, 2003,pp.399-406tion and the capability of unsupervised classifier in [13] Fitzgerald,R.W and Lees,B.G. 1994. As-reducing the impact from spectral variations in ho- sessing the classification accuracy of multi-mogeneous regions. sources remote sensing data. Remote Sens- The Weighted Decision Fusion Method ing of Environment, 47: 362-368.were examined using SAMSON hyperspectal sam- [14]ple data, which shows that the classification accu- www.opticks.org/confluence/display/optickracy and the kappa coefficient is better than that s/sample+datausing MV method.REFERENCES [1] Documention Tutorials on hyprspec by Radall B. Smith [2] Peg Shippert‖ Introduction to Hyperspec- tral Image Analysis‖ Earth Science Appli- cations Specialist Research Systems, Inc [3] http://www.sc.chula.ac.th/courseware/2309 507/Lecture/remote18.htm [4] B. Scholkopf and A. J. Smola, Learning with Kernels. MIT Press, 2002. [5] M. Fauvel, J. Chanussot, and J. A. Benediktsson, ―Kernels Evaluation for 2184 | P a g e

×