Supervised Image Classification Techniques

Supervised persona Classification TechniquesIntroduction In this chapter, a review of Web-Based GIS Technology and beam image variety techniques. piece 2.2 presents a review of Web-Based GIS Technology.in section 2.3 planet images salmagundi techniques atomic descend 18 reviewed.In section 2.4 presents the related work .section 2.5 presents substance ab recitations of web found GIS applications in real world. Section 2.6 presents available commercial web GIS sites. Section 2.7 reviews the types of Geospatial Web Services (OGC)2.3 Image ClassificationImage secernification is a force to automatically categorize all pels in an Image of a terrain into drop cover classes. Normally, multi spiritual information atomic add 18 used to Perform the sort of the ghostlike type present indoors the info for from from each one one picture element is used as the numerical basis for categorization. This concept is dealt under the tolerant subject, that is to say, Pattern Rec ognition. Spectral pattern recognition refers to the Family of mixed bag procedures that utilizes this pixel-by-pixel spectral prepargon as the basis for automated republic cover compartmentalization. Spatial pattern recognition involves the categorization of image pixels on the basis of the spatial birth with pixels surrounding them. Image smorgasbord techniques atomic number 18 grouped into devil types, namely supervised and unsupervised1. The classification process may also include features, such(prenominal) as, land surface elevation and the soil type that ar non derived from the image. Two categories of classification be contained different types of techniques put forward be seen in figFig. 1 F pitiable Chart showing Image Classification12.3 rudimentary steps to apply Supervised ClassificationA supervised classification algorithmic program requires a training sample for each class, that is, a hookup of info points know to afford come from the class of interest . The classification is consequently ground on how obturate a point to be sort out is to each training sample. We shall non attempt to define the word close other than to say that both Geo mensurable and statistical outgo stairs argon used in practical pattern recognition algorithms. The training samples atomic number 18 representative of the known classes of interest to the analyst. Classification methods that relay on use of training patterns argon called supervised classification methods1. The three basic steps (Fig. 2) involved in a typical supervised classification procedure ar as sweep upsFig. 2. Basic steps supervised classification 1(i) learn stage The analyst identifies representative training areas and develops numerical descriptions of the spectral signatures of each land cover type of interest in the scene.(ii) The classification stag(Decision happen)e to each one pixel in the image info set IS categorized into the land cover class it most closely resemb les. If the pixel is insufficiently confusable to any training data set it is usually labeled Unknown.(iii) The produce stage The cores may be used in a number of different paths. Three typical forms of output products are thematic maps, boards and digital data files which become excitant data for GIS. The output of image classification becomes input for GIS for spatial analysis of the terrain. Fig. 2 depicts the flow of operations to be performed during image classification of impertinently sensed data of an area which at long last leads to create database as an input for GIS. Plate 6 shows the land use/ land cover color coded image, which is an output of image2.3.1 Decision conventionality in image classifficationAfter the signatures are define, the pixels of the image are choose into classes based on the signatures by use of a classification finish rein. The conclusiveness swayer is a numerical algorithm that, use data contained in the signature, performs the actu al sorting of pixels into distinct class determine2. There are a number of powerful supervised classifiers based on the statistics, which are frequently, used for various applications. A few of them are a tokenish outmatch to guesss method, average distance method, parallelepiped method, level best likelihood method, modified maximum likelihood method, Baysians method, decision tree classification, and discriminant functions. Decision overtop mess be sort out into 2 types1- Parametric Decision RuleA parametric decision ascertain is trained by the parametric signatures. These signatures are define by the mean vector and covariance intercellular substance for the data file set of the pixels in the signatures. When a parametric decision rule is used, every pixel is charge to a class since the parametric decision space is continuous32-Nonparametric Decision RuleA nonparametric decision rule is non based on statistics therefore, it is independent of the properties of th e data. If a pixel is regain at heart the boundary of a nonparametric signature, then this decision rule assigns the pixel to the signatures class. Basically, a nonparametric decision rule determines whether or not the pixel is located inside of nonparametric signature boundary3 .2.3.2 supervised algorithm for image classifficationThe principles and working algorithms of all these supervised classifiers are derived as follow Parallelepiped ClassificationParallelepiped classification, sometimes also known as box decision rule, or level-slice procedures, are based on the ranges of determine within the training data to define regions within a flat data space. The spectral values of unclassified pixels are projected into data space those that fall within the regions defined by the training data are assign to the appropriate categories 1. In this method a parallelepiped-like (i.e., hyper-rectangle) subspace is defined for each class. Using the training data for each class the limits of the parallelepiped subspace can be defined either by the minimum and maximum pixel values in the abandoned class, or by a certain number of standard deviations on either side of the mean of the training data for the given class . The pixels lying inside the parallelepipeds are tagged to this class. Figure depicts this criterion in cases of two- markal feature space4.Fig. 3. execution of instrument of the parallelepiped classification methodfor three classes utilise two spectral bands, after4. b showline outdo Classificationfor supervised classification, these groups are formed by values of pixels within the training fields defined by the analyst.Each stud can be represented by its centroid, often defined as its mean value. As unassigned pixels are considered for assignment to one of the some(prenominal) classes, the multidimensional distance to each gang centroid is figure, and the pixel is then assigned to the closest cluster. Thus the classification proceeds by always using the minimum distance from a given pixel to a cluster centroid defined by the training data as the spectral verbalism of an informational class. stripped-down distance classifiers are direct in concept and in implementation but are not widely used in remote sensing work. In its simplest form, minimum distance classification is not always accurate there is no provision for accommodating differences in variation of classes, and some classes may overlap at their edges. It is possible to throw much sophisticated versions of the basic speak to just outlined by using different distance measures and different methods of defining cluster centroids.1Fig. 4. Minimum distance classifier1The Euclidean distance is the most common distance metric used in low dimensional data sets. It is also known as the L2 norm. The Euclidean distance is the usual manner in which distance is measured in real world. In this sense, Manhattan distance tends to be more robust to noisy data.Euclidean dista nce = (1)Wherex and y are m-dimensional vectors and denoted by x = (x1, x2, x3 xm) and y = (y1, y2, y3 ym) represent the m attribute values of two classes. 5. While Euclidean metric is useful in low dimensions, it doesnt work well in high dimensions and for categorical variables.Mahalanobis DistanceMahalanobis Distance is similar to Minimum Distance, except that the covariance ground substance is used in the par. Mahalanobis distance is a well-known statistical distance function. Here, a measure of variant can be incorporated into the distance metric directly. Mahalanobis distance is a distance measure between two points in the space defined by two or more correlated variables. That is to say, Mahalanobis distance takes the correlations within a data set between the variable into consideration. If there are two non-correlated variables, the Mahalanobis distance between the points of the variable in a 2D scatter plot is same as Euclidean distance. In mathematical terms, the Maha lanobis distance is compeer to the Euclidean distance when the covariance matrix is the unit matrix. This is incisively the case then if the two columns of the standardized data matrix are orthogonal. The Mahalanobis distance depends on the covariance matrix of the attribute and adequately accounts for the correlations. Here, the covariance matrix is use to correct the cause of cross-covariance between two components of random variable6, 7.D=(X-Mc)T (COVc)-1(X-Mc) ( 2)whereD = Mahalanobis Distance, c = a particular class, X = measurement vector of the aspect pixel Mc = mean vector of the signature of class c, Covc = covariance matrix of the pixels in the signature of class c, Covc-1 = inverse of Covc, T = transposition function3. level best Likelihood ClassificationIn nature the classes that we classify exhibit instinctive variation in their spectral patterns. Further variability is added by the effects of haze, topographic shadowing, system noise, and the effects of mixed pi xels. As a result, remote sensing images seldom record spectrally pure classes more typically, they display a range of brightnesss in each band. The classification strategies considered therefrom far do not consider variation that may be present within spectral categories and do not address problems that arise when frequency distributions of spectral values from separate categories overlap. The maximum likelihood (ML) procedure is the most common supervised method used with remote sensing. It can be set forth as a statistical approach to pattern recognition where the chance of a pixel sounding to each of a predefined set of classes is calculated hence the pixel is assigned to the class with the highest opportunity 4MLC is based on the Bayesian prospect formula.Bayes Classification The MLC decision rule is based on a normalized (Gaussian) estimate of the probability density function of each class 8. Hence, under this assumption and using the mean vector along with the covarian ce matrix, the distribution of a syndicate response pattern can be completely described 9. inclined these parameters, the statistical probability of a given pixel value can be computed for being a share of a particular class. The pixel would be assigned to the class with highest probability value or be labelled unknown if the probability values are all below a threshold set by the user 10.Let the spectral classes for an image be represented byi, i = 1, . . . MWhere, M is the fundamental number of classes. In order to determine the class to which a pixel vector x belongs the conditional probabilities of interest should be followed.P( ix), i = 1, . . . MThe measurement vector x is a column of Digital Numbers (DN) values for the pixel, where its dimension depends on the number of input bands. This vector describes the pixel as a point in multispectral space with co-ordinates defined by the DNs (Figure 2-20).Fig. 4.Feature space and how a feature vector is plotted in the feature spa ce 9The probability p(i x) gives the likelihood that the correct class is i for a pixel at position x. Classification is performed according tox i if p i x p j x) for all j i3i.e., the pixel at x belongs to class i if p(ix) is the puffyst. This general approach is called Bayes classification which works as an intuitive decision for the supreme Likelihood Classifier method 11.From this countersign one may ask how can the available p(xi) can be related from the training data set, to the want p(ix) and the answer is over again found in Bayes theorem 12.From this discussion one may ask how can the available p(xi) can be related from the training data set, to the desired p(ix) and the answer is again found in Bayes theorem 12.p (ix)= p (xi) p (i )/p(x) 4Wherep(i ) is the probability that class i occurs in the image and also called a priori or prior probabilities. And p(x) is the probability of conclusion a pixel from any class at location x.upper limit Likelihood decision rule i s based on the probability that a pixel belongs to a particular class. The basic equation assumes that these probabilities are equal for all classes, and that the input bands have normal distributions as in 13D = ln(ac)-0.5ln(Covc)-0.5(X-Mc)T(Cov-1)(X-Mc) 6WhereD = weighted distance (likelihood),c = a particular class,X = measurement vector of the candidate pixel, Mc =mean vector of the sample of class c,ac =percent probability that any candidate pixel is a instalment ofclass c,(Defaults to 1.0, or is entered from a priori knowledge),Covc = covariance matrix of the pixels in the sample of class c,Covc = deciding(prenominal) of Covariance (matrix algebra),Covc-1 = inverse of Covariance (matrix algebra) ln = natural logarithm function = transposition function (matrix algebra).4- comparing supervised classification techniquesOne of the most important keys to classify land use or land cover using suitable techniques the table showed advantages and disadvantages of each techniques 3 t echniquesadvantagedisadvantageParallelepipedFast and simple, calculations are make, thus cutting processingNot dependent on normal distributions.Since parallelepipeds have corners,pixels that are rattling quite far, spectrally, from the mean of the signature may be classifiedMinimum Distance ClassificationSince every pixel is spectrally immediate to either one sample mean or another, there are no unclassified pixels.Fastest decision rule to compute, except for parallelepipedPixels that should be unclassified,, this problem is alleviated by thresholding out the pixels that are far from the means of their classes.Does not consider class variabilityMahalanobis DistanceTakes the variability of classesinto account, unlike MinimumDistance or ParallelepipedTends to overclassify signatures with relatively large values in the covariance matrix.Slower to compute than Parallelepiped or Minimum DistanceMaximum LikelihoodMost accurate of the classifiersIn classification.Takes the variability of classesinto account by using the covariance matrix, as does Mahalanobis DistanceAn extensive equation that takes a long time to computeMaximum Likelihood is parametric, sum that it relies heavily on anormal distribution of the data in each input band5- trueness assessmentNo classification is complete until its verity has been assessed 10In this context the trueness means the level of conformity between labels assigned by the classifier and class allocation on the ground collected by the user as test data.To research valid conclusions about maps accuracy from some samples of the map the sample must be selected without bias. Failure to stomach these important criteria affects the validity of any further analysis performed using the data because the resulting computer error matrix may over- or under- estimate the true accuracy. The try schemes well determine the distribution of samples across the land scape which allow significantly affect accuracy assessment costs 14When perf orming accuracy assessment for the whole classified image, the known reference data should be another set of data. Different from the set that is used for training the classifier .If training samples as the reference data are used then the result of the accuracy assessment only indicates how the training samples are classified, but does not indicate how the classifier performs elsewhere in scene 10. the chase are two methods commonly used to do the accuracy assessment derived from table .1-the actus reus matrixTable 1.Error matrix15Error matrix (table1 ) is square ,with the same number of information classes that will be assessed as the row and column. Numbers in rows are the classification result and numbers in column are ref-erence data (ground truth ).in this square elements along the main diagonal are pixels that are correctly classified. Error matrix is very effective way to represent map accuracy in that individual accuracies of each category are plainly descried along with both the error of commission and error of omission. Error of commission is defined as including an area into acatogary when it does not belong to that category. Error of omission is defined as excluding that area from the catogary in which it authentically does belong. Every error is an omission from correct category and commission to a wrong category. With error matrix error of omission and commission can be shown clearly and also several accuracy indexes such as boilersuit accuracy, users accuracy and producers accuracy can be assessed .the following is expatiate description about the three accuracy indexes and their calculation methodoverall accuracyOverall accuracy is the portion of all reference pixels, which are classified correctly (in the scene) that assignment of the classifications and of the reference classification agree).it is computed by dividing the total number of correctly classified pixels (the sum of the elements along the main diagonal) by the total number of r eference pixels. According to the error matrix preceding(prenominal) the overall accuracy can be calculated as the followingOA == Overall accuracy is Avery coarse measurement. It gives no information about what classes are classified with good accuracy.producers accuracyproducer accuracy estimates the probability that a pixel, which is of class I in the reference classification is correctly classified . It is estimate with the reference pixels of class I divided by the pixels where classification and reference classification agree in class I . Given the error matrix above, the producers accuracy can be calculated using the following equationPA (class I) =Producer accuracy tells how well the classification agrees with reference classification2.3 users accuracyUsers accuracy is estimated by dividing the number of pixels of the classification results for class I with number of pixels that agree with the reference data in class I.it can be calculated as UA(class I)= Users accuracy pred icts the probability that a pixel classified as class I is actually belonging to class I.2-kappa statisticsThe kappa analysis is discrete multivariate techniques used in accuracy assessment for statistically determining if one error matrix is significantly different than another (bishop).the result of performing of kappa analysis is khat statistics (actually ,an estimate of kappa),which is an- other measure of agreement or accuracy this measure of agreement is based on the difference between the actual agreement in the error matrix(i.e the agreement between the remotely sensed classification and the reference data as indicated by major diagonal) and the chance agreement, which is indicated by the row and column totals(i.e marginal)16A detailed comparison between two data sets, one with near-infrared and three visible and the other with the full 8-bands, was made to emphasize the important role of the new bands for improving the separability measurement and the final classificatio n results 17