n the recent years use of remote sensingsatellite data for urban planning, military,weather forecast, robotics, automatednavigation system, remote surveillance hasincreased by many fold in addition toconventional applications such as naturalresources management. These applications involve acquisition, communication, storage and processing ofterrible number of images of earth surface.This situation is becoming more aggravatedbecause of increased pixel resolution, graylevel resolution, band resolutions and reducedrepetition cycle of satellite. All of these development demands more band width fordownlink lines of satellite in addition to moredisk space for storage.

In communications, data compression techniques under the name hood of imagecoding are widely used to reduce thecommunication bandwidth bottlenecks duringdata communication. For instance, JPEGstandard is used for still image compression [1], MPEG is used for video compression [2]. Also, while communicating data from satellites toground stations some compression methods areused [3].

A typical image processing system isas shown in Figure 1 that is commonlyemployed for remote sensing applications. It isvery common that most of the applicationscientists using original image data for theirprocessing. In majority of remote sensing applications, results of classification are theultimate interest [4].

In this study, we propose to study howthe classification results will vary if we usecompressed image data instead of originalimage data. Usually applications such as landuse classifications assumes samples of a groupwill be having small random variations in theirpixel values while samples of different groupsto be having contrastingly different pixelvalues. Because of the increased pixel and graylevel resolutions, samples of a group may behaving similar pixel values. Moreover, they willbe having high level of spatial auto correlation.Evidently, majority of compression methods exploits this auto correlation to achieve highcompression ratios with acceptable PSNR (PeakSignal to Noise Ratio) values [5].

Our proposed Algorithm is based onfiltering concept [6]. In this Algorithm insteadof sending the original image, we send thefiltered image. In general, the number ofuseful DCT coefficients will become more if 8x8image blocks contain lot of variations in values,otherwise only few DCT coefficients will bemeaningful. If we apply filtering on image itgets smoothened, that is variation of the pixelvalues of a block reduces. It is attractive forpoint of view of compression as the number ofmeaningful DCT coefficients are going toreduce. Thus we are achieving compression benefit. We have compared the compression performance of our algorithm with conventional JPEG algorithm with variety ofmulti band images.

Also in this study, we evaluate theclassification performance of popularclassification algorithms like Maximum Likelihood, Mahalanobis and Euclidian distanceby taking original image data, conventional Our paper work is organized as follows.Section II introduces the standard JPEGalgorithm. Section III explains proposedcompression algorithm. Section IV illustratesthe selected classification algorithms. Section Vincludes details about our experimentationsand results. Section VI contains conclusionsabout our research work II.

JPEG is a well known standardized image compression technique. JPEG loses informationso the decompressed picture is not the same asthe original one. The main reason for use ofJPEG is to reduce the size of image files.Reducing image files is an important procedurefor transmitting files across networks orarchiving libraries. Usually JPEG can removethe less important data before thecompression; hence JPEG is able to compressimages meaningfully, which produces a hugedifference in the transmission time and the diskspace. Figure 2 shows the basic Architecture ofJPEG compression system. Here is a briefoverview of the JPEG compression system. [5] The image is first subdivided into pixelblocks of size 8X8, which are processed left toright, top to bottom. As each 8X8 block or subimage is encountered, its 64 pixels are levelshifted by subtracting the quantity L/2, whereL is the Gray level resolution of the image . The2-D Forward Discrete Cosine Transform (FDCT)(Eq-1) [5]of the block is then computed,quantized using 64 corresponding step sizevalues from the quantization table in Figure 3 [7]. After quantization the DCT coefficientsare rearranged in a zigzag sequence order as shown in the Figure 4. [7] Since the one-dimensional reorderedarray generated under the zigzag pattern ofFigure 3 is qualitatively arranged according toincreasing spatial frequency, the JPEG codingprocedure is designed to take the advantage ofthe long runs of zeros that normally result fromthe reordering. In particular, the nonzero ACcoefficients (the term AC denotes all transformcoefficients with the exception of the zeroth orDC coefficient) are coded using a variablelengthcode that defines the coefficient's valueand number of preceding zeros. The DCcoefficient is difference coded relative to theDC coefficient of the previous sub image. The decompression process performs aninverse procedure. It decodes the Huffmancodes. Then, it makes the inversion of theQuantization step. In this stage, the decoderraises the small numbers by multiplying themby the quantization coefficients. The resultsare not accurate, but they are close to theoriginal numbers of the DCT coefficients. AnInverse Discrete Cosine Transform (IDCT) (Eq.4) [7] is performed on the data received from theprevious step. Finally add L/2 to each subimage. Place the sub images in their correctpositions. (4) The error between the original imageand reconstructed image is calculated in termsof Peak signal to noise ratio (PSNR) = 10 log10 Mean filtering [8] is a simple, intuitiveand easy to implement method of imagesmoothing i.e. reducing the amount of variationbetween one pixel and the next or surroundingpixels. It is often used to reduce noise inimage. The idea of mean filtering is simply toreplace each pixel in an image with the meanvalue of its neighbors including itself. This hasthe effect of eliminating pixel values which areunrepresentative of their surroundings. Usually,3x3 neighborhoods of pixels are consideredwhile calculating mean filtered value of anypixel.

Median filter [9] is normally used to reducenoise in an image like the mean filter.However, it often does a better job than themean filter in preserving useful detail in theimage. Like the mean filter, the median filterconsiders each pixel in the image in turn andlooks at its neighbors to decide whether or notits representative of its surroundings. Insteadof simply replacing the pixel value with themean of neighboring pixel values, it replaces itwith the median of those values.

An outlier [10] is an observation that isnumerically distant from the rest of the data.In an image, a pixel value is very different fromits surrounding pixels, it can be called asoutlier.

From basic statistics, we know that apopulations sample values with someconfidence level can be given as mean ± C*P, where C is weighing factor (critical value) and P is standard deviation of the population. Table-1 shows the commonly used ConfidenceLevels and Corresponding Critical Values [11].In our outlier based algorithms, we take thesesimple confidence limits of normal distributionin deciding whether a pixel is outlier or not. Ifthe pixel is observed to be outlier with thegiven confidence level, we may retain else wemay take its mean filtered or median filtered value. ? Apply mean filtering with a little variation tothe given original image using 3x3 window. Foreach pixel, calculate average and standarddeviation of its neighboring 3x3 pixels. If apixels value is observed to be outlier (not inthe range of Mean ± C*P) then its value is takenas itself else mean is taken as its filtered value. ? Apply DCT on the outliermean filteredimage.

? Apply median filtering with a little variationto the given original image using 3x3 window.For each pixel, calculate median and standarddeviation of its neighboring 3x3 pixels. If apixels value is observed to be outlier (not inthe range of Median ± C*P) then its value istaken as itself else median is taken as itsfiltered value. ? Apply DCT on the outliermedian filteredimage.

Let w 1 ,w 2 , . . . , w m denote m distinctpopulations (classes) with known d-dimensionalprobability density functions p 1 (X), p 2 (X), . . . .. p m (X), respectively. The a priori probabilitiesthat an observation is selected frompopulations w 1 , w 2 , . . . ,w m are denoted by q l ,q 2 , . . . ., q m , respectively [12]. According tothe Bayesian ML classification rule, assumingequal costs for misclassifications, a random ddimensional pixel vector X is classified as classw k (7) Assuming equal a priori probabilities for all theclasses, decision rule (7) becomes:

In Equations ( 7) and (8), the probability densityp k (X) will be given as:

According to this classifier a ddimensionalrandom pixel vector (X) will be assigned to thegroup to which it is nearest [13]. Each group ischaracterized by its mean vector, which iscalculated from training data. Nearness isdetermined by the Mahalanobis distancebetween the group mean and X. Inmathematical terms, the same classificationrule can be represented as: (10) where i = (1, 2, . . . C) groups if d i (X)<d j ( X) forall j U i where (11) and M i s mean vector of i th group indicatesvector should be transposed.

-1 is the inverseof the pooled covariance matrix.

According to this classifier a random ddimensionalpixel vector (X) will be assigned tothe group to which it is nearest [14]. Eachgroup is characterized by its mean vector,which is calculated from training data.Nearness is determined by the Euclideandistance between the group mean and X. Inmathematical terms, the same classificationrule can be represented as: (12) Where i = (1, 2, . . . C) groups if d i (X)<d j (X) forall j U i where (13) and M i is mean vector of i th group. T indicatesvector should be transposed V.

For the purpose of experimental work,Landsat TM data from USGS data base"www.usgs.gov" is used. Confusion matrix [15] is used to assessthe accuracy of an image classification. Thestrength of a confusion matrix is that it identifiesthe nature of the classification errors, as well astheir quantities. In confusion matrix rowscorrespond to classes in the test set, columnscorrespond to classes in the classification result.The diagonal elements in the matrix representthe number of correctly classified pixels of eachclass. The off-diagonal elements representmisclassified pixels. The overall accuracy iscalculated as the total number of correctlyclassified pixels divided by the total number of test pixels.

Another measure which can be extractedfrom a confusion matrix is the kappa coefficient [16] which is a popular measure to estimateagreement in categorical data. The motivation ofthis measure is to extract from the correctlyclassified percentage the actual percentageexpected by chance. Thus, this coefficient iscalculated as (14) cm1,, cm2, ??? cmnare the column 1, 2 ????n marginals rm 1 ,rm 2 , ??? rm n are the row 1, 2 ---------nmarginals n is the total number of test pixels.

The higher the value of kappa, the betterthe classification performance. If all informationclasses are correctly identified, kappa takes thevalue 1. As the offdiagonal entries increase, thevalue of kappa decreases. For classification two data sets were used. Onewas a 1048 X 920 (Owensvalley image) LandsatTM with all 7 bands. The second data setcontained 500 samples of four ground types,Dense scrub, Rock, Forest and Open scrub of thesame scene. This second data set is used to observe the classification accuracy. All the 2000set patterns were classified simultaneously withthe Maximum Likelihood, Mahalanobis andMinimum distance classifiers. Classificationperformance of all the classifiers is displayed intables 4, 5 & 6. Figure 6 shows the Overallaccuracy of all the classifiers. It is observed thatclassification performance on proposedcompression images is almost same as JPEGstandard compression images and original images.