Keywords: JPEG, Image Compression, Image Classification, Remote Sensing

Abstract
Image classification strategies based on the multi-spectral imagery from satellite-based remote sensing have been widely applied in the extraction and classification of land use patterns and land coverage. One of the important elements which will affect the results of classification is the image quality. Recently, due to the large compression ratio achieved by lossy image compression algorithms, such as JPEG, those image compression techniques have gained a lot of attention from the remote sensing discipline. However, as computer based image analysis tools are very sensitive to image quality, small changes in the image content may affect the analysis results. The results after image classification using the compressed multi-spectral images will also be affected due to the deterioration of image quality as a result of increased compression ratio.

This paper describes some experimental investigations into the effect of JPEG compression on the accuracy of image classification using multispectral SPOT image. MLC supervised classification strategy was used for classification with vary compression quality factors. As a result of the investigation, it was found that images with high classification accuracy can be compressed more without too much effects on the overall accuracy of the classification. It is possible to compress a satellite image using a q-factor value of 35 with less than 3% lost of the classification accuracy. However, more experiments need to be conducted using images from different scenes with different land use coverages to support the above remark.

Introduction
Remotely sensing imagery captured from satellites has become one of the most important sources of spatial data for Geographical Information Systems (GIS). Without going through sophisticated processing, most of these images can be used immediately as background images once they are geographically referenced. To extract more information such as digital terrain model (DTM), land use patterns and land coverage, and ground features from the images, sophisticated processes are required to perform the tasks. To extract DTM, a pair of satellite images with overlapping coverage but taken from different locations in space is required. To extract ground features accurately from a satellite image, high resolution satellite image together with the corresponding DTM are needed. For the extraction and classification of land use patterns and land coverage, multi-spectral satellite imagery is required. Image classification strategies based on the multi-spectral imagery from satellite-based remote sensing have been widely applied. These strategies could either be categorized as supervised which the a-prior training data are needed for the discipline-dependence results, or unsupervised which are self-trained for the statistical patterns from the images. One of the important elements which will affect the results of any of the above three processes is the image quality.

Recently, due to the large compression ratio achieved by lossy image compression algorithms, such as JPEG, those image compression techniques have gained a lot of attention from the remote sensing discipline. Most of these lossy compression techniques are mainly designed to exploit human vision system limitations. The image quality of the compressed image is definitely affected but it may not be visible or obvious when examines by human eyes. However, as computer based image analysis tools are very sensitive to image quality, small changes in the image content may affect the analysis results. The results after image classification using the compressed multi-spectral images will also be affected due to the deterioration of image quality as a result of increased compression ratio. The effects of image compression on DTM accuracy were studied by Lam K.W.K. et.al.(1999). It was found that a near linear fall-off in accuracy with decreasing q-factor values (reducing quality settings) from 95 up to the value of 30. Errors increase by 60% when q-factor value is 20. Algarni (1996) conducted an experiment using an unsupervised classification strategy, ISODATA, to study the effect of JPEG compression on the geometric and visual quality of the output of compression using three TM band from LANDSAT. He concluded that only 4.5% pixels in a compressed image were mis-classified but the result could be scene dependent. This paper is to discuss the experimental results conducted on the accuracy of classification using maximum likelihood classifier (MLC) supervised classification strategy using SPOT multispectral images with vary compression quality factors.

Compression Quality Setting in Jpeg
JPEG image compression standard was introduced and developed by the Joint Photographic Experts Group (JPEG) has been working under the auspices of three major international standards organizations - International Organization for Standardization (ISO), the International Telegraph and Telephone Consultative Committee (CCITT), and the International Electrotechical Commission (IEC) - for the purpose of development a standard for color image compression (Pennebake and Mitchell, 1993). This standard provides a guideline for the software developers to implement according to their own specifications. JPEG is "lossy" and is intended for compressing images with little or no obvious changes identifiable by humans. Due to this highly efficient and effective compression techniques, JPEG compression techniques and the JPEG formats have been widely accepted in the remote sensing discipline to optimize data storage and to reduce data transmission time. With the recent announcement by Space Imaging of Denver, Colo. of the successful launch of the world’s first commercial, high-resolution imaging satellite, the importance of image compression for the storage and transmission of remotely sensed images become even more important.

Review of the JPEG compression standard would not be reported in this paper. Comprehensive discussion of the standard can be found in the book written by Pennebaker and Mitchell (1993) and the white papers by Lane (1999). A brief overview can also be found in Lam K.W.K. et. al. (1999). However, it is worth to note that JPEG is more than an algorithm for compression images but also architecture for a set of image compression functions suitable for a wide range of applications involving image compression. A useful property of the architecture of JPEG is to allow adjusting compression parameters (quality factor) to increase the degree of lossiness. Hence, one can trade off file size against image quality. However, as JPEG standard does not specify how quality scales should be implemented, different software developers may use different scales to control the quality factor (q-factor). The scale between 0-100 is the widely accepted scale range but the scale range has nothing to do with percentage of information to be kept in an image.