Effects of JPEG compression on image classfication

With the improvement of spatial resolution, data volume has become an increasingly significant concern, as the shear volume of data is expensive and inefficient in terms of data transmission, processing and storage. As a result, many image compression methods are in use. JPEG is one of the most popular methods. Indeed JPEG has become an industrial standard and has been implemented in many remote sensing image processing systems. This paper aims to experimentally evaluate the effects of JPEG compression on image classification. A scene of SPOT multispectral images was used. The image was compressed by JPEG at various compression levels (or using compression quality factors). All the compressed images are classified using the maximum likelihood classifier (MLC) of supervised classification and ISODATA of unsupervised classification. The classified result using the original (uncompressed) image was used as the benchmark. From the results, it can be found that there could be a significant decrease in image quality when compression is over 35-fold. As a result, the accuracy of image classification is dramatically deteriorated. However, when the compression ratio is smaller than 35-fold, the deterioration of classification accuracy is linear.