The Effect of Lossy Image Compression on Image Classification

The work summarized here was presented at the 1995 IEEE Geoscience And Remote Sensing Symposium in Florence, Italy. This work was supported by a NASA High Performance Computing and Communications grant and also by the Research Institute for Advanced Computer Science (RIACS) at NASA Ames Research Center, Moffet Field, Ca.

The full text can be found in:

++ Paola, J. D., and Schowengerdt, R. A., The Effect of Lossy Image Compression on Image Classification, Proceedings, 15th Annual International Geoscience and Remote Sensing Symposium, Florence, Italy, July 10-14, 1995, pp. 118-120.

Abstract

We have classified four different images, under various levels of JPEG compression, using the following classification algorithms: minimum-distance, maximum- likelihood, and neural network. The training site accuracy and percent difference from the original classification were tabulated for each image compression level, with maximum- likelihood showing the poorest results. In general, as compression ratio increased, the classification retained its overall appearance, but much of the pixel-to-pixel detail was eliminated. We also examined the effect of compression on spatial pattern detection using a neural network.

Experiment

With remote sensing studies becoming more global in nature, and computer processing power increasing, many scientists have been turning to larger and larger data sets. Unfortunately, storage of enormous data sets can be costly, thus making image compression an important consideration in the remote sensing field. For typical earth science imagery, lossless compression will result in about a 2:1 reduction. Lossy compression methods, however, commonly provide 10:1, 20:1, or even higher ratios, while maintaining the visual integrity of the image. The effect of these algorithms on supervised classification is important to consider before any data is archived with lossy compression. In this experiment we have compressed four remotely-sensed multispectral images to varying degrees using JPEG, and have investigated the resulting supervised classifications obtained by the minimum-distance (MD), maximum-likelihood (ML), and three-layer backpropagation neural network classifiers. We have also looked at the effect of compression on spatial pattern detection using a neural network. The four classifications are:

An urban land-use classification of Landsat Thematic Mapper (TM) satellite imagery of Tucson, Arizona, obtained April 1st, 1987.
An urban land-use classification of TM imagery of Oakland, California, obtained August 12th, 1983.
A geologic classification of Airborne Visible Near Infrared Imaging Spectrometer (AVIRIS) aircraft imagery of the Lunar Lake Volcanic Field in central Nevada, obtained September 29, 1989.
A combined temporal AVHRR NDVI (11 bi-weekly composites), spectral AVHRR, and DEM land cover classification of central California, using imagery from January to July, 1992.

Conclusion

Overall, it was found that high quality classifications could be obtained with any of the classifiers for JPEG compression ratios approaching 10:1 or even higher. Qualitatively, the classification retains its overall appearance, but the smoothing effect of high compression tends to eliminate much of the pixel-to-pixel detail. As expected, training on the compressed imagery could raise the training site accuracy, but did not raise the percentage of pixels matching the original classification. The maximum-likelihood classifier degraded faster than the other two because of its specific assumptions about class statistics. The data tends to become uncorrellated with high compression, thus ruining its statistics.

This page created August 4, 1995 by Justin D. Paola

Return to DIAL

home page.

Digital Image Analysis Laboratory
Department of Electrical and Computer Engineering,
University of Arizona,
Tucson, Az 85721
(520) 621-4554