Compression and classification of noisy multichannel remote sensing images
Владимир Васильевич Лукин, Nikolay N. Ponomarenko, Alexander A. Zelensky, Andriy Kurekin, K.V. Lever · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2008
Remote sensing images are commonly formed on-board an observation platform, then transferred via a communication downlink, and finally processed on-land. There are many ways of compressing and then classifying remote sensing images. In this paper we focus on considering two lossy compression techniques under the assumption that the original images are noisy. No pre- or postprocessing is applied. Two classifiers are examined, namely, those based on trained radial basis function neural networks and support vector machines. We study how the parameter that controls the compression ratio of two coders based on the discrete cosine transform influences classification accuracy of these classifiers for a real life three-channel optical image. It is shown that attaining the optimal operation point for both coders is practically equivalent to providing the maximal probability of correct classification of multichannel data. At the same time, the efficiency of image compression characterized in terms of compression ratio, peak signal-to-noise ratio, and probability of correct classification considerably depends upon the coder used. Finally, it is shown that compressing multichannel remote sensing data in the neighborhood of the optimal operation point and near the maximum of the probability of correct classification can be performed in automatic manner.