Classification of Compressed Multichannel Images and Its Improvement

Abstract

A task of classification of multichannel remote sensing images compressed in a lossy manner is considered. It is recalled that lossy compression usually leads to reduction of classification accuracy both in aggregate and for particular classes. Distortions due to compression are characterized by visual quality metric desired values of which can be provided at compression stage. Dependence of probability of correct classification on image quality and compression ratio is analyzed for several widely used classifiers using a test image composed of three component images of Landsat data in visible range. It is shown that different classifiers are sensitive to distortions introduced by lossy compression in sufficiently different degree. We also propose a way to combine classifiers' outputs to improve classification results.