Impact on visual aspect and classification performance of 3D transform coding techniques

Three dimensional transform coding is widely diffused for the compression of multispectral (MS) remotely sensed (RS) images. Two steps are performed in sequence to achieve spectral and spatial decorrelation: first, original data is mapped in the Karhunen-Loeve transformed domain, then a common compression technique (e.g., DCT, VQ, Wavelets) is applied to the “principal components” to reduce spatial correlation. In this paper, on the basis of theoretic considerations and of a large number of tests carried out on MS image data (Airborne-TM) the authors show that the combination of KLT and standard JPEG, in front of an improvement by the visual standpoint can lower classification performance