A noise-robust SVD-ML based classification method for multi-spectral remote sensing images

Research on remote sensing image classification approaches has gained momentum during the past decades, especially since the availability of high resolution and multi/hyper-spectral imagery capabilities. The Maximum Likelihood (ML) based classification methods have been also extensively studied and their effectiveness in this area is clear for researchers. The presented paper is dealing with a novel ML-based multi-spectral classification method which is robustly developed to resist against the huge amounts of additive noises which may infect the remotely sensed data. The proposed approach inventively utilizes a well-organized Singular Value Decomposition (SVD) denoising schema for reducing the effect of noise from the received multi-band data and then the ML classifier is applied to the noise-reduced dataset. Indeed since obtaining adequate number of training samples may be costly or even impossible in many cases and because of the Hughes phenomenon, we also propose to apply PCA for favorably reducing the number of bands to gain a more accurate performance in cases in which few limited training samples are available. The achieved results clearly imply the effectiveness and prominence of the proposed method for multi-spectral image classification especially in the noisy environments as well as in conditions where small amount of training samples is available.