Classification of Pre-Filtered Multichannel Remote Sensing Images

Abstract

It is demonstrated that in most modern applications of multichannel RS noise characteristics deviate from conventional assumption to be additive and i.i.d. Thus, filtering techniques are to be adapted to more sophisticated real-life models. This especially relates to multichannel radar imaging for which it is possible to gain considerably higher efficiency of denoising by taking into account spatial correlation of noise and sufficient correlation of information in component images. New approaches that take into account aforementioned properties are proposed and tested for real life data. It is also shown that filtering is expedient for RS images contaminated by considerably less intensive noise than in radar imaging. Even if noise is practically not seen (noticeable by visual inspection) in original images, its removal by efficient filters can lead to increase of data classification accuracy.