The aim of this in this research work is to present an innovative approach to classification of satellite images based on Markov Random Field, MRF. Markov models are used both on single-band and multi-band images and have the advantage to take into account the spatial context in the process of the classification of multispectral images. This leds to the integration of interactions between different pixels and to extract the maximum information contained in satellite images including textures. In this research work, the classification by Markov Random Fields was applied respectively on the colored composites of the first three principal components of multispectral images Landsat TM from 1986 ,ETM + from 2003 and OLI from 2014 of the department of Sinfra containing respectively 94,7% , 97,4% et 98,4 % of the information. Markov Random Field correctly discriminate the different classes of land use with a Kappa coefficient higher than 0.8 : 0.86 for TM images, 0.91 for ETM + and 0.9 for OLI images.