An approach was proposed for building segmentation from high resolution SAR (Synthetic Aperture Radar) images based on an improved Markov random field (MRF) model.Aiming at the property of low SNR (Signal to Noise Ratio) of SAR images and the complexity of building textures

we obtained the initial segmentation using the maximum likelihood (ML) algorithm based on the multi-scale MRF model and involved the Gabor similarity between pixels based on the traditional MRF potential function

and employed the ICM (Iterative Conditional Model) algorithm to implement the segmentation.The experimental results on several real SAR images show that the proposed approach performs better than traditional methods in the segmentation accuracy

and building boundaries are clearly obtained by the proposed approach.