As one of the main targets in cities， the extraction of buildings is of great importance， and using segmentation methods to separate buildings from the background is the basis for subsequent information extraction. However， traditional Markov Random Field （MRF） model only uses gray information when segmenting buildings in Synthetic Aperture Radar （SAR） images， so the segmentation integrity of gray inhomogeneous targets is poor， and the interrelationship between the two parts of random field energy is not considered， which leads to the results cannot balance regional consistency and edge detail. In order to solve these problems， an improved MRF model for building segmentation in SAR images is proposed. On the one hand， by introducing texture features weighted by the Bhattacharyya Distance into the observed random field， a complete extraction of gray inhomogeneous buildings is achieved； on the other hand， by introducing weights that vary with the number of iterations in the two parts of random field energy， a better noise suppression is achieved while keeping the edges smooth in the dense region， and finally more accurate building segmentation results are obtained. In order to verify the effectiveness and practicability of the proposed algorithm， real SAR images are selected for testing， and the results show that compared with current algorithms， the proposed algorithm has better classification accuracy and Dice coefficient.