Precise and up-to-date spatial information from natural and artificial phenomena are needed for better management of urban and semi-urban areas. For this purpose, extraction of spatial information from buildings, as one of the dominant urban features, has attracted the attention of photogrammetry and remote sensing specialists. Thus achieving an appropriate algorithm for detection and extraction of buildings from aerial and satellite images is crucially important. In this paper, therefore, it has been attempted to propose a method for improving detection of buildings in TerraSAR-X images. In the proposed method, applying integration of classifiers and employing neighbourhood information of each pixel have improved the results of building detection and reduced the amount of noise. Therefore, in this method, initially extraction and selection of the optimum feature is carried out. Then, the capability of various classifiers like Neural Network, Support Vector Machine, Maximum Likelihood and Nearest Neighbour is assessed and compared. Then, for improving the results, the results are integrated in decision level using a moving window. The results of the proposed method with overall accuracy and building detection accuracy of 86.41% and 73.08%, respectively indicates the capability of this method. Also, the proposed integration method has improved the results at least by 5% compared to that of individual usage of each classifier.