Extracting building patterns with multilevel graph partition and building grouping

Building patterns are crucial for urban landscape evaluation, social analyses and multiscale spatial data automatic production. Although many studies have been conducted, there is still lack of satisfying results due to the incomplete typology of building patterns and the ineffective extraction methods. This study aims at providing a typology with four types of building patterns (e.g., collinear patterns, curvilinear patterns, parallel and perpendicular groups, and grid patterns) and presenting four integrated strategies for extracting these patterns effectively and efficiently. First, the multilevel graph partition method is utilized to generate globally optimal building clusters considering area, shape and visual distance similarities. In this step, the weights of similarity measurements are automatically estimated using Relief-F algorithm instead of manual selection, thus building clusters with high quality can be obtained. Second, based on the clusters produced in the first step, the extraction strategies group the buildings from each cluster into patterns according to the criteria of proximity, continuity and directionality. The proposed methods are tested using three datasets. The experimental results indicate that the proposed methods can produce satisfying results, and demonstrate that the F-Histogram model is better than the two widely used models (i.e., centroid model and the Voronoi graph) to represent relative directions for building patterns extraction.