Ensemble learning-based hierarchical retrieval of similar cases for site planning

Abstract

Abstract Site planning not only involves the arrangement of a large number of elements but also considers the complexity of urban systems; hence, the design process involves large workloads and is time consuming. Retrieving similar existing cases helps architects optimize or accelerate the design process. This paper proposes a computational approach that provides valuable references by retrieving similar cases. Three types of attributes are extracted to represent a given parcel: land-use attributes, geometric attributes (area, orientation, etc.), and environmental attributes (greening rate, surrounding facilities, etc.). The complete hierarchical retrieval process is divided into three phases. The first phase selects cases whose land-use attributes are consistent with the target parcel. Then, the similarity distances between the given target parcel and the selected cases are calculated using geometric attributes. The eXtreme Gradient Boosting (XGBoost) classifier is adopted to determine which case is similar to the target parcel. Finally, similarity scores of the retrieved cases are calculated based on the environmental attributes to provide more options during the actual design. In total, 1189 cases with different land-use types in Beijing were collected for the case base. The comparative experimental results confirmed that the proposed ensemble learning-based hierarchical retrieval of similar cases approach improves the accuracy of retrieval results. Furthermore, we use a real-world target parcel to demonstrate the superiority and flexibility of the retrieval process.