Scholarlink

<p>Developing artificial intelligence-based methods for automated site layout generation can significantly reduce the time and manual effort required in urban planning and design processes. To this end, generative adversarial networks (GANs) have been used in many applications. However, conventional GAN-based models have rarely considered the geometric relations between parcels and buildings, and the generated layouts often fail to meet the design requirements due to the poor integration of essential building attributes. To address these issues, this study proposes a model based on a graph constrained GAN (GCGAN), which consists of a graph variational autoencoder (GraphVAE) and a GAN framework. In this model, parcels are represented as tuples, while site layouts within each parcel are encoded as graphs with ring topology to capture spatial and relational structures. GraphVAE is then trained to generate site layout graphs considering parcel attributes and building design parameters (e.g., number of buildings). Furthermore, GAN is trained to generate the layouts of building objects according to the graphs produced by GraphVAE. The GCGAN model is evaluated with a dataset that comprises parcels and their corresponding site layouts in the Guangdong-Hong Kong-Macao Greater Bay Area in southern China. Comparative experiments reveal that GCGAN model outperforms other models such as GANmapper, Pix2Pix, and ESGAN in terms of more realistic building patterns and attributes. With its satisfactory performance, the proposed model has the potential to support the planning and design of urban (re)development by providing reliable simulations of site layouts.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Automated site layout generation for buildings using graph constrained generative adversarial network

  • 期刊论文

  • 发表日期：

    2025年10月9日

查看全文
  • Ming Jiang,
  • Yimin Chen,
  • Xiaoping Liu,
  • Jinding Gao

摘要

Developing artificial intelligence-based methods for automated site layout generation can significantly reduce the time and manual effort required in urban planning and design processes. To this end, generative adversarial networks (GANs) have been used in many applications. However, conventional GAN-based models have rarely considered the geometric relations between parcels and buildings, and the generated layouts often fail to meet the design requirements due to the poor integration of essential building attributes. To address these issues, this study proposes a model based on a graph constrained GAN (GCGAN), which consists of a graph variational autoencoder (GraphVAE) and a GAN framework. In this model, parcels are represented as tuples, while site layouts within each parcel are encoded as graphs with ring topology to capture spatial and relational structures. GraphVAE is then trained to generate site layout graphs considering parcel attributes and building design parameters (e.g., number of buildings). Furthermore, GAN is trained to generate the layouts of building objects according to the graphs produced by GraphVAE. The GCGAN model is evaluated with a dataset that comprises parcels and their corresponding site layouts in the Guangdong-Hong Kong-Macao Greater Bay Area in southern China. Comparative experiments reveal that GCGAN model outperforms other models such as GANmapper, Pix2Pix, and ESGAN in terms of more realistic building patterns and attributes. With its satisfactory performance, the proposed model has the potential to support the planning and design of urban (re)development by providing reliable simulations of site layouts.