The GBA dataset for site layout generation

This dataset contains parcel–building samples from the Guangdong–Hong Kong–Macao Greater Bay Area, processed into graph-structured ".gpickle" files for the GCGAN model. For each land use parcel, the raw parcel polygon, building footprints, and corresponding raster masks are normalized and encoded as a ring-topology graph in which each building node stores its relative location, footprint size, height, and neighbor-based geometric features (radial distance, pairwise spacing, and local turning angle). A super-node represents the parcel center, and parcel-level attributes such as building count, plot ratio, land-use type, and spatial compactness are recorded as graph metadata. Each ".gpickle" file therefore provides a complete sample integrating (1) building geometry, (2) spatial relationships among buildings, and (3) parcel-level planning indicators, along with aligned parcel and building masks in a fixed-resolution grid, enabling deep generative models to learn realistic and attribute-controlled building layout patterns. For more details, please refer to our paper. (Jiang, M., Chen, Y., Liu, X. et al. Automated site layout generation for buildings using graph constrained generative adversarial network. Building Simulation. (2025). https://doi.org/10.1007/s12273-025-1350-7)