Generating parcel-level land use on a cadastral graph: masked discrete diffusion for station areas in Tokyo

Planning and survey practice often needs plausible land-use configurations for an area whose surroundings are known. Generative land-use models work on raster grids or point-of-interest counts, and cellular-automata allocators are calibrated to reproduce a single map. This study treats the task as inpainting on the cadastral graph. Using the 2021 land-use survey of Tokyo's 23 special wards (815,736 parcels, ten classes), we mask all parcels within 500 m of a railway station and train masked discrete diffusion models with graph neural denoisers to regenerate them, in a planning scenario that sees only parcel layout and surrounding use and an update scenario that also sees built form. On 203 station areas in two held-out spatial folds, the update-scenario point estimate reaches parcel accuracies of 0.663 and 0.644 (area-weighted 0.789 and 0.787), against 0.468 and 0.453 for the strongest of four baselines, and reproduces block-level composition (R² 0.811 and 0.859). In the planning scenario point estimates collapse onto dominant classes (R² 0.352 and 0.395), whereas stochastic samples retain composition (0.626 and 0.660) and adjacency structure. Accuracy and pattern realism therefore require different outputs from the same model. Remaining errors concentrate in classes defined by activity inside buildings.