Two-stage generative adversarial networks for urban pocket park design assistance: effects of training-set size and input augmentation

This study develops a two-stage image translation workflow for urban pocket park design assistance, linking rough line drawings, fine line drawings and spatial semantic layouts. To examine how training data and input variation affect design representation, a comparative framework is established using 1,028 spatially aligned image triplets. Four training-set sizes—180, 360, 540 and 720 triplets—and four input conditions are evaluated: original image pairs, geometric rotation augmentation, drawing-style perturbation and local information masking. Each setting is trained independently three times and assessed on a common test set using IoU, Dice, boundary F1 and supplementary image and road-representation metrics. Findings show: (1) increasing the number of original image pairs improves final layout reconstruction, with macro-IoU rising from 0.1222 to 0.1516 and Dice from 0.2053 to 0.2494; (2) augmentation provides task-specific benefits: geometric variation improves line-position agreement, while drawing-style perturbation improves selected road graphic properties, but these gains do not consistently translate into better overall layout reconstruction; and (3) comparing reference and automatically refined drawing inputs identifies distinct weaknesses in the transfer of planting information and the interpretation of water and structures. These findings support selecting augmentation according to the intended representation task and reviewing intermediate drawings alongside final layouts. The study provides an experimental basis for training-data organisation, augmentation selection and staged graphic review in generative design assistance for urban pocket parks.