A national-scale deep-learning classification of apartment layouts and shapes in Korea (1970–2023)

This study presents a morphological analysis of South Korean apartment complexes using deep learning to classify building layouts and shapes across 13,424 complexes (1970–2023). This study developed consolidated typological frameworks: three layout types (Lattice, Right, Distributed) and five shape forms (Rectangle, L-shape, Y-shape, Polygon, Obtuse). YOLOv8-based models achieved 0.87 accuracy for layout classification and 0.917 mAP@0.5 for shape detection. National-scale analysis revealed Distributed layouts as most prevalent (40.6%), followed by Right (38.7%) and Lattice (20.8%). Rectangle forms dominated (68.7%), while L-shape comprised 16.0%. Temporal analysis demonstrated dramatic morphological transitions: Lattice predominated until 1990 (45%–49%), Right dominated 1990s–2000s (48%–66%), and Distributed became prevalent post-2010 (74%–78%). Non-rectangular forms accelerated from 10% (pre-2000) to 70.8% (2020s). Cross-tabulation analysis identified significant associations: large-scale developments (>500 households) preferred Distributed layouts (44%–52%), while tower-type complexes exhibited 90.6% Distributed prevalence. Underground parking strongly correlated with Distributed arrangements (78.5%), contrasting surface parking preference for Lattice (44.8%). These findings provide empirical evidence for residential urban form evolution and inform housing design guidelines and development standards.