Abstract
Migrant architecture reflects selective continuity and reconfiguration across social contexts, but visual form alone cannot establish cultural origin or adaptation. This study develops an auditable human–AI framework for cross-domain architectural-component analysis and demonstrates it on Chinese diaspora buildings in Myanmar. A source corpus of 410 Minnan images with 5110 oriented bounding boxes supported source-only model development, while the target corpus contained 233 photographs from nine Myanmar sites. On a frozen 35-image, 271-component unseen-photo holdout from seven overlapping sites, YOLO11m-P2+C2PSA achieved 45.1% precision, 32.1% recall and 37.5% F1, whereas YOLO26m-P2+C2PSA achieved 56.3%, 18.1% and 27.4%, respectively. A prediction-blind second-reader check on seven selected images yielded 92.8% agreement F1 and median matched IoU of 0.774. Human verification and image-plane analysis identified selective continuity and contrasting component configurations; documentary evidence supported social–institutional embedding more strongly than Myanmar-derived material–morphological adaptation. The framework provides a traceable evidence chain in which detector outputs guide inspection without replacing human verification or independent historical evidence. The Myanmar application is an exploratory case study rather than a population-level estimate.
Funding Information
  • National Natural Science Foundation of China (52578022)
  • Fujian Provincial Department of Science and Technology (2025J01163)
  • Huaqiao University’s Academic Project

This publication has 43 references indexed in Scilit: