Multiscale Quantification of Visual Morphology in Modern Historic Buildings: A Computer Vision Study of the Five Former Concessions in Hankou

Abstract

Modern historic buildings encode exchanges among social development, regional culture, and construction technology. The surviving buildings in the five former concessions of Hankou display diverse facade compositions and window, door, and column forms, but conventional typological analysis depends heavily on expert judgement and is difficult to scale. This study developed a traceable three-level Building–Facade–Component dataset comprising 59 buildings, 77 facades, and 2456 components. At the facade scale, 18 OpenCV-derived morphological descriptors were evaluated first at the facade level and subsequently after multiple facades were aggregated within each Building ID. Building-level permutation tests with false-discovery-rate correction showed that none of the 18 descriptors provided stable evidence for concession-group differences. At the component scale, frozen ResNet50 representations were reduced using principal component analysis and clustered using K-means, while UMAP was used only for exploratory visualization; an archived DINOv2 feature matrix was examined as a secondary representation-sensitivity comparison. The selected partitions yielded three window, two door, and two column cluster-informed visual groupings without component-level reassignment. Independent blinded semantic evaluation showed limited concordance for the three window descriptions (53.4%; κ = 0.296), but stronger concordance for the door (87.3%; κ = 0.750) and column groupings (91.3%; κ = 0.825). Component-level frequencies showed descriptive variation among the sampled concessions; however, building-level composition tests and building-clustered permutation analyses did not support stable concession-related associations. The framework therefore provides complementary, scale-dependent evidence for structured heritage documentation and exploratory morphological comparison while preserving traceability to the original buildings, facades, and component images.