Architectural Style Recognition Using Multisource Data and Deep Learning: A Case of Historical Districts in Shanghai

Architectural style plays a crucial role in urban heritage preservation and urban landscape management. However, the diverse architectural styles shaped by the fusion of Eastern and Western cultures in Shanghai pose challenges for style classification. This study proposes an identification framework that integrates multi-source data and CNNs. First, 11 primary architectural styles in Shanghai were identified based on a review of classical literature. A cross-cultural architectural style dataset comprising 4,584 expert-verified images was constructed using web and social media data, covering 11 categories, such as Lilong-style Houses and Art Deco. A comparative analysis of six CNN architectures, including VGG16, ResNet50, and ConvNeXt-Tiny, revealed that ConvNeXt-Tiny achieved the highest accuracy (84%), outperforming the baseline model by 5.13%. An architectural style map was generated using ConvNeXt-Tiny and 10,313 street view images(SVIs) collected from 12 Historical Districts. This study effectively addresses cross-cultural architectural style recognition and provides insights applicable to other cities.