Digital deconstruction of architectural styles influenced by multiculturalism: A case study of Gulangyu

This study addresses the challenge of quantitatively analyzing hybrid architectural styles in multicultural historic districts. We propose an integrated "typology-AI" framework that combines a novel "L-V-F" style classification model—grounded in typological "prototype-variant" theory—with deep learning and geospatial analysis. A systematic evaluation of four deep neural networks (ConvNeXt, DenseNet, ResNet, Vision Transformer) identified the optimal model for feature extraction, enabling the deconstruction of hybrid styles into a quantifiable formula (V = Σa _i L _i +ΣbⱼFⱼ). Grad-CAM visualization was used to interpret model decisions. Furthermore, an entropy-based metric was introduced to quantify stylistic hybridity, and GIS was employed to map spatial patterns within building clusters. The developed digital deconstruction model provides a scalable, automated workflow from visual recognition to spatial mapping, offering a practical and theoretically grounded tool for the evidence-based conservation and management of complex architectural heritage. This work validates the potential of merging typological theory with AI to derive actionable insights for architectural heritage science.