Preservation and restoration of traditional residential buildings in historical and cultural villages using deep learning and digital twin technology

Preservation and restoration of traditional residential buildings in historical and cultural villages require accurate digital representation of architectural elements and their spatial relationships. To address this need, a deep learning–assisted heritage digital twin framework is was developed using structured point cloud data containing spatial coordinates, geometric descriptors, color attributes, and acquisition metadata, collected through Unmanned Aerial Vehicle (UAV) photogrammetry and Terrestrial Laser Scanning (TLS). Z-score normalization is employed during data preprocessing to standardize heterogeneous numerical features, ensuring balanced model learning. Principal Component Analysis (PCA) is applied for feature extraction to reduce dimensionality while retaining dominant structural and contextual information relevant to heritage elements. A Dynamic Differential Annealed Adaptive Graph Neural Network (DDA-AGNN) is employed for the preservation and restoration of traditional residential buildings. An AGNN is designed to model spatial proximity and relational dependencies among architectural components, enabling accurate identification of heritage elements required for restoration-oriented BIM mapping. To enhance learning stability and parameter selection, DDA Optimization is integrated to adaptively optimize AGNN parameters and avoid premature convergence. The proposed DDA-AGNN achieves a mean Intersection over Union (mIoU) of 95.63%, demonstrating superior capability in capturing structurally consistent heritage representations. All experiments are implemented using Python, demonstrating facilitating the framework’s effectiveness for automated, accurate, and immersive heritage conservation. This holds significant importance for promoting the conservation and restoration of the historical and cultural village environments.