Semantic segmentation of building façade materials and colors for urban conservation

Building façade materials are fundamental to understanding cultural heritage and preserving the historic urban environment. While traditional material analysis requires extensive manual effort, this study introduces an automated methodology of material identification and feature extraction. We construct the Building Façade Material Segmentation (BFMS) database, which surpasses previous datasets in its alignment with architectural standards and the diversity of material categories included. Based on the database, we develop a transformer-based semantic segmentation model that achieves an overall accuracy of 76.5%. Furthermore, the algorithm to extract detailed textures and colors for each material is formulated with K-means clustering and grid statistics. Applied in the Taiping Alley, a historic area in Jingdezhen, the methodology demonstrates robustness and efficiency, yielding valuable guidance for architectural landscape assessment and urban renewal planning. The study promotes the application of cutting-edge deep learning algorithms in urban conservation and contributes to the broader understanding of urban material characteristics.