Research Information System
CONSTRUCTION AND BUILDING MATERIALS, vol.518, 2026 (SCI-Expanded, Scopus)
The accurate identification of timber species used in historic buildings is a crucial yet challenging step in conservation practice, as traditional sampling-based methods involve irreversible loss of original material. This study proposes a non-destructive approach to wood species identification for architectural heritage, with a focus on T & uuml;rkiye's Black Sea region, where timber plays a defining role in vernacular architecture. An original dataset comprising 2435 high-resolution images representing nine locally sourced species was developed under realistic field conditions. These images were captured without surface treatment or laboratory preparation, reflecting the aged and heterogeneous appearance of timber in historic structures. A multi-scale patch attention model based on convolutional neural networks (CNNs) was then trained and evaluated to classify species according to grain orientation and texture features. The model achieved an average accuracy of 96.9% and a Top-3 accuracy of 99.8%. This demonstrates robustness despite dataset imbalance and environmental variation. The framework was validated on the Go & uml;gceli Mosque, Samsun's oldest wooden mosque. It accurately identified oak as the primary species, consistent with dendrochronological data. Beyond high classification performance, the study bridges a critical gap between laboratory-based research and conservation fieldwork. It offers a region-specific, interpretable, and accessible tool for heritage practitioners. The proposed method enhances the efficiency and sustainability of conservation workflows by reducing reliance on destructive tests and laboratory facilities. Ultimately, this research establishes a methodological foundation for developing a national-scale timber dataset and advancing AI-assisted, non-destructive diagnostic practices in architectural conservation.