Deep learning-based in-situ identification of coniferous wood components in heritage architectures of China

Identification of structural wooden components is crucial for heritage architecture conservation as it elucidates the utilization patterns of forest resources and evolution of human civilization. This paper proposes a computer vision-based in situ identification method for wooden components of Chinese heritage architectures, using a dataset comprising 4050 images from 63 components of nine buildings. The optimal algorithm, RepLKNet, trained on coniferous xylarium specimens, achieves 96.67% identification accuracy, with 98.33%, 93.33%, and 90% precision at 50%, 70%, and 90% confidence thresholds, respectively. A minimum sample size of 25 species and 1500 images per genus ensures test accuracy >90%. Impact of structural deterioration (decay and cracks) on accuracy is also evaluated. Cracks significantly affect the wood recognition accuracy of historical components. Performance degrades significantly when cracks span >30% of the image. Latewood integrity is also critical to identification. The proposed method advances structural preservation strategies and preventive maintenance practices in heritage architecture.