Deep learning classification of column scraping photographs used in structural health assessment within the scope of transformation of areas under disaster risk plan

Located on active seismic fault lines, Türkiye is continuously exposed to significant earthquake risks. In the aftermath of the 1999 Marmara Earthquake, the structural integrity and resilience of the country's existing building stock became a major public concern, prompting the implementation of serious regulatory measures. In this context, Law No. 6306 on the Transformation of Areas Under Disaster Risk, which came into effect in 2012, established the legal and technical framework for the identification and renewal of seismically vulnerable buildings. Within the scope of this law, the Principles Regarding the Determination of Risky Structures (PRDRS) aim to evaluate the seismic performance of structures based on scientific and technical criteria. According to PRDRS, a risky structure is defined as a building likely to collapse or suffer severe damage during an earthquake due to material degradation, structural deficiencies, or design flaws. The assessment of such buildings is conducted by institutions licensed by the Directorate of Urban Transformation and involves several experimental and observational methods, including concrete core sampling, reinforcement scraping, Schmidt hammer testing, and rebar detection using cover meters. Among these methods, reinforcement scraping is a destructive inspection technique that enables the direct observation of the internal reinforcement configuration of reinforced concrete elements. This method reveals detailed information such as the stirrup type, diameter, spacing, and hook configuration in the confinement and mid-regions of columns and walls, as well as the longitudinal reinforcement layout and corrosion-related damage. The proper execution of this method is verified by engineers within the Directorate during the review of technical reports prepared by licensed organizations. However, conducting the reinforcement scraping procedure in full compliance with regulatory guidelines presents disadvantages in terms of both time and accuracy, and is prone to human error. In this thesis, the classification performance of deep learning-based convolutional neural network (CNN) models was investigated for the automatic evaluation of column scraping images. In particular, five pre-trained CNN architectures based on transfer learning—MobileNetV2, EfficientNetB0, ResNet50, DenseNet121, and InceptionV3—were trained using various hyperparameter configurations and evaluated comparatively. Each model was assessed using performance metrics such as test accuracy, precision, recall, and F1-score, and the most successful architecture was identified. The results demonstrate that a fast, accurate, and reliable artificial intelligence (AI) supported decision system can be developed as an alternative to traditional inspection methods. This thesis aims to contribute to the practical implementation of AI in the field of structural engineering