PSI - Issue 78

Livio Pedone et al. / Procedia Structural Integrity 78 (2026) 1609–1616

1612

In the first phase of the research, a new dataset consisting of open-access RGB images sourced from di ff erent databases / repositories has been created. The main sources include: (i) the Pacific Earthquake Engineering Research (PEER) Hub ImageNet ( Φ -Net) (Gao and Mosalam, 2020); (ii) the INGV (Italian National Institute of Geophysics and Volcanology) Macroseismic Photographic Database (DFM); (iii) reports from ReLUIS; (iv) available documentation for the Turkey Earthquake 2023 survey; and (v) images from the scientific literature. A total of 4,705 RGB images were selected, and each image has been resized to a uniform resolution of 224 × 224 pixels, maintaining a clear focus on the observed damage to ensure clarity for model training. Moreover, to ensure label accuracy, a rigorous re-labeling process has been conducted by expert structural engineers from our institution: in line with the damage classification criteria adopted in Italy (i.e., the AeDES form; Baggio et al. (2007)), all images have been categorized into four damage levels, namely “Heavy” (H), “Moderate” (M), “Slight” (S), and “Undamaged” (U). The dataset was then split into 3,505 images for training and 1,200 for testing (300 images per damage category). Finally, data augmentation techniques - including horizontal flipping, rotational transformations, and Gaussian noise injection - have also been used to further expand the training set, leading to a final dataset of 19,312 augmented training images. Twelvedi ff erent CNN models have been implemented, and the best performing was the VGG16 model (Simonyan and Zisserman, 2015). This CNN architecture consists of 16 layers, including 13 convolutional layers and 3 fully connected layers. To improve the e ff ectiveness of the model, a transfer learning approach has also been adopted by using the pre-trained version of the VGG16 network (Ebenezer et al., 2021). More details on the training strategy are available in Saquella et al. (2025). Di ff erent widely adopted metrics have been considered to evaluate the accuracy of the numerical results: precision, recall, F1-score, and their weighted averages. The use of di ff erent training sets has also been investigated. An example of results in terms of confusion matrices for two di ff erent datasets is shown in Fig. 2, where “Training set 5” refers to the one enhanced through data augmentation techniques. The results demonstrate the e ff ectiveness of the implemented data augmentation technique. The VGG16 showed an overall accuracy of 89.33% with “Training set 5” (best performing training set). However, some di ffi culties in the damage classification process can be noted, especially in distinguishing between “Slight” and “Moderate” damage. This output would suggest that the uncertainties related to the CNN-based damage classification should be taken into account in the seismic residual capacity assessment.

Original training set

Training set 5

242 11 44 3

268 13 18 1

Undam.

Undam.

200

200

6 226 65 3

15 254 28 3

Slight

Slight

4 2 279 15

4 14 267 15

Moderate

Moderate

100

100

True label

True label

0 1 11 288

0 4 13 283

Heavy

Heavy

0

0

.

.

te

te

Slight

Slight

Heavy

Heavy

Undam

Undam

Modera

Modera

Predicted label

Predicted label

(a)

(b)

Fig. 2: Results in terms of test set confusion matrices using (a) the original training set, and (b) “Training set 5”, i.e., involving data augmentation.

4. Illustrative application

4.1. Description of the case-study RC building

To implement the study, a 3-story RC building is selected. Global dimensions and plan geometry are shown in Fig. 3a. The structural skeleton consists of moment-resistant three-bay frames in one direction and moment-resistant two-bay frames in the orthogonal direction. The case-study is assumed to be located in a high seismicity zone in Italy (B soil type; Peak Ground Acceleration PGA = 0.307 g). This structure is deemed as representative of a pre-1970s