PSI - Issue 78

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

1610

thus increasing the risk of catastrophic consequences in the case of subsequent seismic events such as aftershocks or triggered earthquakes. Therefore, assessing the seismic residual capacity of earthquake-damaged structures (i.e., their ability to sustain subsequent earthquakes) plays a key role in the decision-making on re-occupancy and repair vs. demolition (Pampanin, 2021). Traditionally, in post-earthquake emergency and recovery phases, the evaluation of the safety level of damaged buildings is performed through “tagging” procedures employing only a visual inspection and expert judgment of the damage severity (ATC, 1989; Baggio et al., 2007). However, recent research highlighted that modern machine-learning (ML) techniques, such as Convolutional Neural Networks (CNNs), can represent a supporting tool to automate and standardize this task. From the earliest emergency phases, a CNN-based damage detection tool can be used to process large amounts of photographic documentation collected by non-experts or captured by devices such as cameras and drones. As an example, Gao and Mosalam (2018) investigated the use of deep learning technologies for the recognition of structural damage from images. Musella et al. (2021) discussed the use of artificial intelligence (AI) combined with Building Information Modelling (BIM) for the inspection phase and damage information management, respectively. In Italy, the Italian Department of Civil Protection (DPC) in collaboration with the Network of University Laboratories for Earthquake Engineering (ReLUIS) has been working on image-based damage detection tools based ML techniques in the DPC-ReLUIS project – WP7. Other authors also proposed the use of deep learning to assess the degradation capacity in terms of sti ff ness and strength for RC members according to visible seismic damage (Miao et al., 2023). More recently, Saquella et al. (2025) investigated the use of the VGG16 architecture for damage classification, using both transfer learning and data augmentation techniques. In the past decades, a significant research e ff ort has also been devoted to developing suitable methodologies for a detailed seismic residual capacity assessment of earthquake-damaged buildings. State-of-the-art procedure typically employs nonlinear static analysis and capacity reduction factors to update the nonlinear response of damaged compo nents. In the Federal Emergency Management Agency (FEMA) 306 report (FEMA, 1998), capacity reduction factors in terms of sti ff ness ( λ K ), strength ( λ Q ), and ductility ( λ D ) are adopted to simulate the e ff ects of earthquake-related damage to structural members. A similar approach is also adopted by the Japan Building Disaster Prevention Associ ation (JBDPA) Guideline (JBDPA, 2015). However, in this document, a single capacity reduction factor for structural members is considered, namely the η factor (i.e., the ratio between residual energy dissipation capacity and original energy dissipation capacity). In line with this approach, several past studies focused on the derivation of capacity reduction factors for damaged components (Di Ludovico et al., 2013; Chiu et al., 2021) and on the development of safety and loss assessment frameworks for earthquake-damaged structures (Polese et al., 2013; Cuevas and Pampanin, 2017; Pedone et al., 2023; Rossi et al., 2022; Matteoni et al., 2023). In this context, and in line with recent research (Ji et al., 2023), this paper investigates the possibility of inte grating existing post-earthquake seismic capacity assessment frameworks with ML-based damage detection tool. The proposed methodology employs a CNN-based damage detection tool, trained on 5,000 RGB images sourced from ex isting databases and re-labelled by expert (Saquella et al., 2024, 2025,b). The information on the observed damage is then used to automatically update the nonlinear response of damaged components and assess the seismic performance of the damaged structures through a simplified analytical / mechanical procedure. The paper is structured as follows: Section 2 provides a description of the proposed seismic residual capacity assessment framework; Section 3 details the developed CCN-based damage detection tool; in Section 4, an illustrative application for the proposed methodology is presented; finally, Section 5 provides some concluding remarks.

2. Seismic residual capacity assessment framework

The proposed methodology for vision-based residual capacity assessment of earthquake-damaged buildings is schematically shown in Fig. 1. Each step of the procedure is discussed in more detail below. In Step 1 , a photographic survey is performed to collect data on the observed post-earthquake damage. As previ ously mentioned, in the earliest emergency phases, this information can be obtained as photographic documentation collected by non-experts or captured by drones (mainly for the earthquake damage observable from the outside) and / or other devices such as security cameras. Then, more detailed data are expected to be provided during the post earthquake in-situ inspection phase.