PSI - Issue 78

Marco Martino Rosso et al. / Procedia Structural Integrity 78 (2026) 1382–1388

1384

and construction typology, specifying additional details for masonry buildings, such as presence of curbs, wall texture regularity, floors and roof typology and their connections to the walls. The subsequent three sections and the ninth one are dedicated to damage documentation. Indeed, following EMS-98 standards, the observed damage is categorized according to its extension in one-third, two-thirds, or more than two-thirds, considering both structural and nonstructural elements, resulting in approximately 60 categorical features, and moreover accounting also for external risk factors and site/foundation conditions (e.g., slope stability or ground morphology). In the second-to-last eighth section, the inspector is called to evaluate a usability judgment based on its visual inspection and the documented data in the AeDES form. Its judgment is translated into six categorial usability levels, i.e., A, immediately accessible; B, temporarily unusable (requires minor interventions); C, partially unusable; D, requires further evaluation (considered temporarily unusable to be on the safe side); E, permanently unusable; F, unusable due to external factors only. In order to consider usability classes that are directly interpretable as actual damage classes, i.e., depending solely on the intrinsic characteristics of the observed building conditions, in the current ML predictor, only judges A, B, C, and E have been considered herein. The currently available database is mainly focused on masonry buildings, comprising a set of 13 explanatory features coming from the AeDES forms, i.e. building position, no. of floors, inter-storey height, average floor surface, age, masonry typology, roof typology, masonry wall texture regularity, presence of curbs or chains, presence of isolated columns, hybrid masonry, reinforced masonry, and site morphology. Since they are just related to the observed effects of the Central Italy earthquake events over the area under investigation, besides these 13 features, an additional 9 intensity measures have been attached to the dataset in order to provide also data related to the input event. Specifically, these 9 additional features include the soil category of the municipality, the epicentral distance measured in the municipality centroid, fundamental period estimated from the estimated height of the building, horizontal and vertical peak ground acceleration, horizontal pseudo-spectral acceleration, Arias and Housner intensity, and seismic event duration. The usability classes in the available dataset provide unbalanced classes, since their proportions are respectively 58% for A, 20% for B, 4% for C, and 19% for E. This also evidences the fact that the inspector is also guided by their engineering judgment during the assignment of a certain class, and therefore, he is more prone to assign class E instead of class C, showing their quite subjective decision based on being more conservative. Therefore, the classification task has been reformulated as a binary classification problem merging classes A and B (class AB) for those immediately usable buildings or requiring minor interventions to be usable in a very short time, versus the merged classes CE for buildings considered completely unusable. Despite adopting this procedure, the classification problem remained strongly unbalanced, since the minority class represents only about 22% of the entire dataset. This implies that the problem can be analyzed using imbalance learning strategies such as Principal Component Analysis or the synthetic minority over-sampling technique (SMOTE). In this study, the authors did not consider any imbalance learning strategies, but used more robust metrics for assessing the classification performances, i.e., the balanced accuracy. The training and test set split followed a stratified sampling procedure among the two classes to respect the unbalanced proportion, and used 85% and 15% for the training set and the holdout test set. Two main ML techniques have been adopted as classifiers for building usability assessment, i.e., the support vector machine (SVM) and the random forest algorithm (RF). The standard binary Linear SVM, Cortes & Vapnik (1995), is based on maximization of the optimal separating hyperplane (margin) between two classes in the feature space, and it depends on a single tunable parameter C, acting as a regularization constraint term. The RF model, Breiman (2001), is an ensemble learning method that aggregates the results of multiple weak learners (decision trees) with bootstrap and majority voting mechanisms. Its ensemble nature provides an inherent mitigation of overfitting issues and naturally tends to select the most important features. However, the RF requires choosing the bagging scheme and tuning numerous hyperparameters, such as splitting criterion, maximum depth of trees, maximum leaf nodes, minimum sample split, and the number of weak learner estimators.