PSI - Issue 78

Carpanese Pietro et al. / Procedia Structural Integrity 78 (2026) 536–543

538

While this step is essential to geolocate the buildings and retrieve their footprints, these geometrical features are not enough to carry out risk analyses. In order to perform more accurate risk assessment, additional building characteristics need to be retrieved, such as their use, material, construction period, and height, which are some of the main features that can influence their response to natural events and the impact that may result after a catastrophic event. These features are computed in the input enrichment module of the platform. Firstly, PERIL relies on OSM to obtain tags on the use of buildings, from which it draws to distinguish residential from industrial buildings. This operation allows for the association of appropriate exposure values and of the correct vulnerability models for the building type in question. For residential buildings, the construction material, which can typically be masonry or reinforced concrete, is obtained through newly developed deep learning algorithms that extrapolate this information from the façade image (i.e., from street view imagery). Specifically, Convolutional Neural Networks (CNNs) were trained to perform the automatic recognition of image characteristics. CNNs are a type of neural network designed for image processing, that progressively detect simple to complex shapes through multiple layers (Simonyan and Zisserman 2015). Widely used for object recognition, semantic segmentation, and image classification, they are well-suited for extracting building characteristics from street view images. For a CNN to produce reliable results, however, it is necessary to train it through a robust database of previously catalogued images. For this work, 10,000 images of Italian residential buildings were collected from different database (Dolce et al. 2019, Zuccaro et al. 2023) and labeled according to the characteristics for which the CNNs had to be trained. The algorithm developed for estimating the construction period of residential buildings is similar to the one used for recognizing construction type. The main difference is that there can be multiple construction periods classes, such as the ones chosen in this work: pre-1919, 1919-1945, 1946-1960, 1961-1980, post-1980. For industrial buildings, an alternative algorithm was developed using neural networks trained on satellite imagery of industrial warehouses, considering that street view images are often not available for these buildings. The algorithm was trained on a dataset provided by the Tuscany Region including over 2,000 industrial buildings classified according to the shape of their roof (flat, gabled, or vaulted). Thanks to this dataset, the neural network is able to recognize the type of industrial building roof. Starting from the shape of the roof identified by the algorithm, the same Tuscany dataset was analyzed to develop correlation matrices between the shape of the roof and the material (reinforced concrete, precast concrete, or steel). The construction period of industrial buildings can also be estimated using satellite images. The Italian Geoportale Nazionale (http://www.pcn.minambiente.it/viewer/) provides different maps with orthophotos covering the entire Italian national territory from 1988 to 2012. Comparing different maps, it is possible to infer the year or time range in which the building was built, identifying the first orthophoto in which the building appears. In order to do so in an automated way, an algorithm was developed to retrieve the satellite images of the area of interest, to segment the building (i.e., separate the areas where buildings are present vs where there are no buildings), and check in which photo the analyzed building appears for the first time to estimate its most probable construction year. This feature is essential for determining the reference construction code for the building, and thus its vulnerability. The height of buildings is also a crucial parameter in the input enrichment module, as it allows for the calculation of the number of floors, which gives more accurate estimates of the economic value of the building and of its vulnerability. To do so, PERIL first checks OpenStreetMap for height data, which is reliable but typically absent. Therefore, PERIL queries a LIDAR height dataset provided by Microsoft and Bing Maps (https://www.bing.com/maps). In case even the Bing dataset does not contain information for the buildings of interest, an additional CNN has been implemented to recognize the number of floors of a residential building starting from its street view image, similarly to what has been described for the material and construction period of residential buildings. The last section of the input enrichment module includes the estimation of the exposure, i.e., the economic value of the building. To do so, different components are evaluated based on the intended use of the building, particularly the building itself, which is considered for any intended use, and the content of the building, which is considered relevant only in the case of industrial buildings. The DEI (2023) price list provides an indication of the costs necessary for the construction of buildings with different uses. The construction cost of residential buildings was calculated as the weighted average of the values of different possible types of buildings, leading to an average construction cost of 1575 €/m 2 . For industrial buildings, the average value of the costs reported in the DEI price list