PSI - Issue 44

Gabriella Tocchi et al. / Procedia Structural Integrity 44 (2023) 1972–1979 Gabriella Tocchi et al./ Structural Integrity Procedia 00 (2022) 000–000

1973

2

Nomenclature TC

Town compartment Reinforced concrete Machine learning Population class

RC ML C pop HC

Historical Center mamsl Metres above mean sea level

1. Introduction The improvement of exposure modelling in large scale risk analysis has become a task of great interest in seismic risk studies. The exposure model defines the distribution of buildings within the classes defined by the vulnerability model in each geographic unit of analysis (i.e., the building inventory). Several studies highlighted how a better definition of exposure accounting for regional building typologies may allow a more refined estimate of seismic risk as well social and economic losses due to future earthquakes (Vettore et al. 2020; Tocchi et al. 2022, Polese et al. 2021). For compiling building inventory, the knowledge of the key structural characteristics of exposed building in a region of interest is crucial but it may be not easily obtainable. Even if the level of details needed to associate buildings to vulnerability classes changes from a vulnerability model to another, generally at least information about the construction material and the loads resisting system type are required (Grunthal 1998; Karababa and Pomonis, 2011; Del Gaudio et al. 2020). Models specifically developed for masonry building could also require information about the slabs’ type and the presence of tie roads or tie beams (Braga et al. 1982, Rota et al. 2008, Del Gaudio et al. 2019). The primary source of information usually adopted for compiling building inventory are census data, thanks to their availability and diffusion in whole national territory. However, this kind of data provide only generic information about buildings, such as construction material of the main load resisting structure, the period of construction and the height (Crowley at al. 2014). Therefore, in order to detect more refined structural information, often required by the above-mentioned vulnerability models, time-consuming efforts may be carried out, e.g. implementing building by building survey activities. To avoid the long time and the high costs of such survey campaign, in Italy the Cartis form (Zuccaro et al. 2015) was implemented within Reluis-DPC project. Aimed to provide high quality information for compiling local building inventories, this form allows to rapidly detect ordinary building typologies in sub-area of the town denominated Town Compartments (TC), characterized by homogeneity of the building stock mostly in terms of construction age. Through an interview-based protocol, the prevailing building typologies present in each TC are identified and for each building typology very detailed structural information is retrieved by interviewing an expert technician who has knowledge of building typologies characteristics. Thanks to the Cartis approach, a large number of data on several municipalities in Italy (> 500) were collected, allowing to provide indications for a regionalization of the seismic vulnerability and exposure functions. Moreover, the collected data may allow the compiling of building inventory also adopting more refined vulnerability models (Polese et al. 2019). Despite all the advantages related to this form, nowadays Cartis database cover only the 5% of the whole Italian municipalities. Herein a Machine learning based approach is proposed to define exposure models at local scale so to enhance exploitation of available Cartis data. Data on buildings and population provided by Italian census campaigns (ISTAT) at census tract level are used together with Cartis data to train a supervised learning algorithm aimed to classify municipal areas in homogeneous TCs. The proposed algorithm can be used to identify homogeneous areas in terms of building stock also in municipalities where Cartis form is not compiled yet using the sole information included in ISTAT database. Based on statistical analysis of Cartis data on prevailing building typologies in a given geographic area, the ML method is used to classify areas of municipalities characterized by similar features of the studied area (e.g. based on altimetry) and to facilitate construction of a building inventory at TC level. In this study Campania region’s data are used to train a logistic regression algorithm for the identification of homogeneous