PSI - Issue 44

Carpanese Pietro et al. / Procedia Structural Integrity 44 (2023) 1980–1987 Carpanese Pietro et al./ Structural Integrity Procedia 00 (2022) 000–000

1981

2

1. Introduction Over the last 50 years, earthquakes in Italy caused around 180 billion euros of economic losses in terms of emergency management, recovery, and reconstruction costs, as well as thousands of casualties (DPC 2018). However, mitigating seismic risk and adapting to it is not an easy task, since it is not trivial to assess its components of hazard, vulnerability, and exposure (Dolce et al. 2021, da Porto et al. 2021, Zanini et al. 2019). Particularly, when the aim is to assess the seismic vulnerability of buildings, it can be useful to define fragility curves that relate ground motion intensity to damage states. Among the various methods available in the literature, analytical-mechanical ones are often used to define fragility based on structural models to simulate the seismic behavior of buildings (Donà et al. 2021, Borzi et al. 2021). However, these methods usually require a large amount of data concerning mechanical and geometrical properties to characterize the seismic response correctly. Moreover, especially when dealing with analyses at a territorial scale, i.e., with large neighborhoods or even whole cities, the investigation of these essential parameters often leads to very time-consuming and expensive surveys, thus making risk assessment analyses very challenging (Campostrini et al. 2017, Vettore et al. 2020). For this reason, an efficient and at the same time reliable evaluation of exposure in the area of interest is a key factor for large-scale risk assessments. Artificial intelligence techniques, specifically deep learning, can be taken into consideration to speed up and simplify this investigation process. For example, instance classification techniques have already been used to classify buildings based on street view images (Kang et al. 2018), as well as some of their features such as height (Carpanese et al. 2021, Diaz and Arguello 2016, Yuan and Cheriyadat 2016). Other information about buildings has also been obtained by analyzing morphology, neighborhood features, and urban patterns from topographic maps (Boeing 2021, Fleischmann et al. 2020, Rosser et al. 2019, Wieland and Pittore 2014). The works mentioned above prove that artificial intelligence may be useful in developing algorithms for extracting building information and consequently in performing risk management assessments (Wang et al. 2021). In this paper, special attention is given to the extraction of building footprints from satellite images, as well as to the automatic retrieval of residential building pictures through OpenStreetMap and Google Street View. Then, three Convolutional Neural Networks (CNNs) are developed to predict the height (Low-Rise buildings with 1 or 2 stories or Mid-Rise buildings with 3 or more stories), material (reinforced concrete or masonry), and construction period (Pre 1919, 1919-1945, 1946-1960, 1961-1980 and Post-1980) of each building, given their street-level images. In order to train the CNNs, an ad hoc dataset was previously prepared and labeled according to the three parameters of interest. Finally, the predictions of these features are presented in this paper for a case study area. 2. Recognition of building footprints and retrieval of street view images Building exposure assessment is necessary to understand how urbanized the area is and to identify the different building typologies. To this end, an algorithm was developed to obtain building footprints based on satellite images. Firstly, the script allows two input options to retrieve the desired satellite image: the user can search for a location by city/municipality or enter a pair of coordinates (latitude and longitude) and a distance radius R. The code performs this task using the Mapbox Static Maps service. The second step of the code involves the acquisition of the building footprints found in the satellite image. For this purpose, the OpenStreetMap (OSM) services are used, extracting only the objects identified as “buildings” from the map. Via OSM, it is also possible to store the geometry of building footprints, their centroids (latitude-longitude), and the labels of each building (e.g., “housing”, “commercial,” “religious,” etc.): all the data retrieved can be visualized in a GeoDataFrame, a spreadsheet where each row represents a building, and then saved as a shapefile. Fig. 1 shows the code output for an area centered in (45.965009, 12.653791) with a radius of 500 meters.