Seismic risk assessment represents a major challenge in countries with a considerable seismic hazard and a significantly vulnerable built heritage, such as Italy. Especially when seismic risk needs to be analysed at a large scale, the evaluation of vulnerability may result in very time-consuming and expensive investigations, since it is necessary to define multiple parameters that may influence the seismic response of the buildings.

In order to tackle this problem and carry out faster and more performing risk analyses, artificial intelligence techniques can be taken into consideration. Particularly, satellite images and street view pictures can be used to automatically collect data about buildings, when given as input in Convolutional Neural Networks (CNNs), deep learning algorithms used for image processing and classification.

In this work, satellite images of areas of interest are retrieved, along with preliminary information about the buildings detected, such as building typology and floor area. For each building, street view images are extracted and then used to obtain information remotely. To do so, pre-trained CNNs were selected and fine-tuned to recognize different classes of buildings. Specifically, three CNNs were trained to identify: building material (masonry vs reinforced concrete), height (low-rise buildings vs mid-rise buildings), and construction periods (pre 1919, 1919-1945, 1946-1960, 1961-1980 and post 1980). The CNNs were trained on a dataset of more than 10,000 labeled pictures of Italian residential buildings. 

Through this procedure, parameters can be acquired from images and on-site surveys might become obsolete and unnecessary, with a significant reduction in time and costs. Moreover, a more accurate and efficient investigation of building features can lead to a better vulnerability assessment, thus to more precise seismic risk estimates.