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

1982

3

Fig. 1. Satellite image, building footprints and their labels for an area centered in (45.965009, 12.653791) with a radius of 500 meters.

Since the centroid is known for each building, it is possible to automatically obtain the façade image of the desired buildings through the Google Street View API. The user can choose the output image size, the pitch (vertical angle) and the heading (horizontal angle). In this work, the default size was chosen, i.e., 640x640 pixels. For what concerns the heading, a value was calculated each time to point the camera to the specified location, while an angle of 5° was chosen as pitch, so that taller buildings could fit into the captured image (the default value would be 0°, which means flat horizontal). Fig. 2 shows some examples of street view pictures for the area selected in Fig. 1. The code is able to capture the street view image for all building footprints in the area selected in a short computational time (in the order of minutes).

Fig. 2. Examples of street view images obtained from the coordinates of the building footprints for the area of interest.

3. Convolutional Neural Networks training for the prediction of building features Once the street view pictures are retrieved, some predictions can be made about the height, the material, and the construction period of the target buildings. These three parameters have been chosen since it has been proven that correct identification of these features can lead to a reasonable estimate of vulnerability (Donà et al. 2019, Donà et al. 2020). In order to perform this task, CNNs have been trained and used.