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

1985

6

Table 2. Precision, recall, and F1-score of the three CNNs. Feature Labels Precision

Recall

F 1 -score

Low-Rise Mid-Rise Masonry Pre-1919 1919-1945 1946-1960 1961-1980 Post-1980 R.C.

0.88 0.88 0.80 0.73 0.56 0.23 0.61 0.44 0.27

0.91 0.83 0.93 0.46 0.66 0.05 0.69 0.53 0.11

0.89 0.85 0.86 0.57 0.61 0.08 0.65 0.48 0.16

Height

Material

Construction period

Fig. 3 shows the results of the prediction of building features (i.e., height, material, and construction period) for the area of interest shown in Fig. 1, considering residential buildings only.

Height

Material

Construction period

Fig. 3. Predictions of height, material, and construction period of the residential buildings in the area of interest.

4. Conclusions This paper presented an algorithm that automatically obtains satellite images from a set of coordinates or a place name, returning the building footprints found in the image, along with their coordinates, areas, and category labels. Furthermore, the algorithm allows capturing street view images based on the coordinates of the buildings in that area. Three Convolutional Neural Networks (CNNs) were trained in order to predict the height, the material, and the construction period of each building that can be detected in the area of interest. A specific dataset of images of Italian residential buildings was created to train the CNNs, and the techniques of Transfer Learning and fine-tuning were used during this process, starting from the VGG16 architecture. The CNNs showed good levels of precision, especially for the parameters of height and material. Lastly, an example of predictions of the three chosen features was shown for a case study area. In general, the algorithm proposed in this work allows a fast and remote assessment of the taxonomy of urbanized areas, considering residential buildings only. When these results are combined with seismic vulnerability (e.g., seismic fragility curves) and seismic hazard, they can lead to an automated seismic risk assessment: this allows an effective and rapid estimate of the expected seismic damage as well as risk forecasts (such as economic losses and victims) for