Abstract:
Multiple efforts have contributed to collecting data on the built environment in Beirut.
However, these efforts usually rely on field visits that are time-consuming. Until now
no complete dataset that provides the construction periods and the number of stories of
all the buildings of Beirut exists. Such datasets are crucial to perform seismic risk and
vulnerability assessments.
This work presents a framework that uses artificial intelligence practices for automated
building characterization, with a focus on the construction period of buildings in Beirut.
A total of five construction periods are identified and used in this work which are:
pre1935, 19351-955, 1956-1971, 1972-1990, and post1990. The framework is
implemented into a tool that can predict the period of construction of a building from
clear street-view photographs of its facade. The tool relies on deep learning methods to
analyze the features of the facade image and output the estimated building’s
construction period. Two approaches are used. The first approach only requires as input
a street-view image of the building’s façade, and relies on a deep learning model to
autonomously learn the features in the image and classify it into a construction period.
The second approach requires tabular data as input in addition to a street-view image of
the façade, and relies on a multi-modal machine learning model that uses late fusion to
join the features extracted from two different models: the first is a deep learning model
that uses the street-view image, and the second is a fully connected neural network
model that learns from the tabular input, in particular the number of floors and socioeconomic
background.
To develop all these models, a unified dataset on the building stock of Beirut is
assembled that includes a construction period label and at least one street-view photo
for each building. For this purpose, data is compiled from multiple sources that provide
building street-view photos and/or construction periods. In addition, information on the
number of floors and location of the building is stored in the database when available.
After collecting and pre-processing the data, and validating, correcting or adding
construction period and number of floors label when possible, we end up with a total of
7,198 data points that all contain street-view images and are labeled with a construction
period. Among them, 6,480 also have a number of floors label and contain information
on their location (either exact or approximate, e.g., the name of the neighborhood).
In approach number one, three state of the art models are trained and compared:
DenseNet121, ResNet50, and Swin Transformer v2. The latter is found to be the best
performing model, reaching an accuracy of 75% on the test set. Utilizing additional
tabular data as input in the second approach is found to be beneficial to the overall
performance of the model, with a slight increase in accuracy to 78%. Nonetheless, both
approaches have a weakness when trying to predict buildings in the “1972-1990”
period, with an accuracy less than 50% for this period which can be correlated to the
lack of image examples in the mentioned period. The developed model can be used to
rapidly characterize cultural heritage buildings, and help in urban planning efforts
through the characterization of the built environment. Furthermore, the usefulness of the
proposed tool for seismic risk assessment is illustrated by developing seismic damage
and loss maps for a target neighborhood in Beirut.