You are on page 1of 19

Automation in Construction 132 (2021) 103936

Contents lists available at ScienceDirect

Automation in Construction
journal homepage: www.elsevier.com/locate/autcon

Machine-learning based vulnerability analysis of existing buildings


Sergio Ruggieri *, Angelo Cardellicchio, Valeria Leggieri, Giuseppina Uva
DICATECH Department, Polytechnic University of Bari, Bari, Italy

A R T I C L E I N F O A B S T R A C T

Keywords: The paper presents a machine-learning based framework, named VULMA (VULnerability analysis using MAchine-
Vulnerability analysis learning), for vulnerability analysis of existing buildings. The underlying idea is to provide an indication of the
Existing buildings seismic vulnerability by exploiting available photographs, which can be properly processed to provide some
Machine-learning
input data for empirical vulnerability algorithms. To this scope, a complete processing pipeline has been defined,
which consists in four consecutive modules offering different and specific services. The first module, Street
VULMA, performs the image gathering starting from the raw data; the second module, Data VULMA, provides a
mean for the data labelling and storage; the third module, Bi VULMA, uses the collected data to train several
machine-learning models for image classification; the fourth module, In VULMA, performs a ranking of the
images, their analysis and consequently assigns the vulnerability index. The proposed procedure has been
employed on the existing building portfolio in an extended area of the municipality of Bisceglie, Puglia, Southern
Italy, for which all the modules have been tested and, above all, the machine-learning models of Bi VULMA have
been trained. After, in order to test the efficiency and the reliability of the proposed tools, the entire procedure
has been applied on five case study buildings. The results in terms of vulnerability index have been compared
with the manual computations performed by the authors applying the same algorithm. Despite the proposed tool
could be improved or modified in some of its modules, the obtained results show a good effectiveness of VULMA,
which opens new scenarios in the field of vulnerability assessment procedures and risk mitigation strategies.

1. Introduction of a building taxonomy (characterized by few features such as year of


construction, construction materials, number of storeys, etc.), and data
The development of fast procedures of vulnerability analysis for about damages are then statistically processed to calibrate a class-
seismic actions represents one of the main challenges faced by scientific specific function, able to provide information about seismic vulnera­
community and public institutions in the last years, with the aim to bility. Several examples of empirical methods are provided by the sci­
support requalification strategies of the existing building stock by entific literature, with regard both to reinforced concrete (RC) [2,3] and
developing risk mitigation plans to prioritize the economic resources. To masonry buildings [4,5]. Methods of this type are also often applied on
this end, especially in seismic-prone areas, several approaches for complex buildings, such as masonry churches [6,7], for which the
quantifying the seismic safety of existing building portfolios have been detailed identification of the behaviour under earthquakes becomes
developed, mainly oriented at a large-scale screening on the main hardly feasible. Mechanical methods (also called analytical) consist in
vulnerable parts of the building stock. As recently summarized in Silva the derivation of the vulnerability functions from the analytical estimate
et al. [1], among the wide range of proposals dealing with vulnerability of the seismic behaviour of a building class, by means of a numerical
models at class-level in which the fragility/vulnerability assessment is model (or a sample of numerical models) representative of the focused
performed through a direct derivation of the relationship between typology. Again, there are myriads of scientific works with regard to RC
seismic intensity and damage states/losses, the two fairly popular [8–10] and masonry buildings [11], of which only a few will be here
methods are empirical and mechanical ones. Empirical methods (also mentioned. Still, other methodologies are available, such as hybrids
called observational) consist in the derivation of the vulnerability approaches or rapid visual screening methods, which are often used for
function from the damage observed after the earthquakes occurred in strategic and socially important buildings, such as schools [12] and
the focused area. The damaged buildings are firstly classified by means hospitals [13].

* Corresponding author.
E-mail address: sergio.ruggieri@poliba.it (S. Ruggieri).

https://doi.org/10.1016/j.autcon.2021.103936
Received 31 May 2021; Received in revised form 2 August 2021; Accepted 31 August 2021
Available online 6 September 2021
0926-5805/© 2021 Elsevier B.V. All rights reserved.
S. Ruggieri et al. Automation in Construction 132 (2021) 103936

From this brief introduction, it is quite clear that the options for data. With regard to the accuracy, the availability of several large-scale
vulnerability analysis are manifold, with a continuous growth and methods developed in the last five decades [17] allows to combine
refinement of the methodologies, which from one side improve the several kinds of datasets characterized by different levels of detail.
quality of the results and, on the other side, strongly reduce the Concerning to the management aspects, it is paramount to collect a huge
computational overburden related to the statistical and numerical ana­ amount of data, but these have to be elaborated with the aim to define
lyses. Despite these beneficial aspects, however, it is worth reminding specific classes of buildings under common taxonomies, in order to
the crucial problem in this kind of analyses: the availability of data. As a provide more efficient studies of the seismic vulnerability [18].
matter of fact, for any mechanical or empirical method, the analyst In most approaches currently used in many countries, seismic risk
needs to dispose of a large amount of data to process. In this framework, assessments employ classification schemes based on early scientific
the phase of inventory of the building portfolio is fundamental for the works. An example is the Medvedev-Sponheuer-Karnik scale (MSK–64)
evaluation of different typologies of impact scenarios at a large-scale [19], in which buildings were classified into three distinct vulnerability
and, it represents the basis for identifying the distribution of different classes according to their construction material and type of load-bearing
vulnerability classes in a geographic area that presents buildings having system. Later, this classification scheme evolved into the European
similar features, given a certain taxonomy. Obviously, different degrees Macro-seismic Scale – EMS98 [20] that included three additional clas­
of detail can be achieved, depending on the size of the building stock, the ses, in order to account further structural typologies (e.g. structural
distribution of buildings on the territory and the adopted assessment steel). Another classification example is provided in HAZUS [21], in
model [14]. Additionally, perhaps the most important problem in the which the building stock is represented by typical buildings designed or
data gathering is the possible occurrence of human errors, considering not according to the seismic standards (High-Code, Moderate-Code or
that the input of fragility/vulnerability methods is provided by engi­ Low-Code) and the main parameters detected are the structural system
neering judgments about buildings and their elements. In the end, the and the usage or occupancy class. In HAZUS, the structural system is
availability, accessibility and reliability of data sources for vulnerability assumed to be a key factor in the assessment of the overall building
analysis at large-scale is yet an open issue, despite a consistent state of performance, loss function and fatalities, and 36 model building types
the art about data collection procedures (see Section 2) and several are proposed and used to classify buildings within the general categories
freely accessible web platforms are available (e.g., [15]). of wood, steel, concrete, masonry building systems. Another example is
To face this problem and to improve the data reliability, in this study provided by the RISK-UE projects [22], in which the vulnerability
we are proposing a new machine-learning based method for identifying analysis requires only the age of construction and general information
the key features of buildings in the existing stock, simply starting from about the building typology. In this framework, the analyses of the most
the image of a building. As originally conceived, the tool can be used for recurrent different building types in all European and Mediterranean
several scopes, from the definition of the fast vulnerability response on countries were performed, proposing a matrix of 23 building typologies
the individual buildings to the seismic vulnerability mapping to be having similar performance under seismic events. Recently, the GEM
employed in large-scale risk assessment procedures. Herein, we are Building Taxonomy [19] has described and classified buildings in a
going to train the entire methodology on the existing building portfolio uniform framework for achieving a reliable assessment under seismic
in a specific area and after, we test the results for five different case study actions. The criteria for developing the taxonomy were based on the
buildings but the tool could be directly used for an urban scale appli­ seismic performance of different construction typologies.
cation. The main advantage of the procedure is to provide a supervised Despite the existing several classification schemes, the most critical
classification of images for the identification of a simple vulnerability issue remains the one of data retrieval. Polese et al. [14] have proposed
index, by reducing the bias introduced in the phase of subjective eval­ an exhaustive list of the techniques available in the recent literature,
uation of the building features from the judgmental assignments of starting from those applicable to a wide scale [19,23] to those looking at
domain experts. The proposed tool, which has been named VULMA a more specific level [24]. Accordingly, the data collection phase will
(VULnerability analysis using MAchine-learning) is composed by four take into account the different levels of insight required by the different
consequent modules, which will be defined in the following Sections. application scales. Obviously, the applicability scale is strictly related to
Even if VULMA is still a proof-of-concept, an initial processing pipeline the typology of data, the available source and the scopes of the evalu­
has been established; however, despite in the proposed tool we adopt ation. With regard to the methods for data collection highlighted in [14],
existing scientific-based methods, such as transfer learning in Bi VULMA four different sources for extracting information about the existing
(see Section 3.3) and the vulnerability index proposed by [65] in In building stock are available: (i) Census data; (ii) interview-based sur­
VULMA (see Section 3.4), each of the proposed modules may be subject veys; (iii) Geographic Information Systems (GIS) and remote sensing
to future modifications in order to improve its overall efficiency and methods (high resolution – HR, or very high resolution – VHR satellite
reliability. imagery); (iv) building by building.
Census data are perhaps the simplest source to use, considering that
2. State of the art about data collection of building portfolios each Country periodically performs building Census, and provide basilar
information about buildings (e.g., storey number, age, construction of
2.1. Existing approaches and applications of seismic vulnerability material, state of preservation). Following a similar paradigm, but in a
assessment for class-level purposes more refined way, interview-based surveys characterize the primary
methods adopted for collecting data. These procedures require expert
The main aim of data collection for seismic vulnerability purposes is judgments by specialised technicians, who perform rapid in-situ surveys
the definition of prioritization ranks for seismic risk mitigation strate­ and evaluate the properties and the actual state of the observed build­
gies. The most common methodologies of class-level investigation are ings. In this framework, the main Italian method is represented by the
based on the classification into groups of buildings having common GNDT form [25] developed by the National Group for Defence against
morpho-typological features and that could show overall comparable Earthquake (subsequently adopted and extended by several researches,
performances [16]. The development of schemes of building classifica­ such as Uva et al. [26] or Cara et al. [27]), which provides two different
tion can be improved through the preliminary identification of all levels of detail. The first one requires data about location, geometry,
recurrent geometrical and mechanical features of the buildings usage class, year of construction, structural typology, maintenance and
belonging to a specific area or region and the consequence definition of a damage state, while the second one consists in the collection of 11 pa­
certain number of classes. At this point, it is possible to proceed with the rameters for masonry buildings and 4 parameters for RC buildings [28].
assignment of a class to each building, aiming to achieve a good Eventually, the parameters are used to extract a vulnerability index. A
compromise between the accuracy of results and the management of recent advancement towards the compilation of regional scale

2
S. Ruggieri et al. Automation in Construction 132 (2021) 103936

inventories is provided by Cartis approach [29], implemented in Italy by definition of buildings performance. ML can be defined as a branch of
Civil Protection Department in ReLUIS project. The Cartis survey form Artificial Intelligence (AI), which collects several widely used
has a different approach form the GNDT form, aiming to collect data methodologies.
about the recurrent features of some building classes within homoge­ In general, ML approaches can be divided in two main categories: (a)
neous urban sectors (USs). These units are geographic zones character­ supervised learning; (b) unsupervised learning. The main difference
ized by a homogenous building stock in terms of year of construction, between them lies in the data used to feed the algorithm, as the first type
construction material and structural typology. The procedure is based on of algorithms uses labelled data, while the latter does not. The concept of
the gathering of data through a form organized in 4 sections, and the labelled data is mainly related to domain expertise: the intrinsic
subsequent definition of the recurrent features: meaning of labelled data is determined beforehand by a domain expert.
As such, supervised algorithms try to mimic domain expertise by finding
• Section 0: delimitation of urban sectors; relationships which bind an input (i.e., a new data sample) to an output,
• Section 1: identification of the most common buildings typology in which can be either a class (for classification problems) or a numeric
each urban sector; value (for regression problems). As for unsupervised algorithms, they do
• Section 2: identification of general features of each building typology not rely on previous labelling, therefore trying to gather insight directly
class; from the inner structure of data. A typical example of unsupervised
• Section 3: characterization of the structural elements of all building learning are clustering algorithms, where data are divided according to
typology classes. their relative distance in a n-dimensional space, where n is the number
of features available for each sample.
The Cartis survey form (one form for each town, comprising one or Here, it is necessary to mention that several applications have been
more USs) is compiled by interviewing one or more technicians (e.g. developed in the scientific community in the last years with regard to
engineers or architects) that are local experts with a deep knowledge of structural engineering, and this will be done by referring to some review
the construction features in the area. If properly disaggregated and papers (and references therein) recently released. A comprehensive re­
processed, the information collected on building typologies through the view is proposed by Xie et al. [37], who have explored the possible
Cartis form is potentially useful for applying more refined vulnerability application of ML to earthquake engineering, accounting for the first
models than the ones employable through the simple Census or GNDT two of four traits of interest in the field, as well as ML methodologies,
data. topic area, data resource and scale of analysis. More in detail, they
Turning to remote sensing and GIS-based approaches, the most assessed seven classes of ML approaches applied on four areas of
innovative techniques rely on image processing, taken as a source for application in which ML approaches are used to support seismic hazard
building inventory to rapidly gather widespread geo-referenced infor­ analysis, structural identification and damage detection, seismic
mation about the building stock. Building features can be objectively fragility analysis and structural control. The scientific literature includes
evaluated from high resolution (HR) satellite imagery, through the other review papers about ML methods in structural design and assess­
identification of footprints and size, number of floors, height of floors ment of structure and infrastructure, as recently released by Sun et al.
etc. However, some building features that are crucial to vulnerability [38] and Falcone et al. [39]. In Sun et al. [38] the authors have
assessment, such as construction material or age, cannot be easily formalized all the techniques of supervised and unsupervised learning
established by these methods and, for this reason, they should be com­ and explored the application in building structural design and perfor­
bined with other sources of information. On the other hand, the power of mance assessment by identifying four classes of application: (a) struc­
GIS platforms is the possibility to easily provide access to information, tural response and performance prediction, (b) use and interpretation of
which can be transferred, transformed, overlaid, processed and dis­ experimental data to predict the structural properties of components, (c)
played using numerous software applications and a wide variety of use of the images to gather information; (d) application in structural
formats. In addition, GIS databases are flexible, freely available, open- health monitoring.
source and capable to contain multi-hazard information. Several ex­ Herein, for the scopes of this paper, the problem of data collection,
amples are available in the literature and can be mentioned [30–32], which can represent the limit in the application of any kind of vulner­
which show the use of remote sensing and GIS for collecting big amounts ability analysis method, is of interest. In particular, in order to have
of data with a low burden and allowing a large evaluation of the building efficient predictive models, it become essential to dispose of sufficient
typologies to employ in assessment procedures. and high-quality data that can be investigated. Furthermore, when using
Finally, building-by-building surveys represent the most detailed ML approaches, data should be carefully sampled, mainly to avoid
source of data for individual buildings in a specific area and are aimed to negative effects, such as imbalanced data [40] or overfitting, which is a
perform an accurate vulnerability assessment of the investigated unit. Of common problem in large ML architectures [41]. In addition, one of the
course, detailed investigations at building-level, as required by national main problems is due to the availability of the data for training the ML
and international building codes, are of difficult application for large- algorithms, which for any kind of purpose should be freely accessible
scale purposes, considering the requirement of time, computational ef­ and subjected to quality control measures. Some solutions to this issue
forts and, above all, the necessity of economic resources. For these can be mentioned, where the most significant is represented by the ac­
reasons, this kind of detailed survey is generally applied only during tion of the expert domain, which uses the proper knowledge to reduce
post-earthquake vulnerability and damage survey campaigns [33,34], to the amount of data for achieving good solutions. Still, techniques of
assess buildings that can be yet used or as a benchmark information for transfer learning or generative models (e.g. Monte Carlo simulation,
data mining approaches [35,36]. often employed in structural engineering problem) could be used to
In this broad framework of methods aimed to the collection of increase the starting dataset. Additionally, another concern is repre­
building data for large-scale analyses, it is possible to explore new sented by the data quality, which can heavily affect the results, caused
methods based on more recent concepts that in the last years are gaining by the presence of outliers in the dataset.
ground, such as machine-learning. In the end, considering this brief overview (which is not compre­
hensive of the myriads of applications proposed by the scientific liter­
2.2. Use of machine learning approaches: recent applications and new ature), our aim is to employ ML approaches to develop seismic
perspectives for seismic vulnerability assessment at class-level purposes vulnerability assessment for class-level purposes. For large scale anal­
ysis, few applications are available in the scientific literature. For
The use of the machine learning (ML) in civil engineering application example, Harirchian et al. [42] proposed a prototype of a smartphone
is gaining momentum, especially in the research fields related to the app aimed to collect data and to assess vulnerability on the base of the

3
S. Ruggieri et al. Automation in Construction 132 (2021) 103936

combination among the geometrical features that affect seismic per­


formance of RC buildings. Another example of ML approach is provided
in the work by Mangalathu et al. [43], which adopted ML approaches for
predicting damages on buildings provided by earthquakes on the base of
a dataset of about 2000 buildings and accounting for features as spectral
acceleration, soil category, year of construction, number of storeys, base
area and presence of irregularities. Again, Mangalathu et al. [44] pro­
posed a ML based method for the assessment of damages caused by
earthquakes on bridge portfolios.
Anyway, as mentioned in Section 2.1, the actual practice for large-
scale assessment is based on frameworks that employ large efforts of
technicians and researchers in the definition of the main parameters to
be considered in an analytical/empirical vulnerability model. For these
reasons, VULMA represents a real possibility for overcoming some of the
main issues intrinsically contained in the methods shown in Section 2.1.
Especially for the most employed methodologies (that use Census data
and interview-based surveys), a ML approach could overcome the
subjectivity of the surveyor judgments and, above all, could strongly
reduce the time related to surveys and interviews. Also for the GIS
methodologies, a ML approach could provide data that can be easily
integrated with the existing databases, to complete the set of existing
data to use for a large-scale analysis.

2.3. VULMA: tool definition and organization of modules

The proposed tool is named VULMA, which stands for VULnerability


analysis using MAchine-learning. The aim of VULMA is to provide a
pipeline framework for automatically extracting a vulnerability index
from raw data associated to a building. It is composed by several mod­
ules, each one offering specific features. Each module can be used as an
individual tool for handling a specific use case. Let us assume an overall
flow, concerning each module, where the rationale behind VULMA can
be summarized by the simple flow charts shown in Fig. 1. The relation
between the two flow charts is established according to the adopted
colours, where to each VULMA module (right) corresponds one or more Fig. 1. Abstract flow charts of the interactions among VULMA modules. The
specific tasks (left). Still, in yellow are highlighted the roles of domain colours associate the modules of VULMA (right) and the tasks of each module
experts, as following specified. (left). In addition, domain experts’ roles are highlighted in yellow. (For inter­
The first step is raw data provisioning. Data usually concern a pretation of the references to colour in this figure legend, the reader is referred
building stock within a certain geographical area and are provided by to the web version of this article.)
freely available georeferenced databases (e.g., Technic Regional Car­
tographies, CTRs) that contain georeferenced polygons encompassing Decision Support System. As a matter of fact, VULMA does not neces­
buildings in the area of interest. These data have a great utility, espe­ sarily take autonomous decisions on the status of a specific building, but
cially for the historical centres, where the localization of buildings is instead provides automatic support to domain experts (especially for
provided for masonry aggregate blocks that cannot be considered as large-scale purposes), whose role will be to validate the suggestions
single unit. VULMA accepts data in a standardized GeoJSON format, provided by VULMA.
meaning that it can be adapted to (and used into) several application
scenarios.
In the second step, raw data are sent to a data gathering module, 2.4. Street VULMA
named Street VULMA, which processes them to extract images taken
from online services, such as Google StreetView. Obviously, images Street VULMA is the first module of VULMA and it is able to gather
gathered during this step do not present any kind of labelling. Therefore, data concerning buildings from online services, such as Google Street
the subsequent phase requires a collective effort from domain experts, View, starting from raw data of one/a set of buildings. It is written in
who should identify the main features of interest for each gathered Python [45], and it offers a simplified interface, with two main
image beforehand. To support the time – intensive labelling task, submodules:
domain experts can rely on another module, which allows for data
labelling and storage, named Data VULMA. Once labelled data have • The Fetch submodule, which has the task to take a GeoJSON file as
been stored, they can be used as a base for training a ML model (i.e., a input and to fetch images of buildings included within the borders of
neural network) for image classification. This is possible using another provided data (given such information are available within Street
module provided by VULMA, which is named Bi VULMA. View).
Finally, processed data can be associated with a vulnerability index, • The Clean submodule, which is responsible to remove duplicates
which is provided by the last module included in the VULMA toolset, from data fetched by the Fetch submodule.
named In VULMA.
This brief introduction allows to understand the logical reasoning As for the Fetch module, it acquires images by fixing three different
behind our proposal, which is summarized in the sequence diagram parameters: 1) pitch (i.e., the vertical angle of the camera); 2) field of
shown in Fig. 2. In the next sections, each module will be described in view (i.e., the horizontal angle of the camera, which can be adjusted to
detail. In the end, it is worth highlight the role of VULMA as a real provide a certain zoom effect); 3) heading (i.e., the angle where the

4
S. Ruggieri et al. Automation in Construction 132 (2021) 103936

Fig. 2. Sequence diagram of the operation of the VULMA toolset.

camera is headed). Data are acquired considering a spatial granularity where images fetched from Internet will be downloaded; (iii) using the
(e.g., the distance between two adjacent geographical positions) of 5 m. button Ok, to download photos; (iv) to remove the possible duplicates
After the acquisition of the images, they are automatically converted through the activation of the Clean module. Once that the photos have
in ASCII string format according to a radix-64 representation (Base64). been downloaded, they are stored in a proper database, in order to be
This allows to evaluate the SHA-512 hash function of the Base64 version elaborated in the next module.
of the original image, computing a fixed-length code, which is univo­
cally related to the original image. Hence, by comparing the hashed
2.5. Data VULMA
representation of all images pairs, the Clean submodule is able to
identify, and therefore discard, duplicate images. It is worth noting that
Once that the photos have been stored, it is necessary to switch on
the noise characterizing some images has been considered as an
the second module of VULMA, named Data VULMA. Here, it is required a
“advantage” for the proposed procedure, considering noise as a non-
labelling step, performed by domain experts. This step is particularly
systematic bias for a subset of the whole image dataset (images are
important, because labelled data will be used to build a supervised
supposedly acquired by the same camera, i.e., the one used by the
model for image classification. Obviously, an open issue is the lack of
Google Car, and are mainly related to changes in lighting and weather
objectivity caused by experts while selecting labels [46], which may
conditions). This introduces a variability component in the dataset,
result in biases inserted in the overall model. To overcome it, data will
which leads to a model constituted by better generalization capabilities.
be labelled twice by two different independent domain experts, making
In the end, noisy images are commonly retained in object recognition
the labels less prone to subjective biases. For the case at hand, to assign a
approaches through deep neural networks and data augmentation ap­
unique label to each image, authors used a weighted consensus procedure.
proaches, such as the one used in this work and accounting for the
Namely, after an internal training, two of the authors independently
addition of random noise to original images. An example of images
performed the first labelling (splitting the total amount of photos in two
gathered by Street VULMA is shown in Fig. 3 (a – d). The interface
parts). Afterwards, another author checked the outcomes from the
provided by Street VULMA is shown in Fig. 4, where it is possible to
previous step and, as his/her labelling was assumed to have a higher
upload the GeoJSON file obtained by user from other sources. Once the
weight than the ones from the first domain experts, it was more relevant
GeoJSON is available, the following operations must be performed: (i)
for the final decision. In addition, for the scopes of our application,
using the button Browse, to select the GeoJSON location from the user’s
which aim to define a vulnerability index that usually depends on the
local machine and upload the file into the tool; (ii) to report the folder
subjectivity of surveyors (see Section 2.1), the proposed double data

5
S. Ruggieri et al. Automation in Construction 132 (2021) 103936

Fig. 3. (a – d) – Example of images gathered through Street VULMA.

Table 1
Labels and type of assignment to photos by domain experts (*Typology in­
dicates an established number of options, selectable from a drop-down menu;
**Number indicates an integer number to assign; *** Yes-No is a Boolean
indication to indicate the presence or not of the observed parameter).
Labels Type of assignment

Structural typology Typology*


Number of units Number**
Number of storeys Number
Presence of pilotis floor Yes-No***
Presence of basement floor Yes-No
Presence of superelevation floor Yes-No
Total number of openings Number
Fig. 4. Graphical interface of Street VULMA
Type of roof floor Typology
Presence of vaults Yes-No
labelling could help to improve the key aspect about objectivity, char­ Presence of visible seismic details Yes-No
acterizing this practice. Basing on the abovementioned concepts, Presence of higher ground floor Yes-No
Presence of overhangs Yes-No
domain experts are called upon to elaborate images (one at time) and to
Regularity in-plan Yes-No
assign the proposed set of labels, only basing on the figure observation Regularity in-height Yes-No
and avoiding any kind of mental processing. Firstly, images are cropped
to highlight the relevant content (e.g., the building itself) and reduce the
influence of occlusions. Afterwards, each image is labelled according to the image, avoiding making hypotheses and ensuring to cover all
the criteria defined in Table 1 and according to the proposed types of possible detectable features from the most common typologies of
assignment. It is worth noticing that if two images contain the same buildings. If the domain expert is not sure or does not see the detail, it is
building but displayed from different positions, they are treated as possible avoid assigning the feature. This latter possibility will not in­
different buildings (perhaps with the same labels value). fluence the following phases (see Section 3.3 and subsections), where we
More in detail, all the required labels are defined only by observing can anticipate that the involved ML procedures will be independently

6
S. Ruggieri et al. Automation in Construction 132 (2021) 103936

applied on each label. This means that, for each image, if the value of a 2.6. Bi-VULMA
specific label is not assigned, that picture will not be considered for
training that specific classifier. Still, the same image will be used to train The third module is Bi VULMA, which is used to perform both the
other classifiers, for which a valid label has been assigned by domain training and classification steps shown in Fig. 1. More in detail, Bi
experts. Regarding to labels, the structural typology and the type of roof VULMA allows to train a (deep) Convolutional Neural Network (CNN)
floor can be defined by observing, respectively, the structural material for image classification. Following, two Sections are provided, with a
(RC, masonry, steel) and the kind of cover on the floor (dome, pitches, general introduction to CNN algorithms and the principles at the base of
flat). Still, number of units, storeys and openings can be defined ac­ Bi VULMA.
cording to the count of the features on the photo. Obviously, if the im­
ages have been cropped, number of units is equal to 1 while, for 2.6.1. Convolutional neural network and deep learning
aggregate buildings, the number of units increases. Some properties are CNNs have significantly advanced the state of the art in computer
indicated with a Boolean indication (Yes–No). For example, the pres­ vision, mainly after the seminal contribution given by Krizhevsky et al.
ence of basement floor can be inserted if some openings are visible at the with AlexNet [47], which achieved outstanding results in image classi­
base of the building; the presence of superelevation floor can be inserted fication on the extremely challenging ImageNet dataset [48].
if the picture shows that the colour and size of the last floor are different Since then, the architecture of CNNs has continuously evolved,
from the rest of building; vaults and seismic details can be inserted with embedding a growing representational power through new outbreaking
reference to masonry buildings if there are evident features (e.g. capo­ concepts and intuitions. In the following, we will briefly introduce some
chiave belonging to steel ties); the presence of a higher ground floor can of the most relevant among these advances, on which Bi VULMA has
be inserted if a shop is visible in the image; regularities and irregularities been developed.
(in-plan and in-elevation) can be inserted if the picture shows evident
recesses along the elevation of the focused building. 2.6.1.1. Base layer of a CNN. A typical layer in a CNN is composed by
It is worth mentioning that the proposed features have been estab­ three stages, which are represented by a set of convolutions, a nonlinear
lished on the base of the authors’ opinion, with the aim to satisfy the activation function and a pooling step. Convolutions [49] are the main
following requirements: (1) possible information that can be objectively building block on which CNNs are built. For a two-dimensional signal,
observed in a photo of a building (or of a group of buildings); (2) con­ such as an image I(m,n) where m and n are the number of picture ele­
struction typology of the observed building (or the observed group of ments (pixels) for the width and height of I, convolution is expressed as:
buildings); (3) possible features representing vulnerability sources for a ∑∑
building (or for a group of buildings); (4) possible information to define (i, j) = (I*K)(i, j) = I(m, n)K(i − m, j − n) (1)
a simple vulnerability index based on the morpho-typological features of
m n

a building (or of a group of buildings). Matching the answers provided to where K is a two-dimensional kernel. Convolution layers enable
the above guidelines, authors selected the labels reported in Table 1. Not three important properties of a CNN:
all the detected features will be used but only few of them will be
implemented in this work. At the same time, the set of labels could be • Sparse connectivity: in traditional neural networks, interactions be­
increased in the future and used even for making a mechanical model. tween layers are dense, meaning that every neuron at layer w in­
Nevertheless, in this work we evaluated all the declared features for each teracts with each neuron at layer w-1, for every w. In CNNs, however,
photo with an appropriate model. this is usually not true and is accomplished using a kernel K whose
In order to allow the storage of data labelled by domain experts, Data dimensions are lower with respect to the sizes of I. This reduces the
VULMA has been structured for providing a specific web service having a number of parameters used by the CNNs, allowing for an improved
web architecture as shown in Fig. 5. efficiency in terms of memory requirements and statistical
To ensure future enhancements, Data VULMA provides full support significance;
to two important features: • Parameter sharing means the value of a certain input xi applied to a
specific output si is tied to the value of the input xj applied to the
• geospatial databases, made by means of GeoDjango and PostGIS; output sj. Intuitively, this is achieved by CNNs, as the same kernel is
• containerization through Docker. applied to each pixel of I. Parameter sharing therefore allows for
memory efficiency, as fewer parameters are learned and stored;
As for usability, Data VULMA provides a user-friendly interface for • Equivariance to translations: this is directly derived from parameter
adding new data, which can be easily used by domain experts who may sharing and causes a shift to the output which is directly related to a
not have a strong ICT background, as shown in Fig. 6. The graphical shift applied to the input. That is, if g is a function which shifts the
interface contains all labels in Table 1 that can be set by domain expert, input of the image, and f(xi) is the output of the convolution applied
by means of drop-down menu for assigning a typology, checkboxes for to the input xi, then f(g(xi)) = g(f(xi)).
assigning a Boolean value and textboxes for assigning a number. On the
base of the structural typology, the system allows or not to add some After the convolution step, a nonlinear activation function is applied.
information (e.g., presence of vaults cannot be set in the case of RC This is mainly done to insert nonlinearities within the CNN, which ac­
building). In addition, Data VULMA allows to import images and to count for real-world effects. Historically, sigmoid functions have been
assess the entire database. used [50], however, starting from AlexNet, rectified linear units (ReLU)
have been preferred, due to their capability to achieve same results in a
more efficient way. A ReLU is expressed as follows:
{
xi if xi ≥ 0
f (xi ) = (2)
0 if xi < 0

which means every positive value will be propagated as is to the next


layer, while negative values will be set to zero. The last step is pooling,
which further modifies the output of the layer, by subsampling the input
signal replacing each subset of its pixels with a fixed representation.

Fig. 5. Architecture of Data VULMA.

7
S. Ruggieri et al. Automation in Construction 132 (2021) 103936

Fig. 6. Graphical interface of Data VULMA.

2.6.1.2. Dropout. Dropout layers have achieved increased popularity many convolution layers as possible on top of each other, assuming that
after their introduction by Srivastava et al. in [41], and are now a more complex representation would lead to improved classification
considered as one of the main regularization layers within CNNs. Let us results. However, over a certain limit, the accuracy of the network in
assume the standard feed forward operation for a CNN (and, more terms of classifying out-of-sample data saturates and, eventually, de­
generally, for deep neural networks) can be described as follows: grades. This effect, commonly known as the degradation problem, has
been shown by He et al. in [51]. To overcome this problem, He et al.
(3)
(l+1) (l+1) l (l+1)
si = wi yi + bi introduced the concept of residual blocks.
( ) The rationale behind residual block is to allow deep networks to
y(l+1)
i = f s(l+1)
i (4) match performance of shallower networks; this is achieved by skipping
layers which cause the degradation problem. Let us note that the term
where l ∈ {1, …,L} is the lth hidden layer within the network, sl the inputs residual derives from what the network is trying to learn. As a matter of
into layer l,yl the outputs from layer l, the pair (Wl,bl) weights and biases fact, denoting with H(x) the real output of the network, the residual R(x)
learned by the network at layer l, and f the nonlinearities applied to sl to is given by:
get yl. Dropout regularization weights each output value for a vector of
R(x) = H(x) − x (5)
independent Bernoulli random variables rl, each one with a probability p
to be equals to 1. Obviously, due to the nature of the Bernoulli distri­ which implies:
bution, a reduced output vector ỹ(l) will be produced. This has the effect
H(x) = R(x) + x (6)
to randomly drop nodes during each epoch, therefore allowing the
network to assume different points of view throughout the whole As an intuitive consequence, we can state that the residuals R(x) are
training, causing a mitigation of overfitting phenomena. what is actually being learned by the inner layers of the network.

2.6.1.3. Residual blocks. The first generation of CNNs tried to stack as 2.6.1.4. Inception layers. Inception modules [52] have been introduced

8
S. Ruggieri et al. Automation in Construction 132 (2021) 103936

as an evolution of classic convolution layers. First, let us expand the is mainly provided in a visual form. To this end, a frontend is provided,
concept of image, moving from a m x n image to a m x n x 3 tensor, as implemented using PySimpleGUI, a wrapper on tkinker, one of the most
information for each colour channel (red, green and blue) are stored in a well-known graphic libraries for Python. Bi VULMA has two main
three-dimensional representation. operating modes: the first is training, which allows to train a neural
As previously stated, a single convolution kernel achieves sparse network architecture either from scratch (assuming enough data is
representation and parameter sharing. This implies that correlations provided) or exploiting transfer learning and fine tuning; the second is
between each pixel within a single channel (i.e., spatial correlations) inference, which allows to load a previously trained model to classify an
and between pixels belonging to different channels (i.e., cross-channel input image. The graphical interface of BI VULMA, presented while in
correlations) are learned jointly. The intuition that motivates the training mode, is shown in Fig. 7.
inception module is that such correlations are decoupled. Therefore, an The most interesting section of this graphical interface is probably
inception module tries to independently learn spatial correlations and the middle part, which allows users to define several parameters, as
cross-channel correlations by using a stacked set of convolution kernels. listed below:
Specifically, 1 × 1 convolutions are first used to look at cross-channel
correlations; then, the input of such convolutions is mapped to a small • Batch size, which is the number of samples (that is, images) using per
set of separate spaces (usually three to four), which are independently step during each training epoch.
explored via regular 3 × 3 or 5 × 5 convolution kernels. • Image height and width, which is automatically adapted to selected
neural network architecture (however, the user is still able to edit
2.6.1.5. GPGPU and ImageNet. The advancement highlighted in previ­ that parameter);
ous Sections are only a small part of the overall research which has been • Training epochs, which provides the maximum number of “times”
conducted on CNNs over the last decade. Still, they clearly show how that the weights of the neural network will be refined during
wide has been the adoption of CNNs (and, more generally, deep training;
learning), even though these ideas can be traced back to the 1960s, with • Optimizer, which is used for optimizing network weights during
the work made by Hubel and Wiesel on visual cortex [53]. However, training. Currently, two algorithms are available, that is, Stochastic
until the end of the 2000s, CNNs were not practically feasible due to Gradient Descent (SGD) [59] and ADAM [60];
severe limitations in both computational capacity and available data­ • Loss function and accuracy, which is used to evaluate the perfor­
sets. This is directly related to the number of parameters which must be mance of the network in terms of correct evaluation of the class
learned in a deep CNN, which is often in the magnitude of millions. assigned to each image. Currently, two types of losses and accuracies
Therefore, the most significant advances have been the introduction of are supported, specifically categorical and binary crossentropy/loss.
ImageNet [47], a large dataset composed by several millions labelled
images, and general-purpose computing on graphics processing units It is important to underline that the number of classes is automati­
(GPGPU), which provided enough computational power to train com­ cally inferred by the structure of the dataset itself, and both loss func­
plex deep network architectures. Nowadays, GPGPU is easily accessible, tions and accuracy automatically adjust accordingly. Furthermore, there
due to the availability of cheap, commercial-grade GPUs able to train are currently two callbacks which can be provided to the trainer: (i)
complex networks from scratch. However, gathering an adequate selectable using the Save logs checkbox, which can be enabled to sup­
amount of data for specific use cases is still a significant issue. Hence, port for visualizing the results achieved by the network during training;
data augmentation techniques have been developed [54]; however, (ii) use early stopping, which automatically stops network training when
these may not always prevent overfitting and, consequently, poor clas­ no any improvement is achieved over a certain number of epochs, given
sification performance. Therefore, the use of CNNs in real scenarios with by the value associated to the patience parameter. It is important to
limited datasets is often led by two approaches, named transfer learning underline that early stopping can be useful to prevent overfitting [61].
and fine tuning, which can be combined to achieve state-of-the-art Finally, Bi VULMA gives the option to choose from six base models to
performance. perform transfer learning, that are MobileNetV2 [62], Xception [63],
ResNet152v2, InceptionResNetV2, InceptionV3 [51,52] and NasNet
2.6.1.6. Transfer learning and fine tuning. The idea behind transfer [64]. Fig. 8 shows the graphical interface that appears to the user in the
learning is simple and exploits the internal structure of a deep CNN. As case of inference mode. The inference operating mode allows to load a
the work made by Yosinski [55], CNNs extract different representation pre-trained neural network model, which will be used to classify a given
of the input image at the top of each one of their internal layers. Spe­ image, selected on the left-side menu.
cifically, low level layers (i.e., the ones which are adjacent to the input)
extract generic features, such as edges and basic shapes, while high level 2.7. In VULMA
layers (i.e., the ones next to the classification layer) extract specific
features to the input image. This suggests that training only the pa­ The last module allows to compute the vulnerability index for
rameters of the high-level layers on a dataset relative to a specific buildings on the base of the previous steps. Several methodologies could
problem may lead to optimal results even with a relatively small amount be employed in this step, considering the broad variety of methodologies
of data, as the number of trainable parameters is significantly lower than available in the scientific literature (see Section 2.1). For now, In VULMA
the one from the original architecture. Therefore, the idea is to use a has been developed on the base of the simple approach proposed by
network which has been already trained on generic purpose data (often Frassine and Giovinazzi [65] to define a vulnerability index. The reason
ImageNet), and fix its parameter on a smaller, use-case specific dataset. of this choice is mainly due to the necessity to test In VULMA with a
Optionally, transfer learning can be followed by fine tuning, which methodology already available, which is simple and that allows us to
consists in performing a round of training using a significantly low assess the efficiency of the entire proposed procedure. Of course, in the
learning rate, hence tuning all the parameters of the network on the future developments new methodologies will be proposed and, as above-
small use-case dataset. mentioned, the estimated parameters could be also used to define a
proper mechanical method for more accurately evaluating the seismic
2.6.2. Principle of operation and tool description of Bi-VULMA vulnerability.
Bi VULMA is written in Python 3 [45]. While most of its function­ Coming back to the methodology in [65], it consists in the applica­
alities depend on the state of the art machine learning libraries (spe­ tion of the following formulation:
cifically, TensorFlow [56], Keras [57] and Scikit-Learn [58]), its output ∑
ṼI = VIb + ΔV m (7)

9
S. Ruggieri et al. Automation in Construction 132 (2021) 103936

Fig. 7. Graphical interface of Bi-VULMA.

Fig. 8. Inference mode in Bi VULMA.

where ṼI is the vulnerability index (ranging from 0 to 1, with higher or RC) and the year of construction, while the values of ΔVm are estab­
values that imply a more vulnerable building), computed by summing a lished on the base of some parameters influencing the seismic vulnera­
base vulnerability index (VIb) to some modification coefficients (ΔVm). bility quantification. For our scopes, not all the ΔVm parameters can be
Namely, VIb is defined on the basis of the structural typology (masonry considered in the evaluation and, for this reason, the corresponding

10
S. Ruggieri et al. Automation in Construction 132 (2021) 103936

coefficients will be accounted as null values, as allowed by the original the inhabited centre has been considered. Fr this entire municipality we
method. Tables 2 and 3 report the sets of parameters and the related dispose of a wide database from which the raw data to employ in Street
values associated, respectively, to VIb and ΔVm. VULMA can be extracted:
It is worth noting that masonry buildings are classified according to a
range of the year of construction (according to the release of the Italian 1. Technic Regional Cartographies (CTRs);
building technical standards in the last century), while RC buildings are 2. Census Data, subdivided in Census sections.
identified through the level of seismic design (absent, low, medium). For
these latter, the concept is similar to the one adopted for masonry A CTR is a vector file freely online available, shown in Fig. 10 for the
buildings, considering that a medium level of seismic design can be municipality of Bisceglie and implemented through a GIS environment.
attributed to buildings constructed after the 1971, while absent and low A CTR contains several polygons representing all the urban blocks of
levels of seismic design could be identified, respectively, for buildings buildings belonging to a municipality and including a reduced set of
dating back from 1919 to 1945 and from 1946 to 1970. attributes, such as the use destination of the units enclosed, area and
In order to apply this methodology to In VULMA, it is necessary to height of each block.
firstly know the year of construction for the focused building. Consid­ Regarding to Census Data, a vector file for the desired municipality is
ering that the Bi-VULMA is not able to provide this information, this freely online available and contains several Census sections, each one
issue can be solved with a prior knowledge about the period in which the composed by few buildings grouped in a limited portion of the munic­
focused area (from which the building image is taken) has been built. ipality and for which some specific data are provided. For each Census
This can represent a limit of VULMA and, more in general, of a ML based section, information about structural typology, state of preservation and
vulnerability assessment at large-scale. On the other hand, this kind of number of storeys is available, even if it should be observed that the
information is easily available (for example from the Census data) and it buildings inside a Census Section could not have similar features.
can be automatically integrated in the architecture of the system, thanks In order to train VULMA, we have selected a “complete” portion of
to the availability of the geographical coordinates of the focused the municipality characterized by a similar percentage of RC and ma­
buildings. In addition, for a simple vulnerability index, the year of sonry buildings. The building evolution over time is described in discrete
construction is provided in a range form and hence it has not to be terms according to the year of construction steps shown in Table 2 (for
extremely accurate. RC buildings we considered the design levels “absent”, “low” and “me­
Once that the value of VIb has been defined (based on the structural dium”, respectively in these time range, 1919–1945, 1945–1970,
typology), the values of ΔVm can be easily assigned, according to the >1970). Fig. 11 shows the selected area (red boundary) and the Census
work made in Bi VULMA. More in detail, the modification factors sections, subdivided per year of construction, while Table 4 reports the
associated to number of floors, plan and vertical irregularities, super­ total number of buildings, the percentages of buildings for construction
elevation floor, aseismic devices and structural element details materials and for the year of construction. The total number of building
(considering the presence or not of the pilotis floor) can be evaluated as into the selected area is equal to 817.
defined in Table 3. Instead, the modification factors associated to state of Basing on this quite accurate characterization of information, raw
preservation, structural system, retrofit intervention (for masonry data (GeoJSON format) have been extracted and transferred as an input
buildings), and foundation cannot be accounted for. Finally, the modi­ for Street VULMA. According to the concepts shown in Section 3, about
fication factors associated to the aggregation of buildings (position and 20,000 photos have been downloaded and elaborated (one at time) by
elevation) are considered only in the worst condition (assumed equal to domain experts, who performed the cropping of images, the required
+0.04) if the number of units in the image is greater than one. From the labelling, according to the features in Table 1. After this step, all photos
base set of labelled parameters shown in Table 1, the number of open­ have been stored through the Data VULMA service. The entire ensemble
ings, the presence of overhangs and higher floor have not been used. amounted to about 2500 labelled photos, which represented the input of
Bi VULMA. As expected, for some labels, photos for RC and masonry
3. Application and training of VULMA: the case study of buildings present a usable classification, such as occurs for the number
Bisceglie, Puglia, Southern Italy of storeys. For other labels, only some of the photos displaying RC or
masonry buildings have provided a certain value for that feature, such as
In order to apply the concept and the modules of VULMA to a real occurs for pilotis floor (only for some RC buildings) or steel ties (only for
case study, the municipality of Bisceglie, Puglia, Southern Italy has been some masonry buildings). The consequence of this occurrence is that for
considered. Data of this municipality have been extracted and elabo­ some labels the training dataset was unbalanced (the number of photos
rated with the aim to train VULMA to predict the vulnerability index. without a certain feature is strongly greater than the number of photos
Still, we can anticipate that the validation of the effectiveness of VULMA with that feature). Hence, before to employing the CNN of Bi VULMA,
has been carried out on different buildings out of the selected area. The the authors have balanced the number of photos in the dataset for each
reasons for which we selected this municipality are manifold: (a) it is the label. For this scope, some photos were automatically downloaded and
hometown of the first author; (b) all authors have a deep knowledge of manually selected from Google Images.
the municipality and its building stock; (c) the municipality is charac­ Training has been performed using transfer learning (see Section
terized by RC and masonry buildings, homogenously built during the 3.3.1.6) mainly due to constrains related to the size of the available
time and distributed in the area. dataset, which does not allow to train a neural network model from
As shown in Fig. 9, Bisceglie is located on the Adriatic coast and it scratch. For this study, MobileNetV2 [66] has been chosen as the base
characterized by an urban centre (blue zone) surrounded by a large network, with base parameters weighted on the ImageNet dataset. As for
agricultural area (dark grey). For the scopes of this paper, only a part of the training algorithm, cross-entropy (either binary or categorical,

Table 2
Set of parameters and related values associated to VIb [65].
Typology Masonry Reinforced Concrete

Class I II III IV V VI VII

Year of construction <1919 19–45 46–70 >1971 – – –


Design level – – – – Absent Low Medium
VIb 0.704 0.689 0.669 0.667 0.519 0.434 0.364

11
S. Ruggieri et al. Automation in Construction 132 (2021) 103936

Table 3
Set of parameters and related values associated to ΔVm [65].
Typology Masonry Reinforced Concrete

Class Description I - IV Description V VI VII


State of preservation Good − 0.04 Good – – –
Bad +0.04 Bad +0.04 +0.02 0
Number of floors − 0.08 Low (1–3) − 0.02 − 0.02 − 0.02
Low (1–2)
Medium (3–5) 0 Medium (4–7) 0 0 0
High (>5) High (>7) +0.08 +0.06 +0.04
Structural system Wall thickness
Wall distance +/− 0.04 -
Wall connection
Plan Irregularity Geometry +0.04 Geometry +0.04 +0.02 0
Mass distribution Mass distribution +0.02 +0.01 0
Vertical Irregularity Geometry +0.04 Geometry +0.04 +0.02 0
Mass distribution Mass distribution
Superelevation – +0.04 –
Floor
Roof details Weight, thrust connections +0.04 –
Retrofit intervention ±0.08 –
Aseismic Barbican, foil arches, buttresses − 0.04 –
Devices
Aggregate Building: position Middle − 0.04 0
Corner +0.04 Insufficient aseismic joints +0.04 0
Header +0.06
Aggregate Building: elevation
Staggered floors + 0.04 -
+/− 0.04
Buildings with different height
Foundation Beams − 0.04 0 0
Different level foundations +0.04 Connected beams 0 0 0
Isolated footing +0.04 0 0
Structural elements details Captive column +0.02 +0.01 0
Bow windows +0.04 +0.02 0

Fig. 9. Boundaries of the municipality of Bisceglie, Puglia, Southern Italy.

12
S. Ruggieri et al. Automation in Construction 132 (2021) 103936

Fig. 10. CTR for the municipality of Bisceglie, Puglia, Southern Italy.

Fig. 11. Selected part of the municipality of Bisceglie and indication of Census sections, according to the year of construction.

13
S. Ruggieri et al. Automation in Construction 132 (2021) 103936

Table 4 has been used, preventing the overall accuracy of the training to drop for
Number of buildings and percentages for the buildings of the selected area, more than a patience interval of three consecutive epochs. Finally, the
according to the parameters of the year of construction and construction training/test split of the dataset is in a standard 70/30 percentages.
materials. Once that the parameter is actually present in the picture of the
Parameter Number of Percentage of buildings building, it is identified by a single network, which has been specifically
buildings [%] trained to identify that parameter among the ones shown in Table 3. This
Construction typology RC 435 50.00 results in a set of 15 networks, which can be used as a cascade of models
Construction typology 351 43.00 to determine the overall characteristics of each building. The specific
Masonry choice of this setting (using a cascade of models instead of a single,
Construction typology Other 85 7.00
Year of construction <1919 228 27.91
comprehensive model able to determine each possible combination of
Year of construction 83 10.15 parameters) is mainly related to the small size of the dataset, which
1919–1945 would likely result in an inadequate number of samples for each class if
Year of construction 212 25.95 parameters were to be combined. The results in terms of accuracy on the
1945–1970
test set for some of the trained models are shown in Fig. 12 (indicated
Year of construction >1970 294 35.99
from a) to e) for the labels about structural typology, number of storeys,
irregularity in-plan, irregularity in-height and superelevation floor).
according to the specific number of classes involved in the sub-problem) As it can be seen, even with a small dataset, optimal results can be
has been used as the loss function, while ADAM [60] has been selected as easily achieved using transfer learning, as the accuracy is over 97% for
the optimization algorithm with a learning rate of 0.01. Early stopping each trained model.

Fig. 12. Accuracy after 10 epochs for (a) Structural typology; (b) number of storeys; (c) irregularity in-plan; (d) irregularity in-height; (e) superelevation floor.

14
S. Ruggieri et al. Automation in Construction 132 (2021) 103936

In the end, we avoided applying the last step of VULMA to the same 4. VULMA Validation: evaluation of the vulnerability index
photos used for the training of Bi VULMA, with the aim to validate the
procedure and to compute the index on different cases studies. Obvi­ Once that the ML models have been trained in Bi VULMA, a valida­
ously, once that the ML algorithms are trained, it is possible to compute tion process has been developed. The aim of this step is twofold: (a) to
for each photo the vulnerability index (according to In VULMA), which assess if Bi VULMA is able to recognize the right values of the labels
can be employed in a fast vulnerability analysis at large-scale. shown in Table 1 on a photo about a general building; (b) to evaluate if

Fig. 13. Case study buildings (B1 – B5) for VULMA validation.

15
S. Ruggieri et al. Automation in Construction 132 (2021) 103936

the vulnerability index given by In VULMA is comparable with the same Table 6
one manually computed by authors. To this aim scopes, 5 buildings have Estimated values of VIb, ΔVm and ṼI from the manual computation and In VULMA
been selected and subjected to Bi VULMA. These buildings, named B1- and absolute percentage differences for RC buildings B3, B4 and B5 (the symbol
B5, have been selected from the municipality of Panni, Puglia, South­ “/” indicates that VULMA does not provide that value).
ern Italy, 120 km far from the municipality of Bisceglie. This choice can Label B3 B4 B5
be motivated by the necessity to apply our tool to buildings character­ Manual In Manual In Manual In
ized by features different from the ones selected for the training. In VULMA VULMA VULMA
addition, for the selected buildings, the final value of the vulnerability VIb 0.434 0.434 0.434 0.434 0.434 0.434
index ṼI was already available and could be directly compared with the ΔVm - State of 0.00 / 0.00 / 0.00 /
one provided by In VULMA. preservation
Fig. 13 shows the photos of the 5 buildings B1-B5, where B1 and B2 ΔVm - Number − 0.02 − 0.02 − 0.02 − 0.02 − 0.02 − 0.02
of floors
are masonry building whereas B3-B4 and B5 are RC buildings. All the
ΔVm - 0.00 / 0.00 / 0.00 /
photos have been manually extracted by Google Street View and crop­ Structural
ped in order to be directly subjected to ML algorithms. system
More in detail, B1 is a masonry one-storey building, characterized by ΔVm - 0.00 0.00 0.00 0.00 0.00 0.00
a flat roof with no evident vaults or visible seismic details and dating Irregularity
in-plan
back before the 1919. Also B2 was built before 1919 and is a masonry ΔVm - 0.02 0.00 0.00 0.00 0.00 0.00
building presenting 2 storeys, vaults as detectable over the openings, a Irregularity
pitched roof and no seismic detail. B3 was built in the 60’s and it is the in-height
more peculiar building, since it is a 3-storey RC building with pitched ΔVm - 0.00 0.00 0.00 0.00 0.00 0.00
Aggregate
roof, a lateral part of different constitution (similar to a superelevation
buildings
floor but at an intermediate level), classifiable as irregular in-elevation ΔVm - 0.00 / 0.00 / 0.00 /
and located on a sloping basement. B4 and B5 are simple RC buildings Foundation
made in the 80’s and they present respectively 3 and 2 storeys, pitched ΔVm – 0.00 / 0.00 / 0.00 /
roof and structural regularity. In none of the case study buildings is Structural
elements
present a higher ground floor, while in the RC buildings are always details
visible overhangs. ṼI 0.434 0.414 0.414 0.414 0.414 0.414
From the application of Bi VULMA and In VULMA the vulnerability Absolute 4.60% 0.00% 0.00%
index ṼI has been computed and the results are reported in Table 5 and percentage
differences
Table 6, respectively for masonry (B1 and B2) and RC buildings (B3, B4
[%]
and B5), with the specification of the considered VIb and ΔVm and the
percentage differences (expressed in absolute value) between the
manual and the automatic procedures. From the obtained results it is the irregularity in height (in the higher part of the photo there is an out-
possible to make some observations about the ṼI computed by In of-focus zone).
VULMA. For the masonry buildings B1 and B2, the vulnerability index In addition, in order to apprise the effectiveness of the tool, it is
automatically obtained presents differences of, respectively, of about 10 possible to estimate, for each building, the deviation from the set of
and 20% with respect to those manually computed. These differences parameters (VIb and ΔVm) constituting the automatically computed ṼI
can be addressed to some parameters that have been not considered in (assumed as the predicted values) to the set of parameters (VIb and ΔVm)
the automatic index, such as the state of preservation and the informa­ constituting the manually computed ṼI (assumed as the observed
tion about structural system (e.g., wall thickness, wall distance and wall values). To this scope, a common metric, such as the root mean square
connection), which actually can be detected only through an on-site error (RMSE), can be employed and the results are following reported:
survey. Regarding to RC buildings B3, B4 and B5, the values of the 0.0163 for B1 and B2, 0.0067 for B3 and 0 for B4 and B5. These values
vulnerability indexes are almost completely the same, with a unique suggest a certain degree of effectiveness for the tool and, in this case, a
difference of about 5% for B3, which is due to the wrong identification of reliable prediction of the desired values by Bi VULMA and In VULMA
modules.
Another aspect to clarify is the principle of operation of In VULMA.
Table 5 To compute the final index, the tool considers all the accounted features
Estimated values of VIb, ΔVm and ṼI from the manual computation and In VULMA in Bi VULMA and then, in the automatic computation of ṼI a specific
and absolute percentage differences for masonry buildings B1 and B2 (the
feature cannot be currently excluded as desired by user. This is due to
symbol “/” indicates that VULMA does not provide that value).
the functioning principle of VULMA, where classifiers are based on a
Label B1 B2 cascade of neural networks, each one trained for a certain feature in the
Manual In Manual In current version. Then, using the trained classifiers, the value of each
VULMA VULMA label is automatically assigned by the network itself and the index will
VIb 0.689 0.689 0.704 0.704 always take into account the considered features. This could be a
ΔVm - State of preservation +0.04 / +0.04 / disadvantage for In VULMA in the case of ambiguous values of some
ΔVm - Number of floors − 0.04 − 0.04 − 0.04 − 0.04 labels. On the other hand, considering that the validation accuracy on
ΔVm - Structural system / /
each trained neural network has been estimated to be above 97%, the
+0.04 +0.04
ΔVm - Irregularity in-plan 0.00 0.00 0.00 0.00
ΔVm - Irregularity in-height 0.00 0.00 0.00 0.00 effect on the result could be quantified as a slight variation of the final
ΔVm - Superelevation floor 0.00 0.00 0.00 0.00 value of ṼI, never provoking strongly incorrect estimates.
ΔVm - Roof floor 0.00 0.00 0.00 0.00 The obtained results suggest that the proposed tool is extremely
ΔVm - Retrofit intervention 0.00 / 0.00 /
efficient for RC buildings, while some additional operations are neces­
ΔVm - Aseismic details +0.08 +0.08 +0.08 +0.08
ΔVm - Aggregate buildings 0.00 0.00 0.00 0.00 sary for masonry buildings. Also for building B3 some differences occur,
ΔVm - Foundation 0.00 / 0.00 / considering that the bad quality of the image and the complex evolution
ṼI 0.809 0.729 0.824 0.664 of the building are still open issues for Bi VULMA. Nevertheless, as just
Absolute percentage differences 9.89% 19.42% mentioned above, this is an unborn version of In VULMA, which will be
[%]
subjected to future developments, while Bi VULMA is just highly

16
S. Ruggieri et al. Automation in Construction 132 (2021) 103936

efficient. Lastly, as for the indexes computed by In VULMA, one should each parameter. Second, training can be evaluated using also other
consider that this approach is currently limited by visually available metrics, and both fine tuning and hyper-parameter optimization can be
information and details such as foundation or wall thickness have not used to improve classification accuracy. Finally, index computation by
been considered in this application. In VULMA could be further refined and extended by embedding
contextual information on the building, which cannot be easily extrac­
5. Conclusions and future developments ted by a visual inspection (such as information on foundation).

This study presents a new machine-learning based tool, named Declaration of Competing Interest
VULMA (VULnerability analysis using MAchine-learning), able to capture
the key features of buildings in the existing stock starting from a simple The authors whose names are listed immediately below certify that
photograph of a building. The general framework, presented according they have NO affiliations with or involvement in any organization or
to a processing pipeline, consists in four independent modules, which entity with any financial interest (such as honoraria; educational grants;
can be separately used for different aims. In particular, VULMA is participation in speakers’ bureaus; membership, employment, consul­
composed by: (1) Street VULMA, which processes raw data to extract tancies, stock ownership, or other equity interest; and expert testimony
photos of buildings from a desired area; (2) Data VULMA, which allows or patent-licensing arrangements), or non-financial interest (such as
domain expert to store photos of buildings once that they have been personal or professional relationships, affiliations, knowledge or beliefs)
labelled; (3) Bi VULMA, which is composed by a set of machine-learning in the subject matter or materials discussed in this manuscript.
algorithms (based on convolution neural networks) that, once trained,
are able to capture the labelled features of buildings and (4) In VULMA, References
which provides a simple vulnerability index for a single building.
After a detailed description of each phase, the steps of the general [1] V. Silva, S. Akkar, J. Baker, P. Bazzurro, J.M. Castro, H. Crowley, M. Dolsek,
tool have been applied on the municipality of Bisceglie, Southern Italy, C. Galasso, S. Lagomarsino, R. Monteiro, D. Perrone, K. Pitilakis, D. Vamvatsikos,
Current challenges and future trends in analytical fragility and vulnerability
to provide a trained version of Bi VULMA and In VULMA that in turn modeling, Earthquake Spectra 35 (4) (2019) 1927–1952, https://doi.org/10.1193/
have been validated on 5 case study buildings belonging to another 042418EQS101O.
municipality. The results of the application show that, despite the [2] C. Del Gaudio, M. Di Ludovico, M. Polese, G. Manfredi, A. Prota, P. Ricci, G.
M. Verderame, Seismic fragility for Italian RC buildings based on damage data of
vulnerability index employed is very simple, the tool provide reliable
the last 50 years, Bull. Earthq. Eng. 18 (5) (2020) 2023–2059, https://doi.org/
estimates. 10.1007/s10518-019-00762-6.
For concluding, we wish to highlight the new scenarios that a tool [3] M. Dolce, A. Goretti, Building damage assessment after the 2009 Abruzzi
earthquake, Bull. Earthq. Eng. 13 (8) (2015) 2241–2264, https://doi.org/10.1007/
like VULMA can provide in the existing fast vulnerability assessment
s10518-015-9723-4.
procedures to develop for risk mitigation strategies: [4] G. Zuccaro, F.L. Perelli, D. De Gregorio, F. Cacace, Empirical vulnerability curves
for Italian mansory buildings: evolution of vulnerability model from the DPM to
1. VULMA provides a method that reduces the time effort related to curves as a function of acceleration, Bull. Earthq. Eng. 19 (8) (2021) 3077–3097,
https://doi.org/10.1007/s10518-020-00954-5.
surveys and interviews, usually involved in the fast vulnerability [5] A. Rosti, C. Del Gaudio, M. Rota, P. Ricci, M. Di Ludovico, A. Penna, G.
assessment procedures (e.g., Cartis). As a matter of fact, it allows to M. Verderame, Empirical fragility curves for Italian residential RC buildings, Bull.
define a vulnerability index only by identifying buildings in an area Earthq. Eng. 19 (8) (2021) 3165–3183, https://doi.org/10.1007/s10518-020-
00971-4.
and by using their photos; [6] S. Ruggieri, C. Tosto, G. Rosati, G. Uva, G.A. Ferro, Seismic Vulnerability Analysis
2. VULMA provides a supervised classification of images and, if of Masonry Churches in Piemonte after 2003 Valle Scrivia earthquake: post-event
continuously upgraded, can reduce the usual bias introduced in the screening and situation 17 years later, Int. J. Archit. Heritage 8 (2020), https://doi.
org/10.1080/15583058.2020.1841366.
phase of subjective evaluation of the building features by the judg­ [7] G. De Matteis, G. Brando, V. Corlito, Predictive model for seismic vulnerability
mental assignments of domain experts; assessment of churches based on the 2009 L’Aquila earthquake, Bull. Earthq. Eng.
3. VULMA can be easily integrated with other data, freely online 17 (9) (2019) 4909–4936, https://doi.org/10.1007/s10518-019-00656-7.
[8] V. Silva, H. Crowley, H. Varum, R. Pinho, L. Sousa, Investigation of the
available, such as regional technical maps and Census Data. The characteristics of Portuguese regular moment-frame RC buildings and development
processes of data integration can improve the actual state of the tool of a vulnerability model, Bull. Earthq. Eng. 13 (5) (2015) 1455–1490, https://doi.
and can represent the base for further improvements in the analytical org/10.1007/s10518-014-9669-y.
[9] B. Borzi, R. Pinho, H. Crowley, Simplified pushover-based vulnerability analysis for
vulnerability estimates;
large-scale assessment of RC buildings, Eng. Struct. 30 (3) (2008) 804–820,
4. VULMA structure is conceived in order to be continuously evolved in https://doi.org/10.1016/j.engstruct.2007.05.021.
each of the proposed steps. In particular, especially for In VULMA, [10] S. Ruggieri, F. Porco, G. Uva, D. Vamvatsikos, Two frugal options to assess class
the identify parameters can be processed in analytical/mechanical fragility and seismic safety for low-rise reinforced concrete school buildings in
southern Italy, Bull. Earthq. Eng. 19 (2021) 1415–1439, https://doi.org/10.1007/
methods, with the aim of computing fragility and vulnerability s10518-020-01033-5.
curves for the single photos or for the entire class; [11] B. Borzi, H. Crowley, R. Pinho, Simplified pushover-based earthquake loss
5. VULMA can assume the role of Decision Support System, as an assessment (SP-BELA) method for masonry buildings, Int. J. Archit. Heritage 2 (4)
(2008) 353–376, https://doi.org/10.1080/15583050701828178.
automatic support to domain experts in large-scale vulnerability [12] S. Ruggieri, D. Perrone, M. Leone, G. Uva, M.A. Aiello, A prioritization RVS
analysis. methodology for the seismic risk assessment of RC school buildings, Int. J. Dis. Risk
Reduct. 51 (2020) 101807, https://doi.org/10.1016/j.ijdrr.2020.101807.
[13] D. Perrone, M.A. Aiello, M. Pecce, F. Rossi, Rapid visual screening for seismic
Obviously, some improvements are required for each module. evaluation of RC hospital buildings, Structures 3 (2015) 57–70, https://doi.org/
Among the possible developments for In VULMA, we have to consider 10.1016/j.istruc.2015.03.002.
that the actual state of the involved labelling, even with the double- [14] M. Polese, M.G. d’Aragona, A. Prota, Simplified approach for building inventory
and seismic damage assessment at the territorial scale: an application for a town in
check procedure described in Section 3, could present a subjective southern Italy, Soil Dyn. Earthq. Eng. 121 (2019) 405–420, https://doi.org/
bias that may be still an open issue. In addition, also the photos may be 10.1016/j.soildyn.2019.03.028.
automatically elaborated, avoiding the direct operations made by [15] M. Dolce, E. Speranza, F. Giordano, B. Borzi, F. Bocchi, C. Conte, A. Di Meo,
M. Faravelli, V. Pascale, Observed damage database of past Italian earthquakes: the
domain experts. In the end, VULMA can be also shared with the scientific
Da DO WebGIS, Boll. Geofis. Teor. Appl. 60 (2) (2019) 141–164, https://doi.org/
community, in order to increase the actual datasets and improve the 10.4430/bgta0254.
power of the tool. [16] D.H. Lang, A. Kumar, S. Sulaymanov, A. Meslem, Building typology classification
The training procedure itself could be greatly improved by focusing and earthquake vulnerability scale of central and south Asian building stock,
J. Build. Eng. 15 (2018) 261–277, https://doi.org/10.1016/j.jobe.2017.11.022.
on mainly three aspects: first, improving the overall data quantity, by [17] G.M. Calvi, R. Pinho, G. Magenes, J.J. Bommer, L.F. Restrepo-Velez, H. Crowley,
gathering and labelling an adequate and balanced amount of samples for Development of seismic vulnerability assessment methodologies over the past 30

17
S. Ruggieri et al. Automation in Construction 132 (2021) 103936

years, ISET J. Earthq. Technol. 43 (3) (2006) 75–104. https://pdfs.semanticsch [40] S. Visa, A. Ralescu, Issues in Mining Imbalanced Data Sets-a Review Paper, Sixteen
olar.org/ff47/e72c838a6cba9227fa54d2b622da160b295e.pdf. Midwest Artificial Intelligence and Cognitive Science Conference 2005, 2005,
[18] S. Brzev, C. Scawthorn, A.W. Charleson, L. Allen, M. Greene, K. Jaiswal, V. Silva, pp. 67–73. https://eecs.ceas.uc.edu/~ralescal/PAPERS/VRMaics2005.pdf.
GEM Building Taxonomy Version 2.0, Technical Report, 2013, pp. 1–188. https:// [41] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, R. Salakhutdinov, Dropout: a
issuu.com/gem_wrld/docs/exp-mod-gem-building-taxonomy-20130 (accessed simple way to prevent neural networks from overfitting, J. Mach. Learning Res. 15
March 15, 2021). (1) (2014) 1929–1958. http://jmlr.org/papers/v15/srivastava14a.html.
[19] S.W. Medvedev, W. Sponheuer, V. Karnik, Seismic Intensity Scale Version MSK [42] E. Harirchian, K. Jadhav, V. Kumari, T. Lahmer, ML-EHSAPP: a prototype for
1964, United Nation Educational, Scientific and Cultural Organization, Paris, 1965, machine learning-based earthquake hazard safety assessment of structures by using
p. 7, in: https://academic.microsoft.com/journal/93825082/publication/search? a smartphone app, Eur. J. Environ. Civ. Eng. (2021) 1–21, https://doi.org/
q=Proceedings%20of%20the%20USSR%20Academy%20of%20Sciences&qe=And 10.1080/19648189.2021.1892829.
(Composite(J.JId%253D93825082)%252CTy%253D%270%27)&f=&orderBy=0. [43] S. Mangalathu, H. Sun, C.C. Nweke, Z. Yi, H.V. Burton, Classifying earthquake
[20] G. Grünthal, European Macroseismic Scale, Luxembourg: Cahiers du Center damage to buildings using machine learning, Earthquake Spectra 36 (1) (2020)
Europèen de Gèodynamique et de Seismologie, Conseil de l’Europe 1998 (1998) 183–208, https://doi.org/10.1177/8755293019878137.
99. http://media.gfz-potsdam.de/gfz/sec26/resources/documents/PDF [44] S. Mangalathu, S.H. Hwang, E. Choi, J.S. Jeon, Rapid seismic damage evaluation of
/EMS-98_Original_englisch.pdf. bridge portfolios using machine learning techniques, Eng. Struct. 201 (2019)
[21] FEMA, HAZUS99 Technical Manual. Service Release 2, Federal Emergency 109785, https://doi.org/10.1016/j.engstruct.2019.109785.
Management Agency, Washington, D.C., USA, 2001, p. 42. https://nehrpsearch. [45] Python Software Foundation, Python: A Dynamic, Open Source Programming
nist.gov/static/files/FEMA/PB2008108263.pdf. Language, Python Software Foundation, 2015. https://www.python.org/psf/.
[22] P. Mouroux, E. Bertrand, M. Bour, B. Le Brun, S. Depinois, P. Masure, The European [46] H. Jiang, O. Nachum, Identifying and correcting label bias in machine learning, in:
RISK-UE project: an advanced approach to earthquake risk scenarios, in: International Conference on Artificial Intelligence and Statistics, Proceedings of the
Proceedings of the 13th World Conference on Earthquake Engineering, 2004, Twenty Third International Conference on Artificial Intelligence and Statistics,
p. 3329. https://www.iitk.ac.in/nicee/wcee/article/13_3329.pdf. PMLR 108, 2020, pp. 702–712, in: http://proceedings.mlr.press/v108/jiang20a.
[23] S. Lagomarsino, S. Giovinazzi, Macroseismic and mechanical models for the html.
vulnerability and damage assessment of current buildings, Bull. Earthq. Eng. 4 (4) [47] A. Krizhevsky, I. Sutskever, G.E. Hinton, Imagenet classification with deep
(2006) 415–443, https://doi.org/10.1007/s10518-006-9024-z. convolutional neural networks, Adv. Neural Inf. Proces. Syst. 25 (2012)
[24] K. Pitilakis, P. Franchin, B. Khazai, H. Wenzel, SYNER-G: Systemic Seismic 1097–1105, in: https://proceedings.neurips.cc/paper/2012/file/c399862d3b9d6
Vulnerability and Risk Assessment of Complex Urban, Utility, Lifeline Systems and b76c8436e924a68c45b-Paper.pdf (ISSN: 10495258).
Critical Facilities: Methodology and Applications, 2014, p. 184. https://www.sprin [48] J. Deng, W. Dong, R. Socher, L.J. Li, K. Li, L. Fei-Fei, Imagenet: A large-scale
ger.com/gp/book/9789401788342 (978–94–017–8835-9). hierarchical image database, in: 2009 IEEE Conference on Computer Vision and
[25] GNDT, Gruppo Nazionale Per la Difesa dai Terremoti, Rischio sismico di edifici Pattern Recognition, 2009, pp. 248–255, https://doi.org/10.1109/
pubblici, 1993, pp. 1–134. https://emidius.mi.ingv.it/GNDT2/Pubblicazioni/Bib CVPR.2009.5206848.
lioteca/Risk_ed_pubbl/rischio_sismico_di_edifici_pubblici_parteI.pdf (accessed April [49] I. Goodfellow, Y. Bengio, A. Courville, Y. Bengio, Deep Learning Vol. 1, MIT press,
10, 2021). Cambridge, 2016, p. 800. No. 2, https://books.google.it/books?hl=it&lr
[26] G. Uva, C.A. Sanjust, S. Casolo, M. Mezzina, ANTAEUS project for the regional =&id=omivDQAAQBAJ&oi=fnd&pg=PR5&dq=Goodfellow,++Deep+learning+
vulnerability assessment of the current building stock in historical centers, Int. J. isbn&ots=MNO4iskBOZ&sig=VQxkT4Fwg069MBKQyB0hGB_FLlw#v=onepage
Archit. Heritage 10 (1) (2016) 20–43, https://doi.org/10.1080/ &q=Goodfellow%2C%20%20Deep%20learning%20isbn&f=false (0262035618).
15583058.2014.935983. [50] J. Han, C. Moraga, The influence of the sigmoid function parameters on the speed
[27] S. Cara, A. Aprile, L. Pelà, P. Roca, Seismic Risk Assessment and Mitigation at of backpropagation learning, in: International Workshop on Artificial Neural
Emergency Limit Condition of Historical Buildings along Strategic Urban Networks, 1995, pp. 195–201, https://doi.org/10.1007/3–540-59497-3_175.
Roadways. Application to the “Antiga Esquerra de L’Eixample” Neighborhood of [51] K. He, X. Zhang, S. Ren, J. Sun, Identity mappings in deep residual networks, in:
Barcelona, Int. J. Archit. Heritage 12 (7–8) (2018) 1055–1075, https://doi.org/ European Conference on Computer Vision, 2016, pp. 630–645. https://arxiv.
10.1080/15583058.2018.1503376. org/abs/1603.05027v3.
[28] D. Benedetti, V. Petrini, On the Seismic Vulnerability of Masonry Buildings: A [52] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, Z. Wojna, Rethinking the inception
Method of Evaluation, L’industria delle Costruzioni, 1984, pp. 66–74 (In Italian), architecture for computer vision, in: IEEE Conference on Computer Vision and
http://www.lindustriadellecostruzioni.it/. Pattern Recognition, 2016, pp. 2818–2826. https://arxiv.org/abs/1512.00567.
[29] G. Zuccaro, M. Dolce, D. De Gregorio, E. Speranza, C. Moroni, La Scheda Cartis Per [53] D.H. Hubel, T.N. Wiesel, Receptive fields, binocular interaction and functional
La Caratterizzazione Tipologico - Strutturale Dei Comparti Urbani Costituiti Da architecture in the cat’s visual cortex, J. Physiol. 160 (1) (1962) 106–154, https://
Edifici Ordinari. Valutazione dell’esposizione in analisi di rischio sismico, 34 doi.org/10.1113/jphysiol.1962.sp006837.
Convegno Nazionale GNGTS, Trieste, 2015, pp. 281–287. http://www3.ogs.trieste. [54] C. Shorten, T.M. Khoshgoftaar, A survey on image data augmentation for deep
it/gngts/files/2015/S23/Riassunti/Zuccaro.pdf. learning, J. Big Data 6 (1) (2019) 1–48, https://doi.org/10.1186/s40537-019-
[30] H. Miura, S. Midorikawa, Updating GIS building inventory data using high- 0197-0.
resolution satellite images for earthquake damage assessment: application to metro [55] J. Yosinski, J. Clune, A. Nguyen, T. Fuchs, H. Lipson, Understanding neural
Manila, Philippines, Earthquake Spectra 22 (1) (2006) 151–168, https://doi.org/ networks through deep visualization, in: ICML Deep Learning Workshop, 2015,
10.1193/1.2162940. p. 12. https://arxiv.org/abs/1506.06579.
[31] M. Indirli, Organization of a geographic information system (GIS) database on [56] M. Abadi, P. Barham, J. Chen, Z. Chen, A. Davis, J. Dean, X. Zheng, Tensorflow: A
natural hazards and structural vulnerability for the historic center of san Giuliano system for large-scale machine learning, in: 12th Symposium on Operating Systems
di Puglia (Italy) and the city of Valparaiso (Chile), Int. J. Archit. Heritage 3 (4) Design and Implementation (OSDI 16), 2016, pp. 265–283. https://arxiv.
(2009) 276–315, https://doi.org/10.1080/15583050902803780. org/abs/1605.08695.
[32] E.J. Kim, G. Plessis, J.L. Hubert, J.J. Roux, Urban energy simulation: simplification [57] F. Chollet, Keras, GitHub. https://github.com/fchollet/keras, 2015.
and reduction of building envelope models, Ener. Build. 84 (2014) 193–202, [58] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel,
https://doi.org/10.1016/j.enbuild.2014.07.066. M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos,
[33] F. Braga, M. Dolce, D. Liberatore, A Statistical Study on Damaged Buildings and an D. Cournapeau, M. Brucher, M. Perrot, E. Duchesnay, Scikit-learn: machine
Ensuing Review of the MSK-76 Scale, Seventh European Conference on Earthquake learning in Python, J. Mach. Learning Res. 12 (2011) 2825–2830. https://www.
Engineering, Athens, Greece, 1982, pp. 431–450. http://hdl.handle.net/11573/45 jmlr.org/papers/volume12/pedregosa11a/pedregosa11a.pdf?source=post_page.
9303. [59] S.I. Amari, Backpropagation and stochastic gradient descent method,
[34] M. Dolce, A. Masi, M. Marino, M. Vona, Earthquake damage scenarios of the Neurocomputing 5 (4–5) (1993) 185–196, https://doi.org/10.1016/0925-2312
building stock of Potenza (southern Italy) including site effects, Bull. Earthq. Eng. 1 (93)90006-O.
(1) (2003) 115–140, https://doi.org/10.1023/A:1024809511362. [60] D.P. Kingma, J. Ba, Adam: A Method for Stochastic Optimization, 3rd International
[35] I. Riedel, P. Guéguen, M. Dalla Mura, E. Pathier, T. Leduc, J. Chanussot, Seismic Conference for Learning Representations, San Diego. https://arxiv.org/abs/
vulnerability assessment of urban environments in moderate-to-low seismic hazard 1412.6980, 2015.
regions using association rule learning and support vector machine methods, Nat. [61] W.S. Sarle, Stopped training and other remedies for overfitting, in: Proceedings of
Hazards 76 (2) (2015) 1111–1141, https://doi.org/10.1007/s11069-014-1538-0. the 27th Symposium on the Interface of Computing Science and Statistics, 1995,
[36] I. Riedel, P. Gueguen, F. Dunand, S. Cottaz, Macroscale vulnerability assessment of pp. 1–10. http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.42.3920.
cities using association rule learning, Seismol. Res. Lett. 85 (2) (2014) 295–305, [62] M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, L.C. Chen, Mobilenetv2: Inverted
https://doi.org/10.1785/0220130148. residuals and linear bottlenecks, in: IEEE Conference on Computer Vision and
[37] Y. Xie, M. Ebad Sichani, J.E. Padgett, R. DesRoches, The promise of implementing Pattern Recognition, 2018, pp. 4510–4520. https://openaccess.thecvf.com/cont
machine learning in earthquake engineering: a state-of-the-art review, Earthquake ent_cvpr_2018/html/Sandler_MobileNetV2_Inverted_Residuals_CVPR_2018_paper.
Spectra 36 (4) (2020) 1769–1801, https://doi.org/10.1177/8755293020919419. html.
[38] H. Sun, H.V. Burton, H. Huang, Machine learning applications for building [63] F. Chollet, Xception: Deep learning with depthwise separable convolutions, in:
structural design and performance assessment: state-of-the-art review, J. Build. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition,
Eng. 33 (2020) 101816, https://doi.org/10.1016/j.jobe.2020.101816. 2017, pp. 1251–1258. https://arxiv.org/abs/1610.02357.
[39] R. Falcone, C. Lima, E. Martinelli, Soft computing techniques in structural and [64] B. Zoph, V. Vasudevan, J. Shlens, Q.V. Le, Learning transferable architectures for
earthquake engineering: a literature review, Eng. Struct. 207 (2020) 110269, scalable image recognition, in: Proceedings of the IEEE Conference on Computer
https://doi.org/10.1016/j.engstruct.2020.110269.

18
S. Ruggieri et al. Automation in Construction 132 (2021) 103936

Vision and Pattern Recognition, 2018, pp. 8697–8710. https://arxiv.org/abs/1 [66] M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, L.C. Chen, Mobilenetv2: Inverted
707.07012. residuals and linear bottlenecks, in: Proceedings of the IEEE Conference on
[65] L. Frassine, S. Giovinazzi, Basi di dati a confronto nell’analisi di vulnerabilità Computer Vision and Pattern Recognition, 2018, pp. 4510–4520. https://arxiv.
sismica dell’edilizia residenziale: un’applicazione per la città di Catania, XI org/abs/1801.04381.
Congresso Nazionale “L’ingegneria Sismica in Italia”, Genova, 2004, p. 16 (in
Italian), https://www.anidis.it/convegni-anidis/2004-genova/.

19

You might also like