Crisis Management Systems: Big Data and Machine Learning Approach

Abderrazak Boumahdi, Mahmoud El Hamlaoui and Mahmoud Nassar

IMS Team, ADMIR Laboratory, ENSIAS, Rabat IT Center, Mohammed V University in Rabat, Morocco

Keywords:

Machine Learning, Crisis Management, Big Data.

Abstract:

A crisis is deﬁned as an event that, by its nature or its consequences, poses a threat to the vital national interests

or the basic needs of the population, encourages rapid decision making, and requires coordination between the

various departments and agencies. Hence the need and importance of crisis and disaster management systems.

These crisis and disaster management systems have several phases and techniques, and require many resources

and tactics and needs. Among the needs of these systems are useful and necessary information that can be used

to improve the making of good decisions, such as data on past and current crises. The application of machine

learning and big data technologies in data processing of crises and disasters can yield important results in this

area. In this document, we address in the ﬁrst part the crisis management systems, and the tools of big data

and machine learning that can be used. Then in the second part, we have established a literature review that

includes a state of the art, and a discussion. Then we established a machine learning and big data approach

for crisis management systems, with a description and experimentation, as well as a discussion of results and

future work.

1 INTRODUCTION

The frequency of crises and disasters around the

world is changing and increasing, as well as the

consequences of these different crises and disasters

are numerous and cause several casualties and sev-

eral losses (for example, car accidents, incidents, the

Asian tsunami, the Mumbai attacks, terrorist attacks,

Nepal Disasters, Afghanistan Disaster, etc.)(Boin,

2009). These Crises and disasters can get worse in

some ways, and the ability to cope with some of

these adverse events increases. There may be differ-

ent types of social entities trying to cope with crises.

Among the areas such as individuals, households,

groups and societies, ofﬁcial organizations, both pri-

vate and public (Quarantelli, 1988).

In good crisis management, special tactics and

hard work are used to deal with eventualities of the

present situation or that arise during an emergency.

Also crisis management relies to a large extent on the

application of tactics speciﬁcally adapted to unfore-

seen situations of a given community disaster. The

change of adversity in this area poses new challenges

for many organizations and individuals, such as gov-

ernments, crisis management organizations, social

workers, etc. It also creates a new research agenda for

students and researchers. This also opens the doors

and gives opportunities to apply several technical and

theoretical approaches of several scientiﬁc ﬁelds, in

order to solve some problems in the sector of crisis

and disaster management. Among these areas that can

be applied are the tools and technologies of big data

and machine learning.

Big data is now developing rapidly in all areas of

science and engineering. Learning this massive data

offers signiﬁcant opportunities and transformative po-

tential for various sectors (Qiu et al., 2016). Espe-

cially with the use of these massive data with big data

techniques, in machine learning algorithms. The ﬁeld

of crisis management systems is among the different

ﬁelds of application of big data and machine learning

(Alpaydin, 2009) (Carbonell et al., 1983), and it still

need several improvements.

Problematic. Although the ﬂow of information

within a crisis and disaster management system is

much greater, and decisions during times of crises

may be more difﬁcult to take. By taking into account

all possible information on previous crises and dis-

asters, we can improve the decisions that need to be

made in case of crises or disasters to help decision

makers in the ﬁeld of crisis management.

Research Questions. During this work, we have

several research questions that can be asked, and we

want to try to answer them, among these questions:

What are the difﬁculties that can be encountered in

making decisions about crisis and disaster manage-

Boumahdi, A., El Hamlaoui, M. and Nassar, M.

Crisis Management Systems: Big Data and Machine Learning Approach.

DOI: 10.5220/0009790406030610

In Proceedings of the 15th International Conference on Evaluation of Novel Approaches to Software Engineering (ENASE 2020), pages 603-610

ISBN: 978-989-758-421-3

603

ment?

How can we improve the decisions of crisis manage-

ment systems using machine learning and big data?

What are the machine learning and big data ap-

proaches proposed in the ﬁeld of crisis management?

how can we improve them?

What data are available on crisis management? How

can we use them to improve the decisions of crisis

management systems?

Methodology. In this work, a literature review

was conducted to build a state of the art in the ﬁeld of

machine learning and big data in crisis management

systems. Then we made a discussion based on this

state of the art in this area. Then, and at the base of the

literature review, we try to propose a machine learn-

ing and big data approach in this ﬁeld, with a detailed

description and an experiment. Then the analysis of

the results and suggestion of the next works.

Collection of Papers. In order to collect docu-

ments for this work, we conducted a search in paper

databases and forums for crisis management and in-

formation systems, and databases on machine learn-

ing and big data. With the documents of the interna-

tional conferences in this ﬁeld. In addition, we also

conducted web searches to identify relevant informa-

tion that did not appear in academic papers.

These documents were chosen according to their

originality and their contributions in the ﬁeld of crisis

management systems, taking into consideration the

application of the tools of machine learning and big

data. In terms of the appearance of machine learning

in the ﬁeld of crisis management systems, we have

taken the documents published since 2005 to ﬁnd out

the evolution of this ﬁeld. But in terms of the results

obtained, we have taken recently published articles

and documents, for example from 2017 to the present,

in order to be able to understand and compare these

results obtained

2 STATE OF THE ART

2.1 Big Data in Crisis Management

The era of Big Data has begun to cover all aspects of

our life. With the wide attention in the ﬁeld of Big

Data, a large number of new technologies have be-

gun to emerge, and these technologies will become

important tools for the acquisition, storage, and anal-

ysis of Big Data. According to the theory of Crisis

Management, Big Data can lead the direction of a po-

tential crisis management and has created opportuni-

ties for it to improve and control through the analysis

of crisis information. Scientists and analysts are fac-

ing one of the biggest challenges of managing large

volumes of data generated at times of disasters and

crisis. Therefore, the role of big data in disaster and

crisis management has been evolving.

In (Bellomo et al., 2016) N. Bellomoa et al. pro-

pose an essay concerning the understanding of hu-

man behaviours and crisis management of crowds in

extreme situations using Big Data to deal with De-

cision Making toward Information Management and

Crisis Response. They said that the overview on

crowd dynamics and safety problems presented has

shown that the literature in this ﬁeld can give valu-

able contributions to the crisis management of human

crowds in evacuation situations. In (Watson et al.,

2017) Hayley Watson et al. present ﬁndings from a

case study of big data roadmap, and supports ﬁndings

from other studies in that big data is able to contribute

to crisis response efforts. They said that the litera-

ture has demonstrated how the increased use of dis-

parate datasets can positively affect preparedness and

response to crises and disasters, in particular, the use

of big data. So that can aid decision making within re-

sponse, activities emerges as an important beneﬁt of

big data in this sector.

In (Ma and Zhang, 2017) Yefeng Maa and Hui

Zhangb study the problem of information man-

agement at China’s Emergency Operations Center

(EOC), and aim to propose a data-driven knowledge

management system (KMS) supporting decision mak-

ing, coordination, and collaboration within EOCs and

with the public, whose the big data analytics is em-

ployed. They said that Big data analytics is incor-

porated to knowledge management process in order

to improve the capability of data processing and cri-

sis situations. Their case study shows that the pro-

posed knowledge management solution is helpful for

improving situation awareness and decision making

when dealing with social security incident. In (Doka

et al., 2017) Katerina Doka et al. have presented a

storage and processing platform, that is able to sup-

port applications and services that use the power of

Big Data produced by mobile and social network

users to detect and manage emergencies. This sys-

tem uses collection and analysis of mobile and social

networking data before, during and after a disaster.

They focused on the scalability of the Big Data frame-

works.

In (Akter and Wamba, 2017) Shahriar Akter and

Samuel FossoWamba examine big data in Disas-

ter Management to present main contributions, gaps,

challenges and future research agenda, based on a sys-

tematic literature review. Their study aims to con-

tribute to a better understanding of the importance

ENASE 2020 - 15th International Conference on Evaluation of Novel Approaches to Software Engineering

604

of big data in disaster management. They said that

decision makers need to address various challenges,

such as crisis analytics platform, data governance,

data quality, analytics capabilities, etc... As limita-

tions of their work, they said that they considered only

papers that satisﬁed some speciﬁc criteria, they con-

sidered coherent and inherent with Disaster Manage-

ment. For example, they did not include unpublished

works, book chapters and conference proceedings.

2.2 Machine Learning in Crisis

Management

Technology changes exponentially, whereas organi-

zations change logarithmically (Martec’s law). That

is, the technology in the ﬁeld of crisis management is

evolving very quickly and these changes seem to be

accelerating. While the change of an organization, its

way of thinking and behaving is always difﬁcult and

slow (Dugdale et al., 2019). There are many more

other works to review and use in this ﬁeld, but the

most important are the most relevant ones that can

give us an idea about how to improve crisis man-

agement systems by using the big data and machine

learning, based on the established work. We will still

need to cite some criteria that can help us to com-

pare, or at least understand the machine learning and

big data models proposed. These criteria will also be

inspired by some works established in this context,

namely (Zagorecki et al., 2013) and (Lin et al., 2011)

and others.

In (van Someren et al., 2005) Robert de Hoog

and Guido Bruinsma address the problem of select-

ing and distributing information to users in function

of their characteristics. In particular, they propose a

trainable system for information distribution and ex-

pect that this will reduce problems due to informa-

tion overload, and will give more effective collabora-

tion between all actors in the crisis management sys-

tems environment. The idea was to study the use of

machine learning methods to automatically construct

context-speciﬁc task proﬁles, and use a description of

user activities to guide information distribution and

system training. In (Schulz et al., 2012) Axel Schulz

et al. show an approach for turning massive amounts

of unstructured citizen-generated in making better in-

formed decisions, and aims the utility of Linked Open

Data and crowd sourcing for processing data from so-

cial media. Their goal was to minimize the manual

efforts for ﬁltering information by introducing ma-

chine learning methods such as clustering and trained

classiﬁers, by taking into consideration three steps:

Information collection, information classiﬁcation and

information enrichment.

In (Zagorecki et al., 2013) Adam T. Zagorecki et

al. have discussed the application of data mining and

machine learning techniques to support the decision

making processes for the disaster and crisis manage-

ment. They said that the challenge of applying DM

and ML methods to disaster and crisis management

starts at the point of the identiﬁcation of useful data

that can be exploited by the used methods, this data

could be static when it is collected prior to the dis-

aster event or dynamic when it is a real-time event

that is produced during the disaster. Moreover, they

have discussed the problem of using simulated data

instead of using real data during the processing. In

(Schulz et al., 2013) Axel Schulz at al. have talked

about using tweets with emotions for crisis manage-

ment, by showing the systematic evaluation of an ap-

proach for sentiment analysis on micro-posts that al-

lows detecting seven emotion classes. They noted a

remark from results that the accuracy and the recall of

an approach using tweets labeled with a 7-class senti-

ment classiﬁer are better than simply using tweet with

negative sentiment using a 3-class sentiment classi-

ﬁer. In (Nguyen et al., 2016) Dat Tien Nguyen et al.

have proposed to use Deep Neural Network (DNN) to

identify informative tweets and classify them into top-

ical classes by using a new online algorithm based on

stochastic gradient descent to train DNNs in an online

fashion during disaster situations. The results showed

that the performance of the model in Binary Classiﬁ-

cation model is quite inconsistent as the size of the

in-event training data varies and dropped when the

training size is between 2200 to 3900 tweets. In ad-

dition, the performance of the model in Multi-Class

Classiﬁcation run provides very low accuracy at the

ﬁrst training and continue to drop until a good num-

ber of training examples.

In (Lanfranchi, 2017) Vitaveska Lanfranchi

presents an analysis of the ethical risks and implica-

tions of using automated system that learn from social

media data, to provide intelligence in crisis manage-

ment, and shows a short overview on the use of social

media data in crisis management and its implication

of machine learning and social media data by using

an example of a scenario. They reported that the most

used sources of information for fast and agile cri-

sis information are probably social media or crowd-

sourced data, and the combination of social data and

machine learning algorithms to understand and ﬁlter

the data has high ethical implications. They said also

that the system will be pretrained on an existing cor-

pus of Twitter data related to Crisis and Emergency

Management, by performing sentiment analysis, per-

forming social network analysis, use a network visu-

alization and starting following users that post useful

Crisis Management Systems: Big Data and Machine Learning Approach

605

contents.

In (Lagerstrom et al., 2016) Ryan Lagerstrom et

al. have proposed an approach for image classiﬁca-

tion using low-level features built on pretrained clas-

siﬁers, in the context of ﬁre emergency in the Aus-

tralian state. They used a convolutional neural net-

works for feature extraction and Random Forests for

classiﬁcation to detect ﬁres on images to help man-

aging emergencies. They showed that these method-

ologies could classify images into ﬁre and not not-ﬁre

classes with a good accuracy. In (Giannakeris et al.,

2018) Panagiotis Giannakeris et al. presented a detec-

tion approach for classifying objects (ﬂood, ﬁre, etc.)

in disaster scenarios in order to build a warning sys-

tem framework for detecting people and vehicles in

danger. They used transfer learning to classify im-

age in this context, and they have achieved a good

accuracy. In (Muhammad et al., 2018) Khan Muham-

mad et al. proposed an early ﬁre detection framework

using convolutional neural networks for surveillance

cameras to detect ﬁre in the indoor and outdoor en-

vironments. They also used transfer learning based

on AlexNet to classify images and detect ﬁre disaster,

which raised a good accuracy.

In (Arru et al., 2019) Maude Arru et al. present

a method of data analysis that helps crisis decision

makers to determine whether or not to alert the popu-

lation in a likely crisis situation. The work describes

four step decision support process, with the use of de-

cision trees, which will help provide decision makers

with an indication of the likely behavior of a popula-

tion in response to an alert. By using this informa-

tion, and by focusing on the users and analyzing the

behavior of the population in the event of a crisis, they

said that they can then decide whether or not to trigger

an alert of crisis. In (Buettner and Baumgartl, 2019)

Ricardo Buettner and Hermann Baumgarti examine

how deep learning can be applied to evacuation situa-

tions. They show how artiﬁcial agents can recognize

exhaust panels, such as doors and stairs for evacuation

route planning. In this context, they used a network of

convolutional neurons for image recognition in emer-

gency situations, with an interesting result and a high

accuracy of recognition, which surpasses the current

methods.

2.3 Discussion

During the review of the work established in the ﬁeld

of application of machine learning in crisis manage-

ment, and taking into consideration the criteria that

we have talked about, we can say that there is a lot

of effort and many ideas applied in this area. At the

level of models and approaches built by researchers,

we see that the most common machine learning mod-

els in this ﬁeld are binary and multi-class classiﬁca-

tion methods (using NBB, NBM, etc..), and modern

deep learning techniques such as DNN, CNN and Bi-

LSTM, with other multiple techniques such as NLP,

DM and role-task framework. It appears that the most

frequent works are those that use a classiﬁcation (bi-

nary or multi-class) and one of deep learning tech-

niques (DNN, CNN, ...). These methods and models

have been suggested in this area just in recent years,

since some of them have proposed models with their

experiments and results, and others have just pro-

posed models without doing an experiment, to make

them in future work.

For works that have proposed machine learning

models with their experiments and results, we note

that the measurements used in the experiments are not

uniﬁed, standard performance measures such as preci-

sion recall, ROC and others are always present, along-

side other measures such as F-Measure, AUC and F1-

score, as well as general measures like average per-

formance, Completeness, etc.. For the sources of the

collected data, it appears that the most frequent source

is Twitter, and this by using the tweets of the users

during the crises. Also the intervention of the Volun-

teers and human-participation during the labelization

of the data. It should also be noted that most of the

datasets used in these works are not very wide, espe-

cially those containing only labeled tweets, and that

large datasets usually contain simulated data and not

real data. This may lead us to think about dataset

types and data quality used in experiments. Also,

there are works that have reported that their models

and experiments can be improved, either by the good

choice of datasets or by the improvement of the learn-

ing of the model or by the change and the modiﬁca-

tion of the parameters.

3 BIG DATA AND MACHINE

LEARNING APPROACH

ADOPTED

Seeing the approaches mentioned in these previous

studies, and taking into consideration the possibility

of improving each approach with the nature of cri-

sis management systems, we can propose an approach

that essentially serves to classify decisions (knowing

their types previously). The datasets used in the train-

ing should preferably be the reports generated from

previous real crises, with the possibility of using in-

formative data of Twitter users for example.

In general, crisis management systems generate

ENASE 2020 - 15th International Conference on Evaluation of Novel Approaches to Software Engineering

606

reports that give more information about the method

of crisis management and the plan that was used, and

also gives an opinion on this management method

(good, bad, ...). This information includes: the level

of the crisis (Response Level), the management time,

the percentage of resources used, etc. This approach

(Figure 1) can be rectiﬁed later in this work or in

subsequent works, depending on the experimental re-

sults. The major problem that can surely be posed is

the existence of a good real dataset that can help the

model to have a good learning.

Figure 1: Machine Learning Approach.

3.1 Which Datasets to Use?

The phase of choosing the database is among the most

important phases in a machine learning approach or

project. It is enough that the absence of a useful

database can prevent the success of the whole ap-

proach, and using a good dataset can provide an op-

portunity to get good results and have useful interpre-

tations especially when the data is real. In our case,

it will be preferable that the dataset has large dimen-

sions and will be compatible with the characteristics

of big data that we know. Also the databases used

in this type of domain must contain real information

on crisis management systems, such as crisis manage-

ment plans and reports, and even decisions made by

these systems.

3.2 Extracting and Selecting Features

Feature extraction is a universal problem in machine

learning, because it is very critical to understand what

are the important and necessary aspects of the prob-

lem to solve. In another way, we need to know what

are the criteria and the essential components of the

data to use and to take into account during the imple-

mentation and exploitation of the data to have a good

learning. Features extraction depends heavily on the

chosen database, generally, we take at the beginning

the majority of the database components as features

and sometimes we take them all, especially when the

components are not much. In our case, where the

databases can be reports from crisis management sys-

tem or crisis data from the users, the features depend

on the information and criteria given by these reports,

so they can be like crisis management time, resources,

crisis levels, type of crises, informative data, etc.

When extracting features, this extraction is more

or less correct to make a good learning, because these

features contain information on the data used but

sometimes they contain information that is not nec-

essarily important. Whence comes the idea that it

is not necessary to take all extracted features. The

most important thing is to take as much information

as possible, and it will be better to do that with a few

features. The idea of this step is to make a selection

of a subset of input variables of a learning algorithm

while ignoring the rest. In other words, we can talk

about a reduction in dimension. Because the exper-

iments shows that when we use data with very high

dimension, models usually choke because the time of

training increase exponentially with a high number of

features, and models begin to have a risk of overﬁting.

In our case, the features extracted from the database

usually depend on the information given by the cri-

sis management reports, these features are related to

the management time, the resources needed, the crisis

levels, the types, and other variables. The reduction

of the dimensions of these variables can give us only

the most useful features in learning,and can give us

the types of output of the model and the components

of the decisions.

3.3 Learning Models?

Taking into consideration the deﬁnitions and proper-

ties of machine learning, its types and characteristics,

doesn’t necessarily tell us which machine learning

model to use. However, it give us an intuition on how

these models work which may leave us in the hassle

of choosing the suitable model for our problem.

In our case, we can say that they exist several

kinds of making a model from this machine learn-

ing approach. However, the closest model is to make

a classiﬁcation model. Because we are talking here

about generation of decisions, and the method of gen-

erating these decisions, until now in our work, con-

sists in knowing the classes of certain decisions or

the components of these decisions. Provided that we

know beforehand the types of the classes of the deci-

sions (for example the level, the priority, ..). We can

also do a logistic regression model, to estimate the

percentage of some decisions (for example the per-

centage of resources needed to manage a given crisis).

Crisis Management Systems: Big Data and Machine Learning Approach

607

4 MODEL EXPERIMENTATION

4.1 Dataset

During our search for datasets available in this area,

we took into consideration the preferred type of data

on crises, like reports on real crises. This type of data

can be generated by crisis management organizations,

and is not always free.

There are many data on crises, either natural or

man-made crises. As an example, we used a database

on natural crises in the country of Afghanistan in

a given period of time (2016-2018 for example).

These information includes assessment ﬁgures from

Red Crescent Societies, national NGOs, international

NGOs, and others, and are based on the reports re-

ceived. This dataset provides information on natural

crises in Afghanistan, such as the type of crisis, the

number of people killed by each crisis, the number of

homes destroyed, and so on.

Another dataset that may be useful to us, in the

same context of the natural crises in the country of

Afghanistan, and more precisely on the response to

the ﬂoods. This dataset

presents information on the

degree of recovery, including the level and type of

coping strategies used by both assisted and unaided

households since the 2014 ﬂoods.

A third dataset that we use in the same con-

text is ”Nepal - Disaster data from 1991-2010”

witch includes the disasters that occurred in Nepal

published by Open Database License (ODC-ODbL).

This dataset gives some information on these crises

in Nepal, as the type of the crises, the human victims

(deaths, missing, injured ..), the destroyed houses, etc.

4.2 Generating Features, Choosing a

Model

First, a classiﬁcation model was used to classify the

type of a crisis using the dataset that provides infor-

mation on natural crises in Afghanistan. For exam-

ple, we chose to classify two types of crisis using the

information given. Knowing the type of crisis from

certain information (such as the number of victims

and houses affected) allows us to have an idea about

the priority of managing this crisis, and also to have

an idea about the resources needed to use during the

management of this crisis. The interest of this classi-

ﬁcation is to be able to extract a component of a de-

cision towards a crisis. This component is the type of

a crisis, according to some information about natural

https://data.humdata.org/organization/reach-initiative

http://www.desinventar.net/DesInventar/report.jsp

crises such as people killed, affected people, affected

families, damaged homes, destroyed homes, etc. And

it is up to us to choose beforehand the classes of the

types to work with (of the ﬁrst case we choose two

types of classes to carry out a binary classiﬁcation).

Then, another classiﬁcation was made based on

the numeric features of the second dataset. This clas-

siﬁcation aims to classify the level of a natural crisis

(i.e. low community damage, and high community

damage). This classiﬁcation of crisis levels allows to

have an idea about the priority of such a crisis. Also,

and according to the available data, it allows to give

an opinion and a judgment on the intervention of the

systems of crisis management towards a given crisis.

The interest of this classiﬁcation is to obtain another

component of a decision about a crisis. This compo-

nent is to know the degree of damage of a crisis on a

given community, based on information obtained dur-

ing an evaluation carried out on the communities af-

fected by the crisis. Knowing the degree of damage

of a crisis from some information on this crisis (such

as demographics, vulnerable groups, expenses, etc.)

will allow us to take into consideration the degree of

damage to know the priority of each case of a crisis,

and to improve reactions in a similar situation.

4.3 Results

We have chosen features, then train the model using

these features and modify the choice of these accord-

ing to the results obtained. The ﬁrst choice, which

is obvious, is to use all the numeric features. The

choices that follow are based on the results obtained

and the importance of the information contained in

each feature in relation to the desired decisions. We

chose the accuracy to measure the performance of the

classiﬁcation model, because it is simple and gives

us the desired interpretation of the degree of perfor-

mance, and it serves to know how many samples did

the model correctly label among all the samples. So

we can summarize the results of the classiﬁcation in

two stages. One for the component of the type of

crises (table 1), and the other for the classiﬁcation of

the degree of damage on the community (table 2).

4.4 Discussion

Based on the results of the experimental part of the

model, we can draw some essential remarks. The

choice of the two decision components used to man-

age crises was or classify the type of crisis, and the

degree of damage to the community, for the reasons

given in the previous sections. The results obtained

during the training of the models are acceptable and

ENASE 2020 - 15th International Conference on Evaluation of Novel Approaches to Software Engineering

608

Table 1.

Dataset Features Accuracy

All numeric features 0.3101

Dataset 1 Personskilled 0.2658

Houses-destroyed

Families-displaced

Families-affected 0.7468

Individuals-affected

Deaths , Injured ,Missing 0.7760

Dataset 2 ,Victims , Affected

Deaths ,Victims ,Affected 0.7634

Table 2.

Dataset Features Accuracy

All numeric features 0.6539

Dataset 3 expenditures-health 0.6477

demographics-household

expenditures-transport

are different meanings. In the case where we have

chosen to classify the type of crises using all numer-

ical features, we notice a very low accuracy of the

model. This is due to the large number of columns

used in learning, which interferes with the model

learning and that is proven in the ﬁeld of machine

learning.

When we chose to use a smaller number of fea-

tures (which summarizes the largest amount of data)

such as crisis-affected families and crisis-affected in-

dividuals in the ﬁrst dataset (disasters in Afghanistan)

we had acceptable results. Also in the second dataset

(disasters in Nepal), when we did a learning with

a small number of features that are relevant, like

Deaths, Injured, Missing, Victims, Affected, we got

almost the same results. So it can be said that even the

two datasets used in the classiﬁcation of the type of

crisis are different and represent information on crises

in different countries, the type classiﬁcation of natu-

ral crises can be done using almost the same features.

These features represent information about the indi-

viduals and families affected by these crises and the

victims of these crises.

In the classiﬁcation of the other component of de-

cisions, which is the degree of damage to the com-

munity, the use during the learning of all the numeric

features, and the use of a small number of these fea-

tures, give almost the same results. This can be ex-

plained by the small number of columns used (which

does not give a big difference even if we perform a

dimension reduction), and also the large number of

records that are larger than the ﬁrst two datasets.

5 CONCLUSION

The application of big data tools and machine learn-

ing techniques and algorithms in several ﬁelds gives

in most cases good results that are useful in each area,

which give something more, and especially who is

able to be improved every time. In this work, we

have seen and shown that the area of crisis and dis-

aster management is not an exception in this context.

As there is a lot of work going on in this area to apply

big data and machine learning, using different types

of data and different data sources. These works also

have the possibility to be improved to obtain better

results, this is according to the researchers who pub-

lished in this ﬁeld, and according to the results that

they obtained.

According to our knowledge in the ﬁeld of ma-

chine learning, and according to the machine learning

work published in this ﬁeld, the results obtained by

this approach remain acceptable, especially as a ﬁrst

attempt. And the proposed approach remains ﬂexible

to be modiﬁed and adapted to the conditions of the

crisis management area. also the approach and this

work in general has many prospects to be carried out

as works of future. Although the datasets used in this

work are real and present speciﬁc crises and disas-

ters, we need, in the context of perspectives and future

work, to use several other datasets on other crises and

disasters to improve the accuracy rate of these mod-

els in order to be able to extract good decisions. And

also using datasets of images of crises during training

to get a good performance.

The biggest perspective in our work is to continue

to look into the ﬁeld of machine learning applied to

crisis management systems, and to the ﬁeld of crises

in general. They remain a lot of things to apply at the

level of management and also the prediction of crises

and minimization of risks. Especially when there is

the possibility to exploit dynamic and varied data such

as data from social networks. In addition, there still

many things to apply from big data techniques, such

as massive data storage techniques, parallel and real-

time processing and calculations, and massive data

quality measurements.

The ﬁeld of studies of car accidents and road ac-

cidents is a very interesting area to apply approaches

based on machine learning and big data. This is due

to several factors, such as the problem of the increase

of road accidents, and also the necessity to exploit

the existing data on these types of crises in several

countries. we can say the same things about the ﬁelds

of local and global epidemics, as well as earthquakes

and ﬂoods and forest ﬁres. It will also be interesting

and useful to use these types of data and other types

Crisis Management Systems: Big Data and Machine Learning Approach

609

of data, to make a predictive model that can predict

crises before they occur.

REFERENCES

Akter, S. and Wamba, S. F. (2017). Big data and disaster

management: a systematic review and agenda for fu-

ture research. Annals of Operations Research, pages

1–21.

Alpaydin, E. (2009). Introduction to machine learning.

MIT press.

Arru, M., Negre, E., and Rosenthal-Sabroux, C. (2019). To

alert or not to alert? that is the question. In Proceed-

ings of the 52nd Hawaii International Conference on

System Sciences.

Bellomo, N., Clarke, D., Gibelli, L., Townsend, P., and

Vreugdenhil, B. (2016). Human behaviours in evacu-

ation crowd dynamics: from modelling to “big data”

toward crisis management. Physics of life reviews,

18:1–21.

Boin, A. (2009). The new world of crises and crisis man-

agement: Implications for policymaking and research.

Review of Policy research, 26(4):367–377.

Buettner, R. and Baumgartl, H. (2019). A highly effective

deep learning based escape route recognition module

for autonomous robots in crisis and emergency situ-

ations. In Proceedings of the 52nd Hawaii Interna-

tional Conference on System Sciences.

Carbonell, J. G., Michalski, R. S., and Mitchell, T. M.

(1983). An overview of machine learning. In Machine

learning, pages 3–23. Elsevier.

Doka, K., Mytilinis, I., Giannakopoulos, I., Konstanti-

nou, I., Tsitsigkos, D., Terrovitis, M., and Koziris,

N. (2017). Exploiting social networking and mobile

data for crisis detection and management. In Interna-

tional Conference on Information Systems for Crisis

Response and Management in Mediterranean Coun-

tries, pages 28–40. Springer.

Dugdale, J., Negre, E., and Turoff, M. (2019). Introduction

to the minitrack on ict and artiﬁcial intelligence for

crisis and emergency management. In Proceedings of

the 52nd Hawaii International Conference on System

Sciences.

Giannakeris, P., Avgerinakis, K., Karakostas, A., Vrochidis,

S., and Kompatsiaris, I. (2018). People and vehicles

in danger-a ﬁre and ﬂood detection system in social

media. In 2018 IEEE 13th Image, Video, and Mul-

tidimensional Signal Processing Workshop (IVMSP),

pages 1–5. IEEE.

Lagerstrom, R., Arzhaeva, Y., Szul, P., Obst, O., Power, R.,

Robinson, B., and Bednarz, T. (2016). Image clas-

siﬁcation to support emergency situation awareness.

Frontiers in Robotics and AI, 3:54.

Lanfranchi, V. (2017). Machine learning and social media

in crisis management: Agility vs ethics. In Proceed-

ings of the 14th International Conference on Informa-

tion Systems for Crisis Response and Management.

IMT Mines Albi-Carmaux (

Ecole Mines-T

ecom).

Lin, W.-Y., Hu, Y.-H., and Tsai, C.-F. (2011). Machine

learning in ﬁnancial crisis prediction: a survey. IEEE

Transactions on Systems, Man, and Cybernetics, Part

C (Applications and Reviews), 42(4):421–436.

Ma, Y. and Zhang, H. (2017). Enhancing knowledge man-

agement and decision-making capability of china’s

emergency operations center using big data. Intelli-

gent Automation & Soft Computing, pages 1–8.

Muhammad, K., Ahmad, J., and Baik, S. W. (2018). Early

ﬁre detection using convolutional neural networks

during surveillance for effective disaster management.

Neurocomputing, 288:30–42.

Nguyen, D. T., Joty, S., Imran, M., Sajjad, H., and Mitra,

P. (2016). Applications of online deep learning for

crisis response using social media information. arXiv

preprint arXiv:1610.01030.

Qiu, J., Wu, Q., Ding, G., Xu, Y., and Feng, S. (2016). A

survey of machine learning for big data processing.

EURASIP Journal on Advances in Signal Processing,

2016(1):67.

Quarantelli, E. L. (1988). Disaster crisis management: A

summary of research ﬁndings. Journal of manage-

ment studies, 25(4):373–385.

Schulz, A., Paulheim, H., and Probst, F. (2012). Crisis in-

formation management in the web 3.0 age. Proceed-

ings of ISCRAM.

Schulz, A., Thanh, T. D., Paulheim, H., and Schweizer, I.

(2013). A ﬁne-grained sentiment analysis approach

for detecting crisis related microposts. In ISCRAM.

van Someren, M., Netten, N., Evers, V., Cramer, H.,

de Hoog, R., Bruinsma, G., et al. (2005). A trainable

information distribution system to support crisis man-

agement. In 2nd International Conference on Infor-

mation Systems for Crisis Response and Management

(ISCRAM). Citeseer.

Watson, H., Finn, R. L., and Wadhwa, K. (2017). Organi-

zational and societal impacts of big data in crisis man-

agement. Journal of Contingencies and Crisis Man-

agement, 25(1):15–22.

Zagorecki, A. T., Johnson, D. E., and Ristvej, J. (2013).

Data mining and machine learning in the context of

disaster and crisis management. International Journal

of Emergency Management, 9(4):351–365.

ENASE 2020 - 15th International Conference on Evaluation of Novel Approaches to Software Engineering

610