International Journal on Computational Science Applications (IJCSA)

Scope & Topics

Computational science is interdisciplinary fields in which mathematical models are combined with scientific computing methods to study of a wide range of problems in science and engineering. Modeling and simulation tools find increasing applications not only in fundamental research, but also in real-world design and industry applications.

Topics of interest include but are not limited to, the following

  • Modeling, Algorithms and Simulations
  • Problem Solving
  • Scientific Computing
  • Computational Sciences (Biology, Chemistry, mathematics, physics, forensics, etc.)
  • High Performance Computing
  • Machine Learning
  • Network Analysis
  • Computer Architecture
  • Visualization and Virtual Reality as Applied to Computational Science
  • Architectures and computation models, compiler, hardware and OS issues
  • Large Scale Scientific Instruments
  • Memory system, I/O, tools, programming environment and language supports

Paper Submission

Authors are invited to submit papers for this journal through Submission system. Submissions must be original and should not have been published previously or be under consideration for publication while being evaluated for this Journal.

Important Dates

  • Submission Deadline    :  March 25, 2023
  • Notification                  :  April 25, 2023
  • Final Manuscript Due   : April 28, 2023
  • Publication Date  : Determined by the Editor-in-Chief

International Journal on Computational Science Applications (IJCSA)

Scope & Topics

Computational science is interdisciplinary fields in which mathematical models are combined with scientific computing methods to study of a wide range of problems in science and engineering. Modeling and simulation tools find increasing applications not only in fundamental research, but also in real-world design and industry applications.

Topics of interest include but are not limited to, the following

  • Modeling, Algorithms and Simulations
  • Problem Solving
  • Scientific Computing
  • Computational Sciences (Biology, Chemistry, mathematics, physics, forensics, etc.)
  • High Performance Computing
  • Machine Learning
  • Network Analysis
  • Computer Architecture
  • Visualization and Virtual Reality as Applied to Computational Science
  • Architectures and computation models, compiler, hardware and OS issues
  • Large Scale Scientific Instruments
  • Memory system, I/O, tools, programming environment and language supports

Paper Submission

Authors are invited to submit papers for this journal through Submission system. Submissions must be original and should not have been published previously or be under consideration for publication while being evaluated for this Journal.

Important Dates

  • Submission Deadline    :  March 11, 2023
  • Notification                  :  April 11, 2023
  • Final Manuscript Due   : April 19, 2023
  • Publication Date  : Determined by the Editor-in-Chief


Jasleen Kour1, Saboor Koul2 and Prince Zahid3

1Assistant Professor, Department of CSE, BGSBU, Rajouri, J&K, India
2,3UG Scholars, Department of CSE, BGSBU, Rajouri, J&K, India



Quantum cryptography is based on quantum mechanics to guarantee secure communication. It allows two parties to produce a shared random bit string known only to them. These random bits can be used as a key to encrypt and decrypt messages. The most important and unique property of quantum cryptography is the ability of the two communicating users to detect the presence of any third party trying to gain knowledge of the key. It is based on fundamental aspects of quantum mechanics. By using quantum entanglement or quantum super positions and transmitting information in quantum states, a communication system can be implemented which detects eavesdropping. Quantum cryptography is used to produce and distribute a key, not to transmit any message data. This key along with certain encryption algorithm, is used to encrypt (and decrypt) a message, which can then be transmitted over a standard communication channel. This paper concentrates on comparison between classical and quantum cryptography as well as survey on various quantum key distribution protocols used to generate and distribute the key among communicating parties.


Quantum Cryptography, QKD, bits, photons and qubits.


Data communication security can be defined as complex process that implies networks, users and applications, all of these connected by a set of modern technologies. So, Information systems are very vulnerable to attacks and illegitimate penetrations, to data incidental or intended data. Cryptography is a concept to protect the information transmission over such networks.

Cryptography is the science where the use of mathematics occurs to encrypt and decrypt data. It enables user to store sensitive information or transmit it across insecure networks (like the Internet) so that it cannot be interpreted by anyone other than the intended receiver. To achieve such level of security different algorithms are used for secure transmissions which unite the message with additional information to produce a cryptogram. These algorithms are known as Cipher and the additional information is known as the key. This method is termed as encryption. Whereas cryptanalysis is the science of analyzing and breaking secure communication without knowing the encryption technique. Classical cryptanalysis involves an interesting combination of analytical reasoning, pattern finding, patience, determination, and luck. Quantum cryptographic devices generally make use of individual photons of light and take benefit of Heisenberg’s Uncertainty Principle, according to this principle cryptographic protocols can invoke up streams of random bits whose values will remain unknown to third parties. When we use these bits as key material for Vernam ciphers, we can get Shannon’s ideal of perfect secrecy—cheaply and easily. The development of quantum cryptography was inspired by some limitations of classical cryptography methods. In classical cryptography, communicating parties share a secret sequence of random numbers, the key, that is exchanged by some physical mean and thus open to security loopholes. The classical cryptography does not detect eavesdropping like quantum cryptography, also with increase in computing power and new computational techniques are developed, the numerical keys will no longer be able to provide satisfactory levels of secure communications. These weaknesses led to the development of quantum cryptography, whose security basis is quantum mechanics. This paper presents the comparison of quantum and classical cryptography on several background, quantum cryptography key protocols and real world application of quantum cryptography.


Both quantum cryptography and classical cryptography can be compared on following

2.1. Fundamental Dimension

In theory, any classical private channel can be easily monitored inertly, without the knowledge to sender or receiver that the eavesdropping has been done. Classical physics is the theory of macroscopic bodies and phenomena such as radio signals that allows a physical property of an object to be measured without disturbing other properties. Cryptographic key like information is encoded in computable physical properties of some object or signal. Thus there is open possibility of passive eavesdropping in classical cryptography.

Quantum theory which is basis of quantum cryptography is believed to direct all objects, but its consequences are mainly noticeable in individual atoms or subatomic particles like microscopic systems. As far as classical cryptography is concerned there is frequent requirement of using longer keys as computational power doubles in every 18 months and cost of computation is reducing rapidly with time [moors law]. Thus an algorithm using k bit key which is secure may not be secure in future, i.e. it needs regular updating. On the other hand, security in quantum cryptography is based on the basic principles of quantum mechanics, so the possibilities of major changes requirements for future are almost negligible.

2.2. Commercial dimensions

Commercial solutions for QC that already exist; they are only suitable for point-to-point connections. On the other hand, crypto chip made by the Siemens and Graz technical university makes possible the creation of networks with many participants, and cost of €100,000 per unit, the system is very expensive and requires a lot of work. On other hand classical cryptography can be implemented in software and its cost for consumer is almost zero. Also, cryptographic system based on classical cryptography can be implemented on small hardware component like smart card , but this is major issue in case of quantum cryptography shrinkage to such a level require too much development.

2.3. Technological dimensions

Chinese scientists accomplished the world`s most long-distance of quantum communication transmission (teleportation), or as “instant matter transmission technology” technology. From the China University of Technology and researchers at Tsinghua University, Hefei National Laboratory in their free-space quantum communication experiments, and effectively enlarges the communication distance to 10 miles [9]. But classical cryptography can be used to communication distance of several million miles. According to the latest research, Toshiba achieve new record bit rate for quantum key distribution, that is, 1 Mbit/s on average [10]. On the other hand the bit rate of classical cryptography depends on the computational power largely.

2.4. Other dimensions

Communication medium is not an issue in classical cryptography because its security depends only on the computational complexity. Thus, this removes the need for excessively secure channels. On the other hand communication of quantum cryptography require a quantum channel like optical fiber or through air (wireless), also, there is constantly a likelihood of modification in polarization of photon due to Birefringence effect or rough paths that cause change in refractive index due to damage sometimes. Also, an n-bit classical register can store at any moment exactly one n-bit string. Whereas an n-qubit quantum register can store at any moment a superposition of
all 2n n-bit strings.

Quantum cryptography is based on mixture of concepts from quantum physics and information theory. The security standard in QC is based on theorems in classical information theory and on the Heisenberg’s uncertainty principle. Experiments have demonstrated that keys can be exchanged over distances of a few miles at low bit rate. Its combination with classical secret key cryptographic algorithms permits increasing the confidentiality of data transmissions to an extraordinary high level. From comparison, it’s obvious that quantum cryptography (QC) is having more advantage than Classical Cryptography (CC) though some issues are yet to be solved. This is mainly due to the implementation problems but in future there exist possibilities that most of the problems in quantum cryptography will get resolved.


Quantum key distribution is a key establishment protocol which generates symmetric key material by using quantum properties of light to transfer information from one Client to another Client in a manner which uses the results of quantum mechanics. By using the quantum properties of light, current lasers, fibre-optics and free space transmission technology can be used for QKD (Quantum key distribution), so that many observers claiming security can be based on the law of quantum physics only. Based upon the necessary principles of Quantum mechanics the QKD protocols are divided into two categories some are based on Heisenberg Uncertainty Principles and others are based on quantum entanglement.

3.1. Protocols based on Heisenberg Uncertainty Principle:

3.1.1. BB84 protocol

In 1984 Charles Bennet and Gilles Brassard for the first time proposed a protocol known as BB84 protocol which depends upon the Heisenberg Uncertainty principle. Quantum key Distribution (QKD) is used in quantum cryptography for generating a secret key shared between two parties using a quantum channel and an authenticated classical channel. The private key obtained then used to encrypt message that are sent over an insecure channel (such as a conventional internet connection) as shown fig. below .A bit can be represented by polarising the photon in either of the two bases i.e. Rectilinear base(R) and Diagonal base (B).Binary 0 represents the polarisation of 0° degree in rectilinear base or 45° degree in diagonal base. Similarly binary 1 represents the polarisation of 90° degree in the rectilinear base or 135° degree in diagonal base. [1][5][15][16]


There are two steps involved in key distribution for BB84 protocol, as explained below:
a) One way communication channel (via quantum channel).

Step i) User A (Alice) randomly chosen polarized photon and send it to the user B (Bob) over Quantum channel.

Step ii) In this, user B receives photons using random basis either rectilinear or random

b) Two way communication (via classical channel).
Step i) User A will use classical channel to inform user B about the polarisation A chose for every bit sent to B without disclosing the bit value.

Step ii) Now user will compare the polarisation sequence he receives from user A with the sequence he generated.

Step iii) Bits of same orientation of those two sequences can be used as secret key.

3.1.2. BBM92 protocol

It is the modified version of BB84 protocol which uses only two states instead of four states which were used in BB84 protocol. Charles Bennett, Brassard and Mermin devised another, the so-called BBM92 protocol. They realized that it was not necessary to use two orthogonal bases for encoding and decoding. It turns out that a single non-orthogonal basis can be used instead, without affecting the security of the protocol against eavesdropping. This idea is used in the BB92 protocol, which is otherwise identical to BB84 protocol.


As shown in fig 1B. BB92 above, 0 represents 0°degree in the rectilinear basis and 1 represents 45°degree in the diagonal basis. Client A transmits string of photons to client B which were encrypted by randomly chosen bits just like in BB84 protocol. But here client A chooses the bits by an authoritative rule to which base client B must use. But still Client B chooses randomly a basis by which to measure but if Client B chooses wrong base he will not measure anything; a condition in quantum mechanics which is known as an erasure. After every step Client B tells Client A that the bits end by client A whether or not he measured it correctly.[2][3][7]

3.1.3. SARG04 protocol

SARG04 protocol was proposed by Scarani in 2004. In this protocol the four states of BB84 protocol is used with different information such a new protocol is developed which is capable of performing without failure under a wide range of conditions when attenuated laser pulses instead of single photon source.

Its first phase is similar to BB84 protocol. In the second Phase Client A instead of directly
announcing her base to Client B, it announces a pair of non-orthogonal states one of which client A uses it to encode its bit. Client B measures the correct state if it has used the correct base. Client B will not be able to determine the bit or will not measure the correct states of client A if he chooses the wrong base. The length the key will remain ¼ of the raw key after shifting stage if there are no errors.

The SARG04 protocol has almost same security to BB84 in perfect single-photon implementations, If the quantum channel is of a given visibility (i.e. with losses) then the QBER of SARG04 is twice that of BB84 protocol, and is more sensitive to losses.SARG04 is more secure than BB84 protocol in presence of PNS attacks.[1][4][5][14]


3.2. Protocols based on Quantum Entanglement.

3.2.1 E91 protocol

Ekert, in 1992, performed the process of key distribution through entanglement of photons in a quantum channel. He proposed a method of harnessing Bell`s inequalities. In this method any of the three, A (Alice), B (Bob), or the third party could produce entangled photons. Separation in each pair is such that the communicants, A and B could receive one of each pair. Quantum entanglement means to define the quantum states of one object without referencing the quantum states of another object far away from it. The fact that entangled states are used conceals the information about the key from the eavesdroppers, hence more secure method. The states of particle are not collapsed until the moment of measurement, so trying to access the system is as looking for something that doesn`t exist yet.

Both A and B choose randomly and independently from two different orientations of their analysers to measure the polarization of photons. A typically physical set-up is shown in fig3A: given below, using active polarization rotators (PR), polarizing beam splitter (PBS) and avalanche photodiodes (APD). [8][11][13]


3.2.2. COW protocol

A new protocol, given by Nicolas Gisin et al in 2004, was proposed for QKD based on the weak coherent pulses at high bit rates. The protocol was termed as Coherent One-Way protocol l(COW protocol). The main feature of the method was the setup being experimentally simple and resistant to interference visibility and to photon numbers splitting attacks, hence more efficient in terms of distilled secret bits per qubit.


The figure 4A above represents the COW protocol. The message is encoded in time. Alice sends Coherent pulses that are either empty or have a mean photon number μ < 1. Each logical bit of information is encoded by sequences of two pulses, μ-0 for a logical ―0‖ or 0-μ for a logical ―1‖.

Alice sends decoy sequences μ-μ for security needs. Bob measures the arrival time of the photon on his data-line, detector DB to obtain its key. Bob randomly measure the coherence between successive non-empty pulses, bit sequence ―1 -0‖ or decoy sequence, with interferometer and detectors DM1 and DM2. If wavelength of the laser and the phase in the interferometer are well aligned, we have all detection on DM1 and no detection on DM2. A loss of coherence and therefore a reduction of the visibility reveal the presence of an eavesdropper, in which case the key is simply discarded, hence no information will be lost.[1][11][12][7]

3.2.3. DPS protocol

Differential –phase-shift QKD (DPS-QKD) is a new quantum key distribution scheme that was proposed by K. Inoue et al. Figure 5A mentioned below shows the setup of the DPS-QKD scheme.


From the Alice site, a pulse train of weak coherent states is phase-modulated randomly by {0, π} for every pulse and then to Bob with an average photon number less than one per pulse. From the Bob`s site, the phase difference is measured between the two sequential pulses using one bit delay. Mach-Zender interferometer and photon detectors record the photon arrival time and which detector clicked. Bob, after transmission of the optical pulse train, tells Alice the time at which photon was counted. Alice comes to know from this time information and her modulation data about the detector clicked at Bob`s site. Under an agreement that a click by detector 1 denotes ―0‖ and click by detector 2 denotes ―1‖, for example Alice and Bob obtain an identical bit

The DPS-QKD scheme has certain advantageous features including a simple configuration, efficient time domain use, and robustness against photon number splitting attack. [9][10][11]


The study of quantum cryptography is heavily application-oriented. Therefore, a fundamental question is: Who really needs it rather than classical cryptography? The most significant advantage of quantum cryptography is the forward security. That is, if a classified message is appropriately encrypted with quantum cryptography, its distribution will stay secure as long asquantum mechanics is valid. In contrast, a classified message that is encrypted with classical cryptographic algorithm may stay safe only for a certain period of time. The length of this ―secure period‖ is predictable only if the increase of computational power is predictable. Quantum cryptography can be a favored choice for applications that require long-term information security. There can be a long list of potential (or maybe present) clients. Here we raise a few examples.

4.1. Government agencies.

This includes intelligence, diplomatic, and military agencies. Often under the name of national interest, some information (like some pictures taken in Guantanamo Bay detention camp) is expected to be kept confidential for decades, during which such confidential information may be extensively distributed among different government agencies. Note that the Canadian census data are kept secret for 92 years. Therefore, if we conducted a census in 2009, the data would not be released until 2101, which is the next century! Quantum cryptography can help keep these sensitive data secure during transmission.

4.2. Financial institutes.

Financial information is very sensitive and needs long-time confidentiality. Quantum encrypted links between financial institutes can substantially reduce the risk of leaking the clients’ information during communication.

4.3. Health care providers.

Health care records are being digitized gradually. Digital records of patients are often distributed between different health care providers facilitate medical treatments. The distribution of a patient’s health record may have to be kept secure for the life span of the patient, and quantum cryptography can certainly be of help. Note that quantum cryptography is not the only method to guarantee unconditional communication security. It is not even the only solution to the key distribution problem.


The techniques adapted from classical computer science are applicable to quantum key
distribution protocols is an appropriate sign that quantum cryptography is a rousing new area of research work. In this paper we endeavored to introduce quantum cryptography, QKD protocols and QC applications. It explains about different quantum key distribution protocols. These protocols can be used along with encryption technique to achieve higher level of security. There is much more to describe about quantum cryptography


[1] Abhishek Parakh, (2015), ―New Protocol for Quantum Public Key Cryptography‖, IEEE ANTS 2015 1570203267 .

[2] Otilia Cangea, Carmen Silvia Oprina, Mihai-Octavian Dima, (2016), ―Implementing Quantum Cryptography Algorithms for Data Security‖ ,ECAI 2016 – International Conference – 8th Edition Electronics, Computers and Artificial Intelligence 30 June -02 July, 2016, Ploiesti, ROMÂNIA.

[3] V. Padamvathi, B. Vishnu Vardhan, A.V.N. Krishna,(2016), ―Quantum Cryptography and Quantum Key Distribution Protocols: A Survey‖, Advanced Computing (IACC), 2016 IEEE 6th International Conference.

[4] H. Bennett, (May 1992),‖Quantum cryptography using any two no orthogonal states‖,‖ Phys. Rev. Lett., vol. 68, no. 21, pp. 3121–3124,.

[5] Tamaki, K., M. Koashi, and N. Imoto, (2003), ―Unconditionally secure key distribution based on two non orthogonal states‖,‖ Physical Review Letters 90, 167904 [preprint quant-ph/0210162].

[6] M. Christandl, R. Renner, and A. Ekert, (Mar. 2004 ), ―A generic security proof for quantum key distribution‖,‖ arXiv:quant-ph/0402131.

[7] M. A. Nielsen and I. L. Chuang, (2000), ―Quantum Computation and Quantum Information‖, Cambridge, UK: Cambridge University Press.

[8] Eduin H. Serna, (2009), ―Quantum Key Distribution Protocol with Private-Public Key‖‖, Quantum Physics (quant-ph),

[9] E. Biham, T. Mor, (1997), ―Security of quantum cryptography against collective attacks‖, Physical Review Letters 78 (11) 2256–2259.

[10] Ekert, Artur K.,( August 1991), ―Quantum cryptography based on Bell’s theorem, Physical Review Letters‖, Vol. 67, No. 6, 5, pp 661 – 663.

[11] Inoue K, Woks E and Yamamoto Y, (2002), ―Differential phase shift quantum key distribution‖, Phys. Rev. Lett. 89037902.

[12] M. Lucamarini, S. Mancini, (2005), ―Secure deterministic communication without entanglement, Physics Review Letters‖, 94140501.

[13] M. Elboukhari, M. Azizi, A. Azizi, (2009), ―Implementation of secure key distribution based on quantum cryptography‖‖, in Proc. IEEE Int. Conf Multimedia Computing and Systems (ICMCS’09), pp 361 – 365.

[14] M. Elboukhari, Mostafa Azizi, and Abdelmalek Azizi, (2009), ―Integration of Quantum Key Distribution in the TLS Protocol‖‖, IJCSNS, Vol. 9 No. 12 , pp.21-28,.

[15] N. Lutkenhaus, (1996), ―Security against eavesdropping in quantum cryptography, Physical Review‖, A 54 (1) 97–111.

[16] P. W. Shor and J. Preskill,( July 2000), ―Simple proof of security of the BB84 quantum key distribution protocol‖, Phys. Rev. Lett, vol. 85, no. 2, pp. 441– 444.

[17] R.Hughes,J.Nordholt,D.Derkacs,C.Peterson, (2002), ―Practical free-space quantum key distribution over 10km in daylight and at night‖, New journal of physics 4 (2002)43.1-43.14.URL:

[18] C. H. Bennett and G. Brassard, (1984), ―Quantum cryptography: Public key distribution and coin tossing,‖ in Proc. IEEE Intl. Conf. on Computers, Systems, and Signal Processing, pp. 175–179.

[19] U. Vazirani and T. Vidick, (Sep 2014), ―Fully device-independent quantum key distribution.‖, Phys. Rev. Lett., 113:140501,.

[20] S.-A. Wang and C.-Y. Lu.( Aug 2013), ―Quantum secure direct communication network In Nanotechnology (IEEE-NANO)‖, 13th IEEE Conference, pp 752–755.

[21] Verma Deepankar, Kour Jasleen, (June 2014), ―Image Steganography Based on Hybrid Cryptography  Algorithm Designed by making use of Blowfish, AES, RSA Algorithm‖, IJARCSSE, Vol. 4, Issue 6.


Mohammad Rezwanul Huq, Abdullah-Al-Mosharraf and Khadiza Rahman

Department of Computer Science and Engineering

East West University, Dhaka, Bangladesh


Social media becomes a communicating channel during a natural disaster for detecting disaster events because people share their opinions, feelings, activity during the disaster through the Twitter. Twitter is not simply a platform for broadcasting information, but one of informational interaction. So, we use this platform for mining various disaster relevant tweets during a natural disaster. We examine more than 4,500 tweets during crisis moment. In this paper, we propose a classifier for classifying the disaster phases using social data and identify these types of phases. We use KNN, a machine learning classification algorithm for classifying the disaster relevant tweets. By knowing different phases of a disaster, response teams can detect where disaster will happen; the medical enterprise can be prepared to mitigate the damage after disaster and neighborhood area may also be alert to face the disaster. We classify the disaster-related tweets into three phases that are: pre (preparedness before the disaster event), on (during disaster event), post (impact and recovery after the disaster).We also take the geolocation with latitude and longitude of the disaster event for visualizing it using an earth map which can be useful to emergency response teams and also increase social awareness of the disaster.


Social Media, Natural Disaster, Phase Detection, Machine Learning, Geolocation, Awareness


Social data now has become the popular media for extracting valuable information and a source that may contribute to situational awareness [23]. During the natural calamities people share their experience using tweet, and by seeing this tweet, we can understand that disaster is occurring in a particular area. Whenever Twitter users say rain is occurring or cyclone is acting to this area, by collecting this type of tweet we can detect in which area the natural disaster is occurring. Many researchers examined the disaster-related tweet for the sake of people and relied on a four-phase categorization (preparedness, response, impact, recovery) to mitigate their sufferings [2, 7]. So we  think social network Twitter can be an excellent source for collecting twitter data during a natural disaster. We try to collect disaster relevant data from the Twitter and in previous studies focus on relevant data in disaster phases and extracted relevant tweets into many categories [25]. So the relevant tweet is extracted based on disaster keyword. Our main challenge is building a classifier for different phases and to classify the disaster relevant tweet and to identify various phases during a natural disaster. In this paper, we build a model for classifying the Twitter into three phases (pre, on, post). Moreover, we also show the geographical map of the affected are by visualizing and pointing out that type of different phases [19]. The main objective of our work is to help and inform the Medicine Company, people, and disaster response team about the impact of the catastrophe and to alert them. As we pointed out these phases on the Earth map, the response team can easily understand where and when this disaster is happening by seeing the map, and they can take immediate actions.

1.1. Tweets reflect phases during natural disaster

Twitter users tweet before the disaster for making awareness and during the natural disaster they tweet about the condition of nature, and after the disaster has occurred, they tweet for remediation. Here many types of tweet we find. Certain kinds of the tweets are like “be prepared for the Cyclone Debbie,” “Cyclone Debbie is coming,” and by seeing this kind of tweet, we can say it is a pre-status tweet. During a disaster event, some tweets are like, “it is raining” or “cyclone is occurring, ” and from these tweets, we can say that this place is current disaster zone area. After the disaster, people share their sufferings, impacts and damages happened by the disaster, and the tweets are like “food”, “help”, “injured”, “death”. By seeing these kinds of the tweet, we can say that for remediation food, fund or any kinds of help is needed in that area for the mitigation of damage. Besides tweet also reflect the geographical status of the disaster affected area, and by collecting this location with the GPS, we also pointed out these three phases on an earth map which will help to make situational awareness.

1.2. Motivation

Our main challenge is identifying different phases of the natural disaster using social data. For this reason, we try to classify the disaster relevant data into some phases. This classification provides a framework to predict the pre-status before the disaster event, the on-status during a disaster and the post-status after the disaster has occurred. We select social data for our work because there could be many informative tweets can be found from them. For understanding the disaster event, we classify the disaster data into three phases which are pre (preparedness, public alert, and awareness), on-time (during the disaster event), and post (impact, damage, remediation, recovery). By detecting pre-phase, emergency response team can say that this area is in the disaster zone. So they can make awareness among the people that a catastrophe is coming, so to get prepared. Doctor, nurse and medical enterprise may get ready for curing the affected people. They also make awareness between the neighborhood areas because there is a high probability of a disaster occurring in that area. Thus awareness can be performed before the disaster takes place and that is on the pre-status phase. When a disaster is happening in a particular area, people become rushed to fight against the disaster, and during a disaster event, electricity loss happens in most of the cases, and internet connection may cut down. So, during a disaster, only a few tweets can be found. Most of the time this type of tweet is about the impact of the weather or nature during the disaster, and that is in the on-time-status phase. By detecting the post-status phase, response team can know about the damage caused by the disaster and take quick action to recover the damage. Remediation is needed to help the affected disaster zone after the disaster has occurred. By detecting the post phase emergency, the concerned people should be transferred to the medical center and take immediate medical care and proper medicine to cure them. Besides by seeing the geographical map of that type of phases, people can get alerted, and response team can create a situational awareness.

1.3. Research Questions

We have carefully considered the following research issues to accomplish our goal. In fact, the answer to these questions contains the core concept behind our work.

1.3.1. Why do we select a keyword from Twitter?

During a disaster event, many tweets can be tracked from the Twitter. We use this social media for tracking disaster relevant data, and this data is needed for classifying into three phases which we were determined in our work. Besides tracking public data is very easy. Without extracting this disaster data, we could not be able to train our algorithm, and therefore we use Twitter data

1.3.2. Why do we need classifying disaster phases?

We need a classifier for classifying the disaster data into three classes and detecting each class of a tweet that it is in a particular phase. We can detect the phases by comparing the training data with the test data. So classifying disaster phases is needed for detecting disaster phases of a tweet

1.3.3. Why do we assign the weight of a tweet?

There may have some extra word in the keyword from which we cannot understand that what type of tweet is it. So basically ignoring the last letter of a word we match every word of a tweet with the relevant keyword which we manually extracted. Based on the matching keyword we assign a weight of a tweet. For declaring the phases of a tweet weight is needed.

1.3.4. Why do we use KNN algorithm?

KNN algorithm is easy to understand and easy to implement. This classification algorithm helps to predict the data with better accuracy. So we use this algorithm so that our result can be much better. As this method contributes to predict data, we compare this predicted data with our actual data and can measure whether our result is right or wrong.

1.4. Overview of the Proposed Solution

Our main goal is detecting disaster phases during a natural disaster. So we trained a classification algorithm for classifying disaster data into three phases that are pre-status (before the disaster), on-time-status (during a disaster) and post-status (after a disaster has occurred). So to fulfill our aim, first of all, we extract data from social media based on hashtag using Twitter 4j API. We also extract data using time-boundary and the bounding box (collecting data with latitude and longitude).As there were many noisy data, so we clean the data using C++ programming based on the disaster hashtags. For the data cleaning process, we match a tweet word with a relevant keyword (manually extracted) and if any word matches then it was considered as a relevant data otherwise noisy data. For keyword matching, we check and compare the tweet word and related keyword in three steps (matching the tweet word with relevant keyword directly,  matching the tweet word with keyword considering the last letter of a word as a relevant keyword, also consider the increasing length of 1 of a word as relevant).According to this process, we assign a weight of a tweet. We also normalize the time into minutes based on the posting time of a file.

We use KNN algorithm to classify this relevant data into three phases using four descriptive features (pre-status weight, on-status weight, post-status weight, time in minutes) and test features as like 1,0,0 (indicates pre-status high); 0,1,0 (on-status high); 0,0,1 (post-status high); 1,1,0 (prestatus and on-status both are same and high); In this case confusion occur because we cannot detect the phases. We split the file of data into training data and test data according to 2:1 ratio that means the training data is 0.67 (67%) and the test data is 0.33 (33%). Then we trained our algorithm by calculating the Euclidian distance between each test data with all of the training data and take the nearest distances 5 data as the value of K we assign 5 and take the majority voted data from it. From that majority voted data we can predict the disaster phases.

We extract disaster relevant data manually and compare the actual data with the predicted data which we get after classifying and show how accurate our result is. Then we get the accuracy of our result based on the correct prediction. We also visualize and demonstrate the percentage of three phases using a pie-chart and graphically represent the data with the posted time in minutes using axis-chart. After that, we create a map using the latitude and longitude of the affected area. In an earth map, we try to point out the particular area under each disaster phase.

After completing this classification step, we evaluate our result by calculating the precision, recall, F1-measure of each phase and the overall accuracy of our paper.


We extract social data using Twitter 4j API and assign a weight of a status. For calculating the weight we need max time and min time of tweets (in minutes) which is in the file, also need a ratio which we identified using a ternary search [14, 21]. We classify the disaster data into three phases. For classifying disaster phases, we use a machine learning algorithm that is KNN to predict the different phases and also use confusion matrix for calculating the accuracy of our experiment. Then we use matplotlib library for visualizing our work and create pie-chart, axischart, and Earth-map.

2.1. Twitter 4j API

Twitter4J is a Java library for the Twitter API [12]. With Twitter4J, we can easily use our Java application with the Twitter service. The Twitter Platform connects our website or application with the global conversation happening on Twitter. To get the recent tweets based on hashtags, we can use streaming API through which we can extract the data.

2.2. Ternary Search

For finding a maximum or minimum point in U-shape graph, the ternary search is the best choice. A ternary search [14, 21] is an example of a divide and conquer algorithm. A ternary search determines either that the minimum or maximum cannot be in the first third of the domain or that it cannot be in the last third of the domain, then repeats on the remaining two-thirds.

2.3. K-nearest Neighbor (KNN)

A K-nearest neighbor is a machine learning classification algorithm [8]. It is a similarity base learning. Whenever we have a new point to classify, we find its K-nearest neighbor from the training data and the new point is assigned from the majority of classes. The distance is calculated by using the following measures: Euclidean, Minkowski, Manhattan. In classification problems, KNN is most commonly used. The main drawback of KNN is the complexity in searching the nearest neighbor for each sample. We implement this algorithm in Python [20].

2.4. Matplotlib

Matplotlib [13] is a plotting library for the Python programming language and its numerical mathematics extension NumPy. It provides an object-oriented API for embedding plots into applications using general-purpose GUI toolkits. Pyplot is a matplotlib module which provides an MATLAB-like interface. Matplotlib is designed to be as usable as MATLAB, with the ability to use Python, with the advantage that it is free. We also use multiple maps using subplots [19].

2.5. Confusion matrix

Confusion matrix contains information about actual and predicted classifications done by a classification system and describes the performance of a classifier model [9]. For evaluating the performance [18] of such systems, we have to use the data in the matrix.


Many researchers provide disaster-related information during a natural disaster using social data. Ashktorab et al. [1] introduce tweedr, a twitter-mining tool, which extracts actionable information during a natural disaster for helping disaster responder team. The tweedr pipeline consists of three main parts: classification, clustering, and extraction. Some researchers also make an automatic method for extracting information nuggets [3] so that they can help professional emergency responders and the Twitter data has been categorized into different features such as Caution, Advice, Fatality, Injury, Offers of Help, Missing and General Population Information. Recently it has been found that for helping emergency responders to act quickly for disaster response and disaster relief, actionable data can be extracted from social media. Here [4] a tweet tracker tool has been proposed. During disaster, by tracking and monitoring disaster-related twitter data researchers help Humanitarian Aid and Disaster Relief (HADR) respondents for gaining valuable insights and situational awareness in the affected area

Many studies have applied to work with geographic locations of tweets of the disaster event to help the emergency response team to make situational awareness. Social media messages have been separated into four categories (preparedness, response, impact, and recovery) [2] to understand the disaster event, and this framework has been done with the relevant tweet to take action quickly and efficiently in the impacted communities. Other researchers also describe the use of disaster phases (mitigation, preparedness, emergency response, and recovery) [7] which has assisted both disaster researchers and managers. In this paper, it has been suggested that the use of disaster phases can improve the theoretical and applied dimension of the field during disaster periods and also mentioned that disaster researchers have used disaster phases to organize significant findings and recommendations about disasters.

Big Data can help in all four phases of disaster management [22]. However, significant big data research challenges arise because of disaster management requirements for quality of service (e.g., highly available real-time response) and quality of information (e.g., reliable communications on resource availability for the victim). Besides, social data is a source of extracting valuable information to make situational awareness [23]. The goal of this paper is to identify and measure features that could support technology in analyzing mass emergency situations. The contribution of this article is considering “situational update” information that is communicated by people through micro blogging in mass emergency situations.

It has been described that annotating social media data with geographic coordinate is more valuable for quickly finding out the area under the victim [5]. Twitter user locations are being estimated by propagating the locations of GPS-known users across a Twitter social network. A method has been invented to locate the overwhelming majority of active Twitter users by examining their locations. The algorithm assigns a location to a user based on the locations of their friends by calculating the min, median and max distance between their friends. Moreover, the automatic geolocation of social media messages is beneficial [10]. They analyze that since different people, in different locations write messages at different times; these factors can significantly vary the performance of a geo-location system over time.

In this paper, they consider the task of tweet geo-location, where a system identifies the location where a single tweet was written. Here [11] it has been investigated and improved on the task of text-based geo-location prediction of Twitter users. They present an integrated geo-location prediction framework and investigate what factors impact on prediction accuracy. Here they evaluate the impact of temporal variance on model generalization and discuss how users differ regarding their geo-locatability.

The contributions of this paper [6] are to introduce AIDR (Artificial Intelligence for Disaster Response), a platform designed to perform automatic classification of crisis-related social data. The objective of AIDR is to classify messages that people post during disasters into a set of userdefined categories of information (e.g., “needs”, “damage”) For this purpose, the system continuously ingests data from Twitter, processes it (i.e., using machine learning classification techniques). Imran et al. [25], extracted tweets into several categories, caution and advice, casualty and damage, donation and offer, and information source during a natural disaster. A platform AIDR is also described here, which collects human annotations over time to create and maintain automatic supervised classifiers for social media messages.

The contribution of this paper [15] is that conducting a systematic comparative analysis of nine state-of-the-art network-based methods for performing geolocation inference at the global scale. The analysis identifies a significant performance disparity between that reported in the literature and that seen in real-world conditions. All implementations have been released in an open source geo-inference package to aid reproducibility and future comparison.

Lie et al. [16] declare that Twitter data contains valuable information that has the potential to help improve the speed, quality, and efficiency of disaster response. However, supervised learning algorithms require labeled data to learn accurate classifiers. Experimental results suggest that, for some tasks, source data itself can be useful for classifying target data. However, for tasks specific to a particular disaster, domain adaptation approaches that use target unlabeled data in addition to source labeled data are superior.

This paper [17] presents a CyberGIS framework that can automatically synthesize multi-sourced data, such as social media and socioeconomic data, to track disaster events, to produce maps, and to perform statistical analysis for disaster management. In this framework, Apache Hive, Hadoop, and Mahout are used as scalable distributed storage, computing environment, and machine learning library to store, process and mine massive social media data. Several attempts have taken to illustrate volunteered geographical data (VGI) [24]. Volunteered data obtained through Google news, videos and photos are added to modify the contour regions. A new methodology for the generation of flood hazard maps is presented using remote sensing and volunteered geographical data.


Our solution consists of several computational steps. Figure 1 shows the flow chart of the
proposed solution. In this section, we describe each computational step in detail.

4.1. Collecting Data

We collect tweet from Twitter using Twitter4j API in JAVA based on hashtags. Using those hashtags, we collect randomly 4,500 tweets without any bounding-box and no timebound.

4.2. Keyword Selection

From 4500 tweets, we select a total number of 65 keywords and classify them into three different phases – Pre-Status, On-time-Status, and Post-Status. Table 1 shows these keywords. Sometimes some keywords are classified under one phase, but due to the presence of another word, the actual meaning of the tweet could be divided into a different phase. As for example, consider tweets – “Response teams are coming” and “response teams are working”. Here the former belongs to “Pre-Status” and the latter belongs to “Post-Status”. Therefore, we keep another checking to solve this type of problem. Table 2 shows the list of this type of keywords.Capture



4.3. Assigning Weight of a Tweet

We formulate an equation to assign a weight of a tweet. The equation is:


In Eq. 1, time is measured in minutes. Number of matches is the number of keyword matches  with the words of a tweet. We consider three rules for counting matches.

i. Directly and equally the keyword and a word of a tweet is matched.
Example: (“rain”[Keyword] and “rain”[A word of tweet]).

ii. If there is any last letter like „.‟(dot), „,‟(comma), „?‟(interrogation symbol and
„!‟(Exclamatory symbol) then we ignore it and check without it, Example:
(“rain”[Keyword] and “rain.”[A word of tweet]).

iii. If a keyword matches completely but the word_of_tweet_length – Keyword_length = 1
then we accept that as a match. (“wind”[Keyword] and “winds”[A word of a tweet]).
The ratio is found by using ternary search. Lower bound and upper bound of ternary search is 0.0 and 1.0. The actual value is calculated by taking the average accuracy of five random iterations of each ratio point. Figure 2 shows the justification of taking ratio = 0.75.

4.4. Providing the Ground Value

We provide the ground value of all tweets manually by reading all cleaned tweets and determine their original value.


4.5. Predicting Phase of the Tweet

Now we use KNN [K=5] to predict the phase of a tweet. 67% data are using as a training set and 33% used as a test set of a data file. There are four characteristic features in the dataset (e.g. PreStatus, On-time-Status, Post-Status, Time) and three target features (Pre-Status, On-time-Status, Post-Status). As an example, the weight of the Pre-status is higher than the other two target features the value of 3 target features would be [1, 0, 0].

Now program selects five nearest training dataset by using Euclidian distance. The three phases point is counted from the target features of these 5 data and returns the phase which gets the maximum number of point and predicts that as the phase of that tweet Now we compare the predicted value with the actual ground value. We build a confusion matrix based on the average of 5 iterations.

We take the last 10% data whose predicted value and the actual value is equal. Afterward, we count which phase occurs most from these tweets and then select that phase.

4.6. Demonstration of the Proposed Solution

In Queensland, Australia recently a Cyclone named Debbie has occurred. This disaster hits the Queensland in the evening of 28th March 2017. We collect tweets in three steps based on two hashtags (e.g. #Cyconedebbie, #disaster). Table 3 shows the overview of this data collection.


There are 4 files of dataset, file1 has step-1 data (pre-status), file2 has step-2 data (on-timestatus), file3 has step-3 data (post-status) and file4 has all data (three-status).

Then we give the actual value of all these tweets manually trained the KNN algorithm to predict the phase of these data.

4.6.1. Visualization of file1

In Figure 3, we visually represent the percentage of the actual value and the predicted value of phase1 using a pie chart. From here we can see that in pre-status phase the percentage of data is 85.11 (actual) and 77.08 (predicted), in on-status phase the percentage are 14.89 (actual) and 22.92 (predicted), whereas in post-status phase there is 0.00% data in both cases. As in pre-status phase in both actual and predicted the percentage is higher than another two phases so we can say that this chart indicates that this data is under pre-status phase. When we collect data from cyclone-Debbie there, we see that whenever people tweet for the awareness before the cyclone in that time cyclone has been beginning and so during the pre-status phase, people were posting an
on-status tweet. So we get tweets in on-time-status.



In Figure 4, the actual data and predicted data of phase1 are graphically represented using axis diagram. Here we visualize three phases using three colors (green for pre, red for on, yellow for the post). In this figure, the x-axis represents the Time in minutes, and the y-axis is the Number of Tweets. The circled point indicates the data of each phase including time represented by the color. By comparing and observing the actual and predicted data we can understand that in pre-status phase there are many data and in on-status phase, there are little bit data and here is no yellow data so post-status phase is empty. By observing the data, we can understand that this is pre-status data because the number of tweets is higher in pre-status phase in both cases.

Figure 5 shows two tiny green stars pointing out the affected area [both for actual and predicted data]. As a green star represents as Pre-Status, we can say that this is under Pre-Status phase. By seeing the visualization of file1, we can observe that our solution declares step1 (26th -27th , March 2017) was the Pre-Status phase.


4.6.2. Visualization of file2

Figure 6 describes the visual representation of the actual and predicted value of phase2 using piechart. By comparing figures of actual and predicted values, we can see that the percentage of data in pre-status phase is 9.38 (actual) and 9.09) (predicted), in on-status phase percentage is 56.25(actual) 57.58 (predicted), whereas in post-status phase it is 34.38 (actual) 33.33 (predicted). We could say that in on-status phase the percentage is higher than the two. So we could say that we correctly predict that this is in on-status phase because actual and predicted value is almost same.


In Figure 7, the actual and predicted data of phase2 is graphically represented using axis diagram. Here we visualize three phases. By comparing the actual and predicted value, we can understand that in on-status phase there is a higher data than the two. By observing the data, we can understand that this is in on-status phase though there are post-status data because of the flooding during the cyclone.

From Figure 8 we can see that a tiny red circle is pointing out the affected area. As a red circle represents as On-time-Status, we can say that this is under On-time-Status phase. By seeing the visualization of file2, we can observe that our solution declares step2 (28th
-29th , March 2017) was the On-Time-Status phase.


4.6.3. Visualization of file3


From this figure (Figure 9) of actual and predicted value by comparing the percentage of three statuses we see that the percentage of pre-status and the on-status phase is same (10.17) for both cases and in post-status phase, 79.66 (actual and predicted). As the percentage of post-status

phase is higher than another two phases so we can say that this chart indicates that this data is under post-status phase. A little bit of pre-status and on-status phases also here.


The actual and predicted data of phase3 is graphically represented using axis diagram. From this chart, we can see that yellow circle point indicates the higher data than the red and green data. As the yellow color represents the post-status phase, so by observing the data, we can understand that this is in post-status phase. So we can predict well that this is post-status data.


From this figure (Figure 11), we can see that two tiny red circles & two yellow triangles pointing out the affected area. As red circle represents as On-time-Status, and yellow triangle represent Post-Status. So we can say that this is under On-time-Status and Post-Status phase.

By seeing the visualization of file3, we can observe that our solution declares step3 (30th -31th , March 2017) was the Post-Status phase.


Here we calculate some comparison results in all three file using confusion matrix and find Precision, Recall and F1-measure from 5 random iterations of those files.

5.1. Confusion Matrix of file1

In this confusion matrix as shown in Table 4, from the 48.2 actual pre-status tweets, the program can predict that 43 were pre-status tweets and 5.2 were on-status tweets, and from the actual 4.4 on-status tweets, the program can predict 4.4 on-status tweets correctly and here is no post-status tweets. The accuracy of file1 we get is 90.11%.


5.2. Confusion Matrix of file2

In this confusion matrix as depicted in Table 5, from the 15.6 actual on-time-status tweets, the program can predict that 14.8 were on-time-status tweets, 0.6 were pre-status phases, and 0.2 were post-status tweets, and from the actual 6.8 post-status tweets, the system can predict 6.2 post-status tweets. The accuracy of file2 we get is 95.70%.


5.3. Confusion Matrix of file3

In this confusion matrix (see Table 6), from the 57.4 actual post-status tweets, the system can predict that 49.2 were post-status tweets, 5.4 were pre-status tweets, and 2.8 were on-time-status tweets, and from the actual 0.8 post-status tweets, the system can predict 0.8 post-status tweets. The accuracy of file 3 is 85.90%.


5.4. Performance Evaluation using Precision, Recall and F1-measure


Figure 12 shows precision, recall, and F1-measure of tweets of file 1. We get precision (1.000), recall (0.892) and F1-measure (0.942) of pre-status tweets. In on-status it is, precision (0.458), recall (1.000) and F1-measure (0.682). Similarly, for post-status,


In Figure 13, we get precision (0.692), recall (1.000) and F1-measure (0.817) of pre-status tweets. In on-status, precision (1.000), recall (0.949) and F1-measure (0.973). In post-status, precision, recall, and F1-measure value are same (0.968). This measurement indicates that this is on-timestatus.


The result of file3 is shown in Figure 14. We get no precision, recall, and F1-measure is
ambiguous of pre-status tweets. In on-status, precision (0.222), recall (1.000) and F1-measure (0.360). In post-status, precision (1.000), recall (0.857) and F1-measure (0.922). From this measurement, we can say that this is post-status tweets.


Twitter is a popular social media, and during natural disaster or any other crisis moment this source is beneficial for tracking the disaster data to alert people or helps the affected people. In our work, we track this twitter data with latitude and longitude using text mining technique during a disaster, Cyclone Debbie in Queensland, Australia (26 March to 31 March 2017, three steps). A machine learning classifier technique is introduced here to categorize these messages. Our aim was to measure and classify these tweets into three phases (pre, on, post) and detect the phases. We also visualize the actual and predicted value of each category phases using pie chart and axis chart and seeing this graph we can easily compare the actual data and the predicted data and the percentage of each phase and also visualize our experimental result by showing the earth map and pointing out these three phases on that map. Our result indicates that there is an increasing number of classification accuracy and we calculate this using confusion matrix. We also take the screenshot of the accuracy of each category phases and the all the phases to show our result how accurate it is.

In future, the most damaged area of a disaster can be pointing out by using our work so that by seeing the pointed map affected people can get help, food, and donation by the response team. Our paper is also helpful for examining the disaster tweet with their geo-location by mapping. In future, the more risky zone of a disaster can also be finding out so that people can alert and make a recovery plan or leave the dangerous area for reducing the damage. Also, it can be possible to predict the nature and the danger level during a natural disaster. It is also feasible to visualize the live disaster moment in future and will be very helpful for disaster responder. By visualizing the live disaster event, it will be straightforward to find out the area under the disaster event.


[1] Z. Ashktorab, C. Brown, M. Nandi and A. Culotta, Tweedr: Mining Twitter to inform disaster response. In: Proceedings of the 11th International ISCRAM Conference, USA, 2014.

[2] Qunying Huang and Yu Xiao, Geographic Situational Awareness: Mining Tweets for Disaster Preparedness, Emergency Response, Impact, and Recovery. In: ISPRS Int. J. Geo-Information, 2015, 4(3), 1549-1568.

[3] P. Meier, C. Castillo, M. Imran, S. M. Elbassuoni, and F. Diaz, Extracting information nuggets from Disaster-related messages in social media. In: 10th International Conference on Information Systems for Crisis Response and Management, 2013.

[4] S. Kumar, G. Barbier, M. A. Abbasi, and H. Liu, Tweet Tracker: an analysis tool for humanitarian and disaster relief. In: ICWSM, 2011.

[5] Ryan Compton, David Jurgens and David Allen, Geotagging One Hundred Million Twitter Accounts with Total Variation Minimization. In: International Conference on Big Data (BigData), IEEE, 2014, 393-401.

[6] M. Imran, S. Elbassuoni, C. Castillo, F. Diaz, and P. Meier, Practical extraction of disaster-relevant information from social media. In: Proceedings of the 22nd international conference on World Wide Web companion, Republic and Canton of Geneva, Switzerland, 2013, 1021-1024.

[7] D. M. Neal, Reconsidering the phases of disaster. In: Int. J. Mass Emerg. Disasters, 15, 1997, 239– 264.

[8] Siddharth Deo Kar, Lecture Notes on K-nearest Neighbor Algorithm.

[9] (accessed on Feb 02 2017)

[10] Mark Dredze, Miles Osborne, and Prabhanjan Kambadur, Geolocation for Twitter: Timing Matters. In: Proceedings of NAACL-HLT, 2016, 1064-1069.

[11] Bo Han, Paul Cook and Timothy Baldwin, Text-Based Twitter User Geolocation Prediction. In: Journal of Artificial Intelligence Research, 49, 2014, 451-500.

[12] Twitter 4j API: (accessed on Jan 11 2017)

[13] Matplotlib library: (accessed on Jan 11 2017)

[14] Ternary Search: (accessed on Jan 01 2017)

[15] D. Jurgens, T. Finnethy, J. McCorriston, Y.T. Xu, D. Ruths, Geolocation Prediction in Twitter Using Social Networks: A Critical Analysis and Review of Current Practice. In: ICWSM, 2015, 188-197.

[16] H. Li et al. Twitter Mining for Disaster Response: A Domain Adaptation Approach, In: 12th International Conference on Information Systems for Crisis Response and Management, 2015.

[17] Q.Huang et al. DisasterMapper: A CyberGIS framework for disaster management using social media data In: Proceedings of the 4th International ACM SIGSPATIAL Workshop on Analytics for Big Geospatial Data, ACM, 2015, 1-6.

[18] S. Visa et al. Confusion Matrix-based Feature Selection, In: MAICS, 2011, 120-127.

[19] (accessed on Jan 11 2017)

[20] (accessed on Jan 13 2017)

[21] Ternary search algorithm: (accessed on Jan 01 2017)

[22] C. Pu, M. Kitsuregawa, Big Data and Disaster Management: A Report from the JST/NSF Joint Workshop. Georgia Institute of Technology, CERCS, 2013.

[23] S. Vieweg, A.L. Hughes, K. Starbird, L. Palen, Microblogging during two natural hazards events: What Twitter may contribute to situational awareness. In: Proceedings of the 2010 SIGCHI Conference on Human Factors in Computing Systems, USA, 2010, 10–15.

[24] E. Schnebele, G. Cervone, Improving remote sensing flood assessment using volunteered geographical data. In: Nat. Hazards Earth Syst. Sci. 13, 2013, 669–677.

[25] M. Imran, S. Elbassuoni, C. Castillo, F. Diaz, P. Meier, Practical extraction of disaster-relevant information from social media. In: Proceedings of the 22nd International Conference on World Wide Web Companion, Brazil, 2013