1

Disaster and Pandemic Management Using MachineLearning: A Survey

Vinay Chamola, Senior Member, IEEE, Vikas Hassija, Sakshi Gupta, Adit Goyal, Mohsen Guizani, Fellow, IEEEand Biplab Sikdar, Senior Member, IEEE

Abstract—This paper provides a literature review ofstate-of-the-art machine learning algorithms for disaster andpandemic management. Most nations are concerned aboutdisasters and pandemics, which, in general, are highly unlikelyevents. To date, various technologies such as IoT, object sensing,UAV, 5G and cellular networks, smartphone-based system,and satellite-based systems have been used for disaster andpandemic management. Machine learning (ML) algorithms canhandle multi-dimensional, large volumes of data that occurnaturally in environments related to disaster and pandemicmanagement and are particularly well suited for importantrelated tasks such as recognition and classification. Machinelearning algorithms are useful for predicting disasters andassisting in disaster management tasks such as determiningcrowd evacuation routes, analyzing social media posts, andhandling the post-disaster situation. Machine learning algorithmsalso find great application in pandemic management scenariossuch as predicting pandemics, monitoring pandemic spread,disease diagnosis etc. This paper first presents a tutorial onmachine learning algorithms. It then presents a detailed reviewof several machine learning algorithms and how we can combinethese algorithms with other technologies to address disaster andpandemic management. It also discusses various challenges, openissues and, directions for future research.

Index Terms—Machine Learning, Disaster management,Pandemic management, Healthcare, Crowd evacuation, Socialdistancing

I. INTRODUCTION

Over the last decade, more than 2.6 billion humans havesuffered from catastrophic disaster outbreaks such as tsunamis,floods, earthquakes, cyclones and landslides, and variouspandemics. Disasters have been the cause of several fatalitiesin the past, one of the deadliest disasters was an earthquakein New Guinea which left around 58, 300 people displacedaccording to the displacement tracking matrix (DTM) [1]. Thefloods that took place in China in July 1931, caused 4, 000, 000deaths that are yet the highest number of deaths from a naturaldisaster. Disasters are usually physical environmental changes,

This research was supported by the Ministry of Education, SingaporeVinay Chamola is with the Department of Electrical and Electronics

Engineering & APPCAIR, BITS-Pilani, Pilani Campus, India 333031 (e-mail:vinay.chamola@pilani.bits-pilani.ac.in).

Vikas Hassija, Sakshi Gupta and Adit Goyal are with the Department ofComputer Science and IT, Jaypee Institute of Information Technology, Noida,India 201304 (e-mail: vikas.hassija@jiit.ac.in, sakshigupta5000@gmail.com,aditgoyal@hotmail.com).

Mohsen Guizani is with Department of Computer Science and Engineering,Qatar University, Qatar (e-mail: mguizani@ieee.org ).

Biplab Sikdar is with the Department of Electrical and ComputerEngineering, National University of Singapore, Singapore 119077, Singapore(e-mail: bsikdar@nus.edu.sg)

Digital Object Identifier: XXXXXXXXXXXX

whereas pandemics refer to the rapid spread of a disease overa wide area. There also have been several pandemic outbreaksacross the world. To name a few, there’s the Americanplague (16th century), yellow fever in Philadelphia (1793),H1N1 swine flu (2009-2010), Ebola pandemic (2014-2016),and the recent COVID-19 (2019-Present). The deadliest onewas Black death (14th century) that had spread from Asiato Europe, causing many deaths. Recently, disaster andpandemic management has become one of the hotspot areas forresearch. There are some recent and important works done forprevention and management of COVID-19. The authors of [2]present a detailed review of applications of machine learningand artificial intelligence for tackling this pandemic. Thepaper focuses on screening, predicting, forecasting, contacttracing, and drug development for the COVID-19 pandemic.Disasters can be either natural or man-made [3]. Suchsituations are spontaneous and complex, risking human lives,the environment, and the economy of a country. Therefore,any nation would like to opt for the most efficient andaccurate algorithms to control such ordeals. The methodologyused to predict the foreseen consequences of a disaster orpandemic plays a significant role in its management. Withmore accurate predictions and understanding, we can utilizeour resources more efficiently. The advances in computerscience have made available a large volume of data for disastermanagement authorities. Such data is often unstructured,making it challenging to clean and process such high volumesof data.

To date, many people suffered greatly because of the lack ofa proper disaster and pandemic management system. A properprediction of a disaster could not be done, and victims were notevacuated on time from the disaster outbreak area. People werenot provided with mitigation measures post-disaster. Also,during the pandemic, efficient steps could not be followedto prevent further spread of the outbreak. To address suchissues, this paper provides a detailed review of all the existingprocedures and techniques that can be employed during thepost, and pre-disaster period to minimize the losses as muchas possible.

The systems developed for assisting with disaster predictionneed to be robust enough to handle the challenges that canaffect a disaster management system [4]. For instance, incase of a sandstorm or a hurricane, the vision of a systemcan be lowered due to the presence of dust particles, or, inhazy conditions, autonomous driving system should be safe[5]. Another challenge could be the loss of communicationduring a disaster. Furthermore, other challenging tasks include

2

Table I: List Of Abbreviations.

Abbreviation DescriptionADC Analog-to-Digital ConverterAIP Application Infrastructure ProviderANN Artificial Neural Network/Artificial Potential FieldAPF Attraction Potential FieldAPM Ardupilot MegaASN Active Sensor NetworkBB-SVM Buffering and bagging SVMCART Classification and Regression treesCEOS Committee on Earth Observation SatellitesCIoT Cellular IoTCLOTHO Crowd Lives Oriented Track and Help OptimizationDJI Da-Jiang InnovationsDMS Document Management ServicesDTM Displacement Tracking MatrixERESS Emergency Rescue Evacuation Support SystemFANETS Flying Ad-Hoc NetworksGCI Ground Controlled InterceptionGCS Ground Control StationGEOSS Global Earth Observation System of SystemsGIS Geographic Information SystemGPRS General Packet Radio ServiceGPS Global Positioning SystemGSM Global System for MobileLiDAR Light Detection and RangingLTE Long-Term EvolutionMANET Mobile Ad Hoc NetworkML Machine LearningNOIR Network Optimization Indicator ReportOSM OpenStreetMapPCA Principal Component AnalysisRADARSAT Radar SatelliteROS Robotic Operating SystemRSS Radio Signal StrengthSAR Synthetic-Aperture RadarWSN Wireless Sensor Network

maximizing the number of people protected during a disasteror a pandemic, evacuating people at the right time, identifyingthe vulnerable areas for the spread of a pandemic, reaching themost affected people/areas and providing them with sufficientresources, evaluating the loss to the economy, and many more[6]. Decision-makers are often provided with large volumesof data and need to make predictions and decisions as quicklyas possible [7], [8]. Deep learning techniques can be usedfor image classification and 3-D segmentation for medicalpurposes [9].

Machine learning (ML) has recently emerged as one ofthe key computing technologies and is increasingly beingapplied in day-to-day life, and various industrial domains[10]. ML is an application of artificial intelligence (AI)that uses algorithms that work on characteristics of availabledata to make further predictions. Nowadays, in the era ofvarious other emerging technologies such as UAV (UnmannedAerial vehicles), IoT (Internet of things), and satellite-basedtechnology, the network is becoming more autonomous. Suchsystems require several local decisions to be made, such asbandwidth selection, data rate selection, power control, anduser association to a base station. We can use ML algorithms

to address these issues and lower human intervention inuncertain and stochastic environments. To summarize, MLalgorithms have the following advantages over other existingtechnologies:

• ML algorithms can easily process a high volume of dataand can use it to identify trends. Moreover, the MLalgorithms help in easily analyzing various types of data.The application of ML in day-to-day life, such as trafficpredictions, video surveillance, online customer support,has also increased its popularity.

• Rule-based technologies in ML can help in detectingfake messages. The use of ML algorithms reduces theneed for human intervention and decision making. Suchtechniques help prevent rumours, especially in the caseof man-made disasters.

• The performance of ML algorithms tends to improveas the data increases. For example, in an earthquakeprediction model, when the data increases, the algorithm’sability to predict also increases.

• ML algorithms can handle multi-dimensional data anddetect outliers in the data-set. In situations dealingwith extreme hazards, outlier analysis is an importanttechnique. Rather than removing all these outliers, specialattention should be given to them when we are trying topredict highly unlikely events like disasters or pandemics.

Various machine learning algorithms can be employed insuch cases to make fast and reliable decisions. The furthersubsections of this paper discuss how these algorithms canbe applied to make decisions with better accuracy. Thesealgorithms have a wide variety of applications and can help indecreasing human intervention [11]. Moreover, certain trendscan also be identified to make better predictions based onprevious data. These algorithms can also be applied to detectand break the chain of the spread of diseases.

Although there are several surveys related to the use ofML algorithms, very few focus on the applications of MLin disaster and pandemic management. The authors of [12]present a survey of ML models and big data analytics inthe healthcare field. The authors of [13] present the useof ML algorithms to detect plant disease. The authors of[14] give an overview of the use of artificial intelligence formanagement of the COVID-19 pandemic by using it for imageprocessing, data analytics, etc. Similarly, the authors of [15]discuss comprehensive studies that use IoT in smart healthcare,artificial intelligence, and big data analytics with a prime focuson COVID-19 pandemic. To the best of our knowledge, thereis no detailed survey that exclusively focuses on the use ofML algorithms for disaster and pandemic management. Inthis paper, we present a basic tutorial of ML algorithms anda comprehensive literature review on the applications of MLalgorithms for disaster and pandemic management. The maincontributions of this work are as follows:

• We discuss the various ML algorithms that can be usedin different phases of disaster management to constructdeployable models.• We present a detailed review of various technologiesused in disaster and pandemic management and how ML

3

Figure 1: Structure of this paper.

algorithms can be used by them for this end.• We discuss how different ML models integrated withother technologies can be deployed at various phases ofdisaster and pandemic management.• We carry out an assessment of the open issues,challenges, and future research directions for ML-baseddisaster and pandemic management.

The rest of this paper is organized as follows. Section IIpresents a tutorial on machine learning algorithms. SectionIII reviews the applications of ML in pre-disaster scenarios.Section IV presents the applications of ML to determine crowdevacuation routes. Section V reviews ML related works inpost-disaster management. Section VI presents applications ofML in miscellaneous issues related to disaster management.Challenges, open issues, and future research directions arediscussed in Section VII. Finally, Section VIII presents theconclusion of the paper. Fig. 1 elaborates on the structure ofthe paper. Furthermore, Table I presents the list of commonabbreviations used in the paper.

II. MACHINE LEARNING OVERVIEW

In this section, we present a basic overview of variousML algorithms that can be used for disaster and pandemicmanagement. These algorithms can be categorized intosupervised learning, unsupervised learning, and reinforcementlearning.

A. Supervised learningIn supervised machine learning algorithms, the training data

provided to the computer is labelled, and a set of expected

output results are provided. We expect the machine to learnthe pattern from this data and predict the output values for newdata inputs. Supervised machine learning includes two majorprocesses:

1) Classification techniques: Classification techniques areused in the estimation of membership of the communityfor data instances, and classify a data item into one ofmany predefined classes [16]. While classification in machinelearning is a well-established methodology, it struggles withissues such as managing missing data. In both trainingand classification processes, missing data set values maytrigger problems. This issue can be solved by approacheslike overlooking the omitted data or swapping outliers withprobable data [17]. The following are the most commonclassification techniques that are being used in different phasesof disaster and pandemic management.

• K-nearest neighbours (KNN) - KNN is a location-basedapproach used for classification and sometimesregression. This algorithm assumes that similarobjects occur next to each other and evaluates itsk nearest neighbours for similarity. Training examplesare multidimensional element space vectors, eachwith a class name. The algorithm’s training processconsists merely of holding the test sample featurevectors and class labels. k is a user-defined constantin the classification process, and an unlabeled vector(a question or checkpoint) is identified by assigningthe mark that is most popular among the k trainingsamples closest to that question point. The best choice

4

of k depends on the dataset. Typically, higher k valuesreduce the impact of noise on classification but renderthe boundaries between classes to be less distinctive.When k is increased, the forecasts are more reliable andaccurate due to majority or average voting.To make the KNN method more accurate, weights may beassigned to the nearest neighbours before classification.This algorithm is easy to apply and has low timecomplexity. There is no need to construct a model,change several parameters, or create supplementaryassumptions. The algorithm can be used for bothpurposes-classification and regression. The key downsideto KNN is that it tends to become slower as data volumeincreases. The authors of [18] have compared KNN withother algorithms to detect Acute Respiratory Infections(ISPA). Similarly, [19] has used KNN to detect Influenzain patients. Also, authors of [20] have used KNN onthe location data from user’s smartphones to predictthe movement of infected people (that helps in socialdistancing).

• Support Vector Machine (SVM) - This algorithmworks by identifying a hyper-plane (refer to Fig. 2) thatclassifies the data points. Several different hyper-planesmay be used to distinguish the two types of datapoints. SVM aims to find a plan with the maximumrange, i.e., the maximum gap between the data pointsin both groups. Maximizing the gap from the marginsgives some clarification such that potential data pointscan be identified with better accuracy. Hyper-planes areboundaries for decision making and help to distinguishdata points. Data points that fall on either side of thehyperplane can be assigned to different groups. Often,the hyper-plane dimension depends on the number offeatures. The hyper-plane is only a line until the numberof input features exceeds 2. Whenever the number offeatures of the input is 3, then the hyper-plane wouldbecome a 2-D plane. It gets hard to picture when thenumber of features is greater than 3.To reduce costs, authors of [21] have used SVM tochoose the best representative crowd reflecting pilgrims’behaviour for its further processing through fuzzy logic.SVMs can model the limits of nonlinear judgments,so there are several kernels to pick from. Also, theyare relatively resilient against overfitting, especially inhigh-dimensional space.SVMs are often used for confirmation of a disaster,as in [22]. We can use SVM to classify EH’s (EdgeHistogram) state of emergency through sensor info. Infact, in the model in [18], SVM showed the highestaccuracy in comparison to other algorithms for detectingwhether a person is suffering from ISPA. Also, in [23],it is found that SVM has better accuracy for locationverification of a user than wireless networks that requirechannel characteristics data to operate. Furthermore, theauthors of [24] have proposed the use of SVM to classifyaerial images into flood-affected and non-flood affectedareas. Also, authors of [25] have used SVM to determinedisaster outbreak in an area. We can use deep SVMs with

IoT enabled convolutional neural networks for COVID-19diagnosis and classification [26].

• Naive Bayes - Naive Bayes classifiers are a set of Bayes’Theorem-based classification algorithms. It is called naivebecause all the features that are classified are assumed tobe independent of each other, which is quite unrealisticin real life. The data is split into two components,feature matrix and response vector. The feature matrixincludes the whole data collection in the form of vectors(rows) where each row represents the relative variabletype, whereas, in a response vector, each row representsan outcome class. The authors of [22] have proposednaive Bayes combined with UAV technology to assesspost-disaster damage. Naive Bayes’ acts as a classifierfor classifying areas from data collected by UAVs intosafe and under-threat areas. Also, authors of [27] havecombined Deep Belief Networks (DBN) with naive Bayesto detect the user’s location.In models mentioned in [28] and [29], naive Bayesoutperformed all other classifiers to classify tweets thatcan help in managing social networking issues duringdisaster or pandemic periods.2) Regression techniques: The regression technique is apredictive learning feature that maps a data object to apredictive variable with real meaning [30]. The followingare the most common regression techniques that canbe used in different phases of disaster and pandemicmanagement.

• Logistic Regression - Logistic regression is based onprobability and uses sigmoid as its cost function. Thevalue of the cost function is assumed between 0 and 1.Authors of [28] and [29] have used logistic regressionto extract useful post-disaster information from tweets.In their cost function, 1 represents the relevance of atweet to disaster, whereas 0 represents its non-relevance.Furthermore, authors of [31] have used logistic regressionto determine the survival rate of people during a disaster.The relationship between the survival rate and the extentof the disaster is calculated.

• Decision Trees - In a decision tree, features arerepresented by internal nodes, decision rules by branches,and outcomes by leaf nodes. Many measures can be usedto classify, but entropy and Gini index is widely used.Authors of [32] have used decision trees for sandstormdetection. Two classes are defined in this work -sandstorm or no-sandstorm. Furthermore, the authors of[20] have used decision trees to determine the locationof the user during a pandemic.

• Bayesian Ridge Regression - A Bayesian approach isa probabilistic method for estimating statistical models.Bayesian regression is typically useful when we havea poor source of data, one that is insufficient orpoorly distributed. Ridge regression is similar to linearregression, but the variance of the estimates is reducedby using a regularized version of the model througha cost function. The response of this regression modelis obtained from a probability (Gaussian or normal)distribution, unlike linear regression, where the output

5

Figure 2: Basic SVM classification.

is estimated as a single value. An application of thismodel is given in [33], where the authors have usedBayesian ridge regression to predict the number of peoplein an area. By predicting the number of people, socialdistancing measures can be implemented.

• Random Forest - Decision trees are responsive to theparticular data that is used to train them. If the trainingdata is updated, the outcomes of the decision tree canbe quite different. They are also computationally costly,bring a risk of overfitting and tend to find local optimabecause they can’t go back after splitting. Random forestsare used to fix these limitations. Random forest is anensemble (multiple models combined) model techniquein which multiple decision trees are trained together toproduce one output. This merging of decision trees istermed as bagging. Authors of [34] have used a randomforest to detect changes post-disaster. Similarly, authorsof [35] have used them to detect damages to buildings.Apart from the classification of land cover performedwith the help of random forests in [36], they can also beused for the prediction of floods [37]. The random forestmodel in [38] outperformed the rest with 95% accuracy inpredicting the number of people infected with influenzain public places. Due to its advantage in combining theoutput of all decision trees, it results in high accuracy.Also, the results from [39] show that the random forestclassifier outperformed KNN, SVM, and decision treewith 77.8% accuracy.

• Gradient Boosting - Boosting is a method by whichweak learners are converted into strong learners. Gradientboosting slowly and sequentially trains several models. Itis also categorized as an ensemble technique and worksby optimizing a loss function.Authors of [38] have used gradient boosting to predictthe number of people infected by Influenza. Also, authorsof [41] have used gradient boosting to predict returningpatterns for people who left from the disaster-affectedarea. This classifier outperformed all other classifiers withan accuracy of 86.4%.

• Artificial Neural Networks - Artificial Neural Networks

(ANN) are fully-connected, multi-layer neural networksthat consist of an input layer, several hidden layers, andan output layer. All the nodes in one layer are connectedto all the other nodes in the subsequent layer. The inputsare multiplied by their weights and passed through anactivation function. The output of one node becomes theinput for another, and later continuing in this manner,we obtain the final output. The input for a given node iscalculated as (where b is the bias and usually taken as 1):

Input = f(b+ y · x) = f

b+ m∑j=1

yjxj

(1)

where:

y ∈ d1×m, x ∈ dm×1, b ∈ d1×1, Input ∈ d1×1. (2)

The authors of [42] have proposed the use of ANN toidentify or calculate a storm’s intensity. Also, ANN, inintegration with IoT technology, can be used to detectthe user’s location [43]. In this way, the location ofan infected user can be identified, and social distancingmeasures can be followed in a better way. Similarly,neural networks can be used to verify the user’s location[23]. The results in [23] show that ANN can be used forclassification when the data is insufficient for locationverification by wireless networks. Also, the authors of[33] suggest the use of ANN to increase the accuracy oftheir model to predict the number of people in an area.

• Deep Neural Networks- Neural networks lack creativity,whereas DNN is a better and adapted model thatcan be used for various applications that requirecreative conclusions. DNNs are opaque in nature, wherehigh-level is not explicitly programmed, but complexbehaviour emerges from the interaction between millionsof simple computational units. Further, it also uses alarge number of nodes to predict results. DNN consistsof an input layer, some hidden layers, and an outputlayer. The nodes of previous layers in this model transferinformation to the nodes of the next layer. The keyadvantage of DNNs is that they can progressively learnfeatures and modify their results on a certain basis. Theauthors of [44] have used DNN to determine a crowdevacuation route, and the model achieved 78% accuracy.Moreover, the authors of [45] propose a deep neuralarchitecture which can be used for crowd evacuation withthe help of UAVs. Also, authors of [46] and [47] haveused DNN to predict the number of people in an area.Next, we discuss a specific type of DNN known asConvolutional Neural Network (CNN), which are widelyused for analyzing images and videos.

• Convolutional Neural Networks - ANN sometimeslose the orientation of pixels in an image. This canbe a hindrance in classifying aerial images collectedby UAVs. In contrast to ANN, CNN can classify withbetter accuracy and are better at capturing orientation.Authors of [48] have proposed CNN to highlight the riskin areas after a disaster. CNN models are a sequenceof layers, including input, convolution (which contains

6

Figure 3: Basic CNN model with two convolution layers (C1 and C2), two pooling layers (P1 and P2) and two fully connectedlayers (FC1 and FC2) which is used to detect changes in a disaster area [40].

Figure 4: a) Convolution layer architecture b) Pooling layer architecture [40].

several tensor filters), pooling, and fully connected layers.• Convolution layerThis layer connects all the nodes in the lower layer tothe upper layer, as illustrated in Fig. 4(a). In Fig. 3,C1 and C2 are two convolution layers and are directlyconnected to the input. This layer’s calculations arebased on its node value and edge weight. This layerperforms the same function as passing an imagethrough a filter. The edge weights define the featuresto be focused on in an image. To be precise, this layerextracts features from an image.• Pooling layerThis pooling layer (P1 and P2 in Fig. 3) helps theoutput of the CL to achieve translational steadiness.This layer’s lower layer is connected to the upperlayer differently from convolution, i.e., the maximum

value is determined from all nodes in the lower layer,and that is passed to the upper layer, as illustrated inFig. 4(b). This layer tends to affect less during minortransformations in an image because it only considersthe maximum value.• Fully connected layerAll the nodes in this layer are connected to the nodesin the lower layer. This layer calculates the scores foreach class and determines the class for the image. InFig. 3, FC1 and FC2 are fully connected layers whereFC2 determines if there has been a change in an imageor not.

Authors of [40] use an approach based on CNN toidentify the areas affected by disaster and achieve 81-90%accuracy. Also, authors of [49] have used CNN toevaluate future disaster risk. Furthermore, the authors

7

of [50] have used a CNN-based model for internet ofhealthcare things (IoHT). We can use it for diagnosis andtreatment of people affected by a pandemic.

B. Unsupervised Learning

Unlike supervised machine learning algorithms, the datafed to the computer in unsupervised learning is unlabeled.The user has no insights into the nature or pattern of thedata. These algorithms try different techniques to detectpatterns or explore the structure of information. The mostsignificant type of unsupervised learning is clustering.1) Clustering: In clustering, a common descriptivefunction [51] in which a finite set of clusters are used toclassify the data is pursued. Clusters are usually generatedby considering that patterns in the same cluster should besimilar, whereas patterns in different clusters should notbe.

• K-means - It is one of the common techniques to clustera dataset into several clusters. In K-means clustering,initially, the number of clusters (k) are initialized.Centroids are selected for N points randomly by totteringthe dataset. Then the centroids are updated by taking themean of all points within one cluster. The iterations mustcontinue until the clusters stop changing. For calculatingthe similarity, Euclidean distance or cosine similarity isgenerally used.Authors of [52] used K-means to detect which areas weredamaged and affected by flood by clustering the dataset.Authors of [53] have used K-means for clustering crowdbehavior. Similarly, it is used to classify the crowd intodifferent classes by the authors of [54]. Authors of [55]used it to predict the spread of cholera disease.

• K-medoids - It is another clustering technique that isbased on dissimilarities between data points. It workswell with outliers. First, k data points are randomlyselected as medoids. All the remaining points are thenassigned to these medoids based on minimum distance.The dissimilarity between two data points Xj and Yj iscalculated as:

Dis =∑Xj

∑Y j∈Xj

|Y j −Xj|. (3)

This might provide us with different clusters in severalruns because of the initial randomness of medoids.Authors of [56] have used the K-medoids algorithm toplan an evacuation route.

• Fuzzy C-means - Similar to K-Medoids, this also beginsby randomly assigning centroids and initializing the datapoints randomly to clusters. This algorithm uses a specialparameter αjk and fuzzy parameter F . αjk is the extentof the associativity of a data point to a cluster, and F isgreater than 1. When F = 1, it becomes the k-meansmethod, and on increasing its value, it tends to fuzzinessmodel.For a specific number of iterations, medoids arecalculated and updated for every point in the cluster.Authors of [57] have proposed a hybrid model of Fuzzy

C-means with a neural network to classify Landsatimages, that can be used in predicting disasters. Thismodel achieves 96.87% accuracy by outperformingsimple Fuzzy C-means, which achieves 91.80% accuracy.Fuzzy clustering is used by the authors of [58]providing healthcare-as-a-service, which can be helpfulin post-disaster management.

C. Reinforcement learning

Reinforcement learning is a type of machine learning thatcan be regarded as a self-sustainable system, learningiteratively. A common algorithm used for this is explainedbelow.Q-learning - Decision making happens sequentially here.The output depends on the current input giving the bestsolution, and the next input depends on the previousoutput. It is widely used in applications where the aimis to maximize or minimize an objective function. Fπ(p)is the objective function, π represents the policy and pdefines a particular state [59].

Fπ(p) = Gπ

[ ∞∑t=0

γrt (st, bt) |p0 = p

]= Gπ [rt (pt, bt) + γFπ (pt+1) |p0 = p]

(4)

Q-learning uses model-free learning, where the agentwill not try to learn explicit models of the statetransition and objective functions. However, it directlyderives an optimal policy from the interactions with theenvironment.Authors of [56] and [60] have used reinforcementlearning to determine an evacuation route. The objectivefunction in this scenario can be defined as to minimizethe risk on a route.

Summary: In this section, we have presented the basicsof various machine learning algorithms which are furtherclassified as supervised learning, unsupervised learning,and reinforcement learning. All these ML algorithms canbe applied in various phases of disaster and pandemicmanagement. The next section focuses on the detailed reviewof works that use ML models in different stages of disastermanagement.

III. APPLICATIONS OF MACHINE LEARNING MODELS INPRE-DISASTER MANAGEMENT

Disaster management aims to reduce the impact of a disasterand save as many lives as possible. This section presentsa comprehensive review of ML models in association withother technologies such as IoT, UAVs, geodesics, satellite,remote-sensing, and smartphone-based that are used forpredicting a disaster, crowd evacuation, and post-disasterscenarios.

A. Predicting a disaster

If a disaster is predicted in time, warning signals can befloated to people, and they can take the necessary safetymeasures. The accurate prediction of a disaster lies in

8

analyzing spatial and temporal data of an area and predictingthe characteristics of a disaster, such as the water level of aflood or the magnitude of an earthquake. In the subsequentsections, we will discuss the use of different technologies indisaster prediction and how various ML-based models can helpin enhancing the efficiency of other methods in predictingdisasters.

1) IoT and ML-based Models for Disaster Prediction: IoTis an arrangement of interconnected machines used to collectdata over a network without human interference [61]. It hasenabled the deployment of a large number of applicationsin various fields. Authors of [62] present the use of sensorson trees for predicting fire outbreaks in forests. The sensorsmeasure features such as temperature, CO levels, greenhousegases, and moisture. Similarly, microwave sensors can beused to analyze the earth’s movements for the prediction ofearthquakes.

ML algorithms can be used to process the data collected byIoT devices and provide better accuracy. Flood managementraises a variety of obstacles for IoT based strategies. It involvesa complex set of parameters with multiple interdependencies,including rainfall, pressure, and rate of flow. The sensornetwork has to be configured according to the parametersselected. The number of sensors and their interconnectionrelies significantly on the river’s measurements. Authors of[63] present an approach in which ML is used for detecting aflood. It uses Hadoop MapReduce for the removal of duplicatevalues. Then the rules are generated based on four attributes,namely, rain sensor, humidity, water flow sensor, and waterlevel sensor. The rules are further used in a CNN. Theproposed neural network classifies the data into positive andnegative for the occurrence of floods.

The model in [64] uses multiple sensors connected toADC (Analog to Digital Converter) like temperature, moisture,water level, and CO level. The data is sent to a raspberry-pidevice in digital form. Then, this data is analyzed with theQ-learning algorithm based on penalty and reward. Q-valuesare used to analyze the risk of any activity in mines. Similarly,the model in [65] uses ANN and logistic regression forrainfall prediction. An integrated framework focused on IoTand ML is proposed in [66] to forecast flood risk in ariver basin. To capture data, the software uses a revampedmesh network interface via ZigBee to the WSN and a GPRSmodule to transfer data over the Internet. The data sets arethen analyzed using an ANN. The findings of the studyindicate a significant change in the approaches already in use.The selection of sensor network deployable areas is also abig challenge. Additionally, certain QoS parameters, such aslatency, instability, severity thresholds, etc., should be checkedbefore implementation. Authors of [67] have explained howvarious ML algorithms can be implemented to address thesecurity concerns of IoT models.

The recent development of UAVs (Unmanned AerialVehicles) has solved the accessibility issue for areas thatare difficult to reach and monitor because of human andlogistic constraints [68]. They amplify satellite images withobservation gaps and are a scalable and portable solutionthat can provide high-resolution images. However, the cost

of UAVs makes their deployment an obstacle for developingcountries. Authors of [69], [70], and [71] present cost-effectivedrones that have a longer range and bigger payload andcan be effectively used for mapping real-time data andmonitoring nearby activities. Also, conservationdrones.org, anon-profit organization, is actively building low-cost UAVswithout compromising any of its features. Authors of [72]present a Smart-eye model wherein UAVs use LTE tocommunicate with Smart-eye centers. This model uses imagestitching technology in which it compares the real-time imageswith the previous original images of an area. It uses aCCTV (Closed-circuit Television) camera, which makes itcost-effective. The minimum altitude and velocity of UAVsneed to be always regulated in this model. Further, this modellacks a larger flight time. The model in [73] presents a solutionto tackle these parameters using a web-based platform inthe case of Rotolan landslides. Further, the model in [74]uses mobile nodes to provide resilience when a node failsduring a disaster. A sensor node, connected to UAV andacting as a mobile node is deployed in [75] and is capableof communicating with WSN in the area. The limitation ofthis model is that it cannot survive during extreme weatherconditions, and its low battery life is still a challenge. Though,after utilizing a mobile anchor node for localizing unknownnodes, WSN provides us with a better fault tolerance than otherUAV models [76]. It focuses on predicting floods and deploysa pressure sensor that is used to easily estimate the height ofthe water level. Flooding is predicted based on observationof a sudden increase in water level. Also, this model needsless aerial space and can easily land on irregular grounds dueto its rotary wings as compared to conventional fixed wings.With the use of UAVs, there is a threat to the privacy of theuser’s data because it is deployed in an open atmosphere. Theauthors of [77] propose that we use blockchain for ensuringthe data security.

Remote sensing images are extensively utilized in diversefields. After the collection of images from drones, they needto be classified efficiently. ML algorithms are well suitedfor this task. In the work proposed in [78], water featuresare distinguished from non-water features to identify floodedareas from Landsat 7 (L7) images incorporated with DigitalElevation Model (DEM) data.

Unlike [78], the model presented in [85] uses passiveLandsat 7 and active RadarSat images to analyze floodsand develop a flood hazard map. The classification of waterand land areas are easy in RadarSat images, and resultswere compared with Landsat images. High-resolution Landsatimages could be used to provide information about floods,except for the monsoon season due to the cloud cover.However, RadarSat images can analyze during the monsoonseason also. A map highlighting all the information about therisk of floods was developed using the above and GIS data.

Authors of [24] use a hybrid of SVM and K-means todetect flooded areas. It classifies aerial images collected bydrones into flood-affected and non-flood affected. This modelachieved an accuracy of 92%. This model may also bedeployed on drone sensors so that they can automaticallydetect the areas. Authors of [57] aim at forming a three-band

9

Table II: Summary of related works for disaster prediction and early signals (Acronyms used in the table- CEOS: Committeeon Earth Observation Satellites, DEM: Digital Elevation Model, NOIR: Network Optimization Indicator Report).

Category Reference Target issue Technology used Hardware/API used Case studies

[63], 2020 Floods prediction Convolution neuralnetwork Hadoop MapReduce Surat, India

[65], 2019 Rainfall prediction ANN and Logisticregression LoraWan UEM CampusIoT and ML

[66], 2016 Forecast flood risk Artificial NeuralNetworks ZigBee,WSN None

[79], 2019 Prediction Geofencing UAVs Base station,Flight controller Surat, India

[80], 2017 Assessing areas Adhoc network asaerial mesh network

Raspberry-pi withNOIR pi camera UEM Campus

UAV[81], 2019

Disasterassessment SWIFTERS DJI UAV, ROS library,

Map server None

[82], 2019Disaster

assessment

Star algorithm, Tabusearch, Gradient

descentMulti-UAVs Jiuzhai valley earthquakes

ML and UAV [24], 2019 Detecting floodareas

SVM, K-meansclustering and PCA Drones Aerial images

[83], 2010 Better predictionCentre surrounded

filters with gaussianweighted mean centers

Air-borne LiDAR 2010 Haiti earthquake

Geodesics based

[84], 2013 Better prediction GEOSS and CEOSapproach

Spatio-temporalinfrastructure

2008 earthquake inSichuan, China andNamibia flood plot

Satellite [78], 2002Water and non-water features

Landsat 7 Thematicmapper and DEM data None Floods in Pitt County,

North Carolina

Remote sensing [85], 2011Develop floodhazard maps

Supervised MCLclassification

RADARSAT remotesensing data, GIS data

and ground dataMaghna river basin

ML and Remotesensing [36], 2012 Classification of

land coverRandom forest

classifierLandsat-5 Thematic

Mapper data Province of Granada

Data mining [86], 2005Tropical cyclone

intensity Apriori-base None Atlantic basin

Android based [87], 2017 Alert signals Partition basedtrajectory distance JSON file Haiti

ML andObject sensing

[37], 2011Prediction of

floodsRandom forest

classifier NoneMomance River — Haitiand Wenchuan town —

China

[88], 2017Classification of

land coverRandom forest

classifier and SVM None Scopus databases

[32], 2018Sandstormdetection

CART decision tree,Naïve Bayes and

Logistic RegressionNone Riyadh, Dammam, and

Jeddah

[40], 2017Flood and

landslide detectionConvolutional Neural

Network None Japan and Thailand

[42], 2014 Storm intensity

Symbolic AggregateApproximation (SAX)and Artificial Neural

Network (ANN)

Satellite-image data Typhoon and Tropicalcyclones

ML

[25], 2016Disaster

recognition BB-SVM ERESS Kansai University

image with higher resolution, also known as a fused image,using Brovey transform. PCA (Principal Component Analysis)is used to reduce this to a 1-D image. Then, a neuro-fuzzyapproach that uses KCN and FLCIM algorithms are used. TheFCM model depends on the value of m (fuzziness index) andleads to wrong results. The use of this hybrid approach, calledFuzzy Kohonen Local Information C-Means (FKLICM),overcomes the limitations of both these models. This modelalso shows a higher accuracy than the FCM approach. Thehigh accuracy and non-parametric nature of random forest

classifier can be used for land cover classification [36]. The RFclassifier, due to its advantages mentioned in Section II in thispaper, shows better accuracy. The key downside of the modelis that because of the numerous classification trees producedby the resampling of the same dataset, it becomes challengingto grasp the rules used to produce the final classification.

This has contributed to the advent of object-based imageprocessing (OBIA) methods that have been developed toresolve these problems. The remote sensing communityhas made significant efforts for almost twenty years now

10

Figure 5: Object oriented image analysis integrated with random forests [88], [37].

to encourage the usage of object-based technologies forland-cover mapping [89]. Supervised algorithms combinedwith object-based detection methods have always beenan important part of land-cover mapping remote sensingwork since 2010 [90]. Many choices such as segmentationsystem, accuracy evaluation, classification algorithm, trainingdata, input characteristics, and objective groups mustbe chosen for object-based classification processes. Toaddress these requirements, many authors have proposedsupervised object-based classification methods particularlyaccommodated to each use case. Authors of [88] highlightthat Landsat series remote sensing images are mostly used incontrolled object-based detection due to their strong qualityand usability. Moreover, the blurry rule-based classificationstrategy meets a plateau in object-based classification, whilesupervised object-based classification reaches a productionheight. RF shows the best results in an object-based grouping,followed by SVM. In fact, NN (neural networks) remainunsuitable for more comprehensive usage in the descriptionof objects. However, this approach still needs to be verifiedfor urban areas.

The model presented in [37] uses a random forest combinedwith object-oriented image analysis (OOA) for landslidedetection tested on sample datasets of Haiti, Italy, Chinaand France, which were recently affected by landslides.Integrated variable measurement and selection processes, andhigh-quality program applications that are freely available,allow random forests to integrate with OOA, as illustratedin Fig. 5. The authors found a clear over-predictionof landslide areas for all situations if a class-balancedsample was utilized in training. Over-prediction was moreprominent for Barcelonnette (France) and Messina (Italy),where visual inspection of the photographs already indicateda large difference in class when compared to the other tworegions. More work is required to refine the segmentation

process, which is currently focused solely on spectral details.Therefore, a preliminary sample-based estimation of thevariable value can be an important method for determiningwhich additional layers would be included in the segmentation.

The authors of [91] performed an analysis to observeintensity guidance accuracy of tropical cyclones over a periodof two decades. The models proposed to forecast the strengthof the hurricane up to 120 hours advance, but forecasts ofseverity are less reliable relative to forecasting the path ofthe storm [91]. The Weather Researched Forecast (WRF) isa regional predicting framework that monitors sandstormsdependent on the previous day’s environmental conditions.Authors of [32] predict whether, in the following hours, a sandstorm would occur or not, by utilizing ML algorithms such asCART (Classification And Regression Tree) analysis, logisticregression, and naive Bayes classifiers. The system classifiessandstorm or no-sandstorm events and sends alert signals topeople. The results show that CART analysis performed betterthan naive Bayes and regression-based classifiers.

CNNs are resilient to darkness, and capable of sufficientlyand most critically obtaining the catastrophe attribute toresolve misleading by variables, which would impact theefficacy of disaster prediction. By trimming and scaling aerialimagery gathered from Google Earth Aerial Imagery, authorsof [40] generated training data patches for before and afterdisasters. Both patches are qualified using CNN to isolate thearea which is recognized as a catastrophe zone. The accuracywas 80-90% for this system. The authors of [86] presentan apriori-based association rule algorithm for detecting theintensity of tropical cyclones. According to the associativityalgorithm, if items X and Y are selected at random, we canbe certain that item Z was also selected. For the associationrule, the user specifies minimum support. The issue with thismodel is that we need to decrease the minimum support andmulti-term associations need to be selected. To solve this issue,

11

authors of [42] use Symbolic Aggregate Approximation (SAX)combined with ANN to approximate the intensity of stormsby using 8 features. These features were reduced by PiecewiseAggregate Approximation (PAA) and were given as inputs toneural networks. The F-measure for this model was 0.93.

With the recent development of airborne-LiDAR, it isextensively used in diverse fields such as medical, military,agriculture, pollution, and architecture [92]. It provides 3Dhigh-resolution samples of earth’s topographical features andother objects. These detected features may be used to predictdisasters such as floods, landslides, earthquakes, etc. withbetter accuracy. Geodesic curves are used to estimate the meancurvature for each vertex [83]. It can detect terrain featuresaccording to a variety of surfaces (i) Pass (ii) Channel (iii) Pit(iv) Ridge (v) Peak. Then, the Gaussian-weighted average iscomputed for those curvatures. The currently used method in[83] can be sped up by using a multi-resolution mesh. Also,more features could be taken into account such as normals andprincipal curvatures.

Earth observational evidence in mapping disastermanagement is also used by Group on Earth Observations(GEO) and the Committee on Earth Observation Satellites(CEOS). The AIP (Application Infrastructure Provider)process of GEO has evolved and recently adopted theReference Model of Open Distributed Processing (RM-ODP)[93]. The GEOSS (Global Earth Observation System ofSystems) (2005-2012) [94] aims at disaster reduction andmanagement. By combining the earth’s observations withother information, it aims at spreading early-warning signals,better preparedness, and risk assessment for rebuildinghouses post-disaster. CEOS is working on increasing the useand raising awareness of EO (Earth’s observation) data fordisaster management. It has undertaken a project to reducethe use of ad-hoc arrangements in the management systemthat have become an obstacle in processing EO data. Authorsof [95] present us with a solution for these obstacles, inwhich research and improved EO to help tackle the issuesencountered by remote agencies in processing satellite data.However, some things remain unclear, like the resourcesshared, their interdependencies, the effect of shape datapolicies, and how new users can access services [84]. Theauthors of [84] present a CEOS GA.4.D architecture withan emphasis on using GCI (Ground Controlled Interception)to address issues such as user authorization and combiningsocio-economic data with EO for better observations. It alsoidentifies gaps in services and focuses on limited resources forvictims. However, this architecture requires more unrestrictedaccess to disaster data. Attempts to use this architecture atregional and local levels are in progress.

Emergency departments are usually equipped withsyndromic surveillance systems that generally use statisticaldetection algorithms like CUmulative SUM (CUSUM) [96],auto-regressive integrated moving average (ARIMA), theHolt-Winters algorithm, and the Early Aberration ReportingSystem (EARS). The performance of these algorithmsdepends on a trade-off among false-positive rate, sensitivity,and timeliness. The work by authors of [97] shows that themodels mentioned above require intensive parameter tuning,

Figure 6: Disaster prediction by BB-SVM [25].

and simply using the preset parameters is not always optimalfor a given algorithm or data set. Moreover, they showedthat data needs to be examined regularly to fine-tune theparameters in these algorithms according to the data used.These observations limit the abilities of small syndromicsystems to predict a widespread outbreak.

B. Early signals

If early signals are delivered to people in an area at the righttime, numerous causalities can be prevented. Such systemscan be installed in offices, homes, or other public places towarn people in time. Sensor networks are recommended forthe systems to track disasters (e.g. the Building AdministrationSystem) and address this issue [98], [99]. Such devices need awide range of separate sensors such as smoke, heat, radiation,etc., installed in advance in houses, and such technologies mayhave significant costs to be incurred.

The authors of [100] have proposed a program to suggesta model that is specialized for weather detection. However,we can only utilize such systems in limited instances wherethe sensor grid has formerly been installed. Even then, ifthe system cannot operate due to power loss, sensor damage,broken transmission lines, and so on, then this sensor networkcannot function properly. The model in [25] aims to improvea previous model ERESS (Emergency Response SupportSystem) based on MANET (Mobile Ad-hoc Network) usingML [101]. ERESS Mobile Terminal (EMT) accumulates datafrom its sensors. Every EMT uses its personal sensors, whichallows it to operate everywhere. The disaster-recognitionalgorithm method is described as follows: (1) An EMT usesSVM for detecting some temporary risk; (2) If the EMT sensesa temporary risk, it floods the details to those in the area; (3)Each EMT regularly measures temporary risk identificationaccumulation; (4) Every EMT matches the logistic model witha regression curve; (5) A disaster is determined if the curveis convergent to the values of all EMTs within a building ina period of 1 minute or less. This method of detection alsomakes the system autonomous.

The BB-SVM (Buffering and Bagging SVM) model’s EMT[25] has two buffers: (1) State judgment buffer and (2) disaster

12

detection buffer. The model comprises of two phases. Thefirst phase is shown in Fig. 6. First, each EMT generatesinputs from all sensors and transfers them to a SVM model.When the input data is determined to be that of a crisis,the EMT amasses the outcome in its state judgment buffer.The outcome stays in the buffer before Time to Live (TTL)runs out. If the reserve for the state judgment is reached tothe limit, the EMT must transmit about it being in a crisis.Second, the EMT accumulates the temporary risk in its disasterdetection buffer, ignoring other EMTs. Often, the temporaryrisk stays in the buffer before TTL runs out. If the bufferfor disaster prediction accumulates to the amount of half thenumber of EMTs, a disaster outbreak is determined by theEMT. The TTL duration of a buffer may be adjusted throughbagging learning. For this purpose, the BB-SVM changes thetemporary warning or disaster outbreak parameters accordingto specific circumstances with less data than the traditionalapproach. The EMT in this model changes the judgment of thedisaster outbreak repetitively. Also, there is the issue that EMTrelies on the activity of just the individual to determine theoccurrence of a catastrophe. Finally, ERESS cannot distinguishthe incident caused by an unusual sound or light.

The authors of [87] present another approach that uses amobile application to alert people. It uses a partition basedtrajectory distance to find the nearest shelter place duringevacuation. SQLite database is used, and this applicationprocesses the JSON (JavaScript Object Notation) file anddelivers messages through notification via the app. This modellacks the ability to deliver early signals during immediateeffect disasters such as cyclones, floods, etc.

Summary: In this section, we reviewed the variousapplications of ML models in disaster management. Thevarious algorithms and models used for disaster predictionand communication are presented. The approaches using MLalgorithms for this are summarized in Table II.

IV. APPLICATIONS OF MACHINE LEARNING MODELS INCROWD EVACUATION

A large number of people around the globe have sufferedfrom natural disasters since ancient times. Delays in theevacuation during a disaster often lead to increased casualties.This often happens because the evacuees do not recognizethe routes of catastrophe occurrence and evacuation. Tominimize the number of casualties caused by such incidents,it is necessary to quickly identify catastrophe locations andto figure out as soon as possible any appropriate routesfor evacuation. However, it is challenging to identify thecatastrophe occurrence and in very limited time, eg. 30−40seconds during earthquakes, to direct individuals to suitableescape routes. Moreover, traditional approaches have not beenable to assist evacuation immediately following the onset ofa catastrophe. More rapid evacuation is needed for suddendisasters. For example, carbon monoxide can spread quicklyduring fires in houses [102]. Thus, it becomes vital to evacuatejust after the occurrence of the disaster.

APF (Artificial Potential Field) can be used to developa crowd lives oriented track and help optimization system

(CLOTHO) for crowd evacuation [103]. This app uses amobile terminal (IoT side) for collecting data and cloudfor storing data. This model has four main modules: datacollection, network transmission, cloud, and user platform.Attraction and repulsion potential field from a disaster pointare calculated. The APF calculated is further controlled bythe shelter distance threshold. The resultant force is used todetermine the direction of evacuation. Unlike [103], the modelmentioned in [104] involves two layers: sensor layer, which isused to collect data and uses the MTS400 CC sensor, and theIoT layer, in which real-time data is processed using resourcescheduling and banker’s algorithm. The limitation of thismodel is that developments are still needed in a heterogeneousscenario. ML algorithms can be used to process a large amountof data in these models.

Mobile phones are an important component used tocommunicate with victims during a disaster [105]. Thedevelopment of GPS has made it easier to track the locationof victims. During a disaster, the network is often congesteddue to a lot of SMS messages, and thus, cell broadcastingservice is used to contact victims [65]. This further reducespanic. For evacuation of people, the authors of [106] presenta GSM (Global System for Mobile Communications) alarmdevice to be deployed in the nearby area for early-warnings.This device collects its data from the weather office, but thisdoes not provide the residents with a quick way for evacuation.An area mail disaster information system for tsunami warningand evacuation system to assist fishermen is presented by theauthors of [107] and [108]. This system tracks fishing boatlocations using GPS and provides them with warning signalsthrough mobile phones. This is still a faster approach than theprevious one.

Different from [106], in [109], the authors have proposeda smartphone app that notifies people about the evacuationroutes. They have used two parameters to evaluate the locationof the user and proximity-relations using Bluetooth. The datais processed through a SVM model that outperforms the linearregression and Gaussian models. However, this model requirestesting with real-world data during a disaster. The authorsof [56] have proposed an improved reinforcement learningapproach for crowd evacuation. The main challenges whichwere covered were modifications in the model when thereis an increase in people density and effective communicationbetween authorities. For the initial clustering, the K-medoidsalgorithm is used. Deep learning algorithms could also beimplemented for the same model to improve its accuracyfurther.

With the recent increase in the usage of smartphones,several applications have also been deployed for crowdevacuations. The authors of [111] use an app employing OSM(OpenStreetMap), DMS (Document Management Service),and Google C2DM server. The limitation of this app is thatthe data must be uploaded on OSM. This model failed whentested in Bangladesh because of the lack of uploaded data.When a user logs into the application, he/she is provided withan unique ID that is then sent to the DMS. If the applicationdetects a risk in the user’s area, it notifies him/her. In caseof risk, it also provides users with the shortest route to the

13

Table III: Summary of works for crowd evacuation (Acronyms used in the table- CIoT: Cellular IoT, COCO: Common ObjectsIn Context, LSTM: Long-Short-Term-Memory, RCNN: Recurrent CNN, SSIM: Structural Similarity Index).

Category Reference Target issue Technology used Hardware/APIused

MLinvolved Case studies

[103],2018

Crowd Evacuation CLOTHO (APAand APA-RF)

None No Chemical plant inNanjing, China

IoT[110],2018

CIoT UAVs None No Snow Avalanches

[111],2012

Evacuationprogram OpenStreetMap DMS No Bangladesh, India

Androidbased

[112],2014

Rescue teamtracking

Algorithm basedon location GPS-Receiver No Indonesia

[113],2016

Finding a safezoneroute OpenStreetMap PostGIS database No Makati, Phillipines

(Earthquakes)[114],2019

Clustering (Birdflocking) FANETS No None

UAV[82], 2019

Priority based routeselection

Graph theory,Euler cycle and

integerprogramming

None No Hospital Pavia Arecibo

[115],2015

ActivatingContraflows

Decision trees Weka (Datamining software) Yes Hurricanes

[109], 2013 Prevent crowddisasters

SVM, Linearregression and

Gaussian modelGPS Sensors Yes San Francisco

[25], 2016

BB-SVM,Dijkstra’s

algorithm andDepth first search

(DFS)

ERESS Yes Kansai university

[44], 2019Deep Neural

networksRaspberry Pi-3 andspectrum analyzer Yes Ritsumeikan university

[56], 2019K-medoids andReinforcement

learningNone Yes Office scenario

[116],2017

LSTM model Spatio andtemporal features Yes Kumamoto

earthquakes[117],2014

Determine anevacuationroute

Naive Bayes None Yes Fire hazard

ML

[21], 2018SVM and Fuzzy

logic None Yes Hajj (A Muslimpilgrimage event)

[53], 2018K-means and

hierarchialclustering

Alarms Yes An office building

[39], 2016Deep CNN andRandom Forest None Yes UMN, UCSD and

Pets2009

[54], 2019

Classification ofcrowd situation CNN classifier and

K-means None Yes Video data

[60], 2018Planning

evacuation routeReinforcement

learning None Yes Hong Kong fireoutbreak

shelter. The authors of [113] also use OSM and store datain a PostGIS database. One of its functions, which is tocalculate the minimum distance between two geographicalplaces, is used. A parameter called building risk factor (BRF)is calculated, which denotes the danger that a building posesto the road during an earthquake. Then, the Dijkstra algorithmis used to find a safe route. This model only uses BRF as theparameter to decide the safest route. Further, this model did nottake into account the building’s height. The taller a building,the more danger it would pose to the road. An application

that covers all these limitations using Bayesian networks isproposed by authors of [118].

The use of cameras for collecting data for crowd evacuationaffects people’s privacy. Thus, models are needed that work ona mechanism different from analyzing and processing cameraimages so that the public’s privacy is not violated. The authorsof [44] have proposed a model that analyzes radio wavesof a user’s smartphone. The strength of waves is measuredusing sensors. Due to the complexity of the fluctuations of thedevice’s waves, an ML-based model is necessary. The density

14

of the crowd in an area is determined by using a Deep NeuralNetwork (DNN), as illustrated in Fig. 7. This model achievedan accuracy of 78%. However, some other ML algorithms canbe used that require a smaller training dataset in comparisonto this model.

Instead of Micropilot MP-Vision and other components, theauthors of [119] use a custom UAV designed with resourcesto achieve lower cost and maintain the quality level. Thismodel’s air-frame uses a Super SkySurfer fixed-wing EPO(Extended Prediction Orbit) foam frame, ArduPilot Mega(APM) autopilot system, airspeed sensor, camera payload, andbatteries. Radio telemetry is used to establish communicationbetween the UAV and GCS (Ground Control Station). TheGCS uses APM Mission Planner that is employed to controland generate routes for the UAV during its flight. This UAVhas several advantages over others: increased flight time (30-50minutes), and large coverage of area (4 sq km). Also, theautopilot system is open-source, which enables independentdevelopers to be engaged in identifying issues and updatingsoftware. The disadvantage of this model is the loss of UAVdue to GPS signal loss, and it lacks a safety mechanism duringfailure. Further, it is more prone to hardware errors (assemblyand wiring of parts) and bugs in code. A mechanism is alsorequired to be added in the model to avoid risks during bellylandings. An approach based on Q-learning is proposed tocover these limitations in [120]. This approach targets robustradio-network signals by not compromising on the UAV’sfeatures such as its flight time, cost, and coverage area. Theauthors of [45] propose a deep neural network architecture forfacilitating navigation of UAV in indoor environments, whereGPS fails in precise localization and navigation.

The use of lane contraflow reversals is important foremergency evacuations. Cities or areas with emergencyevacuation measures in effect may utilize contraflow lanes toincrease the number of roads necessary for traffic to evacuate.Contraflow lane reversals reverse traffic paths along escaperoutes from their usual directions. Hurricane Katrina haddevastated southeast Louisiana in August 2005. An evacuationplan which successfully evacuated an estimated 1.1 millioncitizens was implemented before landfall [115]. The reversal ofthe contraflow lanes was in effect for 25 hours and resulted insignificantly decreased traffic delays. However, ML algorithmscan be implemented for locating new parts of the contraflowand establishing evacuation measures in other sensitive areas.Research in [121] explores a method focused on existingtraffic patterns and path availability to enable contraflowlane reversals. To model the contraflow plan, the evacuationroutes and local traffic situation graphs are used. A pathassessment algorithm is used to evaluate if existing trafficpatterns need contraflow lane activation. Decision trees areused to determine when the current state of the evacuationmodel should activate all contraflow lanes. The next stepsin the proposed methodology is to expand the number offeatures that the training set takes into consideration, changethe methodology to define specific contraflow segments, anduse simulation tools to construct training sets and validateoutcomes. The authors of [25] present a BB-SVM based modelfor predicting disasters. Then an evacuation path is chosen

based on Depth First Search (DFS) and Dijkstra’s algorithm.Tracing safe routes can also be done using a Multi-Objective

Genetic Algorithm (MOGA) [122]. This process involves threephases. The collecting phase involves collecting GPS andacceleration data from the user’s smartphone to the cloud.The analyzing phase involves calculating the walking distancebetween nodes, walking speed (average), and pedestrian trafficto determine safety evaluations. With safety evaluations, timetaken, and the distance between the initial and final nodesare calculated, and the maps are provided to the users. Thismarks the third phase. The user has a choice to choose mapsaccording to their desired characteristics (shortest time orshortest distance or safety). The characteristics of the routesget upgraded continuously, which is a unique feature of thismethod compared to other works. Moreover, new maps aregenerated using selection and cross-over techniques that helpin providing multiple evacuation route maps. The model hasnot yet been validated with real values of average speed andfor pedestrian traffic (the authors here assumed it to be a setvalue in calculating safety evaluations). Also, alpha, beta, andgamma selection for the calculation of safety evaluation valuehave to be chosen optimistically, which remains a challenge fordifferent types of places. The authors of [123] propose a hybridof genetic algorithm and ML algorithms such as SVM, naiveBayes, and random forest to process UAV images for disastermanagement. This was successfully deployed on a wildlifereserve in Namibia. The model achieved 80% accuracy withSVM and random forest outperforming the rest.

Tracing safe routes can also be implemented througha similar approach based on Virtual Voting and AdaptivePricing Algorithm [124]. Authors of [125] have used K-meansclustering and K-nearest neighbor regression algorithms forthe same. Data is first pre-processed and categorized into crimeand accident datasets. The place is divided into clusters bylatitude and longitude by K-means clustering based on crimesand accidents. ’Direction service’ is a class of Google APIthat gives all possible routes from source to destination, andway-points are assigned for every 2 km. The risk score of theway-points is calculated based on K-nearest clusters aroundit. A risk score is simply the summation of accident scoresand crime scores determined by assigning scores to each typeof crime and accident. The value of K is determined by thelowest root mean square error. KNN scores for the way-pointsare analyzed using the R2 score, which is a simple measure ofdistance from the regression line. Experiments were conductedfor Manhattan Borough in New York City, and it successfullyshowed alternate safe routes that are different from the defaultroute shown by Google Maps based on the shortest distance.Previous works failed to categorize small areas and onlydid larger area classification with either SVM or other MLalgorithms. However, this work has found safe routes coveringsmall areas. KNN analysis showed high R2 scores of 0.910(accident score) and 0.974 (crime score), which implies thatthe classification has been done well and is distant from theregression boundary. Moreover, previous works in this fieldwere highly subjective, whereas this work is not. However,more factors have to be taken into account in calculating thesafety of the route (risk score). Also, this method consumes

15

Figure 7: Deep Neural network model to determine crowddensity [44].

more time in pre-processing, which can be improved (60-80%of the entire time).

Traditional neural networks encompass some constraints,such as the inability to vary the length and work on differenttraining data. To target this issue, the authors of [116] haveproposed an LSTM (Long Short-Term Memory) model forplanning crowd evacuation. The CNN layer can be utilized asa filter in this model that processes spatial data and leaves theLSTM with only temporal data. Since LSTM can be trainedusing complex real-time data, it achieved 95% accuracy in thereported experiments. Different approaches using supervisedand unsupervised machine learning algorithms were presentedin [117] and [21]. The authors of [117] have used a naiveBayes model. This model is deployed in a fire outbreakscenario. Using a Bayesian model, it determines the routesthat are safe to use for evacuation. It assigns probability-basedphysical and emotional factors to every route which shows thesafety probability of the route to be used.

The authors of [21] have proposed a model that combinesSVM with fuzzy logic. SVM works on choosing the bestrepresentative crowd group by integrating both the physicalproperties of an area and the characteristics of a crowd. Thefuzzy rules are then created on the basis of this output of theSVM model. The authors reported accuracy of 98.77% forthe model. However, the effect of psychological factors is notconsidered in this model.

One way to plan evacuation routes successfully can bepre-detection of the behaviour of crowds so that optimal stepscan be taken. The authors of [53] have initially detected acluster of crowds, built a feature matrix, and then used MLalgorithms. The behaviour of crowds is divided into threeclusters: (1) the ones assembling in groups, (2) the onesapproaching exit doors, and (3) the ones following leadersor responsible authority. K-means and hierarchical clusteringare used for the same. Similarly, other behaviours can also bestudied to plan evacuation routes.

During a fire outbreak, people generally rush towards theexit, which can create chaos and result in more casualties.

To tackle this issue, the authors of [60] have proposed anevacuation plan based on reinforcement learning. On the basisof the previous movement of individuals, their next step ismodified in every iteration. The objective function is definedas minimizing the evacuation time. However, situations wherepeople wait for their friends or relatives or follow their paths,are not taken into account in this model. Also, the effectsof harmful smoke generated from the fire are not taken intoaccount.

The authors of [39] have proposed a model which analysesvideo of the crowd in an area. It uses a deep CNN modelto classify video features such as the speed of the crowd andphysical features of the environment to classify the crowd’sbehavior as a normal situation or an alert situation. Thisachieves an accuracy of 77.82%. Then, three random forestclassifiers were used to identify abnormal behaviour in themotion of the crowd. Further, multiple cameras can also beused to speed up the process. To improve the accuracy ofthe video-based model, the authors of [54] have proposeda new framework. The collected video data is classified inseveral situations by using a CNN model. This model alsocovers a mechanism to process distorted images by comparingthem with the original image’s pixels. Further, different crowdsituations are determined based on entropy techniques.

Summary: This section has covered a survey of ML-basedmodels to determine a crowd evacuation route. Apart fromML-based models, UAV, IoT, and android-based models arealso discussed. All the models are summarized in Table III.The common requirement for all ML-based models was aninitial clustering algorithm. K-means and CNNs were found tobe used more commonly as compared to other ML algorithms.

V. APPLICATIONS OF ML MODELS IN POST-DISASTERMANAGEMENT

A. Detecting ChangesThe changes post-disaster are important to analyze in order

to calculate the loss to the economy, to plan specific measuresfor the recovery from a disaster, and rehabilitation measures.First, in order to perform rescue operations, cellular networkshave to be regenerated in the affected area for communication[127], [128]. For this purpose, a framework for networksassisted by three different types of UAVs: (1) tetheredBackhaul Drone (tBD), (2) untethered Powering Drone (PD)and (3) untethered communication Drone Base-station (cDBS)is created. Further, algorithms like ant-foraging are commonlyused to drive UAVs to locations with a high density ofsurvivors. GIS analysis and remote sensing (RS) are thenused to plan and coordinate recovery. This is done byanalyzing spatial differences and then providing the requisiteinformation to those working on the ground. In the RS-basedassessment, most of the developed methods focus on physicalrecovery. Their indicators are mostly at the building-level,and thus collecting information manually becomes difficultand expensive. Further, the majority of UAVs have verylimited flight time, providing insufficient data for manualdifferentiation. Thus, incorporating techniques like machinelearning through which machines can be trained becomeextremely important.

16

(a) UAV-ML model for analyzing threats in an area.(b) When SVM model notifies UAV, it updates anddivides its search based on the information provided.

Figure 8: UAV-ML model for disaster relief in a city [126].

A vision-based disaster detection system using machinelearning algorithms to detect the affected area from aerialimages in real-time needs to be developed. Here, the imagedata would be collected by sensors mounted on UAVs, andthen processed by deep learning algorithms such as CNNsto detect the presence of a disaster such as fire, flood, andlandslide in real-time, and also the number of structuraldamages to the buildings [48]. A similar methodology usingthe algorithm of random forests was used in detecting damagesto buildings during the 2010 Haiti earthquake [35].

The authors of [126] have proposed a model in which UAVscollect the data, and it is further processed by a model builtusing structure-variable discrete dynamic Bayesian networkand SVM. In this model, UAV collects data at a particulartime interval from varying distances. As it collects new data,a new node is added in the Bayesian network with a particularprobability, as shown in Fig. 8(a). This data is fed to a SVMclassifier for identifying any threats in the area. If the SVMdetects any threat, it signals a message to the UAV, and itstime interval of collecting data is further divided into smalltime intervals so that the area can be evaluated accurately, asshown in Fig. 8(b). This leads to a better evaluation of an area.

ML techniques have the advantage of immediately filteringimages, which would have required months to be sortedmanually. Temporary settlements can also be detected,indicating areas of survivors in need. Machine learningapproaches can be used to combine data, remove unreliabledata, and identify informative sources to finally generate a heatmap, which identifies areas of urgent need. By application ofCNNs, fire and smoke have been successfully detected fromraw RGB images with an accuracy of 97.9 % [48]. Also, dueto their high accuracy, they prove to be a much better approachas compared to other techniques that are in popular use likeSVM and random forests. CNNs have shown advantages overother ML methods. However, its hidden layer is a black boxcreating uncertainty and dependency on the training data.

The authors of [129] use GIS analysis data to detect changesin an area post-disaster. It uses OpenStreetMap data andsatellite images. Enhanced change detection index (ECDI)is used as a parameter to differentiate between pre-disaster

and post-disaster images. This model is deployable in a largearea. Since ECDI calculates edges, gradient, and texture, thiscan be further used to analyze effects in buildings. However,the issue lies in changes in the solar zenith that manipulatesthe data. To address these issues, we can incorporate ML todifferentiate between the same. These models have a highreliance on real-time datasets and geo-referencing of data. MLalgorithms can work by only requiring post-disaster imagesand can result in better accuracy. The work in [52] showshow K-means can be used to detect areas that are affectedby a disaster. It achieved an accuracy of 85% with respect tolow-clarity and cloudy images dataset. However, the K-meansalgorithm mixed the damaged and reconstructed areas duringits initial clustering. The accuracy can be further improved ifhigh-resolution images are used.

Unlike the K-means approach used in [52], authors of [34]use the Simple Linear Iterative Clustering algorithm (SLIC) forsegmentation purposes. This method reduces the work furtheras the user only needs to specify the number of super-pixels.By applying the Random forest classifier, an accuracy of 90%can be obtained with the use of 400 n_trees. Since it takes themean of all decision trees, it tends to show better accuracy.With the help of these results, rescue teams can prioritize theirservices to areas based on the level of damage.

The authors of [130] use land cover information to evaluatepost-disaster areas. This information has been widely used inremote-sensing technology [131]. SVM classifier is used toclassify images into positive and negative recovery. Positiverecovery refers to the areas which have recovered to theiroriginal state. SVM achieved 90.8% accuracy in this case.The CNN model, which is also competitive, is not used herebecause of the limited amount of training data (CNN modelsrequire a large amount of training data).

B. Extracting Useful Information from Social Media Sites

Social media has become a powerful tool in today’s world.Data generated on social media sites every second is huge,and this data can be employed to derive useful results.Incorporating this data can help post-disaster managementteams to make better and more informed decisions. Studies

17

have used Twitter data to concentrate on the role ofdevelopment and exchange of knowledge on social networkingsites to increase consciousness about the situation as wellas to evaluate post-conditions. Technology is currentlyconcentrating on using automatic approaches to obtainmeaningful data during disaster times from Twitter data.The authors of [132] have differentiated data into initialand secondary from previous sources, where individualsnot only shared information about their own situation butalso the accessible and relevant knowledge that had beenre-tweeted. The authors of [133] observed that the tweetscontain contextual news that creates awareness of the disaster.The authors base their conclusions on a study of the 2009wildfires in Texas.

The authors of [134] use ML to evaluate people affectedby a disaster, their information, and their comparison withthose who were not affected. All disaster-related tweets areidentified from the Twitter dataset [Comment, Date, Location]using Dynamic Query Expansion (DQE). The location fromwhere a user tweeted more was estimated as his/her homelocation. To analyze the trending topics, Dirichlet regressionwas used. The topics were classified according to factors suchas age, income, education, etc., and a correlation between themwas observed. A similar study was conducted for people whowere not affected to study the impacts on them. However, itis a bit difficult to extract reliable patterns from such a largedataset.

Extracting useful information from tweets is also a helpfultask. After the pre-processing of data by removing all theextras and representing it in a vector form using TermFrequency Inverse Document Frequency (TF-IDF) and Bagof Words representation, classification was performed in [28].SVM (73.66%) and logistic regression (74.58%) achieved thehighest accuracy among all the classifiers. The voting classifierwas also proposed, which showed a competitive accuracy(74.16%) However, a mechanism needs to be added that canverify tweets and cross-validate the location of the disasterso that there is no misguidance later for the rescuers. Theauthors of [29] propose a different approach to classify tweetsby checking whether they can be used as disaster-relatedinformation or not. Natural Language Processing (NLP) isused to clean data collected from Twitter. Count and TF-IDFvectorizer was used to convert tweets to a vector. Then, thedata was fed to several classifiers. Logistic regression (99.7%)and naive Bayes (98.8%) outperformed all other classifiers.This model can also be expanded to classify tweets on thebasis of the nature of users.

Another area of research can be predicting returning patternsfor the people who left their homes because of a disaster.Returning patterns need to be estimated so that furtherplanning for resource allotment can be done. The governmentand disaster management authorities need to be ready to facea change in demand for vital public services such as electricpower, gas, and water. The authors of [41] use sentimentssuch as the psychological state of users, the magnitude ofthe tweeted tragedy, and success in the recovery area tounderstand returning patterns. The assumption in this modelis that sentiment, a manifestation of different dimensions and

restoration levels, could be exploited to predict evacuee returnbehavior. Each tweet w is allotted a sentiment score S(w),by connecting each word in the tweet to a dictionary andallocating a weight highlighting its emotional effect. Then,the time at which the evacuees left, the distance, and thedestination they have traveled to are calculated. In additionto these parameters, it also extracts parameters from threeother networks: (1) offline spatial network, (2) online personalnetwork, and (3) online agency network. The Gradient Boostclassifier is used to predict the returning behavior, whichis defined as 0 for a negative return and 1 for a positivereturn. It outperformed all other classifiers with an accuracy of86.4%. This is because gradient boosts can effectively modelnon-linear dependence between the parameters in comparisonto SVM and logistic regression. However, to use this modelon other disasters, certain parameters need to be changed.

C. Minimizing Future Disaster RiskFuture disaster risk minimization is a crucial task that needs

to be performed. A GIS-based emergency response databaseserver is used to help the emergency rescue authorities getprepared for life-saving operations in a small amount of timeby predicting extreme climatic conditions [135]. This databaseserver provides the climatic information to the relevantauthorities via a satellite link. The satellite has two functions,first being remote sensing and weather forecasting, and secondis to transmit the extreme input climatic parameters receivedfrom GIS database servers to rescue and medical authoritiesto prepare them for life-saving procedures. However, theaforementioned technology is not well integrated with moderntechniques of artificial intelligence. The forwarded informationis analyzed by humans, a tedious and error-prone process thatcan be replaced by more accurate machines.

Drought prediction systems using Deep Belief Network(DBN), which can calculate and predict different droughtindexes such as the Standardized Precipitation Index (SPI)with much higher precision is proposed in [136]. Similarly,machine learning techniques such as pattern recognitionneural networks, recurrent neural networks, and randomforest models can be used to determine relationshipsbetween calculated seismic parameters and future earthquakeoccurrences. Moreover, patch-wise object detection techniquesbased on CNNs can be utilized for automatic disaster detectionsuch as floods and landslides [137]. Image classificationbased on the deep learning algorithms has been found tohave much higher accuracy for landslide recognition thanpreviously used methods. In addition, by integrating suchprocesses, it is possible to respond to the evolving disastersituations [49]. The accuracy of landslide detection by CNNis calculated with a precision of 0.93, recall of 0.94, andF-measure of 0.93. While predicting the disaster, the majorconcern is to save human lives. This objective can be furtheraccomplished by involving technologies such as IoT in theprocessing. This could turn out to be a quick and alternativemeans of communication in the disaster-struck region, whereIoT-enabled devices (battery-powered wireless devices) canbe used to provide data network resilience during disastersituations [138].

18

Table IV: Summary of works for post disaster and to minimize future disaster risk (Acronyms used in the table- RS: RemoteSensing).

Category Reference Target issue Technology used Hardware/APIused

MLinvolved Case studies

RS and UAVs [127], 2018Collect

information aboutsurvivors

Ant foraging GIS analysis No 800*800 area

[126],2011

Assessing area forthreats

Naive Bayes andSVM None Yes Not mentioned

ML and UAVs[48], 2017 Highlighting areas

with risk

CNN and MarkovRandom field to

process datacollected by UAVs

None Yes Earhquake, flood andlandslides

RS[129],2017

Detect changespost-disaster

Enhanced changedetection index

(ECDI)

Openstreet mapand satellite

imagesNo Van,Turkey

(Earthquake)[35], 2018 Detecting damages

to buildings Random forest None Yes Haiti (Earthquake)

[49], 2018Evaluating future

disaster risk CNN None Yes Floods and Landslides

[34], 2017 Detect changespost-disaster

Simple LinearIterative Clusteringalgorithm (SLIC)

and Random forest

Aerial images byGeoEye1 Yes Japan (Earthquake and

Tsunami)ML

[130], 2019Classify areas into

positive andnegative recovery

SVM Land cover data Yes Tacloban,Phillipines(Typhoons)

[52], 2019Detect which areaswere damaged andaffected by flood

K-means clustering Optical ASTERimages Yes Tohoku (Tsunami)

[28], 2018 Extracting usefulinformation

Logisticregression, SVM

and Votingclassifier

Yes Chennai (Rainfall)

[41], 2019Predict returningpattern for people

who leftGradient Boosting

Twitter data

Yes New Jersey (HurricaneSandy)

Social media

[29], 2019Identifying whichtweets could beused for useful

information

Logisticregressionand Naive bayes

Yes Hurricane Florence andHurricane Michael

[134], 2019Relation of tweetsbetween disasteraffected and notaffected people

Dirichletregression andDynamic QueryExpansion (DQE)

Yes New York City(Hurricanes)

Summary: This section covered different approaches fordetecting changes post-disaster. In today’s world, social mediagenerates huge amounts of data every second and it can beextremely helpful if relevant information can be extracted fromit. This section presented various models to make social mediauseful in disaster management. This section further reviewedML algorithms for minimizing future disaster risks. Table IVsummarizes all the works for post-disaster and to minimizethe future disaster risk.

VI. APPLICATIONS OF MACHINE LEARNING MODELS INPANDEMIC MANAGEMENT

A. Prediction and diagnosis of a pandemic

To prevent the outbreak of a pandemic, an early diagnosisis a must. Technologies for early diagnosis of ISPA (AcuteRespiratory Infections) disease are required to reduce its

adverse effects on babies and prevent it from turning fatal.The authors of [18] carried out a test to foretell anddiagnose, whether the person is affected with ISPA or not,which is done by evaluating the evidence acquired withthe usage of the ML algorithms such as SVM, neuralnetworks, KNN, and naive Bayes. Various symptoms, suchas fever, headache, flu, etc., were used by the classifiers.The confusion matrix was used to calculate the accuracy,and SVM outperformed all the algorithms. The authors of[19] predict antigenic variants of H1N1 influenza structuredon a stacking model. In the design of a stacking modelbased on epidemics and pandemics, three separate featureengineering methods were implemented to check its viabilityand universal adherence, namely, residue-based methods, tenregional band-based methods, and five epitope region-basedmethods. This stacking model employs a concept similarto k-fold cross-validation to build out-of-sample predictions

19

that function for small to medium-size datasets. Accordingly,logistic regression, naive Bayes, and neural network methodswere used to design the Level 2 models to get better resultsthan single models. Also, random forest and gradient boostingwere applied to create the Level 2 versions.

Such Level 2 models made projections for the outcomeof all results, which were then put into the second stageof the training data, identified as Y (n ∗ n). Parameter naccounts for the number of new functions in the dataset. TheLevel 3 model was developed by the classifier of logisticregression and evaluated on Y (n∗n) to generate the resultingeffects of antigenic variant predictions. Because of the rapidantigenic shift or drift that cause antigenic variants to occur,the effects of specific mutations on antigenicity are not clearlyknown. Outputs based on various function vectors showed nodifference, suggesting that producing new characteristics fromresidual sites was possibly not the key reason for affectingprediction. For the proper modelling of the H1N1 variationmechanism, several more factors such as climate and thehuman immune system, need to be considered, which is outof focus of this model.

To solve this feature issue, authors of [55] use rainfall as avariable to predict the spread of cholera in Haiti. The rainfallpattern is treated as a marked Poisson process. The continuousrainfall events are processed as discrete events where the depthof rainfall is presented by the mark. K-means clustering wasfurther used to predict the spread of disease. The limitationof this model is the lack of ability to accurately predictthe evolution of new cases. The improvement in the currentsituation, i.e., real-time data is also not taken into account.

Pioneering work in using Bayesian networks and NLPfor influenza detection is presented in [139]. Data wascollected from free-text emergency department (ED) medicalreports comprising of 468 reports on PCR (Polymerase ChainReaction) positive influenza patients from the period January1, 2008 to August 31, 2010 and 29, 004 reports of patientsnot associated with a positive PCR test from the period July1, 2010, to August 31, 2010. The classification architecturecomprised of the NLP-finding- extraction component withTopaz used as NLP parser to encode findings into oneof three values: acute, non-acute, and missing, followedby a classifier network. The authors studied seven MLclassifiers and compared their diagnostic capabilities againstan expert-built Bayesian classifier. Three configurations wereused based on methods to deal with missing data forperformance evaluation. The authors used AUROC and BrierSkill Score as a measure of classification performance whileNLP-finding-extraction performance was measured usingaccuracy, recall, and precision.

The authors observed a tie between naive Bayes, logisticregression, SVM, and ANN with the highest AUC being 0.93in the configuration where all missing values were assigned‘False’. The NB showed superiority (BSS: 0.35, AUC: 0.93)in terms of least training time and the ability to treatmissing values without preprocessing. The authors were ableto show that ML classifiers performed better than Expert-madeBayesian classifiers, for a given NLP extraction system and alarge amount of training data. Similar work was reported by

authors of [140] on the same dataset with nine combinationsof finding-extraction methods and Bayesian classifiers. Thehighest AUROC achieved by them was 0.79 for a combinationof expert-finding with a BN-EM-topaz classifier. Moreover,they only considered AUC for classification evaluation whichis susceptible to bias due to inclination of the test data with oneclass, potentially leading to an unskilled classifier problem.The authors of [140] only considered Bayesian networks asclassifiers without a comparative study between other machinelearning models.

The authors of [141] have proposed a method of genomesequencing to identify SARS-CoV-2 (COVID-19). The methodis based on the additional sequencing of viral complementaryDNA. Sequencing data can be retrieved from the PCR of theoriginal viral RNA which then can be used along with cDNAto identify COVID-19. Alignment methods like FASTA andBLAST are used for classification done from viral sequencingtechniques. However, the major issue with this method isthat it necessarily requires base sequences for the detectionor classification. A detection test has been proposed thatcombines molecular testing with deep learning. This workapplies state of the art deep convolutional neural networkswhich can automatically create features starting with genomesequencing of the coronavirus. Samples are divided into 9:1for training and testing using a 10-fold cross-validation scheme[142]. The data used was genome sequence data from therepository 2019 Novel Coronavirus Resource [143] and NCBIrepository [144].

The results of the experiments show that the aboveapproach correctly classifies SARS-CoV-2 and distinguishesit from other coronaviruses. The proposed method identifiesCOVID-19 with an accuracy of 98%, and can classifydifferent coronaviruses with an accuracy of 98.75% [142].There are several prediction models proposed for COVID-19detection based on CT scan images of the chest using deeplearning which have accuracy less than the above method(around 82%), and chest radiography may miss patients withpneumonia in early phase [141], [145], [146]. However, theabove model has been trained on very few samples due tolimited data available. Also, limited genome sequences areavailable and considered in this model. The authors of [147]also propose a COVID-19 prediction model, for Brazil, usinga time-sliding window algorithm.

The authors of [148] integrated current surveillance datawith Internet search queries to forecast influenza epidemics.Data were obtained from the China National Scientific DataCenter for Public Health along with a public Baidu searchengine database based on Baidu Index for queries related toinfluenza. Regression was performed using Support VectorMachines, and hyperparameters were tuned by leave-one-outcross-validation. Performance metrics used for this task wereRoot Mean Square Error (RMSE), Mean Absolute PercentageError (MAE), and Root Mean Square Percentage Error(RMSPE). Correlation analysis was performed on data withdifferent lag times and search queries with a significantcorrelation coefficient. The authors of [149] mention thefeatures that were chosen for the SVM regression model.Application of internet search query data for infectious

20

diseases provides real-time surveillance of epidemics andovercomes the shortcomings posed by lag-time in conventionalflu surveillance. A strong correlation between search termsand influenza cases was found in the study. Given the shortincubation period of influenza, most of the search queriesrelated to symptoms and medication were closely correlatedwith cases in the same month. The authors found thatthe models based on ensembled data performed better andwere more robust than the models based on other singledata sources. Similar work was done by Google Flu Trends(GFT) [150]. However, it was concluded by authors of [148]that GFT data could not replace conventional surveillancemethods to be a reliable pandemic surveillance system. Thisis due to the negligence of epidemiological factors such asgeographical location, illness complaints, the age distributionof patients, or clinical manifestations in the Internet data. Thecorrelation coefficient of the search keywords relies heavilyon existing vocabulary data. However, due to the changes inthe social media environment, many new search vocabulariesare produced. The new vocabularies were not considered inthe model.

One of the most significant aspects of pandemicmanagement is close monitoring of the people in the affectedareas. The authors of [151] provide a comprehensive surveythat includes the use of several machine learning algorithms forsmart decision support systems for healthcare. For determiningthe re-occurrence of a pandemic, traditional models utilize anauto-regressive integrated moving average (ARIMA) model forprediction [152]. The model construction trades off betweenparameter estimation, diagnostic checking, and identification.Examples of usage of IoT and big data analytics through smartwearables to collect health-related information of citizens havealso been prevalent [15], [153]. Conventionally deployed flusurveillance systems are based on hospital and laboratorydata, thus limiting their real-time surveillance due to lag-time[154]. The IoT based data collection techniques assume mostcitizens to possess a smart wearable or sensor-enabled device.However, in developing countries, that are generally the worsthit by an epidemic, this remains far-fetched.

B. Social distancing

COVID-19, a major pandemic outbreak, has caused largescale devastation in the world. Within only nine months(January to September, 2020), this has spread in 213 countriesalready with approximately 32 million cases [155] overall.Also, this figure is increasing at a high pace everyday.The authors of [156] highlight the role of deep learningalgorithms during COVID-19 in various phases such as diseasesurveillance, risk prediction, medical diagnosis, screening, andvirus modeling and analysis. Since the spread of influenza,SARS, and then COVID-19, social distancing has become animportant tool to handle outbreaks [157], [158]. The work in[159] presents simple methods to maintain distance betweenpeople to prevent virus transfers. Social distancing is definedas living in such a way that physical contact with people isavoided. This can be divided into two categories: communityand personal. Community social distancing involves avoiding

any crowd gatherings in public, such as schools, colleges,and places of worship, minimizing traveling, and sealingthe areas where people have already been affected. Personalsocial distancing involves quarantining individuals to theirhomes, leaving one’s home only for important purposes, andmotivating others also to follow the same.

Various technologies have been used for social distancing.GPS is a technology that can assist to detect the locationsof people. The authors of [169] propose indoor wirelesspositioning, which does not require the use of GPS andcan thus work in areas where GPS cannot. Bluetooth is atechnology that works well in such cases, as it allows multipleconnections at the same time. This feature used by the modelproposed in [160] to evaluate the dependency between distanceand RSS (Radio Signal Strength) of any two devices. If thedistance detected is less than 1.4 meters, it generates a warningmessage to enable the users to be aware and maintain adistance. However, this model tends to fail in some scenariosdue to inaccuracies in the user’s location, and there is nomeasure proposed yet to address it [170]. Also, for this modelto work, Bluetooth should always be on. Thus, there are stillchallenges for this technology to work upon. The authors of[171] have proposed a novel ML algorithm integrated withBluetooth technology to calculate the risk of COVID-19 innearby areas.

For directing people about public social distancing, authorsof [162] propose a model that provides a count of peoplein an area. This model works in a complex network and iscomputationally effective, which various models using Zigbeeand WiFi lack. The model works by picking up variousclusters, and the one with the largest size is selected. Then, amaximum likelihood equation is derived from a probabilitydensity function plotted. Finally, the number of people isestimated based on the equation. In the same manner, thework in [163] provides a mathematical design with lowercomputational complexity to count the number of people. Thismodel uses a method based on detecting energy. These modelscan be combined with various ML algorithms to improve theiraccuracy.

Due to their varied applications, UAVs have proved fruitfulin many situations. The authors of [166] have used thecombination of recurrent and convolutional neural networksto process data collected from UAVs. These UAVs and theirdata are then used to predict the traffic density in an area. Thisinformation can be provided to users and can be used to signalthem about the overcrowded areas, as shown in Fig. 9(c).The authors of [173] have proposed a model for estimatingtraffic jams in an area based on blockchain and deep neuralnetworks. Also, authors of [174] have proposed a model forinter-vehicle communication using a directed acyclic graphand game theory.

A portable device is proposed that collects data via a radarsensor and a camera and provides it as input to neural networksin [164]. The ANN is trained constantly with data to identifywhether the object near it is a vehicle, a stationary object,or a human being. This model’s accuracy improves with useand gets trained when the user is near an object. This modelcan also keep a track of the users who are violating social

21

Table V: Summary of works for pandemic management (Acronyms used in table- COCO: Common Objects In Context, RSS:Radio Signal Strength).

Category Reference Target issue Technology used Hardware/APIused

MLinvolved Case studies

[18], 2020 Detecting ISPASVM, KNN, NeuralNetworks and Naive

BayesNone Yes Indonesia

ML [19], 2018

Predictingantigenic variantsof H1N1 influenza

virus

SVM, KNN, NeuralNetworks and Naive

BayesStacking model Yes WHO

[55], 2013Predicting spreadof cholera disease K-means None Yes Haiti

Android-based [160],2013

To detect distanceand signal the user Bluetooth, RSS

Wirelesspositioning and

GPSNo Not mentioned

Radio-waves [161],2013

To detect distanceand signal the user Ultrasound A Transmitter and

A detector No A hospital

ML inandroid based

app[20], 2019

Movement ofinfected people

KNN and DecisionTree Service Provider Yes Not mentioned

Sensor based [162],2017

Count people in anarea

Maximum likelihoodequation Radar sensor No Novelda, Norway

Radar-based [163],2017

Count people in anarea Posteriori algorithm None No Not mentioned

ML andRadar

Sensors[164],2020

Detect distancebetween people ina crowded area

ANN and Kalmanfilters

Radar sensor and acamera Yes Not mentioned

Sensor based [165],2012

Detect humanmovements indoor

Improved headingestimation model

Smartphonesensors No iPhone 4s

ML and IoT [43], 2014Detect user’s

locationANN and Radial

functionsFT-6200 kit, Zig

bee Yes Third floor, ChungHua university

ML and UAV [166],2020

Predict traffic level

Recurrent Neuralnetworks andConvolutional

networks

UAV Yes Not mentioned

[27], 2016Detect user’s

location

Deep BeliefNetworks(DBN) and

Naive Bayes’None Yes

Gowalla networkand Brightkite

network

[23], 2019Verify user’s

location ANN and SVM Wireless networks Yes Momentumproject, GermanyML

[38], 2020Random Forest andGradient Boosting

Microphone andThermal Cameras Yes Microsoft COCO

dataset[167],2020

Predict number ofinfected people DNN and Gaussian

process None Yes Indoor shoppingmall

[47], 2018 RNN and DNN Simpy simulator Yes Big data challenge,Italia

[168],2019

LSTM NeuralNetwork None Yes Geolife project

ML and5G cellularnetworks [33], 2018

Predict number ofpeople in an area Bayesian regressor,

Random Forestregressor and Gaussian

regressor

None Yes San Francisco

[46], 2020 DNN and CNN GPRS Yes

A mobile networkdata, Youtube,Snapchat and

Facebook

22

Figure 9: Application of ML in social distancing - (a) Infected people movement prediction (b) Quarantined people locationprediction (c) Road traffic prediction (d) Sickness trend prediction [172].

distancing. However, different from [164], the authors of [27]propose another approach for detecting a users’ locations thatis based on their most influential friends in a social network.Initially, a deep belief network (DBN) was implemented bychoosing M top influential friends and evaluating their socialinfluence on a user. The model showed an approximately16.87% gap in its theoretical and practical accuracy becauseit randomly selects some top friends which does not seemto be an effective way. Then, a naive Bayes model was used,and it showed approximately 19.82% accuracy increase. NaiveBayes is comparatively more scalable, and its predictions arebased on probability which makes it more reliable.

The authors of [20] have combined ML with smartphones’locations to warn the users regarding their safety from infectedneighbours in the area. KNN and decision trees are used for

determining the location of a user, and a Markov model isused for location prediction. By training the model with thereal-time dataset of the area, the movement of infected peoplecan be predicted. People can get alert notifications on theirsmartphones, as shown in Fig. 9(a).

ANN is combined with a robot-based model to increaselocation accuracy in [43]. With this integration, the modelbecomes more scalable, cost-effective, and user-friendly, andit showed an error of only 2.7m. Also, radial basis functionnetworks are used with Zigbee which has an advantage overWi-Fi systems. Wireless networks have low accuracy whileverifying a user’s location and require channel characteristicsdata to be available. The authors of [23] have proposed a modelusing SVM and ANN which outperformed the traditionalwireless networks. This model has a promising accuracy,

23

and training the models is the only additional task requiredhere. The authors of [167] present a better model namedthe FedLoc (Federated Localization) framework. This modelis built for verifying the user’s location, as shown in Fig.9(b), without sacrificing their privacy, by using deep neuralnetworks. This was compared with a model based on theGaussian process, where DNN faired better because of theirability to generalize better. FedLoc is based on the integrationof the DNN and Gaussian process models. To protect the user’sprivacy, homomorphic encryption was used. However, to makethe model more effective, the weights could be modified to alower precision level from 64 bits. Furthermore, the authorsof [175] used a deep-learning based framework for monitoringof the social distance protocols during COVID-19 pandemic.The model uses YOLOv3 object recognition paradigm forrecognizing humans in a video footage. The use of transferlearning methodology ensures a higher accuracy for the model.

5G networks have attracted wide attention because of theirlow latency, ability to satisfy user’s potential needs, andrequiring less energy to operate [176], [177]. The authorsof [178] present a detailed idea of how linear regressioncan be used in 5G networks. The authors of [33] presenta Bayesian ridge regression and random forest models topredict the number of people in an area in combinationwith 5G cellular networks. They also aim at using ANN toincrease accuracy. Further, authors of [47] propose to predictthe number of people in an area using DNN and RecurrentNeural Networks (RNN). RNN outperformed DNN as RNNshowed 90% accuracy, whereas DNNs showed 80% accuracy.However, if more training data was available, then the accuracyof the DNN model could be improved further. The authorsof [33] have used ML to predict the number of people inthe user’s proximity. They began with clustering the basestations. Then, the prediction of the number of users is doneby deploying the Bayesian model regressor, random forestregressor, and the Gaussian regressor. In order to predict thenumber of users who are in proximity to a defined cluster, wecan use WiFi-hotspots. This architecture can also be deployedto predict the traffic density in an area so that the users cantravel through areas with fewer vehicles so as to practice socialdistancing.

The authors of [168] present the use of LSTM neuralnetworks integrated with 5G technology to determine theusers’ path based on these historical paths. However, thismodel is very computation-intensive and requires a lot oftraining data for a single user, which is not practically feasible.Thus, a multi-user sequence-to-sequence model was proposedwhich is more user friendly and tackles all the former issues.This sequence-to-sequence LSTM model outperformed SVMand linear regression. This study has open directions tocombine its model with other information about the users. Theauthors of [46] present another model using a mixture of DNNand CNN to build its several layers. A modified loss functionis used to train this model. This is a cost-effective system andalso tackles SLA (Service-Level Agreement) violations.

The authors of [38] have developed a model using ML towarn users in public places by informing them how manypeople near them are infected. This model was built on the

basis of influenza symptoms as its detecting features. Witha microphone array, it analyzes different cough sounds ofusers, and cameras are used to analyze the density of thatarea, as shown in Fig. 9(d). The random forest model showedthe highest accuracy (95%). This model’s accuracy can befurther improved if its training dataset is combined with somereal-time data. This model is highly beneficial in informingpeople about the status of a public place during any phase ofa pandemic.

C. Extent to Which Economy is Affected

In [179], the authors have tried to assess the economicimpact of earthquakes through a hybrid model (SouthernCalifornia Planning Model, SCPM) which follows theGarin-Lowry model and is an integration of state-of-artInput-Output model and spatial allocation model. Thework also encompasses the network model and structureperformance model. The idea behind this model is to estimateeconomic losses by looking at the direct business and structureloss, lifeline network disruption, and spatial impacts thatoccurred due to the disaster. The key technology used fordamage estimation is GIS and drives the state-of-art EPEDAT(Early Post-Earthquake Damage Assessment Tool) model[179]. Although these models estimate the cost of a disasterand its immediate impact on the economy, they do not explainthe important long-term impact of disasters like earthquakeson the economy and growth.

Another method has been proposed which caters to theissues in [179] and estimates the sign and magnitude of bothshort and long-run effects of disaster on growth [180]. Thestudy pursues a comparative event study approach, whichcompares the situation of affected countries post-disaster withthe superficial situation of the country if the disaster would nothave happened and estimates the loss. The superficial state ofthe country is generated by comparison from the built syntheticcontrol group. The data used is the data on natural disastersfrom the EM-DAT (Emergency Events Database), data fromthe Penn World Tables, and real GDP (Gross DomesticProduct) per-capita at PPP (Purchasing Power Parity). Theabove work’s results show that only very large disasters displayan impact on GDP growth in both long and short runs.Also, the results show that these impacts on the economyare driven by the event of radical political changes in thecountry post-earthquake, and the impacts are not significantin the absence of political changes. Results are informativeabout the long-run cost of disasters and the involvement ofpolitical structure in causing damage to the economy and canbe used in further studies. This model works only when a setof regions is affected, and another set is totally unaffected bythe disaster (i.e. control group), and hence cannot be used incase of a global crisis. An approach using naive Bayes can beused to calculate such an effect [181].

D. Classifying Areas into Red, Orange and Green ZonesAccording to Spread During a Pandemic

Classification is generally done through machine learningalgorithms. However, there are technologies like remote

24

sensing that capture sentinel-satellite images where data iscollected across spectrums. However, those data are classifiedthrough ML algorithms [182]. Therefore ML is inevitable.

Classification of areas into green, yellow, and red is basedon certain features: number of infected patients, rate ofinfectivity, rate of recovery, death rate, and number of deaths.These features are extracted from the given statistical data ofa given area and based upon these, the areas are classified. Acomparison of KNN and SVM based classification techniquesis done on textual data to find the type of disease in [183].First, the parameters are evaluated based on training data. Themodel is built using KNN and SVM, and the performance isevaluated using test data. Then, N cross-validations accuracy,model build time, search time, and memory used are comparedbetween the two. Different sample sizes were chosen astraining data, and results were compared. The results showedthat KNN shows decreased accuracy when the size of the dataincreased, whereas SVM works better with a large amount ofdata. In all the varying sample set data, KNN consumed morebuild time and search time than SVM. The difference wassignificantly large as the sample size increased (from 50 to1000). The memory used for both the algorithms was almostconstant. Thus, if the sample size is small, KNN is preferable,and if the sample size is large, SVM is preferable. In our case,if we have to classify zones on a country basis, we would gofor KNN else if it is based on states or cities, we go for SVM.However, the above algorithm comparison was done for textualdata for medical purposes. The actual implementation of theclassification of statistical data on COVID-19 has not beendone yet.

Summary: This section reviewed the applications ofmachine learning models in pandemic management. First,it presented different machine learning models for earlyprediction and diagnosis of a pandemic which can be helpfulin the prevention of an outbreak. Social distancing becomesessential if the spread of virus transfers needs to be controlled.This section also presented how ML algorithms integratedwith various other technologies, like IoT, can be helpful inpractising social distancing. KNN and SVM can be used forappropriate classification purposes, such as to classify the areasinto red, orange, and green zones. Also, ML algorithms such asNaive Bayes can be used to determine the extent to which theeconomy is affected. Table V summarizes all the ML worksfor pandemic management.

VII. CHALLENGES, OPEN ISSUES, AND FUTURERESEARCH DIRECTIONS

The previous sections have discussed different MLalgorithms and how they can be used in various phases duringa disaster or a pandemic. However, some challenges, openissues, and research directions also exist, which are discussedas follows.

A. Challenges

1) Requirement of specific data: ML algorithms requiredata to be cleaned and refined for further processing. Theysometimes also require the extraction of certain features from

data, which is a costly process when the data is large.For example, when CNN is used to isolate the areas intocatastrophe zones, it requires specific forecast features [40].Similarly, the model in [121] requires certain other features inits training dataset to provide a better evacuation system.

2) Accuracy measurement: There exist several parametersfor a model’s accuracy, such as precision, recall, AUC,F-measure, and many more. To accurately measure a model,a non-biased parameter must be chosen according to thescenario. The authors of [140] have only used the AUCmeasure, which is susceptible to bias due to the inclinationof the test data with one class.

Further, certain parameters for many algorithms need to beset to increase accuracy. For example, the number of treesneeds to be specified for better accuracy in random forestmodels [34]. Similarly, in the fuzzy model, parameter m needsto be set appropriately.

3) User’s privacy: The privacy of users must not beviolated by any model [184], [185]. This is a factor which thedeveloper must keep in mind before developing any model.There is a need for mechanisms to protect the user’s privacywhen their data is used by ML algorithms [48]. Some of theML algorithms we discussed above also require the use oftechnologies like IoT, which can breach the privacy of it’susers, and hence, some mitigation strategies are required [186].Moreover, the authors in [187] present a survey discussing thesecurity threats and solution architectures in a supply chainusing ML and other technologies. With an increase in theuse of smart healthcare devices, user’s health can be at riskbecause of the system’s vulnerabilities [188]. Mobile-healthsystem architecture also poses threat to user’s data, and anencryption algorithm is needed to protect such vulnerabilities[189].

B. Open Issues1) Neural networks are not deterministic: Neural networks

are stochastic in nature and not deterministic. A neural networkdoes not understand that smoke density cannot be negative,or that the amount of rainfall in an area cannot be negative.Some constraints to evaluate the inputs need to be added tothese algorithms.

2) Requirement of a large amount of data: The authors of[145] have a limited amount of data to detect COVID-19. TheCNN model requires a large number of samples for training,which are unavailable. Due to a lack of availability of largetraining data, it is often not possible to use the CNN model[131], [47]. The authors of [44] need a ML algorithm thatrequires less amount of data.

3) Inaccurate data: The data fed to AI systems can beinaccurate in some scenarios. Thus, the results produced willalso be inaccurate. There is no mechanism to crosscheck suchlarge amounts of data. Also, since the results are based onprevious trend data, sometimes accurate predictions cannotbe made. For instance, in case of an earthquake, consider aparticular area that always experiences mild earthquakes. Thisarea experiences a strong earthquake in a certain year. Basedon the previous trends, ML algorithms will predict the resultsinaccurately.

25

C. Future Research directionsTo predict a pandemic, ML algorithms, apart from the

symptoms of the disease, need to encompass certain otherfeatures such as climate and human immune systems. Themodel mentioned in [19] lacks both features. However, therainfall pattern as a feature is used to predict the spreadof cholera disease, but the human immune system is stillnot considered in this model [55]. Also, in object-orientedimage analysis (OOA) integrated with random forests, thesegmentation process needs to focus on more details so thatrandom forest’s model accuracy can be increased. Currently,the model is only focused on spectral details. Other featuresmay be included to analyze the areas [37].

1. The pre-processing time for many models must bereduced to save time for their deployment in real scenarios.For example, in the model in [125], the pre-processing timecomprises 60-80% of the total time.

2. Models need to be adaptable to real-world changes inorder to be successful in real disaster scenarios. There is nosuch mechanism for this when the crowd density of the areais increased [56].

3. Accuracy can be increased if high-resolution images areused in ML models [52]. Low-clarity and cloudy images canlead to the under-fitting of the model.

4. Some models which have proven to work well withvirtual laboratory datasets need to be tested on real-time data,for example, the model presented in [154]. Also, we expectthat many existing models that use different technologies canbe combined with ML algorithms to increase their accuracy.Some examples of such models are [43], [23].

VIII. CONCLUSION

In the past few years, natural disasters and pandemics havebecome much more damaging and frequent. This surge in thenumber of disasters and pandemics has caused a strain onthe emergency services, and this is where ML algorithms arerequired to work efficiently and make the best use of existingresources. This paper presented a comprehensive survey of theapplications of ML in disaster and pandemic management.The paper first presented a detailed explanation of machinelearning algorithms. Then, we discussed various phases ofdisasters and pandemics where ML algorithms can be used.The phases are predicting a disaster, detecting a pandemic,delivering early signals, determining crowd evacuation routes,minimizing future disaster risk, social distancing, and othermiscellaneous issues. We have also discussed key issues inseveral technologies and how ML algorithms can overcomethese issues. Finally, we sketched various challenges, openissues, and future research directions.

REFERENCES

•[1] Gianluca Rampolla del Tindaro, “Resident/Humanitarian coordinatorreport on the use of CERF funds, Papua new Guinea rapidresponse earthquake, 2018,” https://cerf.un.org/sites/default/files/resources/18-RR-PNG-29464-NR01_Papua%20New%20Guinea_RCHC.Report.pdf, online; accessed 12 September 2020.

[2] J. H. Samuel Lalmuanawma and L. Chhakchhuak, “Applications ofmachine learning and artificial intelligence for covid-19 (sars-cov-2)pandemic: A review,” Chaos, Solitons Fractals, vol. 139, p. 110059,2020.

[3] Y. Jahir, M. Atiquzzaman, H. H. Refai, A. Paranjothi, and P. G.LoPresti, “Routing protocols and architecture for disaster area network:A survey,” Ad Hoc Networks, vol. 82, pp. 1–14, 2019. [Online].Available: https://doi.org/10.1016/j.adhoc.2018.08.005

[4] X. Du, Y. Xiao, M. Guizani, and H.-H. Chen, “An effectivekey management scheme for heterogeneous sensor networks,” AdHoc Networks, vol. 5, no. 1, pp. 24 – 34, 2007, securityIssues in Sensor and Ad Hoc Networks. [Online]. Available:http://www.sciencedirect.com/science/article/pii/S1570870506000412

[5] A. Mehra, M. Mandal, P. Narang, and V. Chamola, “Reviewnet: Afast and resource optimized network for enabling safe autonomousdriving in hazy weather conditions,” IEEE Transactions on IntelligentTransportation Systems, pp. 1–11, 2020.

[6] T. Yaqoob, H. Abbas, and M. Atiquzzaman, “Security vulnerabilities,attacks, countermeasures, and regulations of networked medicaldevices - A review,” IEEE Commun. Surv. Tutorials, vol. 21, no. 4,pp. 3723–3768, 2019. [Online]. Available: https://doi.org/10.1109/COMST.2019.2914094

[7] L. He, Z. Yan, and M. Atiquzzaman, “LTE/LTE-A network securitydata collection and analysis for security measurement: A survey,”IEEE Access, vol. 6, pp. 4220–4242, 2018. [Online]. Available:https://doi.org/10.1109/ACCESS.2018.2792534

[8] M. Mohammadi, A. Al-Fuqaha, S. Sorour, and M. Guizani, “Deeplearning for iot big data and streaming analytics: A survey,” IEEECommunications Surveys Tutorials, vol. 20, no. 4, pp. 2923–2960,2018.

[9] G. Bansal, V. Chamola, P. Narang, S. Kumar, and S. Raman,“Deep3dscan: Deep residual network and morphological descriptorbased framework for lung cancer classification and 3d segmentation,”IET Image Processing, vol. 14, no. 7, pp. 1240–1247, 2020.

[10] Wikipedia contributors, “Machine learning — Wikipedia, thefree encyclopedia,” 2020, [Online; accessed 30-April-2020].[Online]. Available: https://en.wikipedia.org/w/index.php?title=Machine_learning&oldid=954000510

[11] R. Xie, I. Khalil, S. Badsha, and M. Atiquzzaman, “Collaborativeextreme learning machine with a confidence interval for P2P learningin healthcare,” Comput. Networks, vol. 149, pp. 127–143, 2019.[Online]. Available: https://doi.org/10.1016/j.comnet.2018.11.002

[12] M. Sughasiny and J. Rajeshwari, “Application of machine learningtechniques, big data analytics in health care sector – a literature survey,”in 2018 2nd International Conference on I-SMAC (IoT in Social,Mobile, Analytics and Cloud) (I-SMAC)I-SMAC (IoT in Social, Mobile,Analytics and Cloud) (I-SMAC), 2018 2nd International Conference on,2018, pp. 741–749.

[13] J. Shirahatti, R. Patil, and P. Akulwar, “A survey paper on plantdisease identification using machine learning approach,” in 2018 3rdInternational Conference on Communication and Electronics Systems(ICCES), 2018, pp. 1171–1174.

[14] T. T. Nguyen, “Artificial intelligence in the battle against coronavirus(covid-19): a survey and future research directions,” Preprint, DOI,vol. 10, 2020.

[15] C. Chakraborty, A. Banerjee, L. Garg, and J. J. P. C. Rodrigues,“Internet of medical things for smart healthcare: Covid-19 pandemic,”Springer, Singapore, Series Studies in Big Data, 2020.

[16] A. Soofi and A. Awan, “Classification techniques in machine learning:Applications and issues,” Journal of Basic & Applied Sciences, vol. 13,pp. 459–465, 08 2017.

[17] Mehmed Kantardzic, “Data mining: Concepts, models, methods, andalgorithms,” 2011. [Online]. Available: https://www.academia.edu/18807133/Data_Mining_Concepts_Models_Methods_and_Algorithms

[18] N. L. W. S. R. Ginantra, I. G. A. D. Indradewi, and E. Hartono,“Machine learning approach for acute respiratory infections (ISPA)prediction: Case study indonesia,” Journal of Physics: ConferenceSeries, vol. 1469, p. 012044, feb 2020. [Online]. Available:https://doi.org/10.1088%2F1742-6596%2F1469%2F1%2F012044

[19] R. Yin, V. H. Tran, X. Zhou, J. Zheng, and C. K. Kwoh, “Predictingantigenic variants of h1n1 influenza virus based on epidemics andpandemics using a stacking model,” PLoS ONE, vol. 13, 2018.

[20] S.-B. Cho, “Exploiting machine learning techniques for locationrecognition and prediction with smartphone logs,” Neurocomputing,vol. 176, 05 2015.

[21] A. Namoun, A. Mir, A. Alkhodre, A. Tufail, A. Alrehaili, M. Farquad,M. Alwaqdani, T. Alghamdi, and M. Benaida, “A multi-agentarchitecture for evacuating pilgrims in panic and emergency situations:The hajj scenario,” Journal of Theoretical and Applied InformationTechnology, vol. 96, p. 20, 10 2018.

26

[22] J. Ren and X. Gao, “Situation assessment model for uavdisaster relief in the city,” in 2011 International Workshop onMulti-Platform/Multi-Sensor Remote Sensing and Mapping, 2011, pp.1–6.

[23] A. Brighente, F. Formaggio, G. M. Di Nunzio, and S. Tomasin,“Machine learning for in-region location verification in wirelessnetworks,” IEEE Journal on Selected Areas in Communications,vol. 37, no. 11, pp. 2490–2502, 2019.

[24] J. Akshya and P. L. K. Priyadarsini, “A hybrid machine learningapproach for classifying aerial images of flood-hit areas,” in 2019International Conference on Computational Intelligence in DataScience (ICCIDS), 2019, pp. 1–5.

[25] K. Mori, T. Wada, and K. Ohtsuki, “A new disaster recognitionalgorithm based on svm for eress: Buffering and bagging-svm,” in2016 45th International Conference on Parallel Processing Workshops(ICPPW), 2016, pp. 22–30.

[26] D.-N. Le, V. S. Parvathy, D. Gupta, A. Khanna, J. J. P. C. Rodrigues,and K. Shankar, “Iot enabled depthwise separable convolution neuralnetwork with deep support vector machine for covid-19 diagnosisand classification,” International Journal of Machine Learning andCybernetics, Springer.

[27] Y. Jia, Y. Wang, X. Jin, and X. Cheng, “Location prediction:A temporal-spatial bayesian model,” ACM Trans. Intell. Syst.Technol., vol. 7, no. 3, Feb. 2016. [Online]. Available: https://doi.org/10.1145/2816824

[28] S. Sadhukhan, S. Banerjee, P. Das, and A. K. Sangaiah, “Producingbetter disaster management plan in post-disaster situation using socialmedia mining,” in Computational Intelligence for Multimedia Big Dataon the Cloud with Engineering Applications. Elsevier, 2018, pp.171–183.

[29] N. Assery, Y. Xiaohong, S. Almalki, R. Kaushik, and Q. Xiuli,“Comparing learning-based methods for identifying disaster-relatedtweets,” in 2019 18th IEEE International Conference On MachineLearning And Applications (ICMLA), 2019, pp. 1829–1836.

[30] P. A. Gutiérrez, M. Pérez-Ortiz, J. Sánchez-Monedero,F. Fernández-Navarro, and C. Hervás-Martínez, “Ordinal regressionmethods: Survey and experimental study,” IEEE Transactions onKnowledge and Data Engineering, vol. 28, no. 1, pp. 127–146, 2016.

[31] V.-M. Le, “Machine learning methods for optimization in multi-agentdecision support system : application to sign placement for tsunamievacuation,” Ph.D. dissertation, 12 2016.

[32] H. A. Shaiba, N. S. Alaashoub, and A. A. Alzahrani, “Applyingmachine learning methods for predicting sand storms,” in 20181st International Conference on Computer Applications InformationSecurity (ICCAIS), 2018, pp. 1–5.

[33] M. Polese, R. Jana, V. Kounev, K. Zhang, S. Deb, and M. Zorzi,“Machine learning at the edge: A data-driven architecture withapplications to 5g cellular networks,” IEEE Transactions on MobileComputing, p. 1–1, 2020. [Online]. Available: http://dx.doi.org/10.1109/TMC.2020.2999852

[34] A. R. Joshi, I. Tarte, S. Suresh, and S. G. Koolagudi, “Damageidentification and assessment using image processing on post-disastersatellite imagery,” in 2017 IEEE Global Humanitarian TechnologyConference (GHTC), 2017, pp. 1–7.

[35] A. Cooner, Y. Shao, and J. Campbell, “Detection of urban damageusing remote sensing and machine learning algorithms: Revisiting the2010 haiti earthquake,” Remote Sensing, vol. 8, p. 868, 10 2016.

[36] V. F. Rodriguez-Galiano, B. Ghimire, J. Rogan, M. Chica-Olmo,and J. P. Rigol-Sanchez, “An assessment of the effectiveness of arandom forest classifier for land-cover classification,” ISPRS Journalof Photogrammetry and Remote Sensing, vol. 67, pp. 93–104, 2012.

[37] A. Stumpf and N. Kerle, “Object-oriented mapping of landslidesusing random forests,” Remote Sensing of Environment, vol.115, no. 10, pp. 2564 – 2577, 2011. [Online]. Available:http://www.sciencedirect.com/science/article/pii/S0034425711001969

[38] F. Al Hossain, A. Lover, G. Corey, N. Reich, and T. Rahman, “Flusense:A contactless syndromic surveillance platform for influenza-like illnessin hospital waiting areas,” Proceedings of the ACM on Interactive,Mobile, Wearable and Ubiquitous Technologies, vol. 4, pp. 1–28, 032020.

[39] T. Gupta, V. Nunavath, and S. Roy, “Crowdvas-net: A deep-cnn basedframework to detect abnormal crowd-motion behavior in videos forpredicting crowd disaster,” in 2019 IEEE International Conference onSystems, Man and Cybernetics (SMC), 2019, pp. 2877–2882.

[40] S. N. K. B. Amit and Y. Aoki, “Disaster detection from aerialimagery with convolutional neural network,” in 2017 International

Electronics Symposium on Knowledge Creation and IntelligentComputing (IES-KCIC), 2017, pp. 239–245.

[41] T. Yabe and S. Ukkusuri, “Integrating information from heterogeneousnetworks on social media to predict post-disaster returning behavior,”Journal of Computational Science, vol. 32, 02 2019.

[42] A. Buranasing and A. Prayote, “Storm intensity estimation usingsymbolic aggregate approximation and artificial neural network,” in2014 International Computer Science and Engineering Conference(ICSEC), 2014, pp. 234–237.

[43] L. Luoh, “Zigbee-based intelligent indoor positioning system softcomputing,” Soft Comput., vol. 18, no. 3, p. 443–456, Mar. 2014.[Online]. Available: https://doi.org/10.1007/s00500-013-1067-x

[44] K. Shibata and H. Yamamoto, “People crowd density estimationsystem using deep learning for radio wave sensing of cellularcommunication,” in 2019 International Conference on ArtificialIntelligence in Information and Communication (ICAIIC), 2019, pp.143–148.

[45] P. Chhikara, R. Tekchandani, N. Kumar, V. Chamola, and M. Guizani,“Dcnn-ga: A deep neural net architecture for navigation of uav inindoor environment,” IEEE Internet of Things Journal, pp. 1–1, 2020.

[46] D. Bega, M. Gramaglia, M. Fiore, A. Banchs, and X. Costa-Perez,“Deepcog: Cognitive network management in sliced 5g networks withdeep learning,” in IEEE INFOCOM 2019 - IEEE Conference onComputer Communications, 2019, pp. 280–288.

[47] I. Alawe, A. Ksentini, Y. Hadjadj-Aoul, and P. Bertin, “Improvingtraffic forecasting for 5g core network scalability: A machine learningapproach,” IEEE Network, vol. 32, no. 6, pp. 42–49, 2018.

[48] F. Alidoost and H. Arefi, “Application of deep learning for emergencyresponse and disaster management,” 02 2018.

[49] ——, “Application of deep learning for emergency responseand disaster management,” in Proceedings of the AGSE EighthInternational Summer School and Conference, University of Tehran,Tehran, Iran, 02 2017, pp. 11–17.

[50] S. More, J. Singla, S. Verma, Kavita, U. Ghosh, J. J. P. C. Rodrigues,A. S. M. S. Hosen, and I. Ra, “Security assured cnn-based model forreconstruction of medical images on the internet of healthcare things,”IEEE Access, vol. 8, pp. 126 333–126 346, 2020.

[51] Rui Xu and D. Wunsch, “Survey of clustering algorithms,” IEEETransactions on Neural Networks, vol. 16, no. 3, pp. 645–678, 2005.

[52] J. Sublime and E. Kalinicheva, “Automatic post-disaster damagemapping using deep-learning techniques for change detection: Casestudy of the tohoku tsunami,” Remote Sensing, vol. 11, p. 1123, 052019.

[53] D. Amores, M. Vasardani, and E. Tanin, “Early Detection of HerdingBehaviour during Emergency Evacuations,” in 10th InternationalConference on Geographic Information Science (GIScience 2018), ser.Leibniz International Proceedings in Informatics (LIPIcs), S. Winter,A. Griffin, and M. Sester, Eds., vol. 114. Dagstuhl, Germany: SchlossDagstuhl–Leibniz-Zentrum fuer Informatik, 2018, pp. 1:1–1:15.

[54] G. Wei and Z. Sheng, “Image quality assessment for intelligentemergency application based on deep neural network,” Journalof Visual Communication and Image Representation, vol. 63, p.102581, 2019. [Online]. Available: http://www.sciencedirect.com/science/article/pii/S1047320319301968

[55] L. Righetto, E. Bertuzzo, L. Mari, E. Schild, R. Casagrandi, M. Gatto,I. Rodriguez-Iturbe, and A. Rinaldo, “Rainfall mediations in thespreading of epidemic cholera,” Advances in Water Resources, vol. 60,pp. 34–46, 10 2013.

[56] Q. Wang, H. Liu, K. Gao, and L. Zhang, “Improved multi-agentreinforcement learning for path planning-based crowd simulation,”IEEE Access, vol. 7, pp. 73 841–73 855, 2019.

[57] K. K. Singh, M. J. Nigam, K. Pal, and A. Mehrotra, “A fuzzy kohonenlocal information c-means clustering for remote sensing imagery,” IETETechnical Review, vol. 31, pp. 75–81, 01 2014.

[58] A. Jindal, A. Dua, N. Kumar, A. K. Das, A. V. Vasilakos, and J. J. P. C.Rodrigues, “Providing healthcare-as-a-service using fuzzy rule basedbig data analytics in cloud computing,” IEEE Journal of Biomedicaland Health Informatics, vol. 22, no. 5, pp. 1605–1618, 2018.

[59] M. Buchholz, “Deep reinforcement learning. introduction. deepq network (dqn) algorithm.” Mar 2019. [Online]. Available:https://medium.com/@markus.x.buchholz/deep-reinforcement-learning-introduction-deep-q-network-dqn-algorithm-fb74bf4d6862

[60] K. Tian and S. Jiang, “Reinforcement learning for safe evacuation timeof fire in hong kong-zhuhai-macau immersed tube tunnel,” SystemsScience & Control Engineering, vol. 6, pp. 45–56, 07 2018.

[61] N. Chen, T. Qiu, X. Zhou, K. Li, and M. Atiquzzaman, “Anintelligent robust networking mechanism for the internet of things,”

27

IEEE Communications Magazine, vol. 57, no. 11, pp. 91–95, 2019.[Online]. Available: https://doi.org/10.1109/MCOM.001.1900094

[62] H. N. Saha, S. Auddy, S. Pal, S. Kumar, S. Pandey, R. Singh, A. K.Singh, S. Banerjee, D. Ghosh, and S. Saha, “Disaster managementusing internet of things,” in 2017 8th Annual Industrial Automationand Electromechanical Engineering Conference (IEMECON), 2017,pp. 81–85.

[63] M. Anbarasan, B. Muthu, C. Sivaparthipan, R. Sundarasekar, S. Kadry,S. Krishnamoorthy, D. J. S. R.], and A. A. Dasel, “Detection offlood disaster system based on iot, big data and convolutional deepneural network,” Computer Communications, vol. 150, pp. 150 –157, 2020. [Online]. Available: http://www.sciencedirect.com/science/article/pii/S0140366419310357

[64] Y. S. Lonkar, A. S. Bhagat, and S. A. S. Manjur, “Smart disastermanagement and prevention using reinforcement learning in iotenvironment,” in 2019 3rd International Conference on Trends inElectronics and Informatics (ICOEI), 2019, pp. 35–38.

[65] B. MS and M. R, “Design of disaster management based on artificialneural network and logistic regression.” New York, NY, USA:Association for Computing Machinery, 2019.

[66] I. J. of Science and S. A. B. Research (IJSR), “Reviewpaper on iot based flood prediction model.” [Online]. Available:https://www.ijsr.net/get_abstract.php?paper_id=ART20174866

[67] V. Hassija, V. Chamola, V. Saxena, D. Jain, P. Goyal, and B. Sikdar, “Asurvey on iot security: Application areas, security threats, and solutionarchitectures,” IEEE Access, vol. 7, pp. 82 721–82 743, 2019.

[68] P. Newman, “Uninhabited aerial vehicles: Current and future use,”in Observing Systems for Atmospheric Composition, 01 2007, pp.106–118.

[69] L. Koh and S. Wich, “Dawn of drone ecology: Low-cost autonomousaerial vehicles for conservation,” Tropical Conservation Science, vol. 5,pp. 121–132, 06 2012.

[70] V. Hassija, V. Chamola, S. Garg, D. N. G. Krishna, G. Kaddoum, andD. N. K. Jayakody, “A blockchain-based framework for lightweightdata sharing and energy trading in v2g network,” IEEE Transactionson Vehicular Technology, vol. 69, no. 6, pp. 5799–5812, 2020.

[71] V. Hassija, V. Chamola, D. N. G. Krishna, and M. Guizani, “Adistributed framework for energy trading between uavs and chargingstations for critical applications,” IEEE Transactions on VehicularTechnology, vol. 69, no. 5, pp. 5391–5402, 2020.

[72] J. Yim, J. Bang, S. Kim, S. Jeong, and Y. Lee, “Uav planning tooptimize efficiency of image stitching in disaster monitoring usingsmart-eye platform,” in 2017 International Conference on PlatformTechnology and Service (PlatCon), 2017, pp. 1–4.

[73] S. Frigerio, L. Schenato, G. Bossi, M. Cavalli, M. Mantovani,G. Marcato, and A. Pasuto, “A web-based platform for automatic andcontinuous landslide monitoring: The rotolon (eastern italian alps) casestudy,” Computers & Geosciences, vol. 63, p. 96–105, 02 2014.

[74] J. Ueyama, H. Freitas, B. S. Faical, G. P. R. Filho, P. Fini, G. Pessin,P. H. Gomes, and L. A. Villas, “Exploiting the use of unmannedaerial vehicles to provide resilience in wireless sensor networks,” IEEECommunications Magazine, vol. 52, no. 12, pp. 81–87, 2014.

[75] O. S. Oubbati, M. Atiquzzaman, T. A. Ahanger, and A. Ibrahim,“Softwarization of UAV networks: A survey of applications and futuretrends,” IEEE Access, vol. 8, pp. 98 073–98 125, 2020. [Online].Available: https://doi.org/10.1109/ACCESS.2020.2994494

[76] G. Han, J. Jiang, C. Zhang, T. Q. Duong, M. Guizani, and G. K.Karagiannidis, “A survey on mobile anchor node assisted localizationin wireless sensor networks,” IEEE Communications Surveys Tutorials,vol. 18, no. 3, pp. 2220–2243, 2016.

[77] N. S. Tejasvi Alladi, Vinay Chamola and M. Guizani, “Applicationsof blockchain in unmanned aerial vehicles: A review,” VehicularCommunications, vol. 23, p. 100249, 2020.

[78] Y. Wang, J. D. Colby, and K. A. Mulcahy, “An efficient methodfor mapping flood extent in a coastal floodplain using landsattm and dem data,” International Journal of Remote Sensing,vol. 23, no. 18, pp. 3681–3696, 2002. [Online]. Available:https://doi.org/10.1080/01431160110114484

[79] J. S. Kumar, S. K. Pandey, M. A. Zaveri, and M. Choksi, “Geo-fencingtechnique in unmanned aerial vehicles for post disaster managementin the internet of things,” in 2019 Second International Conference onAdvanced Computational and Communication Paradigms (ICACCP),2019, pp. 1–6.

[80] S. Dey, M. N. Kamal, S. Dutta, A. Tiwari, S. Ray, M. J.Moatasimbillah, N. Saha, N. Adhikary, D. Mukherjee, S. Nayak,R. Dey, and S. Saha, “Ad-hoc networked uavs as aerial mesh networkfor disaster management application and remote sensing: An approach,”

in 2017 8th IEEE Annual Information Technology, Electronics andMobile Communication Conference (IEMCON), 2017, pp. 301–304.

[81] M. Terzi, A. Anastasiou, P. Kolios, C. Panayiotou, and T. Theocharides,“Swifters: A multi-uav platform for disaster management,” in2019 International Conference on Information and CommunicationTechnologies for Disaster Management (ICT-DM), 2019, pp. 1–7.

[82] Z. Zhang, J. Wu, and C. He, “Search method of disaster inspectioncoordinated by multi-uav,” in 2019 Chinese Control Conference (CCC),2019, pp. 2144–2148.

[83] Z. Wang, H. Li, and L. Wu, “Geodesics-based topographical featureextraction from airborne lidar data for disaster management,” in 201018th International Conference on Geoinformatics, 2010, pp. 1–5.

[84] G. S. Percivall, N. S. Alameh, H. Caumont, K. L. Moe, and J. D. Evans,“Improving disaster management using earth observations—geoss andceos activities,” IEEE Journal of Selected Topics in Applied EarthObservations and Remote Sensing, vol. 6, no. 3, pp. 1368–1375, 2013.

[85] R. Hoque, D. Nakayama, H. Matsuyama, and J. Matsumoto, “Floodmonitoring, mapping and assessing capabilities using radarsat remotesensing, gis and ground data for bangladesh,” Natural Hazards, vol. 57,pp. 525–548, 05 2011.

[86] J. Tang, R. Yang, and M. Kafatos, “Data mining for tropical cycloneintensity prediction,” in Sixth Conference on Coastal Atmospheric andOceanic Prediction and Processes, 01 2005.

[87] M. F. Sikder, S. Halder, T. Hasan, M. J. Uddin, and M. K. Baowaly,“Smart disaster notification system,” in 2017 4th InternationalConference on Advances in Electrical Engineering (ICAEE), 2017, pp.658–663.

[88] L. Ma, M. Li, X. Ma, L. Cheng, P. Du, and Y. Liu, “A reviewof supervised object-based land-cover image classification,” ISPRSJournal of Photogrammetry and Remote Sensing, vol. 130, pp. 277 –293, 2017. [Online]. Available: http://www.sciencedirect.com/science/article/pii/S092427161630661X

[89] T. Blaschke, “Object based image analysis for remote sensing,” ISPRSJournal of Photogrammetry and Remote Sensing, vol. 65, no. 1,pp. 2 – 16, 2010. [Online]. Available: http://www.sciencedirect.com/science/article/pii/S0924271609000884

[90] S. W. Myint, P. Gober, A. Brazel, S. Grossman-Clarke, and Q. Weng,“Per-pixel vs. object-based classification of urban land cover extractionusing high spatial resolution imagery,” Remote Sensing of Environment,vol. 115, no. 5, pp. 1145 – 1161, 2011. [Online]. Available:http://www.sciencedirect.com/science/article/pii/S0034425711000034

[91] M. Demaria, C. Sampson, J. Knaff, and K. Musgrave, “Is tropicalcyclone intensity guidance improving?” Bulletin of the AmericanMeteorological Society, vol. 95, pp. 387–398, 03 2014.

[92] OGC, F. (1999) Airborne Laser Scanning Present Status andFuture Expectations. ISPRS Journal of Photogrammetry and RemoteSensing, 1999. [Online]. Available: http://dx.doi.org/10.1016/S0924-2716(99)00009-X

[93] Wikipedia contributors, “Rm-odp — Wikipedia, the free encyclopedia,”2019, [Online; accessed 28-April-2020]. [Online]. Available: https://en.wikipedia.org/w/index.php?title=RM-ODP&oldid=901830988

[94] OGC, GEOSS AIP Architecture—GEOSS Architecture ImplementationPilot (AIP) Feb. 28, 2012, 2005. [Online]. Available: https://www.ogc.org/projects/initiatives/geoss-ogc

[95] ipcc et al., Managing the Risks of Extreme Events andDisasters to Advance Climate Change Adaptation, 2012. [Online].Available: https://www.ipcc.ch/report/managing-the-risks-of-extreme-events-and-disasters-to-advance-climate-change-adaptation/

[96] C. Polanco, J. Castañón-González, A. Macias Hernández,J. L. Samaniego, T. Buhse, and S. Martínez, “Detection ofsevere respiratory disease epidemic outbreaks by cusum-basedovercrowd-severe-respiratory-disease-index model,” Computationaland mathematical methods in medicine, vol. 2013, p. 213206, 082013.

[97] “Bmc public health.” [Online]. Available: https://bmcpublichealth.biomedcentral.com/articles/10.1186/1471-2458-9-483

[98] N. Khan, “N.ahmad,n.riaz,m.hussain, “adhoc wireless sensor networkarchitecture for disaster survivor detection” international journalof advance science and technology, ijast, volume 34, 2011http://www.sersc.org/journals/ijast/vol34/2.pdf.” IJAST, vol. 34, 102011.

[99] X. Du, M. Guizani, Y. Xiao, and H. Chen, “Transactions papersa routing-driven elliptic curve cryptography based key managementscheme for heterogeneous sensor networks,” IEEE Transactions onWireless Communications, vol. 8, no. 3, pp. 1223–1229, 2009.

28

[100] C. Yawut and S. Kilaso, “A wireless sensor network for weather anddisaster alarm systems,” in International Conference on Informationand Electronics Engineering, IPCSIT, vol. 6, 2011, pp. 155–159.

[101] T. Nakamura, K. Kogo, J. Fujimura, K. Tsudaka, T. Wada, K. Ohtsuki,and H. Okada, “Development of emergency rescue evacuation supportsystem (eress) in panic-type disasters: Disaster detection by positioningarea of terminals,” in 2013 42nd International Conference on ParallelProcessing, 2013, pp. 931–936.

[102] W. PITTS and E. Johnsson, “Carbon monoxide formation infires by high-temperature anaerobic wood pyrolysis,” Symposium(international) on Combustion, vol. 25, pp. 1455–1462, 12 1994.

[103] X. Xu, L. Zhang, S. Sotiriadis, E. Asimakopoulou, M. Li, andN. Bessis, “Clotho: A large-scale internet of things-based crowdevacuation planning system for disaster management,” IEEE Internetof Things Journal, vol. 5, no. 5, pp. 3559–3568, 2018.

[104] M. Choksi, H. Gamit, and M. A. Zaveri, “Data sensing and resourcescheduling in clustered environment for post disaster managementusing internet of things,” in 2018 3rd International Conference forConvergence in Technology (I2CT), 2018, pp. 1–5.

[105] L. Abusalah, A. Khokhar, and M. Guizani, “A survey of secure mobilead hoc routing protocols,” IEEE Communications Surveys Tutorials,vol. 10, no. 4, pp. 78–93, 2008.

[106] H. Rossnagel and T. Scherner, “Secure mobile notifications ofcivilians in case of a disaster,” in Proceedings of the 10th IFIP TC-6TC-11 International Conference on Communications and MultimediaSecurity, ser. CMS’06. Berlin, Heidelberg: Springer-Verlag, 2006, p.33–42. [Online]. Available: https://doi.org/10.1007/11909033_4

[107] G. Jayasinghe, F. Fahmy, N. Gajaweera, and D. Dias, “A gsm alarmdevice for disaster early warning,” in First International Conferenceon Industrial and Information Systems, 2006, pp. 383–387.

[108] H. Torii, J. Sawamoto, N. Segawa, E. Sugino, and Y. Nomura,“Tsunami early alert and evacuation support system for fishery workersby mobile phones,” in 2010 IEEE 24th International Conference onAdvanced Information Networking and Applications Workshops, 2010,pp. 704–709.

[109] B. Anzengruber, D. Pianini, J. Nieminen, and A. Ferscha, “Predictingsocial density in mass events to prevent crowd disasters,” vol. 8238,11 2013, pp. 206–215.

[110] F. Sharevski, “Leveraging cellular intemet-of-things for resilient androbust disaster management,” in 2018 IEEE Global CommunicationsConference (GLOBECOM), 2018, pp. 1–5.

[111] K. M. Rahman, T. Alam, and M. Chowdhury, “Location based earlydisaster warning and evacuation system on mobile phones usingopenstreetmap,” in 2012 IEEE Conference on Open Systems, 2012,pp. 1–6.

[112] T. Juhana, “Mobile application design of disaster rescue team tracking,”in 2014 8th International Conference on Telecommunication SystemsServices and Applications (TSSA), 2014, pp. 1–3.

[113] M. R. M. Aldecimo and M. M. De Leon, “Development of anopenstreetmap based safe zone routing system for west valleyfault earthquake disaster in the makati central business district,philippines,” in 2016 3rd International Conference on Informationand Communication Technologies for Disaster Management (ICT-DM),2016, pp. 1–7.

[114] A. Khan, F. Aftab, and Z. Zhang, “Uapm: An urgency-aware packetmanagement for disaster management using flying ad-hoc networks,”China Communications, vol. 16, no. 11, pp. 167–182, 2019.

[115] B. Wolshon, Contraflow for Evacuation Traffic Management. Boston,MA: Springer US, 2008, pp. 165–170. [Online]. Available: https://doi.org/10.1007/978-0-387-35973-1_210

[116] F. Jiang, L. Zhong, K. Thilakarathna, A. Seneviratne, K. Takano,S. Yamada, and Y. Ji, “Supercharging crowd dynamics estimationin disasters via spatio-temporal deep neural network,” in 2017 IEEEInternational Conference on Data Science and Advanced Analytics(DSAA), 2017, pp. 184–192.

[117] J. Radianti and O.-C. Granmo, “A framework for assessing thecondition of crowds exposed to a fire hazard using a probabilisticmodel,” International Journal of Machine Learning and Computing,pp. 14–20, 2014.

[118] P. Haddawy, L. Frommberger, T. Kauppinen, G. De Felice,P. Charkratpahu, S. Saengpao, and P. Kanchanakitsakul, “Situationawareness in crowdsensing for disease surveillance in crisis situations,”05 2015.

[119] C. A. F. Ezequiel, M. Cua, N. C. Libatique, G. L. Tangonan,R. Alampay, R. T. Labuguen, C. M. Favila, J. L. E. Honrado,V. Caños, C. Devaney, A. B. Loreto, J. Bacusmo, and B. Palma, “Uavaerial imaging applications for post-disaster assessment, environmental

management and infrastructure development,” in 2014 InternationalConference on Unmanned Aircraft Systems (ICUAS), 2014, pp.274–283.

[120] P. Valente Klaine, J. Nadas, R. Souza, and M. Imran, “Distributeddrone base station positioning for emergency cellular networks usingreinforcement learning,” Cognitive Computation, vol. 10, 05 2018.

[121] J. W. Burris, R. Shrestha, B. Gautam, and B. Bista, “Machine learningfor the activation of contraflows during hurricane evacuation,” in 2015IEEE Global Humanitarian Technology Conference (GHTC), 2015, pp.254–258.

[122] Y. Ikeda and M. Inoue, “An evacuation route planning for safety routeguidance system after natural disaster using multi-objective geneticalgorithm,” Procedia Computer Science, vol. 96, pp. 1323–1331, 122016.

[123] F. Ofli, P. Meier, M. Imran, C. Castillo, D. Tuia, N. Rey, J. Briant,P. Millet, F. Reinhard, M. Parkan, and S. Joost, “Combining humancomputing and machine learning to make sense of big (aerial) data fordisaster response,” Big data, vol. 4, 03 2016.

[124] V. Hassija, V. Saxena, V. Chamola, and F. Richard Yu, “A parkingslot allocation framework based on virtual voting and adaptive pricingalgorithm,” IEEE Transactions on Vehicular Technology, vol. 69, no. 6,pp. 5945–5957, 2020.

[125] S. Soni, V. Gauri Shankar, and C. Sandeep, “Route-the safe: A robustmodel for safest route prediction using crime and accidental data,”vol. 28, pp. 1415–1428, 12 2019.

[126] J. Ren and X. guang Gao, “Situation assessment model foruav disaster relief in the city,” 2011 International Workshop onMulti-Platform/Multi-Sensor Remote Sensing and Mapping, pp. 1–6,2011.

[127] M. Y. Selim and A. E. Kamal, “Post-disaster 4g/5g networkrehabilitation using drones: Solving battery and backhaul issues,” 2018.

[128] N. Gondaliya and M. Atiquzzaman, “Rttmm: Role based 3-tier mobilitymodel for evaluation of delay tolerant routing protocols in post disastersituation,” in International Conference on Internet of Things and BigData, vol. 2. SCITEPRESS, 2016, pp. 11–20.

[129] D. de Alwis and E. So, “Enhanced change detection index for disasterresponse, recovery assessment and monitoring of accessibility andopen spaces (camp sites),” International Journal of Applied EarthObservation and Geoinformation, vol. 57, pp. 49–60, 05 2017.

[130] M. Sheykhmousa, N. Kerle, M. Kuffer, and S. Ghaffarian,“Post-disaster recovery assessment with machine learning-derived landcover and land use information,” Remote Sensing, vol. 11, p. 24, 052019.

[131] C. J. Van Westen, “Where to build-back-better? analyzing changing riskfor post-disaster reconstruction planning,” in Proceedings of the The 5thInternational Symposium on Mega Earthquake Induced Geo-Disastersand Long Term Effects, Chengdu, China, 2018, pp. 11–16.

[132] K. Starbird and L. Palen, “(how) will the revolution be retweeted?:information diffusion and the 2011 egyptian uprising,” 02 2012, pp.7–16.

[133] S. Vieweg, A. Hughes, K. Starbird, and L. Palen, “Microbloggingduring two natural hazards events: What twitter may contribute tosituational awareness,” vol. 2, 01 2010, pp. 1079–1088.

[134] M. Jamali, A. Nejat, S. Ghosh, and G. Cao, “Social media dataand post-disaster recovery,” International Journal of InformationManagement, vol. 44, pp. 25–37, 09 2018.

[135] N. Khan, “Flood prediction and disaster risk analysis using gis basedwireless sensor networks,a review,” Journel of Basic and AppliedScientific Resaerch, 09 2013.

[136] N. A. Agana and A. Homaifar, “Emd-based predictive deep beliefnetwork for time series prediction: an application to droughtforecasting,” Hydrology, vol. 5, no. 1, p. 18, 2018.

[137] F. Alidoost and H. Arefi, “Application of deep learning for emergencyresponse and disaster management,” 02 2018.

[138] Yang, Li, and Yun, “Big data in natural disaster management: Areview,” May 2018. [Online]. Available: https://www.mdpi.com/2076-3263/8/5/165#cite

[139] A. Lopez Pineda, Y. Ye, S. Visweswaran, G. Cooper, M. Wagner, andF. R. Tsui, “Comparison of machine learning classifiers for influenzadetection from emergency department free-text reports,” Journal ofBiomedical Informatics, 09 2015.

[140] Y. Ye, F. R. Tsui, M. Wagner, J. U. Espino, and Q. Li, “Influenzadetection from emergency department reports using natural languageprocessing and bayesian network classifiers,” 2014.

[141] L. Wynants et al., “Prediction models for diagnosis and prognosisof covid-19: systematic review and critical appraisal,” BMJ, vol.

29

369, 2020. [Online]. Available: https://www.bmj.com/content/369/bmj.m1328

[142] A. Lopez-Rincon, A. Tonda, L. Mendoza-Maldonado, E. Claassen,J. Garssen, and A. D. Kraneveld, “Accurate identification of sars-cov-2from viral genome sequences using deep learning,” 2020.

[143] Beijing Institute of Genomics, Chinese Academy of Science, ChinaNational Center for Bioinformation National Genomics DataCenter, “2019 Novel Coronavirus Resource (2019nCoVR),” https://bigd.big.ac.cn/ncov/?lang=en, online; accessed 15 October 2020.

[144] S. T. Sherry, M.-H. Ward, M. Kholodov, J. Baker, L. Phan, E. M.Smigielski, and K. Sirotkin, “dbsnp: the ncbi database of geneticvariation,” Nucleic acids research, vol. 29, no. 1, pp. 308–311, 2001.

[145] J. Kim, P. Choe, Y. Oh, K. Oh, J. Kim, S. Park, J. Park, H. Na, andM.-d. Oh, “A first case of 2019 novel coronavirus pneumonia importedinto korea from wuhan, china: Implication for infection prevention andcontrol measures,” Journal of Korean Medical Science, vol. 35, 022020.

[146] Y. Chaudhary, M. Mehta, D. G. Gupta, S. Lelis, R. Sharma, A. Khanna,and J. R. Rodrigues, “Efficient-covidnet: Deep learning based covid-19detection from chest x-ray images,” in IEEE Healthcom, December2020.

[147] P. Santos, L. Souza, S. Lelis, H. Ribeiro, F. Borges, R. Silva, A. Filho,F. Araujo, R. R. Rabelo, and J. R. Rodrigues, “Prediction of covid-19using time-sliding window: the case of piaui state - brazil,” in IEEEHealthcom, December 2020.

[148] F. Liang, P. Guan, W. Wu, and D. Huang, “Forecasting influenzaepidemics by integrating internet search queries and traditionalsurveillance data with the support vector machine regression modelin liaoning, from 2011 to 2015,” PeerJ, vol. 6, p. e5134, 06 2018.

[149] D. Olson, K. Konty, M. Paladini, C. Viboud, and L. Simonsen,“Reassessing google flu trends data for detection of seasonal andpandemic influenza: A comparative epidemiological study at threegeographic scales,” PLoS computational biology, vol. 9, p. e1003256,10 2013.

[150] Google Flu Trends and Google Dengue Trends, “Google Flu Trendsand Google Dengue Trends,” https://www.google.org/flutrends/about/,online; accessed 15 October 2020.

[151] M. W. L. Moreira, J. J. P. C. Rodrigues, V. Korotaev, J. Al-Muhtadi,and N. Kumar, “A comprehensive review on smart decision supportsystems for health care,” IEEE Systems Journal, vol. 13, no. 3, pp.3536–3545, 2019.

[152] C. Wang, Y. Li, W. Feng, K. Liu, S. Zhang, F. Hu, S. Jiao, X. Lao,H. Ni, and G. Xu, “Epidemiological features and forecast modelanalysis for the morbidity of influenza in ningbo, china, 2006–2014,”International Journal of Environmental Research and Public Health,vol. 14, p. 559, 05 2017.

[153] S. G. Alonso, I. de la Torre Diez, J. J. Rodrigues, S. Hamrioui, andM. Lopez-Coronado, “A systematic review of techniques and sources ofbig data in the healthcare sector,” Journal of medical systems, vol. 41,no. 11, p. 183, 2017.

[154] G. Manogaran, R. Varatharajan, D. Lopez, P. M K, R. Sundarasekar,and C. Thota, “A new architecture of internet of things and bigdata ecosystem for secured smart healthcare monitoring and alerting,”Future Generation Computer Systems, 11 2017.

[155] “Covid-19 overview,” January 2020. [Online]. Available: https://www.coronatracker.com/analytics

[156] V. Chamola, V. Hassija, V. Gupta, and M. Guizani, “A comprehensivereview of the covid-19 pandemic and the role of iot, drones, ai,blockchain, and 5g in managing its impact,” IEEE Access, vol. 8, pp.90 225–90 265, 2020.

[157] A. Handel, I. M. Longini, and R. Antia, “Intervention strategies for aninfluenza pandemic taking into account secondary bacterial infections,”Epidemics, vol. 1, no. 3, pp. 185 – 195, 2009. [Online]. Available:http://www.sciencedirect.com/science/article/pii/S1755436509000413

[158] C. Fraser, S. Riley, R. Anderson, and N. Ferguson, “Factors thatmake an infectious disease outbreak controllable,” Proceedings of theNational Academy of Sciences of the United States of America, vol.101, pp. 6146–51, 05 2004.

[159] J. Kelso, G. Milne, and H. Kelly, “Simulation suggests that rapidactivation of social distancing can arrest epidemic development dueto a novel strain of influenza,” BMC public health, vol. 9, p. 117, 052009.

[160] “Bluetooth technology.” [Online]. Available: https://www.bluetooth.com/

[161] S. Holm, “Airborne ultrasound data communications: the core of anindoor positioning system,” in IEEE Ultrasonics Symposium, 2005.,vol. 3, 2005, pp. 1801–1804.

[162] J. W. Choi, D. H. Yim, and S. H. Cho, “People counting based onan ir-uwb radar sensor,” IEEE Sensors Journal, vol. 17, no. 17, pp.5717–5727, 2017.

[163] S. Bartoletti, A. Conti, and M. Z. Win, “Device-free countingvia wideband signals,” IEEE Journal on Selected Areas inCommunications, vol. 35, no. 5, pp. 1163–1174, 2017.

[164] S. L. D. C. Johnson, Joseph Jr. Hasan and N. Ahmed,“Social-distancing monitoring using portable electronic devices,”April 17 2020. [Online]. Available: https://www.tdcommons.org/dpubs_series/3158/

[165] W. Kang, S. Nam, Y. Han, and S. Lee, “Improved heading estimationfor smartphone-based indoor positioning systems,” in 2012 IEEE23rd International Symposium on Personal, Indoor and Mobile RadioCommunications - (PIMRC), 2012, pp. 2449–2453.

[166] U. Challita, A. Ferdowsi, M. Chen, and W. Saad, “Machine learningfor wireless connectivity and security of cellular-connected uavs,” IEEEWireless Communications, vol. 26, no. 1, pp. 28–35, 2019.

[167] F. Yin, Z. Lin, Y. Xu, Q. Kong, D. Li, S. Theodoridis, Shuguang, andCui, “Fedloc: Federated learning framework for data-driven cooperativelocalization and location data processing,” 2020.

[168] C. Wang, L. Ma, R. Li, T. S. Durrani, and H. Zhang, “Exploringtrajectory prediction through machine learning methods,” IEEE Access,vol. 7, pp. 101 441–101 452, 2019.

[169] Y. Wang, Xu Yang, Yutian Zhao, Yue Liu, and L. Cuthbert, “Bluetoothpositioning using rssi and triangulation methods,” in 2013 IEEE10th Consumer Communications and Networking Conference (CCNC),2013, pp. 837–842.

[170] S. S. Hassan, S. D. Bibon, M. S. Hossain, and M. Atiquzzaman,“Security threats in bluetooth technology,” Comput. Secur., vol. 74,pp. 308–322, 2018. [Online]. Available: https://doi.org/10.1016/j.cose.2017.03.008

[171] F. Sattler, J. Ma, P. Wagner, D. Neumann, M. Wenzel, R. Schäfer,W. Samek, K.-R. Müller, and T. Wiegand, “Risk estimation ofsars-cov-2 transmission from bluetooth low energy measurements,” 042020.

[172] C. T. N. aand others, “Enabling and emerging technologies for socialdistancing: A comprehensive survey,” 2020.

[173] V. Hassija, V. Gupta, S. Garg, and V. Chamola, “Traffic jam probabilityestimation based on blockchain and deep neural networks,” IEEETransactions on Intelligent Transportation Systems, pp. 1–10, 2020.

[174] V. Hassija, V. Chamola, G. Han, J. J. P. C. Rodrigues, and M. Guizani,“Dagiov: A framework for vehicle to vehicle communication usingdirected acyclic graph and game theory,” IEEE Transactions onVehicular Technology, vol. 69, no. 4, pp. 4182–4191, 2020.

[175] I. Ahmed, M. Ahmad, J. J. Rodrigues, G. Jeon, and S. Din, “A deeplearning-based social distance monitoring framework for covid-19,”Sustainable Cities and Society, p. 102571, 2020.

[176] G. Praveen, V. Chamola, V. Hassija, and N. Kumar, “Blockchain for5g: A prelude to future telecommunication,” IEEE Network, pp. 1–8,2020.

[177] X. Ge, H. Cheng, M. Guizani, and T. Han, “5g wireless backhaulnetworks: challenges and research advances,” IEEE Network, vol. 28,no. 6, pp. 6–11, 2014.

[178] N. Bui, M. Cesana, S. A. Hosseini, Q. Liao, I. Malanchini, andJ. Widmer, “A survey of anticipatory mobile networking: Context-basedclassification, prediction methodologies, and optimization techniques,”IEEE Communications Surveys Tutorials, vol. 19, no. 3, pp.1790–1821, 2017.

[179] S. Cho, P. Gordon, J. Moore, H. Richardson, M. Shinozuka, andS. Chang, “Integrating transportation network and regional economicmodels to estimate the costs of a large urban earthquake,” Journal ofRegional Science, vol. 41, pp. 39–65, 02 2001.

[180] E. Cavallo, S. Galiani, I. Noy, and J. Pantano, “Catastrophic naturaldisasters and economic growth,” University of Hawaii at Manoa,Department of Economics, Working Papers, vol. 95, 01 2010.

[181] F. LI, “The information content of forward-looking statements incorporate filings-a naive bayesian machine learning approach,” Journalof Accounting Research, vol. 48, pp. 1049 – 1102, 12 2010.

[182] T.-N. Phan and M. Kappas, “Comparison of random forest, k-nearestneighbor, and support vector machine classifiers for land coverclassification using sentinel-2 imagery,” Sensors, vol. 18, p. 18, 122017.

[183] J. Raikwal and K. Saxena, “Performance evaluation of svm andk-nearest neighbor algorithm over medical data set,” InternationalJournal of Computer Applications, vol. 50, pp. 35–39, 07 2012.

30

[184] M. A. P. Chamikara, P. Bertók, I. Khalil, D. Liu, S. Camtepe, andM. Atiquzzaman, “Local differential privacy for deep learning,” IEEEInternet Things J., vol. 7, no. 7, pp. 5827–5842, 2020.

[185] ——, “A trustworthy privacy preserving framework for machinelearning in industrial iot systems,” IEEE Trans. Ind. Informatics,vol. 16, no. 9, pp. 6092–6102, 2020.

[186] T. Alladi, V. Chamola, B. Sikdar, and K. R. Choo, “Consumer iot:Security vulnerability case studies and solutions,” IEEE ConsumerElectronics Magazine, vol. 9, no. 2, pp. 17–25, 2020.

[187] V. Hassija, V. Chamola, V. Gupta, S. Jain, and N. Guizani, “A survey onsupply chain security: Application areas, security threats, and solutionarchitectures,” IEEE Internet of Things Journal, vol. PP, 10 2020.

[188] V. Hassija, V. Chamola, B. C. Bajpai, Naren, and S. Zeadally, “Securityissues in implantable medical devices: Fact or fiction?” SustainableCities and Society, p. 102552, 2020.

[189] B. M. C. Silva, J. J. P. C. Rodrigues, F. Canelo, I. M. C. Lopes, andJ. Lloret, “Towards a cooperative security system for mobile-healthapplications.” Electron. Commer. Res., vol. 19, pp. 629–654, 2019.

Vinay Chamola received the B.E. degree inelectrical and electronics engineering and master’sdegree in communication engineering from the BirlaInstitute of Technology and Science, Pilani, India, in2010 and 2013, respectively. He received his Ph.D.degree in electrical and computer engineering fromthe National University of Singapore, Singapore, in2016. In 2015, he was a Visiting Researcher with theAutonomous Networks Research Group (ANRG),University of Southern California, Los Angeles, CA,USA. He also worked as a post-doctoral research

fellow at the National University of Singapore, Singapore. He is currentlyAssistant Professor with the Department of Electrical and ElectronicsEngineering, BITS-Pilani, Pilani where he heads the Internet of ThingsResearch Group / Lab. He has over 50 publications in high ranked SCIJournals including more than 33 IEEE Transaction and Journal articles. Hisresearch interests include IoT Security, Blockchain, UAVs, VANETs, 5G andHealthcare. He is a Guest Editor in Computer Communication, Elsevier; andalso the IET Intelligent Transporation Systems Journal. He serves as an AreaEditor for the Ad Hoc Networks journal, Elsevier. He also serves as anAssociate editor in the IET Quantum Communications, IET Networks andseveral other journals. He is a senior member of the IEEE.

Vikas Hassija received the B.Tech. degree fromMaharshi Dayanand University, Rohtak, India, in2010, and the M.S. degree in telecommunicationsand software engineering from the Birla Instituteof Technology and Science (BITS), Pilani, India, in2014. He is currently pursuing the Ph.D. degree inIoT security and blockchain with the Jaypee Instituteof Information and Technology (JIIT), Noida, wherehe is currently an Assistant Professor. His researchinterests include the IoT security, network security,blockchain, and distributed computing.

Sakshi gupta is currently pursuing the B.Tech.degree with the Jaypee Institute of InformationTechnology (JIIT), Noida. She has completed afew projects in the field of evolutionary algorithms,machine learning, and data analytics. She iscurrently (the summer of 2020) pursuing herresearch internship with the Birla Institute ofTechnology and Science (BITS), Pilani, under Dr. V.Chamola. Her research interests include evolutionarycomputing, machine learning, and deep learning.

Adit Goyal is currently pursuing the B.Tech. degreewith the Computer Science department, JaypeeInstitute of Information Technology (JIIT), Noida.He has completed a few projects in the field ofdata science, machine learning, and big data. He iscurrently pursuing his research internship with theBirla Institute of Technology and Science (BITS),Pilani, under Dr. V. Chamola. His research interestsinclude machine learning, data science, and quantumcomputing.

Mohsen Guizani (S’85–M’89–SM’99–F’09)received the B.S. (with distinction) and M.S.degrees in electrical engineering, the M.S. andPh.D. degrees in computer engineering fromSyracuse University, Syracuse, NY, USA, in 1984,1986, 1987, and 1990, respectively. He is currently aProfessor at the Computer Science and EngineeringDepartment in Qatar University, Qatar. Previously,he served in different academic and administrativepositions at the University of Idaho, WesternMichigan University, University of West Florida,

University of Missouri-Kansas City, University of Colorado-Boulder, andSyracuse University. His research interests include wireless communicationsand mobile computing, computer networks, mobile cloud computing, security,and smart grid. He is the Editor-in-Chief of Wireless Communications andMobile Computing journal (Wiley). He is the author of nine books and morethan 500 publications in refereed journals and conferences. He received the2017 IEEE Communications Society WTC Recognition Award as well as the2018 AdHoc Technical Committee Recognition Award for his contributionto outstanding research in wireless communications and Ad-Hoc Sensornetworks. He was the Chair of the IEEE Communications Society WirelessTechnical Committee and the Chair of the TAOS Technical Committee. Heserved as the IEEE Computer Society Distinguished Speaker and is currentlythe IEEE ComSoc Distinguished Lecturer. He is a Fellow of IEEE and aSenior Member of ACM.

Biplab Sikdar (S’98–M’02–SM’09) received theB.Tech. degree in electronics and communicationengineering from North Eastern Hill University,Shillong, India, in 1996, the M.Tech. degree inelectrical engineering from the Indian Institute ofTechnology Kanpur, Kanpur, India, in 1998, andthe Ph.D. degree in electrical engineering fromRensselaer Polytechnic Institute, Troy, NY, USA, in2001. He was a faculty at the Rensselaer PolytechnicInstitute, from 2001 to 2013, first as an AssistantProfessor and then as an Associate Professor. He

is currently an Associate Professor with the Department of Electrical andComputer Engineering, National University of Singapore, Singapore. Heserves as the Vice dean of Graduate program and as the Area director ofCommunications and Networks lab at NUS. His current research interestsinclude wireless network, and security for Internet of Things and cyberphysical systems. Dr. Sikdar served as an Associate Editor for the IEEETransactions on Communications from 2007 to 2012. He currently servesas an Associate Editor for the IEEE Transactions on Mobile Computing.He has served as a TPC in various conferences such as IEEE LANMAN,GLOBECOM, BROADNETS and ICC to name a few. He is a member of EtaKappa Nu and Tau Beta Pi.