You are on page 1of 28

International Journal of Image and Graphics

Vol. 22, No. 5 (2022) 2250040 (28 pages)


#.c World Scienti¯c Publishing Company
DOI: 10.1142/S0219467822500401

A Review on Natural Disaster Detection in


Social Media and Satellite Imagery Using
Machine Learning and Deep Learning

Swapandeep Kaur* and Sheifali Gupta†


Chitkara University Institute of Engineering and Technology
Chitkara University, Chandigarh-Patiala National Highway (NH-64)
Village, Jansla, Rajpura, Punjab 140401, India
*swapandeep.kaur@chitkara.edu.in

sheifali.gupta@chitkara.edu.in

Swati Singh
Department of Electronics and Communication Engineering
University Institute of Technology
Himachal Pradesh University
Shimla 171005, Himachal Pradesh, India
swatisanjta3@gmail.com

Tanvi Arora
Chandigarh Group of Colleges
Landran, Mohali, Punjab, India
tanvi.coecse@cgc.edu.in

Received 30 May 2021


Revised 21 June 2021
Published 6 October 2021

A disaster is a devastating incident that causes a serious disruption of the functions of a


community. It leads to loss of human life and environmental and ¯nancial losses. Natural
disasters cause damage and privation that could last for months and even years. Immediate
steps need to be taken and social media platforms like Twitter help to provide relief to the
a®ected public. However, it is di±cult to analyze high-volume data obtained from social media
posts. Therefore, the e±ciency and accuracy of useful data extracted from the enormous posts
related to disaster are low. Satellite imagery is gaining popularity because of its ability to cover
large temporal and spatial areas. But, both the social media and satellite imagery require the use
of automated methods to avoid the errors caused by humans. Deep learning and machine
learning have become extremely popular for text and image classi¯cation tasks. In this paper, a
review has been done on natural disaster detection through information obtained from social
media and satellite images using deep learning and machine learning.

Keywords: Disaster; social media; Twitter; satellite; classi¯cation; deep learning; machine
learning.

*Corresponding author.

2250040-1
S. Kaur et al.

1. Introduction
With the increase in change in climate, there is also a rise in the rate of occurrence
of natural disasters along with their severity.1 Global warming along with
growth in population and change in land covers has resulted in an increase in
natural disasters.2 For managing the aftermaths of the disasters, the information
relating to disasters needs to be retrieved and integrated on a timely basis.1
Hurricanes, earthquakes, °oods and wild¯res are some of the natural disasters
that take place on Earth and cause harmful e®ects on humans as well as the
environment.3
Hurricanes or tropical cyclones as shown in Fig. 1(a) are regarded as the most
calamitous among the natural disasters occurring on earth.4 They mostly occur in
the tropical or sub-tropical areas because of the presence of warm seawater. As the
sun heats the seawaters during the summer season, it leads to the creation of huge
clouds that lead to hurricanes. Heavy rainfall, °ood and very high-speed winds at
the rate of 200 miles/h occur during hurricanes causing excessive damage to
property5 and human casualties. Earthquake as shown in Fig. 1(b) is the shaking of
Earth caused due to energy released in the lithosphere of Earth. It is caused due to
the seismic activity which could be either natural or man-made, generating seismic
waves. Earthquakes could also lead to tsunami when there is a displacement of the
seabed. Floods as shown in Fig. 1(c) are basically over°owing of water on dry
ground. River over°ows and heavy rainfall are major causes of °oods. Wild¯re is
an unexpected ¯re that causes burning in natural areas such as grassland, forest
or prairie. They can be caused by natural phenomena like lightning or by
human activity. They can happen at any time or anywhere. This disaster is shown
in Fig. 1(d).
Immediate actions need to be taken in response to the disaster event6 for pri-
oritizing the rescue operations and providing relief to the a®ected people.7 Earlier
windshield survey was conducted which involved volunteers and emergency crews
driving around the areas a®ected. But this method was time-consuming and also
led to errors. People use blogging sites such as Twitter to post text and images
related to emergencies and natural disasters. These posts are immensely useful for
providing relief to the a®ected people by the humanitarian organizations. But
analyzing this high-volume data remains a challenge due to unavailability of the
required tools. Therefore, automated methods like machine learning and deep
learning (DL) need to be used. Images from satellites are also gaining popularity
for damage pattern recognition. Satellites can easily cover huge temporal as well as
spatial areas. But this still relies on visualizing and inspecting by humans which
can be unreliable. Automated processes need to be used and hence, computer
vision comes into picture. Machine learning which is composed of a family of
algorithms allows computers to learn from experience automatically. It also
allows computers to perform better through adaptation to new conditions. DL
which is a subset of machine learning helps to perform these tasks automatically.

2250040-2
A Review on Natural Disaster Detection in Social Media

(a) (b)

(c) (d)

Fig. 1. Satellite images of common disasters: (a) cyclone/hurricane, (b) earthquake, (c) °ood and (d)
wild¯re.3

Successful understanding of images has been achieved by the application of con-


volutional neural network (CNN).8
A CNN is composed of a number of layers that help in data processing. Every
layer in the CNN is basically a ¯lter that helps in detecting features of images. CNNs
are capable of detecting tra±c signs and objects in a much better way than human
beings did.9 CNN is a supervised learning technique composed of mainly hidden
layers that are used for feature detection and the fully connected layers that help in
classifying images.9

2. Computer-Aided Design System for Natural Disaster Detection


Designing of any engineering model is done by a Computer-Aided Design (CAD)
system. It increases productivity and design quality of all the systems. CAD systems
used traditionally are parametric in nature and hence are time-consuming.

2250040-3
S. Kaur et al.

Therefore, arti¯cial intelligence and CAD together came into picture. CAD systems
are being designed for natural disaster detection both using machine learning
and DL.

2.1. Machine learning


Machine learning is a subset of arti¯cial intelligence that enables the system to learn
automatically without being programmed explicitly. Machine learning is further
divided into supervised, unsupervised, semi-supervised and reinforcement learning.
Machine learning is being used for natural disaster detection. Machine learning
techniques have been used for improving the e±ciency of disaster for social media
data classi¯cation. The important requirements like food, transport facility and
housing facility required by the victims have also been analyzed. Integration of
analytics of social media with humanitarian relief needs to be done.4
DL is also getting popularity in the detection and estimation of occurrence of
natural disasters and the damage caused by them. Various adaptations of DL for
disaster damage detection are given in the following.

2.2. CNN
CNN is a neural network specialized to work with two-dimensional images, though
they can also be used for one-dimensional and three-dimensional images data. Image
recognition, video recognition and natural language processing are the major
applications of CNNs. The main building block of CNN is the convolution layer.
Convolution layer consists of ¯lters that are independent in nature and are convolved
with the image to produce feature maps. The next important layer is the pooling
layer that works on individual feature map. Its function is of reducing the spatial size
of the image so that computations are reduced. Recti¯ed Linear Unit (ReLU) layer
adds nonlinearity to the network. The last layer is the fully connected layer in which
every neuron has a connection with the previous layer activations. The fully con-
nected layer helps in predicting the label which is best in describing the image. CNNs
proved to be great at middle and high-level feature extraction by the use of con-
volutional and pooling layers.
CNN's are widely used in natural disaster detection. They have been used for
disaster detection using aerial imagery that were trained using footage from the past
disasters. Training of the model was done using Volan 2018 dataset and 80.69% and
74.48% precision were obtained for high altitude and low altitude, respectively.1

2.3. Recurrent neural network


Recurrent Neural Networks (RNN) is particularly used for Natural Language Pro-
cessing tasks. Recurrent Neural networks help in sequential data analysis. The inputs
and outputs are independent of each other in traditional neural networks. But, this is
not useful in all the cases since if there is a need of prediction of next word in

2250040-4
A Review on Natural Disaster Detection in Social Media

sentences, then words preceding them would be required and hence, RNN comes into
picture. RNNs are called recurrent because same technique is applied on every ele-
ment of a sequence and output is dependent on calculations done previously. They
contain a memory element that keeps record of the previous information. The
commonly used model of RNN is the Long Short Term Memory (LSTM) which help
in classi¯cation and also to predict data based on time series.
For natural disaster detection, progress of hurricane that included preparedness,
response as well as the recovery was found out through the hurricane tweets and
information was disseminated to the appropriate group of people such as government
and rescue missions. Natural language processing was used for the same.12

2.4. Auto-encoders
Auto-encoder is an unsupervised algorithm which uses backpropagation and makes
the target equal to the input values. They are used for reducing the size of inputs into
a representation that is smaller than the original. An auto-encoder does not require
learning the dense layers. It learns through convolutional layers which are better for
video, images and series data. The applications of auto-encoders include image col-
oring which involves conversion of black and white images to a colored image. It is
also used in removing noise and watermarks from the images.
In natural disaster detection, convolutional auto-encoders found out damage
caused by hurricane Sandy in 2012 from aerial imagery and obtained an accuracy of
88.3%.28

2.5. Deep neural network


Since traditional CNN's performance degrades as the depth of the system increases,
deep neural networks allow for training of more deep networks. The accuracy in the
case of traditional networks starts saturating at a particular depth after which the
training error increases due to the degradation of accuracy. Thus, deep neural net-
work helps to simplify the process of training deeper networks. The training error
remains the same as the traditional counterpart; in the case, it does not reduce with
more layers. But, these networks are used as conceptual method for enhancing other
models instead of being used as a di®erent neural network category.
DL is gaining popularity in disaster management through remote sensing and
also through social media.10 In the following sections, the literature survey of the
natural disasters has been presented based on social media and satellite imagery
using DL.

3. Performance Parameters for CAD System


The most important performance parameters that help in ¯nding out which model is
e®ective in determining the damage caused by natural disasters are given in the
following.

2250040-5
S. Kaur et al.

3.1. Evaluation metrics for classi¯cation


(a) Accuracy – It is the ratio of total number of correct predictions upon all the
predictions made as shown in Eq. (1).
True PositiveðTPÞ þ True NegativeðTNÞ
Accuracy ¼ ; ð1Þ
TP þ TN þ FP þ FN
where True Positive (TP) means that the data are correct and are also predicted
correct. True Negative (TN) means that the data are incorrect and are also predicted
incorrect. FP stands for False Positive and FN implies False Negative.
Accuracy is useful in areas where TP and TN are more important. For example,
accuracy is an important metric when one needs to predict the survival status of
people a®ected by disasters from the test set correctly.
(b) Precision – Precision is the ratio of the true positive to the predicted positive as
shown in Eq. (2).
True PositiveðTPÞ
Precision ¼ : ð2Þ
Predicted PositiveðTP þ FPÞ

(c) Recall/Sensitivity – It is found by dividing the true positives by the total


number of positives as shown in Eq. (3).
True PositiveðTPÞ
Recall=Sensitivity ¼ : ð3Þ
Actual PositiveðTP þ FNÞ

Recall calculates what proportion of the positive class got correctly classi¯ed. It
determines what proportion of the actual disaster a®ected people was detected ac-
curately by the system.
(d) Speci¯city – It is found by dividing true negatives by the actual negatives as
shown in Eq. (4).
True NegativeðTNÞ
Specificity ¼ ð4Þ
Actual NegativeðFP þ TNÞ

(e) F1 Score – It is found by taking the harmonic mean of sensitivity (Recall) and
precision as shown in Eq. (5).
2  Precision  Recall
F1 Score ¼ : ð5Þ
Precision þ Recall
F1 Score is less prone to the unpredictability of class imbalance which the metric
accuracy su®ers from. Therefore, if the disaster dataset is imbalanced, F1 score is
used as the performance metric.
(f) Area under Curve (AUC) – AUC–ROC (Receiver Operating Characteristics)
is used to measure the performance of problems of classi¯cation. It helps to

2250040-6
A Review on Natural Disaster Detection in Social Media

distinguish between the classes. If the AUC is high, it means the model is better at
distinguishing between damaged and undamaged buildings from natural disasters.

3.2. Evaluation metrics for regression


Mean Square Error (MSE) – It is one of the most commonly used metric for
regression. MSE ¯nds out the average of the di®erence squared between predicted
and the actual value as shown in Eq. (6).
X n
MSE ¼ 1=n ð^
y i  yiÞ 2 ð6Þ
i¼1

Mean Absolute Error (MAE) – It is the average of absolute di®erence of target


and predicted value of the model as given by Eq. (7).
X n
MAE ¼ 1=n j^
y i  yij ð7Þ
i¼1

R-Square/Adjusted R-Square – It is also known as coe±cient of determination


and is a measure that helps in determining proportion of variance in dependent
variable given by independent variable as shown in Eq. (8).
SSregression
R-Squared ¼ ð8Þ
SStotal
where SSregression is the sum of squares because of regression and SStotal is the total
sum of squares.

3.3. Object detection evaluation metrics


Intersection Over Union (IOU) – An object detection evaluation metric is In-
tersection Over Union that helps in ¯nding di®erence between ground truth and
predicted bounding boxes as given by Eq. (9).
Area of overlap
IOU ¼ ð9Þ
Area of union
Average Precision – It is used for computing the area under the precision-recall
curve for a particular class.
Mean Average Precision (mAP) – This metric helps in comparing ground truth
bounding boxes to the detected box and gives a score based on the comparison.
Average precision is taken out for all the classes that give the mAP for a particular
model. The greater this score, the more accurately the model will help in the detection.

3.4. Segmentation evaluation metrics


Pixel Accuracy – This metric reports the percentage of pixels present in the image
that were classi¯ed correctly. Pixel accuracy is found separately for every class and
for all the classes globally as well.

2250040-7
S. Kaur et al.

4. Publicly Available Datasets for Natural Disaster Images


Data form an integral part of classi¯cation of images especially for machine learning
and DL techniques. Su±cient quantity of data needs to be collected for forming
dataset as it is important for classifying images in DL. Proper construction of the
dataset relies on pro¯ciency in the respective ¯eld and for selection of the required
information. Infrastructure and considerable amount of time are required to bring
the data in the format that is easily understandable. Standard datasets are utilized
generally which are already existing in the domain area. The existing datasets are
bene¯cial since a fair comparison can be made between the various system models.
Commonly used databases in the ¯eld of natural disaster detection are summarized
in Table 1.

5. Related Work
Research work has been done using DL and machine learning separately using
techniques such as Support Vector Machine, CNNs, Decision Trees, Random Forest.
The following sections will discuss the works done previously using these techniques.
The literature review of the various natural disasters like hurricanes, earthquakes
and °oods is divided into two subsections based on DL and machine learning which
are further divided into literature based on social media and literature based on
satellite imagery. A table containing the summary about the literature is presented
in the concluding section.

5.1. Literature of natural disaster detection based on DL techniques


DL is one of the fastest growing ¯elds of present times and has brought about a
massive breakthrough since 2013. It is known as `deep' learning since it is made up of
arti¯cial neural networks consisting of more than two hidden layers. DL is gaining
popularity in disaster management through remote sensing and also through social
media.10 The literature survey of natural disasters based on social media and satellite
imagery using DL is presented in this section.
(a) Literature of Natural Disaster Detection Based on Images from Social
Media
A CNN was developed for detecting °ood photographs from social media.
Training data consisted of 4800 images and the advantage of the system was that the
system was constructed in an iterative manner which could be again trained by a
greater training dataset as and when more °ooding images were available. In the
balanced test set, an accuracy of 93% was obtained for detecting °ood images and
46–63% precision was obtained for the unbalance tweets. Low precision has been
obtained for the imbalanced images obtained from social media.11
DL, spatial temporal analysis and natural language processing had been used for
identifying time-related context and content of social media posts relating to the

2250040-8
Table 1. Database of disaster images.

Reference Database Image size Type of disaster No. of disaster images Classi¯cation label/mask
4 NOAA-AVHRR 66  66 pixels Hurricane 1200 

Satellite images
6 TOMNOD 150  150 pixels Hurricane 23 000 Damage
No damage
10 TOMNOD and 9351  9351 Hurricane 18 474 Damaged
FEMA Non-damaged
11 RIASM 
 Flood 4800 Flooding
Non-°ooding
13 Twitters Spritzer 1  25 088 (°attened Hurricane 1128 Spam
image) Not urgent
Somewhat urgent
Moderately urgent
Highly urgent

2250040-9
14 GIS 
 Flood 1464 Flood
Non-°ood
16 Digital Globe's 161  161 Earthquake 75 468
Worldview
18 NOAA 1920  1080 Hurricane 5041 Undamaged
Damaged ruins
19 WorldView 160  160 
 13 485 

22 ImageNet 299  299 
 16 754
25 HURDAT2 
 Hurricane 48 828 H5
H4
H3
H2
H1
26 WorldView2 256  256 Earthquake 1600 

A Review on Natural Disaster Detection in Social Media
S. Kaur et al.

Table 1. (Continued )

Reference Database Image size Type of disaster No. of disaster images Classi¯cation label/mask
27 LiDAR 1920  1080 Hurricane 700 Undamaged
Debris
Damaged
28 DigitalGlobe 
 Wild¯re 1080 

29 AIDR 224  224 Hurricane 61 000 Severe damage
Mild damage
No damage
43 xBD 
 
 22 068 images and No damage

2250040-10
850 736 building Minor damage
polygons Major damage
Destroyed
44 ISBDA Dataset 
 Hurricane 1030 images sampled Mask R-CNN
from 10 videos
45 Digital Globe 50  50 
 88 million 

46 Digital Globe 1024  1024 
 6226 Grayscale mask
A Review on Natural Disaster Detection in Social Media

natural disasters. The work done in this paper is compared with previous works that
focused to identify disaster events and providing response immediately for the di-
saster whereas this paper provides a rather holistic approach. The work done in this
paper took note and captured the most important information relating to progress of
hurricane that includes preparedness, the response and also the recovery. Tweets
applicable to hurricane were identi¯ed and information was disseminated to the
appropriate group of people such as government and rescue missions. More than 500
million keyword-based tweets were analyzed during the disaster and also before the
disaster. However, this study lacked the in-depth analysis of the information which
was site-speci¯c and did not obtain information individually from each site of the
disaster.12
The paper helps to ¯nd out that whether the machine learning systems that are
handled and trained by humans can help in identifying pertinent images obtained
during the disasters through DL models. Hurricane Harvey that struck Texas, USA
during the period from August to September 2017 was studied by obtaining data
through the tweets on Twitter; 1% random tweets were taken through Twitter and
sample related to the hurricane was extracted. The Tweets were composed of 17 483
images out of which 1128 images were selected at random that were hand-coded by
two authors. For facilitating help and relief to the needy through the natural dis-
asters, images obtained through social media are useful but ¯nding genuine content is
tricky especially when it needs to be found whether image represents one of ¯nding
assistance. The model was composed of VGG-16 CNN and multilayer perceptron
layer and it classi¯ed each image on the basis of its time period and urgency. The
system helped in ¯nding relevant content from the irrelevant social media content.
Since only 1128 images were used, the results were much below than what was
expected even in the case of transfer learning.13
A Gated Recurrent Unit (GRU)-CNN has been used for detecting severity of
°oods. Videos either obtained from CCTV cameras or on social media have been
studied by applying DL to control the hazardous situation that may take place
because of the °oods. It was found that GRU-CNN model performed equivalent to a
Base-CNN model and was more inexpensive computationally than the base model.
The Base-CNN model proved to be ine®ective in mapping the temporal character-
istics of the °ood videos. The advantage of the system is that it helps to forecast and
predict the occurrence of °oods by generating an early warning. Attention
mechanisms and regularization techniques need to be added for improving the
overall e±ciency of the model.14
Twitter which is a blogging site is currently being used by people in a massive
manner. It is also used by people during the times of emergency and natural disaster.
Twitter is useful during disasters but there lies a challenge since the data need to
make sense and also there is unavailability of tools to analyze high-velocity and high-
volume data streams. In this paper, Arti¯cial Neural Network along with Natural
Language Processing has been used for analyzing tweets from Twitter three natural
disasters: Hurricane Harvey, Irma and Maria. Since this system also involved

2250040-11
S. Kaur et al.

humans along with automated methods, there was a need to include the processing
capabilities of humans in order to maintain higher throughput. In the future, this
system could lead to the development of fully automated systems for management of
disasters.15
A multimodal DL model was proposed, for identifying the damage caused by
natural disasters. The data were obtained through the social media posts. The
data were classi¯ed into ¯ve categories of damage including human casualties
and infrastructure damage. Inception model of CNN which was pre-trained and
processed the images, while word embedding model was used for processing text.
An accuracy of 92.62% was obtained. The model proved to be useful to identify
human casualties and damage which helped in allocating resources e±ciently and
saving human lives. Further, better multimodal models could be made for ¯ner and
better categorization.16
Social media images were obtained for four disasters 
 Typhoon Ruby/Hagupit,
Nepal Earthquake, Ecuador Earthquake and Hurricane Matthew  
 and deep CNNs
were employed to ¯nd the severity of damage due to these four disasters. It was found
that VGG16 performed better than BoVW (Bag-of-Visual-Words) and VGG16-fc7
that were chosen as baselines. The paper is one of the ¯rst paper to analyze the level
of damage occurred during disasters. The limitation includes the presence of signif-
icant human-labeling tasks.17
(b) Literature of Natural Disaster Detection Based on Images from
Satellite
A DL network was proposed which studied the di®erent levels of how susceptible
is a °ash °ood in Vietnam. The network was composed of 192 neurons which were
present in three layers. Sigmoid and ReLU were used as transfer and activation
functions, respectively. For the updation and optimization of weights of the DL
neural network, Adaptive Moment Estimation was employed. Database was estab-
lished which included factors such as soil type, rainfall, slope for training and vali-
dation of the model. The hybrid combination of Geographic Information System
(GIS) and DL model provided an accuracy of 92.05% which made this model a good
model for use in ¯nding about °ash °oods and further rescue help. This model
performed better than Support Vector Machine and Multilayer Perceptron Model
which were the benchmarks. The limitation included the non-generalizability to
other disasters.18
Hurricane intensity was estimated automatically by a deep CNN of infrared
images obtained from satellite. A low root mean square error of 8.82 knots was
obtained for the method adopted, Deep PHURIE, which was better than the method
employed previously. Previously adopted method, that is, PHURIE, gave a root
mean square error of 11.2 knots. The model was not only more accurate but also
there was no manual annotation. Further improvement in the system could be
brought about by the use of time-series data used for forecasting.19

2250040-12
A Review on Natural Disaster Detection in Social Media

A CNN was built for automatically detecting damage caused to buildings from
images obtained by satellites. Negative training examples were generated in large
numbers automatically through the data generation pipeline. Two tower subtract
gave the best results as compared to other models. It was shown that the generali-
zation of the model to disasters and new regions was possible if ¯ne-tuning was done
on small example set from those regions. This model's robustness to misalignment in
pre- and post-disaster images could be improved by the addition of translations in
the training images. The model lacked the robustness due to misalignment of pre-
disaster as well as post-disaster images.20
A CNN, modi¯ed U-Net architecture, was built for automatically detecting
damage caused to forests due to storms from images obtained by satellites. The area
of the forests covered by this method was 109 km2 and an accuracy of 92% was
obtained. Predictions on other datasets could be obtained within seconds after
successful training of the network. Since the study was conducted on forest area, if
larger areas get a®ected by storms, aerial surveys would become very di±cult to
conduct as well as time-consuming.21
Obtaining su±cient samples after the disaster is quite complex. To overcome this
problem, single shot multibox detector, a DL technique was developed to determine
damage caused to buildings. Augmentation and pre-training of data were performed.
A convolutional auto-encoder was used to determine the damage caused by Hurri-
cane Sandy in 2012. A VGG-16 model was used and trained by the aerial imagery
obtained after the disaster. mAp and mF1 were improved by 72% and 20%, re-
spectively. The experiment was also performed on dataset of Hurricane Irma and
good results were obtained. Further, this approach needs to be made more gener-
alized for other disasters.22
A CNN has been trained for up/down sampling of images obtained from satellites
and helps in classifying images of damaged buildings. Use of several resolutions along
with the several combinations of these resolutions helped in improving the accuracy
by nearly 4%. Feature maps were added and they helped to capture the most rele-
vant information from the satellite images. But this method could not help in
identifying damage signs which were comparatively smaller in size.23
A CNN was proposed that could ¯nd disaster areas that were a®ected in a severe
way using images obtained from satellites. The features that were extracted were
man-made like the buildings or roads before and after the disaster took place. It
helped to identify the damage caused by two natural disasters, that is, Santa Rosa
¯re and the Harvey Hurricane. F1 score of 81.2% and 83.5% was achieved for the
°ood and ¯re data, respectively. Since the focus was only on buildings and roads,
the study should be made more generalizable for other man-made and natural
structures.24
Assessment of damage after a disaster like hurricanes is of immense importance.
However, the process of evaluation is expensive, prone to errors and even slow. Pre-
processing of data needs to be done before the latest algorithms and methods can be
applied. This paper creates a benchmark dataset for buildings damaged by

2250040-13
S. Kaur et al.

hurricanes. The data were composed of damaged as well as non-damaged buildings


obtained after the hurricane Harvey through aerial and satellite imagery. The data
sources were from TOMNOD project and FEMA. It also shares the dataset publicly
for Greater Houston area after Hurricane Harvey in the year 2017.25
A deep CNN was designed for estimating the intensity of hurricane (tropical
cyclones during the years 1998–2012) using graphical processing unit. Several con-
volution and fully connected layers used regularization methods that made the ex-
traction of features from the images of hurricanes e®ective. The images of satellites
were used and improved accuracy and low root mean square error were obtained.
The estimation of intensity of a new sample could be done within seconds and
required less human e®ort. But, the limitation was that the dataset was not of good
quality and was composed of latitude–longitude coastlines and grids which served as
noise in the dataset.26
A U-net convolutional network which is a DL algorithm was proposed to study
the damage caused by 2011 Tohoku Earthquake-Tsunami using satellite imagery.
A comparison of U-net model was made with deep residual U-net. The U-net model
achieved an accuracy of 70.9%. If pre-processed datasets are available, then this
model will generate damage maps in a time period of 2–15 min. Also, this model could
be used for other disasters as well. The limitation of this paper was that of detection
of a collapsed building by one optical sensor.27
A semi-supervised DL model was built to ¯nd out damage caused by hurricane
Sandy in 2012 from aerial imagery. The approach was composed of three steps, that
is, segmentation, unsupervised training using convolutional auto-encoders and su-
pervised method through CNN. An accuracy of 88.3% was obtained for the semi-
supervised method adopted which was better by 9% of the accuracy obtained from
CNN trained from scratch. This method is useful when there are a few labeled
samples. This model was not generalizable to other disasters.28
DL has been employed to detect damage caused by wild¯res in the year 2017 in
California through satellite images. A CNN was designed that was used for semantic
segmentation of area subjected to ¯res and also mapped burned buildings. With a
small amount of ¯ne-tuning, the model could be generalized to other disasters.29
A DL model was used to classify facilities and objects from the IARPA Functional
Map of the World (fMoW) data into 63 classes. An ensemble of CNN and extraneural
networks was created and those were used to integrate metadata of satellite with the
imagery. The total accuracy was found to be 83%, F1 score was 0.797 and it suc-
cessfully classi¯ed data of 15 of the classes with an accuracy of 95% or more. This
system could predict facilities and objects of interest in the satellite imagery if
combined with a detecting component. Its limitation is that monitoring of satellite
images needs to be done for helping disaster response crew.30
A CNN has been used for automatic detection of landslide and °oods. Satellite
imagery was obtained from Google Earth Aerial Imagery for the countries Japan and
Thailand. Training data were composed of 50 000 patches for both °ood and land-
slide. Accuracy of 80%–90% was obtained for both the disasters. This method

2250040-14
A Review on Natural Disaster Detection in Social Media

automatically detected disaster regions and gave high accuracy. There was also
improvement in F1 score, precision and recall parameters. One of the characteristics
of the dataset was that they had same variation in color as they were taken all on a
sunny day but since this is not always the case, there would be variations in colors
and would be di±cult to process.31
A DL model was built to assess the damage to buildings using only post event
ground-level image data since the data prior to the damage may sometimes be not
available. Three di®erent neural networks have been utilized, that is, one model has
been used for pre-processing data and two networks for extraction of features.32
A deep auto-encoder with wavelet coe±cient (wacDAE) was employed to ¯nd out
about geological disasters especially landslides. Two-dimensional wavelet transfor-
mation was the preprocessing step and was followed by corrupting and denoising.
Finally, the system (wacDAE) was composed of two hidden layers and it was found
that wacDAE performed better than arti¯cial neural network and support vector
machine algorithms when the accuracy, precision and recall were calculated. The
images of the disaster were obtained from Google Earth and the analysis lacked the
veri¯cation on optical remote sensing data.33
A CNN was developed to analyze images obtained from satellites for two natural
disasters that occurred in Thailand and Japan, that is, °ood and landslide, respec-
tively. The network consisted of three convolutional layers. Each convolutional layer
was followed by a max-pooling layer. The network also consisted of two fully con-
nected layers. For both the disasters, an accuracy of 80%–90% was obtained but the
processing became di±cult when there was variation in colors of data since this study
focused only on monochrome data.34
A CNN model has been used to classify images obtained from satellites for
detecting damages in buildings and a comparison has been made with features that
are handcrafted. This method achieved an accuracy of 73% which was comparable to
74% obtained from handcrafted features but CNN model required very less time and
also less memory. Further, imagery obtained from public sources could be used for
detecting damage in buildings.35
DL models were used to determine the damage caused to buildings using remote
sensing images. Both two-dimensional and three-dimensional features were consid-
ered for the earthquake that took place in Sichuan. It was found that high precision
of 90% was obtained for BOW–HoG + color method.36
Multi-layer Neural Network was employed for tracking of movement of cyclones
as well as their intensity in an automatic manner. The interpretation was performed
of NOAA-AVHRR satellite images. For test images accounting to 98%, forecast
about the direction of cyclones was found to be correct. This showed that the model
is generalizable in nature. However, more high-resolution images could be used for
¯nding out more important information.37
Multilayer perceptron which is a common arti¯cial neural network along with
back propagation was proposed for determining a connection between hurricane
appearances and high energy particles °owing from the Sun. The occurrence of

2250040-15
S. Kaur et al.

Hurricanes Katia, Jose and Irma was researched in this paper. The study was per-
formed for the period between 28 September 2017 and 21 December 2017 and it was
predicted that hurricanes could appear 2–4 days before the solar wind breakout.38
In this paper, a study has been done on the assessment of damage caused due to
the Hurricane Dorian using aerial imagery and arti¯cial intelligence. Severity of
damage has been assessed using the stacked CNN model and an accuracy of 61% has
been achieved.39
Disasters induced due to climatic changes cause adverse e®ects on the environ-
ment as well as human beings. In this paper, the DL model has been used for
assessing the spatial and temporal prediction of the disasters. Database of °ood
disaster obtained from Ontario, Canada was used to predict the indices related to the
climatic changes. An ccuracy of around 96% was obtained for the prediction of the
°ood disaster.40
Assessment of damage caused due to hurricanes has been done in this paper to
help the emergency responders. Unsupervised Domain Adaptation has been used to
improve the accuracy along with a Duplex Alignment Network. Accuracy of 98.3%
has been achieved for 9600 images.41
Geographical Information System has been used for ¯nding the susceptance of
°oods which are one of the most dangerous of the natural disasters. Multilayer
Perceptron along with auto-encoder models has been used and an accuracy of 97%
has been achieved.42
xBD is one of the largest datasets that helps in the assessment of damage caused
to buildings and change detection. The dataset consists of 850 736 building anno-
tations that have been spread over 45 362 km2 of imagery. The dataset was composed
of satellite images obtained from 19 natural disasters. ResNet architecture was used
for the classi¯cation process, and precision and F1 scores of 87.70% and 66.31%,
respectively, were obtained from the no-damage class.43
This paper includes the study of assessing the damage caused to buildings after
the occurrence of natural disasters such as ¯re, °oods and hurricanes. The paper
provides two main contributions: the ¯rst being the generation of aerial videos and
the second contribution is the proposal of a new model namely MSNet. This model
achieved state-of-the-art results as compared to methods used previously for the
same dataset.44
Satellite images are an important source of information to assess the damages
caused the various disasters such as hurricane and earthquake. Manual inspection
proves to be impractical because of the enormous amount of data. In this paper,
semi-supervised learning has been used for detecting the damages using satellite
images.45
A segmentation model was proposed for extraction and monitoring of roads au-
tomatically from satellite images. ResNet 34 was used as encoder and U-Net was used
as decoder in the segmentation process.46
Segmentation approach was proposed for generating °ood maps of satellite
images. Multi-temporal, multisensory and multi-resolution images were fused in a

2250040-16
A Review on Natural Disaster Detection in Social Media

CNN. These techniques helped in determining and assessing damage accurately from
post-disaster images. Extremely accurate segmentation maps of the °ood maps were
produced.47
Table 2 summarizes the literature survey of natural disasters based on deep
learning techniques.
Table 2 summarizes the accuracies obtained for various DL techniques for both
social media and satellite imagery for natural disaster detection. Maximum accuracy
of 96.2% is obtained for landslide detection using deep auto-encoders with wavelet
coe±cient (wacDAE) method.

5.2. Literature of natural disaster detection based on machine learning


techniques
Machine learning is focused on developing computer programs that access data and
cause learning to take place by themselves. Machine learning is now widely being
used for disaster detection. The following section includes the literature survey of
natural disaster detection and estimation through both social media and satellite
imagery.
(a) Literature of Natural Disaster Detection Based on Images from Social
Media
Social media posts were assessed to determine the damage caused by natural
disasters. Semantic machine learning techniques were employed along with temporal
and spatial analysis. Earthquake footprints could be easily identi¯ed through this
method. Accuracy of 86.41% was obtained. This technique proved to be better than
previous research since the earlier approaches were keyword-based and did not take
the spatial dimension into account. A challenge that arises is that the models used
have been designed for text that is edited such as blogs and newspaper articles. But,
the data obtained from social media are noisy and irregular comprising of improper
punctuation and slang words.48
Naïve Bayes and Structured Support Vector Machine (SSVM) classi¯cation
machine learning algorithms were applied on tweets of Chennai °oods in 2015. The
machine learning algorithms were applied on 7500 tweets using Matlab software. Out
of 7500 tweets, 5000 Naïve Bayes achieved an accuracy of 78% whereas SSVM
obtained an accuracy of 91%.49
Tweedr, a Twitter tool, was used to ¯nd out relevant information during natural
disasters. It consists of three main models, classi¯cation, clustering and extraction.
Classi¯cation involved machine learning, that is, SVM and logistic regression, for
identifying tweets related to disasters. Clustering phase makes use of ¯lters for
merging the tweets related to each other. In the extraction phase, phrases and tokens
are extracted which report information relating to di®erent classes like damage and
casualties. Validation of tweets was done for 12 crises in the USA since 2006. Logistic
regression gave the best results with an accuracy of 86%. Future work could focus on
exploring con°icting labels like death and missing persons.50

2250040-17
Table 2. Summary of literature survey of natural disaster detection based on deep learning techniques.

Paper ID/year Type of disaster No. of images/videos Technique Accuracy Other performance parameters
39/2021 Hurricane 
 Stacked CNN 61% 

S. Kaur et al.

41/2021 Hurricane 9600 Duplex alignment network 98.3% 



42/2021 Flood Multilayer perceptron and 97% 

auto-encoder
11/2020 Flood 4800 VGG16 93% 

Inception 91% 

Res Net152 91% 

Dense Net 91% 

13/2019 Hurricane 1128 VGG 16 64.61% Validation loss-1.017%
18/2019 Flood 1464 Deep learning neural network 92.05% Positive predictive value-94.55%
Negative predictive value-89.55%
20/2019 Earthquake 75 468 Concatenated only model 
 AUC 0.8008
Channel model post image 
 0.8030
Twin-tower concatenate model 
 0.8120
Twin-tower substract model 
 0.8302
22/2019 Hurricane 5000 Single shot multibox detector 
 F1 score-51.95%

2250040-18
SSD-convolution auto encoder 
 Precision-39.01%
F1 score-56.025%
Precision-44.83%
43/2019 22 068 images and 850 736 ResNet 50 
 F1 score-66.31%
building polygons Precision-87.70%
14/2018 
 13 845 CNN-Multi resolution A 89.8% 

Multi resolution B 92.4% 

Multi resolution C 94.4% 

23/2018 Flood 749 Base CNN 70% 

CNN-GRU (gated recurrent 70% 

unit)
16/2018 Hurricane 16 754 Inception-v3 83.78% 

Inception-v4 83.98% 

VGG 16 82.19% 

Inception resnet-v2 83.61% 

25/2018 Hurricane 18 474 TOMNOD and FEMA data 
 

Table 2. (Continued )

Paper ID/year Type of disaster No. of images/videos Technique Accuracy Other performance parameters
26/2018 Hurricane 48 828 Deep CNN with ¯ve convolu- 77.85% F1 score-0.78
tional layer and three fully RMSE-11.05 kt
connected layers
27/2018 Earthquake 1600 U-net convolutional network 70.9% 

Deep residual U-net model 54.8% 

28/2018 Hurricane 700 Convolutional auto-encoders 88.3% 

29/2018 Wild¯re 1080 Convolutional neural network 
 F1 score 85.9%
17/2017 Hurricane 61 000 VGG-16 74% Precision- 73%
Recall-74%
F1 score-74%
32/2017 
 250 VGG 
 Damage level-0.97

2250040-19
33/2016 Landslide 1200 Deep auto-encoder with wave- 96.20% Precision-97.18%
let coe±cient (wacDAE) Recall-95.26%
35/2016 Earthquake 13 288 Bag of words-surf 
 Precision-73%
BoW-Gabor 
 Recall-71%
BoW-HoG+color 
 Precision-81%
Recall-79%
Precision-90%
Recall-88%
A Review on Natural Disaster Detection in Social Media
S. Kaur et al.

Machine learning and Wireless Sensor Network (WSN) were proposed for early
detection of disasters especially ¯res caused in residential areas. Event detecting
functions of WSNs helped in the detection of natural hazards. For the detection of
the events, the machine learning classi¯er was decision trees due to their simplicity
and ful¯llment of resource limitations imposed by WSNs.51
Machine learning was used to extract important data from microblogging posts
obtained from sites such as Twitter during the times of disasters. The model was
tested on thousands of messages obtained from microblogging sites that were
obtained for a real disaster. It was found that machine learning gave a hit ratio of
close to one in the case of source and number of casualty extractor. This indicates
that most of the good information could be extracted from human extracted data.52
Semi-supervised machine learning was used to extract awareness information
from the social media posts during natural disasters. The algorithm classi¯ed the
tweets successfully with an accuracy of 90% on an average. The dataset included
data obtained during Chennai °oods in 2015.53
(b) Literature of Natural Disaster Detection Based on Images from
Satellite
Computer vision and Machine learning have been applied on satellite images of
wild¯re that took place in California in 2017. These were done to improve the time
e±ciency of handling the disaster situation and also to ¯nd out the buildings that
were damaged. This system made semantic segmentation of regions a®ected by
massive ¯res and helped to map the burned buildings. The system could also be
applied to other disasters. This system could also be used for instance segmentation
which helps to distinguish and count the di®erent objects.54
Machine learning algorithms were used to ¯nd the damage occurred during the
2011 Japan earthquake and Christchurch earthquake. Satellite images were used;
1500 data regions were used and an accuracy of 90.4% was obtained by the Random
Forest classi¯er. Six damage levels were considered which ranged from one to six with
one representing no damage and six representing maximum damage. Since the work
was done on classifying into binary classes, an improvement could be obtained using
multiclass classi¯cation for assessing the damage of disaster-struck regions based
only on the imagery obtained after the disaster.55
Data fusion techniques comprising of machine learning and statistical techniques
were proposed for disaster management through data obtained through satellites
RADARSAT-2. Fusion was done of Probabilistic Neural Network and Support
Vector Machine. This approach was composed of feature-level and pixel-level data
along with machine learning techniques. It was found that fusion techniques out-
performed the non-fusion methods and an accuracy of 82.8% and 72% was achieved
for levee landslide detection and tornado detection. Machine learning techniques
provided better binary classi¯cation as compared to linear techniques like maximum
likelihood. Further investigation could be done on whether vegetation has any cor-
relation with the disasters.56

2250040-20
A Review on Natural Disaster Detection in Social Media

Machine learning techniques were employed to detect °oods using optical satellite
images. The satellite images of 2010 Ljubljana Moor °oods were taken into consid-
eration. Machine learning exposed the features and threshold values de¯ning boundary
between un°ooded and °ooded cases easily which reduces the time to determine °oods.
J48 algorithm provided the best results. This algorithm reduced the time of deter-
mining °oods and provided accurate maps immediately essential for disaster man-
agement. An improvement was required during the urban areas classi¯cation.57
Machine learning algorithms were applied on satellite images to determine the
damage caused by °ash °ood in 2007. It was found that J48 machine learning al-
gorithm gave the highest accuracy of 88% for the training set and 95% for the testing
set. J48 algorithm proved to be an optimal algorithm to perform binary classi¯ca-
tion. Other methods such as ensemble techniques were also used but they did not
provide a better classi¯cation. The aspects that need to be addressed are the prob-
lematic areas where there is di±culty in distinguishing between °ooded and not
°ooded regions such as ¯elds and built-up areas.58
Genetic algorithm was proposed to generate a map for damage through textural
analysis of pre- and post-event satellite images. The conditions were measured of
1133 buildings after their analysis. An accuracy of 74.4% and kappa coe±cient of
0.63 were obtained.59
Three methods that are machine learning, object-based classi¯cation and pixel-
based water delineation were applied to detect the °ooded areas in Slovenia. Medium
resolution radar images that are RADARSAT-2 and ENVISAT were used. Neither
of the methods used gave good performance overall. The three methods need to be
selected according to a particular situation. Pixel-based method came out to be the
fastest while object-based method gave the best de¯ned objects. Machine learning
algorithms can also be applied to data other than satellites. But, de¯ciencies were
found in images obtained from radar to detect areas of water in various landscapes.60
Floods cause immense destruction to human beings and property. Timely evac-
uation is very essential. In this paper, a WSN system has been used for detecting
°oods. This is done through observation of changes in weather. A LoRaWAN net-
work was used and implementation was done in Arduino and Raspberry Pi. Support
Vector Machine algorithm was used for classifying into °ood or no °ood category
with an accuracy of 98%.61
Table 3 summarizes the literature survey of natural disasters based on machine
learning techniques.
Table 3 summarizes the accuracies obtained for various machine learning techniques
for both social media and satellite imagery for natural disaster detection. Maximum
accuracy of 93% is obtained for wild¯re detection using Decision Trees algorithm.

6. Discussion
Social media has been emerging as an important method of disseminating informa-
tion and communication during disaster events. Data in the form of text or images

2250040-21
S. Kaur et al.

Table 3. Summary of literature survey of natural disaster detection based on machine learning techniques.

Paper ID/year Type of disaster No. of images/videos Technique Accuracy Other performance parameters
54/2018 Wild¯re 1080 Semantic segmentation 
 F1 score (burned buildings)-85.9%
F1 score (unburned buildings)-81.8%
55/2017 Earthquake 1500 Decision Tree 83.44% F1 score-87.75%
Naive Bayes Support Vector Machine 84.77% F1 score-87.22%
Random Forest Voting Classi¯er 66.23% F1 score-79.68%
90.07% F1 score-92.23%
89.40% F1 score-91.53%
56/2017 Landslide 7139 Support Vector Machine 82.25% 

49/2016 Flood 7500 Naïve Bayes 78.51% Error rate-21.49 %
Recall-81%
Precision-77%
SSVM 91% Error rate-8.99%

2250040-22
Recall-95%
Precision-91%
57/2013 Flood 201 J48 89% 

JRip 88% 

Random Forest 82% 

58/2013 Flood 255 °ood images JRip 85% 

Bagging J48 92% 

Random Forest C45 92% 

53/2010 Wild¯re 1400 Decision trees 93% 

59/2008 Earthquake 1133 Genetic algorithm 74.4% Kappa coe±cient 0.63
A Review on Natural Disaster Detection in Social Media

which have been captured and also shared by social-media users are usually avail-
able. But, whether the data are relevant and authentic still remains a big challenge.
Selecting relevant data from the enormous amount of unstructured data obtained
from social media is quite di±cult. The uploaded data may include images of car-
toons, posters, text, advertisements and modi¯ed photographs.62 Also, social media
lacks public datasets and uses only self-collected datasets for example Twitter posts.
Recently, satellite imagery and deep and machine learning are being used to
locate areas which have been a®ected by a natural disaster. This is done by keeping
man-made things as reference, for example buildings. The change in these man-made
features before and after the disaster helps to determine areas where the relief e®orts
should be concentrated. The models based on CNN are trained for detecting man-
made features such as buildings and roads from satellite images.63 Satellite imagery is
becoming more prevalent for disaster detection because the images can be captured
and processed in a short span of time. Satellite imagery is found to be useful for
recognizing the pattern of damage caused by the disaster.62 But, the satellite images
may also be obscured by clouds during the natural disasters.
DL is widely used in speech recognition, natural language processing and com-
puter vision. It is particularly used in areas where large number of images and text
have to be analyzed and intelligence similar to humans is involved.64,65 DL use in
disaster detection is now getting widely used.66
Supervised machine learning algorithms67 are used for actuating rescue e®orts and
disaster response.68 Machine learning algorithms69 can be used for extracting useful
information from noisy satellite imagery and social media data.70
Since both the social media images and satellite imagery have their own advan-
tages and disadvantages, images from both these sources could be used for
achievement of better results in terms of accuracy, precision, recall and F1 score.

7. Conclusion
It is imperative to ¯nd out the natural disaster impact and the damage extent
accurately and on timely basis for the disaster management to be successful. Tra-
ditionally, the damage assessment was done through ground survey but that process
was time-consuming and also involved a large workforce. Data in the form of images
and text obtained through social media are easily available but the relevance and
authenticity of the data is a great challenge. The data may consist of cartoons,
advertisements, posters, text, etc. So, ¯nding out the useful data from the abundant
data available is quite di±cult. Satellite imagery is gaining importance for disaster
detection since they provide coverage of large spatial and temporal areas. But, the
satellite images may also be covered with clouds during the disaster. Since human
error is a possibility, both social media posts and satellite imagery require the use of
automated methods. Hence, DL and machine learning come into picture.71,72 This
paper provides a survey of use of DL and machine learning for disaster prediction,
estimation and detection through data obtained through social media and satellites.

2250040-23
S. Kaur et al.

References
1. Y. Pi, N. D. Nath and A. H. Behzadan, \Convolutional neural networks for object de-
tection in aerial imagery for disaster response and recovery," Adv. Eng. Inform. 43,
101009 (2020).

2. D. T. Bui, N. D. Hoang, F. Martínez-Alvarez, P. T. T. Ngo, P. V. Hoa, T. D. Pham, P.
Samui and R. Costache, \A novel deep learning neural network approach for predicting
°ash °ood susceptibility: A case study at a high frequency tropical storm area," Sci. Total
Environ. 701, 134413 (2020).
3. K. Ahmad, M. Riegler, K. Pogorelov, N. Conci, P. Halvorsen and F. De Natale, \JORD:
A system for collecting information and monitoring natural disasters by linking social
media with satellite imagery," in Proc. 15th Int. Workshop on Content-Based Multimedia
Indexing (2017), pp. 1–6.
4. R. Kovordanyi and C. Roy, \Cyclone track forecasting based on satellite images using
arti¯cial neural networks," ISPRS J. Photogramm. Remote Sens. 64(1), 513–521 (2009).
5. M. Dawood and A. Asif, \Deep-PHURIE: Deep learning based hurricane intensity esti-
mation from infrared satellite imagery," Neural Comput. Appl. 32(13), 1–9 (2019).
6. Q. D. Cao and Y. Choe, \Building damage annotation on post-hurricane satellite imagery
based on convolutional neural networks," Nat. Hazards 103(3), 1–20 (2020).
7. J. Doshi, S. Basu and G. Pang, From satellite imagery to disaster insights, preprint
(2018), arXiv:1812.07033.
8. M. Pritt and G. Chern, \Satellite image classi¯cation with deep learning," in 2017 IEEE
Applied Imagery Pattern Recognition Workshop (AIPR) (IEEE, 2017), pp. 1–7.
9. T. J. Brinker, A. Hekler, J. S. Utikal, N. Grabe, D. Schadendorf, J. Klode, C. Berking, T.
Steeb, A. H. Enk and C. von Kalle, \Skin cancer classi¯cation using convolutional neural
networks: Systematic review," J. Med. Int. Res. 20(10), e11936 (2018).
10. S. A. Chen, A. Escay, C. Haberland, T. Schneider, V. Staneva and Y. Choe, \Benchmark
dataset for automatic damaged building detection from post-hurricane remotely sensed
imagery," preprint (2018), arXiv:1812.05581.
11. H. Ning, Z. Li and M. E. Hodgson, \Prototyping a social media °ooding photo screening
system based on deep learning," ISPRS Int. J. Geo-Inform. 9(2), 104 (2020).
12. M. A. Sit, C. Koylu and I. Demir, \Identifying disaster-related tweets and their semantic,
spatial and temporal context using deep learning, natural language processing and spatial
analysis: A case study of Hurricane Irma," Int. J. Digital Earth 12(11), 1205–1229 (2019).
13. B. W. Robertson, M. Johnson, D. Murthy, W. R. Smith and K. K. Stephens, \Using a
combination of human insights and `deep learning' for real-time disaster communication,"
Prog. Disaster Sci. 2, 100030 (2019).
14. K. Lohumi and S. Roy, \Automatic detection of °ood severity level from °ood videos
using deep learning models," in 2018 5th Int. Conf. Information and Communication
Technologies for Disaster Management (ICT-DM) (IEEE, 2018), pp. 1–7.
15. F. Alam, F. O°i, M. Imran and M. Aupetit, \A Twitter tale of three hurricanes: Harvey,
Irma, and Maria," preprint (2018), arXiv:1805.05144.
16. H. Mouzannar, Y. Rizk and M. Awad, \Damage identi¯cation in social media posts using
multimodal deep learning," in ISCRAM (Rochester, NY, USA, 2018).
17. D. T. Nguyen, F. O°i, M. Imran and P. Mitra, \Damage assessment from social media
imagery data during disasters," in Proc. 2017 IEEE/ACM Int. Conf. Advances in Social
Networks Analysis and Mining 2017 (Sydney, Australia, 2017), pp. 569–576.

18. D. T. Bui, N. D. Hoang, F. Martínez-Alvarez, P. T. T. Ngo, P. V. Hoa, T. D. Pham, P.
Samui and R. Costache, \A novel deep learning neural network approach for predicting
°ash °ood susceptibility: A case study at a high frequency tropical storm area," Sci. Total
Environ. 701, 134413 (2020).

2250040-24
A Review on Natural Disaster Detection in Social Media


19. D. T. Bui, N. D. Hoang, F. Martínez-Alvarez, P. T. T. Ngo, P. V. Hoa, T. D. Pham, P.
Samui and R. Costache, \A novel deep learning neural network approach for predicting
°ash °ood susceptibility: A case study at a high frequency tropical storm area," Sci. Total
Environ. 701, 134413 (2020).
20. J. Z. Xu, W. Lu, Z. Li, P. Khaitan and V. Zaytseva, \Building damage detection in
satellite imagery using convolutional neural networks," preprint (2019),
arXiv:1910.06444.
21. Z. M. Hamdi, M. Brandmeier and C. Straub, \Forest damage assessment using deep
learning on high resolution remote sensing data," Remote Sens. 11(17), 1976 (2019).
22. Y. Li, W. Hu, H. Dong and X. Zhang, \Building damage detection from post-event aerial
imagery using single shot multibox detector," Appl. Sci. 9(1), 1128 (2019).
23. D. Duarte, F. Nex, N. Kerle and G. Vosselman, \Satellite image classi¯cation of building
damages using airborne and satellite image samples in a deep learning approach," ISPRS
Ann. Photogramm. Remote Sens. Spat. Inform. Sci. 4(2), (2018).
24. J. Doshi, S. Basu and G. Pang, \From satellite imagery to disaster insights," preprint
(2018), arXiv:1812.07033.
25. S. A. Chen, A. Escay, C. Haberland, T. Schneider, V. Staneva and Y. Choe, \Benchmark
dataset for automatic damaged building detection from post-hurricane remotely sensed
imagery," preprint (2018), arXiv:1812.05581.
26. R. Pradhan, R. S. Aygun, M. Maskey, R. Ramachandran and D. J. Cecil, \Tropical
cyclone intensity estimation using a deep convolutional neural network," IEEE Trans.
Image Process. 27(2), 692–702 (2017).
27. Y. Bai, E. Mas and S. Koshimura, \Towards operational satellite-based damage-mapping
using U-net convolutional network: A case study of 2011 Tohoku earthquake-tsunami,"
Remote Sens. 10(10), 1626 (2017).
28. Y. Li, S. Ye and I. Bartoli, \Semisupervised classi¯cation of hurricane damage from
postevent aerial imagery using deep learning," J. Appl. Remote Sens. 12(4), 045008 (2018).
29. G. Novikov, A. Trekin, G. Potapov, V. Ignatiev and E. Burnaev, \Satellite imagery
analysis for operational damage assessment in emergency situations," in Int. Conf.
Business Information Systems (Springer, Cham, 2018), pp. 347–358.
30. M. Pritt and G. Chern, \Satellite image classi¯cation with deep learning," in 2017 IEEE
Applied Imagery Pattern Recognition Workshop (AIPR) (IEEE, 2017), pp. 1–7.
31. S. N. K. B. Amit and Y. Aoki, \Disaster detection from aerial imagery with convolutional
neural network," in 2017 Int. Electronics Symp. Knowledge Creation and Intelligent
Computing (IES-KCIC) (IEEE, 2017), pp. 239–245.
32. F. Alam, M. Imran and F. O°i, \Image4act: Online social media image processing for
disaster response," in Proc. 2017 IEEE/ACM Int. Conf. Advances in Social Networks
Analysis and Mining 2017 (2017), pp. 601–604.
33. Y. Liu and L. Wu, \Geological disaster recognition on optical remote sensing images using
deep learning," Proc. Comput. Sci. 91, 566–575 (2016).
34. S. N. K. B. Amit, S. Shiraishi, T. Inoshita and Y. Aoki, \Analysis of satellite images for
disaster detection," in 2016 IEEE Int. Geoscience and Remote Sensing Symp. (IGARSS)
(IEEE, 2016), pp. 5189–5192.
35. M. Dawood and A. Asif, \Deep-PHURIE: Deep learning based hurricane intensity esti-
mation from infrared satellite imagery," Neural Comput. Appl. 1–9 (2019).
36. R. Kovordanyi and C. Roy, \Cyclone track forecasting based on satellite images using
arti¯cial neural networks," ISPRS J. Photogramm. Remote Sens. 64(1), 513–521 (2009).
37. B. Gokaraju, R. A. Nobrega, D. A. Doss, A. C. Turlapaty and R. C. Tesiero, \Data fusion
of multi-source satellite data sets for cost-e®ective disaster management studies," in
SoutheastCon 2017 (IEEE, 2017), pp. 1–5.

2250040-25
S. Kaur et al.

38. B. Resch, F. Usländer and C. Havas, \Combining machine-learning topic models and
spatiotemporal analysis of social media data for disaster footprint and damage
assessment," Cartogr. Geogr. Inf. Sci. 45(4), 362–376 (2018).
39. C. S. Cheng, A. H. Behzadan and A. Noshadravan, \Deep learning for post-hurricane
aerial damage assessment of buildings," Computer-Aided Civil Infrastruct. Eng. 36(6),
(2021).
40. M. Haggag, A. S. Siam, W. El-Dakhakhni, P. Coulibaly and E. Hassini, \A deep learning
model for predicting climate-induced disasters," Nat. Hazards 107, 1–26 (2021).
41. C. Lin, Y. Li, Y. Liu, X. Wang and S. Geng, \Building damage assessment from post-
hurricane imageries using unsupervised domain adaptation with enhanced feature dis-
crimination," IEEE Trans. Geosci. Remote Sens. (2021).
42. M. Ahmadlou, A. K. Al-Fugara, A. R. Al-Shabeeb, A. Arora, R. Al-Adamat, Q. B. Pham,
N. Al-Ansari, N. T. T. Linh and H. Sajedi, \Flood susceptibility mapping and assessment
using a novel deep learning model combining multilayer perceptron and autoencoder
neural networks," J. Flood Risk Manage. 14(1), e12683 (2021).
43. R. Gupta, R. Hosfelt, S. Sajeev, N. Patel, B. Goodman, J. Doshi, E. Heim, H. Choset and
M. Gaston, \xBD: A dataset for assessing building damage from satellite imagery,"
preprint (2019), arXiv:1911.09296.
44. X. Zhu, J. Liang and A. Hauptmann, \MSNet: A multilevel instance segmentation net-
work for natural disaster damage assessment in aerial videos," in Proc. IEEE/CVF
Winter Conf. Applications of Computer Vision (2021), pp. 2023–2032.
45. L. Gueguen and R. Hamid, \Large-scale damage detection using satellite imagery," in
Proc. IEEE Conf. Computer Vision and Pattern Recognition (2015), pp. 1321–1328.
46. A. Buslaev, S. Seferbekov, V. Iglovikov and A. Shvets, \Fully convolutional network for
automatic road extraction from satellite imagery," in Proc. IEEE Conf. Computer Vision
and Pattern Recognition Workshops (2018), pp. 207–210.
47. T. G. Rudner, M. Rußwurm, J. Fil, R. Pelich, B. Bischke, V. Kopačkov a and P. Biliński,
\Multi3Net: Segmenting °ooded buildings via fusion of multiresolution, multisensor, and
multitemporal satellite imagery," Proc. AAAI Conf. Artif.l Intell. 33(1), 702–709 (2019).
48. B. Anbalagan and C. Valliyammai, \#ChennaiFloods: Leveraging human and machine
learning for crisis mapping during disasters using social media," in 2016 IEEE 23rd Int.
Conf. High Performance Computing Workshops (HiPCW) (IEEE, 2016), pp. 50–59.
49. N. Pandey and S. Natarajan, \How social media can contribute during disaster events?
Case study of Chennai °oods 2015," in 2016 Int. Conf. Advances in Computing Com-
munications and Informatics (ICACCI) (IEEE, 2016), pp. 1352–1356.
50. Z. Ashktorab, C. Brown, M. Nandi and A. Culotta, \Tweedr: Mining twitter to inform
disaster response," in ISCRAM (2014), pp. 269–272.
51. M. Bahrepour, N. Meratnia, M. Poel, Z. Taghikhaki and P. J. Havinga, \Distributed
event detection in wireless sensor networks for disaster management," in 2010 Int. Conf.
Intelligent Networking and Collaborative Systems (IEEE, 2010), pp. 507–512.
52. M. Imran, S. Elbassuoni, C. Castillo, F. Diaz and P. Meier, \Extracting information
nuggets from disaster-related messages in social media," in ISCRAM (2013).
53. G. Novikov, A. Trekin, G. Potapov, V. Ignatiev and E. Burnaev, \Satellite imagery
analysis for operational damage assessment in emergency situations," in Int. Conf.
Business Information Systems (Springer, Cham, 2018), pp. 347–358.
54. F. Samadzadegan, H. Rastiveisi and W. G. Viii, \Automatic detection and classi¯cation
of damaged buildings, using high resolution satellite imagery and vector data," Int.
Archiv. Photogramm. Remote Sens. Spatial Inform. Sci. 37(Part B8), (2008) 415–420.

2250040-26
A Review on Natural Disaster Detection in Social Media

55. K. R. Nia and G. Mori, \Building damage assessment using deep learning and ground-
level image data," in 2017 14th Conf. Computer and Robot Vision (CRV) (IEEE, 2017),
pp. 95–102.
56. L. Peter, M. Matjaž and O. Krištof, \Detection of °ooded areas using machine learning
techniques: Case study of the Ljubljana moor °oods in 2010," Disaster Adv. 6(2), 4–11
(2013).
57. P. Lamovec, T. Velkanovski, M. Mikos and K. Osir, \Detecting °ooded areas with ma-
chine learning techniques: Case study of the Selška Sora river °ash °ood in September
2007," J. Appl. Remote Sens. 7(1), 073564 (2013).
58. M. Bahrepour, N. Meratnia, M. Poel, Z. Taghikhaki and P. J. Havinga, \Distributed
event detection in wireless sensor networks for disaster management," in 2010 Int. Conf.
Intelligent Networking and Collaborative Systems (IEEE, 2010), pp. 507–512.
59. B. Gokaraju, R. A. Nobrega, D. A. Doss, A. C. Turlapaty and R. C. Tesiero, \Data fusion
of multi-source satellite data sets for cost-e®ective disaster management studies," in
SoutheastCon 2017 (IEEE, 2017), pp. 1–5.
60. T. Veljanovski, P. Lamovec, P. Pehani and K. Oštir, \Comparison of three techniques for
detection of °ooded areas on ENVISAT and RADARSAT-2 satellite images," in V:
Geoinformation for Disaster Management, Gi4DM 2011 (2011).
61. J. Al Qundus, K. Dabbour, S. Gupta, R. Meissonier and A. Paschke, \Wireless sensor
network for AI-based °ood disaster detection," Ann. Oper. Res. 8, 1–23 (2020).
62. S. A. Chen, A. Escay, C. Haberland, T. Schneider, V. Staneva and Y. Choe, \Benchmark
dataset for automatic damaged building detection from post-hurricane remotely sensed
imagery," preprint (2018), arXiv:1812.05581.
63. J. Doshi, S. Basu and G. Pang, \From satellite imagery to disaster insights," preprint
(2018), arXiv:1812.07033.
64. S. Kaur, S. Gupta, S. Singh and I. Gupta, \Detection of Alzheimer's disease using deep
convolutional neural network," Int. J. Image Graph. 21(1), 2140012 (2021).
65. K. M. Sagayam, A. D. Andrushia, A. Ghosh, O. Deperlioglu and A. A. Elngar,
\Recognition of hand gesture image using deep convolutional neural network," Int. J.
Image Graph. 21(1), 2140008 (2021).
66. Y. Liu and L. Wu, \Geological disaster recognition on optical remote sensing images using
deep learning," Proc. Comput. Sci. 91, 566–575 (2016).
67. S. B. Kotsiantis, I. D. Zaharakis and P. E. Pintelas, \Machine learning: A review of
classi¯cation and combining techniques," Artif. Intell. Rev. 26(3), 159–190 (2006).
68. M. Bahrepour, N. Meratnia, M. Poel, Z. Taghikhaki and P. J. Havinga, \Distributed
event detection in wireless sensor networks for disaster management," in 2010 Int. Conf.
Intelligent Networking and Collaborative Systems (IEEE, 2010), pp. 507–512.
69. M. Verma, D. Kumar and T. Arora, \A review of machine learning techniques for feature
based classi¯cation of autism spectrum disorder" Adv. Appl. Math. Sci. 19(5), 387–396
(2020).
70. F. Samadzadegan, H. Rastiveisi and W. G. Viii, \Automatic detection and classi¯cation
of damaged buildings, using high resolution satellite imagery and vector data," Int.
Archiv. Photogramm. Remote Sens. Spatial Inform. Sci. 37(Part B8), 415–420 (2008).
71. I. Gupta and S. Gupta, \Breast cancer detection using deep learning and machine
learning techniques: A review," JARDCS 11, 1487–1498.
72. T. Arora, \Classi¯cation of human metaspread images using convolutional neural
networks," Int. J. Image Graph 21(3), 2150033 (2021).

2250040-27
S. Kaur et al.

Swapandeep Kaur has been working as an Assistant Professor


at Chitkara University since 2016. She did her M.Tech in Elec-
tronics and Communication Engineering with specialization in
Optical Engineering in 2016. She did in BE in Electronics and
Communication Engineering in 2012. Currently, she is pursuing
her Ph.D. in Machine Learning and Deep Learning. She has
published seven papers.

Sheifali Gupta is a Professor in the Chitkara University Research


and Innovation Network (CURIN) at Chitkara University, Punjab
Campus, India. She specializes in the area of Digital Image Pro-
cessing, Pattern Recognition, Machine Intelligence, Biomedical
Image Processing, agriculture-based image processing and deep
learning. She has published more than 60 research papers and
articles in reputed national and international journals and con-
ferences. She has ¯led 15 patents in the ¯eld of image processing
and mechatronics. She has conducted di®erent workshops based on Modeling Real-
Time Video Processing System in Simulink using the video and image processing
blockset. She has received a prestigious IRDP Award 2018 for remarkable achieve-
ments in Teaching, Research and Publications. Presently, she is working on an
agriculture-based project in collaboration with Dr. Y S Parmar University of Hor-
ticulture and Forestry Nauni, Solan. She is presently guiding postgraduate and
doctoral students.

Swati Singh is an Assistant Professor at the University Institute


of Technology, Himachal Pradesh University, Shimla, Himachal
Pradesh, India. She has done her B.Tech, M.Tech and Ph.D. in
Electronics and Communication Engineering. Her specialization
includes Image Processing, Machine Learning, Deep Learning,
Networking and Digital Electronics. She has 8 years of teaching
experience and has published more than 15 papers and ¯led 3
patents.

Tanvi Arora received her B.Tech degree in Computer Science


and Engineering from Punjab Technical University, Jalandhar,
Punjab, India in 2002, M.Tech degree in Information Technology
from Punjab Technical University, Jalandhar, Punjab, India in
the year 2007 and completed her Ph.D. in Computer Science &
Engineering from Dr. B. R. Ambedkar National Institute of
Technology, Jalandhar, Punjab, India in 2018. Her teaching and
research interests include Image Processing, Pattern Recognition,
Machine Learning, Data Mining and Network Security.

2250040-28

You might also like