Review on Disaster Defect Management in

Machine Learning
Dr J T THIRUKRISHNA1, M R SUPRIYA 1 , MAYAKUNTLA ROHINI 1 ,
NISHMITHA 2, NITHYA SHREE M 2
1
Department of Information Science and Engineering, Dayananda Sagar Academy of Technology and Management, Bangalore 560082, India
2
Department of Information Science and Engineering, Dayananda Sagar Academy of Technology and Management, Bangalore 560082, India
Corresponding author: M R SUPRIYA([email protected])

ABSTRACT - Internet data performs a crucial part in calamity response during disaster events. The disaster
might be natural or man-made, Internet community have proven to be the most productive transmission and
combine among afflicted communities. Research has manifested utilizing methods for machine learning to
recognize the applicable posts on web,net and other sources for disaster reciprocation. Communication
breakdowns and association failures are common issues during disaster response, leading to delayed or
inefficient aid distribution. In this paper, we are going to evaluate the effect caused due to calamities. This
research is categorized into two proportions: Response after the disaster, identifying the damage caused, and
allowing for faster response. The findings of this study may help experimenters in the future identify relevant
areas for disaster assessment.

Keywords: Disaster Management, Machine Learning, Internet Data.

I. INTRODUCTION during emergencies. These updates encapsulate
diverse information, spanning from details on
Every year, millions of individuals worldwide fatalities and The accessibility of resources to the
confront the far-reaching consequences of natural pressing requirements of affected and injured
and man-made disasters, resulting in profound individuals. Following a disaster witnesses a surge
implications for their lives. Tragically, these events in user-generated content on internet community,
often lead to the loss of human lives, alongside contributing valuable insights into the prevailing
substantial damage to infrastructure and properties. situation, In essence, forums and Internet data not
The orchestration of handling disasters operations only serve as dynamic repositories of real-time
becomes crucial in the phases preceding, during, information both during and following emergencies
and following such calamities, with the overarching but also as conduits for the global outpouring of
objectives of averting human casualties, emotional responses and support from individuals
safeguarding both individuals and infrastructure, across the digital landscape. During natural and
mitigating economic repercussions, and restoring a human-made calamities, a substantial influx of
sense of normalcy. The intricate nature of disasters, content floods microblogging services,
coupled with the critical and intricate aspects encompassing a variety of data formats such as
response to disasters efforts, necessitate resilient audio, photos, text, and videos. Notably, platforms
decision-making processes, further fortified by the like Twitter have emerged as prominent channels
integration of information technology. The for the exchange of news and updates. Twitter,
imperative for effective and well-informed disaster recognized as a pivotal microblogging site during
management has become increasingly evident, crises, facilitates the transmission of copious
given the scale and repercussions of such events. In information through tweets. Whether informative or
recent times, notable strides in machine learning not, chirps in the midst of emergencies consistently
(ML) have been instrumental in enhancing the feature disaster-related terminology. During periods
capabilities of disaster management. Various types of crisis, valuable tweets contribute specifics crucial
of disasters, ranging from hurricanes, earthquakes, for victims, humanitarian groups, and responders,
floods, and wildfires, to landslides, necessitate covering aspects such as the needs of affected
adept management strategies. Furthermore, the individuals, casualties, infrastructure damage, and
application of advanced technologies, particularly resource availability. Twitter's unique feature,
in the realm of AI, has proven beneficial in allowing users to follow others without mutual
addressing the complexity and scale of these consent, extends its utility to non-registered
challenges. In the last ten years, a great deal of individuals, enabling them to access updates. This
focus has been placed on the use of web platforms inclusive approach empowers community members
as vital channels for providing updates in real time to monitor critical data in real-time, distinguishing
Twitter as an unparalleled platform that aggregates pertinent information during a catastrophe a
and maintains a substantial repository of potentially challenging task. This study delves into the
life-saving disaster-related information. In contrast, categorization of damage assessment tweets,
non-informative tweets lack substantive details addressing both human
about the disaster, making the identification of

and infrastructure damage, as a binary and multi- preparedness, response, and recovery, but lacks a
class classification challenge. Recognizing the systematic examination of the defects and
significance of diverse categorical damage shortcomings within
information in disaster scenarios, we introduce an
innovative approach. This approach involves the these frameworks. This research seeks to address
development of weighted features through the this gap by conducting a thorough analysis of
application of linear regression and SVR techniques current practices in disaster defect management,
for both binary and multi-class classification. These encompassing the identification, analysis, and
features incorporate low-level lexical and syntactic remediation of shortcomings when responding to
attributes, along with the inclusion of the most disasters systems.
frequently used words. The experimentation phase
involves testing various classifiers to determine the III. METHODOLOGY
most suitable one for the proposed features. The In disaster defect management, Convolutional
chosen classifier is then applied to different disaster Neural Networks (CNNs) are often employed for
datasets, and its performance is evaluated using image analysis and damage assessment. CNNs are
diverse parameters. The key contributions of this particularly effective at capturing spatial hierarchies
study can be outlined as follows: and patterns within visual data, making them well-
suited for duties like identifying structural defects
 We present a groundbreaking method that
or damages in the following a disaster. Gather
leverages low-level lexical and syntactic
relevant image or video data related to the disaster-
features, along with the incorporation of
affected areas. This can include footage, or
top-frequency words weighted using SVR
photographs of internet data before, as well as
and linear regression algorithms. Notably,
following the disaster. Preprocess the data by
our proposed method is vocabulary-
resizing images, normalizing pixel values, and
independent, ensuring accurate tweet
augmenting the dataset to enhance the model's
identification even when the training data
ability to generalize. Divide the dataset into sets for
for the model comes from distinct disaster
validation and training The model is trained to
datasets.
identify patterns and characteristics linked to flaws
 A comparative analysis is conducted with or damage using the training set. In order to reduce
various state-of-the-art methods across the discrepancy between expected and actual
different datasets.. It excels in identifying results, the CNN learns to modify its internal
damage assessment during disasters, parameters (weights and biases) during
showcasing superior performance in both training.Train the CNN to detect defects or damage
binary and multi-class classification in images. The network learns to identify visual
scenarios for both in-domain and cross- patterns indicative of structural issues, such as
domain contexts. cracks, deformations, or other visible damage. The
CNN could be fine-tuned to classify the severity of
II. PROBLEM DEFINITION defects or to locate specific types of damage within
images. Combine the CNN with Geographic
The proliferation of disasters has been increasing Information System (GIS) data to enhance spatial
day by day, there exists a necessity to manage the analysis. This integration allows for mapping the
effects caused since, the calamities that occur. detected defects onto geographical locations and
Natural disasters pose significant threats to assessing the overall influence on the affected area.
communities and infrastructure, often resulting in Deploy the trained CNN for real-time monitoring of
widespread devastation and the need for prompt and disaster-affected regions. After the disaster, use the
effective response strategies. In the realm of CNN to analyze historical data alongside new data
disaster management, the identification and to evaluate the evolution of defects over time. t's
rectification of defects in existing disaster response crucial to remember that careful assessment of the
systems are critical for minimizing casualties and quality of data is necessary for successful
mitigating long-term impacts. The existing body of deployment. A Siamese Network is a neural
knowledge largely focuses on general disaster network that consists of two identical
subnetworks meaning that they contain the a Convolutional Neural Network. Two figures that
same parameters and weights. Each subnetwork are either comparable (positive example) or not
can be any neural network designed for images like similar (negative example) are fed into the network.

During training, we pass the images through the

subnetworks, and we get as output two feature

vectors, one for each image. We want these two

vectors to be as close to each other as feasible if the
input pairs are similar, and the other way around.
To achieve this, we use the contrastive loss function
that takes a pair of vectors (x i, xj) and minimizes
their Euclidean distance when they come from
similar images while maximizing the distance
otherwise:

L = (1 - y) * ||xi – xj||^2 + y * max(0, m - ||xi – xj||^2)

where y=0 if the images are similar and y=1

otherwise. Also, m is a hyperparameter, defining
the lower bound distance between images that
are not similar.

Fig.2. Temple at Chennai after the cyclone effect

K-means clustering is another considerable

algorithm in machine learning for disaster defect
management. Define the features that are pertinent
to the analysis. These could be structural attributes,
geographical coordinates, or other measurable
characteristics related to potential defects or
damage. Standardize or normalize the
characteristics to guarantee that they are on a
similar scale. This is important for the K-Means
algorithm, as it is distance-based. Decide on the
numeral of clusters (K) that the algorithm should
identify. This could be based on the features of the
data or domain knowledge. Sorting data points into
K clusters according to similarity is the aim. Run
Fig.1. Temple at Chennai before the cyclone effect the K-Means algorithm on the preprocessed data.
The closest cluster center, or centroid, is iteratively
assigned data points by the algorithm, which then
updates the centroids according to the average of
the assigned points. The result is K clusters, and
Every cluster denotes a collection of related data
points.Use geographical information to map the
identified clusters onto the affected area. This can
provide insights into the spatial distribution of
potential defects or areas at higher risk. Combine
the results from K-Means clustering with other
relevant data sources, such as historical data on
disasters, socio-economic data, or real-time sensor
data, to enrich the analysis and provide a more
comprehensive understanding.
The probabilistic machine learning algorithm Naive features and classes (e.g., defective or non-
Bayes is primarily employed for classification defective). Naive Bayes provides probability
tasks. While it may not be the most typical choice estimates for each class. This can be useful in
for disaster defect management since it is disaster defect management for understanding the
simplifying assumptions, it can still be applied in confidence degree of the model's predictions.
certain contexts. Determine which features are Probability estimates can inform decision-makers
pertinent to the process of classification. These about the uncertainty associated with classifying
could include geographical data, structural features, certain structures or areas. Depending on The
or other quantifiable qualities. Divide the dataset accessibility of new data, the Naive Bayes model
into sets for testing and training. The training set is can be updated to adapt to changing conditions.
utilized as to teach the Naive Bayes model to Continuous learning allows the model to improve
understand the probabilistic relationships between its accuracy over time.

IV. RESULTS AND DISCUSSION natural disasters. The majority of the effects of
natural disasters are being exacerbated by man-
The effect of natural calamities has been atrocious, made practices like deforestation and
these natural calamities are among the reasons for industrialization. Millions of families were worse
migration and mitigation of lives. Several off due to the spike in the demand for factories.
experiments were conducted to understand the Controlling these activities is necessary to lessen
underlined reason for cause of natural disasters the effects of nature. The intricacy of disasters has
those experiments are the reason to some extent that produced an expanding application of Machine
today we can prevent calamities and also could be Learning(ML) advances. The taxonomy offers a
retained from those effects. Algorithms are used, as useful method for categorizing upcoming seismic
the article mentions, to predict the damage a events to direct the distribution of resources for
disaster will cause and to facilitate a quicker disaster management according to
response. The study indicates that the negative particular requirements. This paper provides the
effects of disasters are getting worse every year. In findings of a review study that looked into the ways
the previous 12 months, more than 2.5 million machine learning approaches have been applied to
humans have been replaced due to natural disasters. different aspects of disaster management to support
Not just human replacement, there exists a loss and enhance those operations. The reality that some
occurring to infrastructure and properties also.Fig.1. disaster studies have been written in languages
depicts the temple, before the effect of natural other than English could be significant to
disasters. Similarly Fig.2. depicts the temple, after this area of study. When users communicate
the effect of natural disasters with the Siamese through platforms, there may be variations in the
network algorithm comparison between the two response dynamics and coordination.. It will only
figures is done, which results positive if both the be feasible to conduct additional research on these
images match else results negative if any mismatch newest Internet community platforms once data
with regards to pixels or the sight of the figures. extraction API support is in place. The other
Data from the internet community has proven to be reviewed research included a wide range of topics
a valuable resource for disaster management. including case studies, applications, early warning
defects and this article is mainly based on how systems, and risk and vulnerability assessment. This
effectively Internet data is assessed in disaster review paper includes another spread of topics such
management. This Internet data can be acquired as damage assessment, faster response post-disaster,
utilizing the Application Programming and hazard prediction. Future studies should focus
Interface(API). The other findings concerning on utilizing machine learning to increase the
disaster management suggest that inward a gap in efficiency of disaster recovery operations. Since
completely achieving 100% accuracy within the disaster recovery operations have to be long-lasting,
framework of disaster control. This study also helps studies should concentrate on applying machine
to understand how effectively internet data could be learning to improve mitigation efforts and lower
used which saves time than assessing the scenes vulnerabilities. Robust and validated ML solutions
from the catastrophe by physically presenting at the are necessary since, the complexity and criticality
moment. In this article, we tried to achieve that of disaster operations. Since disaster operations
failure with the aspect of handling the disasters. have an impact on human life, the developed
models. For decision-makers and domain experts to
V. CONCLUSION understand it, it should also be explicable.
One of the major reasons for the loss of human Additionally, research ought to concentrate on
lives and damage to property and infrastructure is enhancing data quality creating innovative methods
for gathering data, and utilizing crowdsourcing to integrated convolutional neural network and
increase the efficiency of disaster management sorting algorithm for image classification for
strategies operations based on ML. The figures in efficient flood disaster management,2023
this article depict the differences by comparing the
input images and it involves the use of Siamese 4. Abdullah Kurkcu, Fan Zuo, Jingqin Gao,
network algorithm. Internet Data performs a crucial Ender Faruk Morgul and Kaan Ozbay,
role in locating affected individuals, learning about Crowdsourcing Incident Information for
their current situation, and gathering information Disaster Response using Twitter ,2017
about different rescue operations carried out during
natural and man-made disasters. Our goal in this 5. Rania Rizki Arinta and Andi W.R. Emanuel,
dissertation was to evolve a machine-learning Natural Disaster Application on Big Data
method for managing disaster assessment data. The and Machine Learning, 2019
primary drawback with this paper is it provides a 6. Adel Alshehri, A Machine Learning
restricted number of outcomes. Natural disasters Approach to Predicting Community
can vary greatly in scale and intensity, ranging from Engagement on Social Media During
localized events to widespread catastrophes. The Disasters,2019
sheer magnitude of some disasters can overwhelm
local, regional, and even national response 7. Lokabiram Dwarakanath , Amiruddin
capacities. Systems that use machine learning Kamsin ,Rasheed Abubakar Rasheed, Anitha
models like CNN, as well as systems that use image Anandhan and Liyana Suhib, Automated
processing techniques like edge detection, Machine Learning Approaches for
segmentation, and pixel analysis, are the main tools Emergency Response and Coordination via
used to address the problems associated with flood Social Media in the Aftermath of a
management. The most widely utilized methods for Disaster,2021
acquiring images are UAV imaging, Remote
sensing, and SAR. The methods currently in use 8. Vasileios Linardos , Maria Drakaki ,
from both images The domains of processing and Panagiotis Tzionas and Yannis L. Karnavas,
machine learning typically concentrate on both pre- Machine Learning in Disaster Management:
and phases after a disaster. When it comes to Recent Developments in Methods and
identifying Internet images about disasters, images Applications,2022
are essential. Nevertheless, a number of the highly
rated Internet images featuring photos had nothing 9. Maria Drakaki,Machine learning to analyze
to do with emergency services. Future studies could social media data for disaster
incorporate the taxonomy into a sophisticated management,2022
system, like a drone designed to collect data by
10. Sreenivasulu Madichetty and Sridevi M, A
scouting areas affected by Natural disasters.
Novel Method for Identifying the Damage
enabling natural disasters to be classified in real
Assessment Tweets during Disaster, 2020
time before the usage of rescue and aid supplies.
11. Anna Kruspe , Jens Kersten and Friederike
VI. REFERENCES
Klan, Review article: Detection of
1. Hafiz Suliman Munawar , Ahmed W.A. informative tweets in crisis events,2020
Hammad and S. Travis Waller , A review on
12. Donald Douglas Atsa'am , Terlumun
flood management technologies related to
Gbaden and Ruth Wario, A machine
image processing and machine
learning approach to formation of
learning ,2021
earthquake categories using hierarchies of
2. Donald Douglas Atsa'am , Terlumun magnitude and consequence to guide
Gbaden and Ruth Wario, A machine emergency management ,2020
learning approach to formation of
13. Hafiz Suliman Munawar , Ahmed W.A.
earthquake categories using hierarchies of
Hammad and S. Travis Waller, A review on
magnitude and consequence to guide
flood management technologies related to
emergency management,2023
image processing and machine
3. Md Azharul Islam , Shawkh Ibne Rashid , learning ,2021
Niamat Ullah Ibne Hossain , Robert
Fleming and Alexandr Sokolov , An