11 II February 2023

https://doi.org/10.22214/ijraset.2023.49175
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 11 Issue II Feb 2023- Available at www.ijraset.com

Detecting Natural Disasters Using Deep Learning

B. Keshav1, M. Cherryl1, P. Naveen Kumar3, Dr. G. Murugan4
1, 2, 3 4
Student, Professor, Computer Science and Engineering, Vardhaman College of Engineering, Hyderabad, India

Abstract: Natural disasters cannot be stopped, but they can be spotted, allowing people valuable time to flee to safety, as is often
noted. One strategy is to utilize computer vision to supplement current sensors, which improves the accuracy of natural
catastrophe detectors and, more significantly, enables people to take preparations, stay safe, and prevent/reduce the number of
fatalities and injuries caused by these disasters. As a result, responding to natural disasters like earthquakes, floods, and
wildfires requires extensive work by emergency responders and analysts who are on the ground. A low-latency data source for
understanding crisis conditions has arisen in social media. While the majority of social media research just uses text, photos
provide additional insight into accident and disaster situations.
Keywords: VGG16, CNN, CLR, Learning rate finder, Overfitting

I. INTRODUCTION
For response organizations, unexpected onset events like earthquakes, flash floods, and car accidents must be quickly identified.
However, gathering information in an emergency is time-consuming and expensive because it frequently calls for manual data
processing and professional evaluation.
There have been attempts to use computer vision algorithms on synthetic aperture radar, satellite photography, and other remote
sensing data to reduce these laborious efforts. Unfortunately, these methods are still expensive to use and insufficiently reliable to
gather pertinent data in emergencies.
Additionally, satellite imagery only offers an above perspective of the disaster-affected area and is subject to noise like clouds and
smoke (i.e., common images during storms and wildfires).
According to studies, social media posts in the form of text messages, pictures, and videos can be accessible right away when a
disaster strikes and can provide crucial information for disaster response, including reports of infrastructure damage and the
immediate needs of those who have been affected. Social media imaging is still underutilized, nonetheless, in contrast to other data
sources (such as satellites), mostly due to two significant difficulties.
First off, social media picture streams are notoriously noisy, and disasters are no exception. Sizable chunks of social media
photographs are irrelevant to particular disaster categories even after applying a text-based filter. Second, although deep learning
models, the industry standard for image classification, are data-hungry, there is currently no large-scale ground-level picture dataset
available for the development of robust computational models.
In this work, we address these issues and look into the detection of accidents, damage, and natural disasters in photos. The large-
scale Incidents Dataset, which comprises 4,428 scene-centric photos and is classified into four classes—cyclones, earthquakes,
floods, and wildfires—is presented first. Our model uses these pictures as the training and testing datasets.

II. ABOUT NATURAL DISASTERS

Natural disasters are tragic occurrences brought on by global natural or natural processes. The number of lives lost, the amount of
money lost, and the population’s capacity to recover gauges disaster intensity. Natural catastrophes in general bring harm and loss to
the local population as well as the environment. Floods, earthquakes, tsunamis, landslides, volcanic eruptions, and storms are
examples of natural catastrophes.
Three categories further distinguish the severity and scope of the harm.
Minor disasters are those that cover more than 50 km or up to 100 km., Fires are considered minor catastrophes.
1) Disasters of Medium Size: These catastrophes have a radius of 100 to 500 kilometers. These cause greater harm than minor
disasters. Medium-scale disasters include landslides, tornadoes, erosion, and more.
2) Disasters: These calamities affect a region larger than 1000 km. These produce the most catastrophic environmental impact. At
a high level, these catastrophes may potentially affect the entire nation. Large-scale disasters include earthquakes, floods,
tsunamis, and other natural calamities.

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 1042
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 11 Issue II Feb 2023- Available at www.ijraset.com

A. How to Detect Natural Disasters

We employ sensors to keep an eye out for natural disasters all over the world.
Seismic sensors (seismometers) and vibration sensors (seismoscopes) are used to keep an eye out for earthquakes (and downstream
tsunamis). To identify a tornado's distinctive "hook echo," Radar maps are used (i.e., a hook that extends from the radar echo).
Water level sensors track the height of water along a river, stream, etc. While Flood sensors gauge moisture levels.
Although wildfire sensors are still developing, it is expected that they will eventually be able to identify minute amounts of smoke
and fire. The purpose of each of these sensors is to early detect a natural disaster, warn people, and enable them to flee to safety.
By combining existing sensors with computer vision, we can improve the accuracy of natural disaster detectors and, most
importantly, enable people to take protective measures, stay safe, and prevent/reduce the number of fatalities and injuries caused by
these disasters.

B. Objective
The primary goal of this project is to create a cutting-edge Convolutional Neural Network (CNN) model for classifying natural
disaster images and videos into different disaster kinds. On the dataset, the model is trained and tested. The system should accept
images and input and provide output on the probability of natural disasters occurring, the goal is to predict them at an early stage.

Abbreviation and Acronyms

1) CNN: Convolution Neural Network
2) CV: Computer Vision
3) NN: Neural Networks
4) VGG: Visual Geometry Group
5) CLR: Cyclical Learning rate
6) LRF: Learning Rate Finder

III. RESEARCH METHODOLOGY

It is well said, “Natural disasters cannot be prevented — but they can be detected, giving people precious time to get to safety. One
strategy is to utilize computer vision to supplement current sensors, which improves the accuracy of natural catastrophe detectors
and, more significantly, enables people to take preparations, stay safe, and prevent/reduce the number of
Fatalities and injuries caused by these disasters. In our research, we demonstrate the automatic detection of natural disasters in
photos and video feeds using computer vision and deep learning techniques. We have thought after that, we will go over our dataset
of four classes related to natural disasters.

Then, we created a series of tests that will:

1) Aid us in optimizing our database's VGG16 (which was trained on ImageNet).
2) Discover the ideal learning rates like cyclical learning rates.
In the work of Ethan Weber and Ferda Ofl, they provided the Incidences Dataset in this study, which consists of 446,684 human-
annotated photos covering 43 incidents across various scenes.
On millions of social media photos from Flickr and Twitter, they do image filtering tests and use a baseline classification model that
reduces false-positive errors. With the help of these tests, they demonstrate how the Incidents Dataset may be used to find pictures
of incidents in the field. [1]
Considering Albertus Joko Santoso; Findra Kartika Sari Dewi; Thomas Adi Purnomo Sidhi[2].The technology can analyze
sequences of satellite imagery taken before and during a natural catastrophe to identify patterns. The goal of this project is to select
the best wavelet to compress the satellite image sequences and to use an artificial neural network to recognize natural disaster
patterns. Satellite imagery sequences of tornadoes and hurricanes are used in this investigation. [2]

A. Population and Sample

The data is collected from different sources; some are social media platforms like Twitter, Facebook, etc. 428 images are collected
through google images, and these images are sampled according to their respective classes of natural disasters (cyclones, Floods,
Earthquakes, and Wildfires).

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 1043
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 11 Issue II Feb 2023- Available at www.ijraset.com

B. Data and Sources of Data

We have utilized Google Photos' capabilities to quickly collect training images and so reduce the amount of time it takes to create a
dataset. The dataset includes images belonging to four classes of natural disasters (cyclones, Earthquakes, floods, and wildfires).
The distribution of image count is as below: -

Table 1: Images Count

Here is the link for accessing the dataset, the dataset is stored in google drive and accessed when required for execution: -
https://drive.google.com/drive/folders/139H6Nmf9gBbP15BXSRCD6MLHTAwem1mt?usp=sharing

C. Theoretical Framework
This section includes the pre-processing of data.
1) Data Cleaning and Feature Extraction: Data pre-processing, often known as data cleansing, is the first step in data extraction.
The goal of data cleaning is to simplify the dataset so that it is easier to deal with. One observation per row and one variable per
column are two traits of a clean/tidy dataset.
2) Model: The machine learning process is carried out using a deep learning artificial neural network that has a hierarchy of levels.
The model is built on deep networks, in which the information flow starts at the initial level. The model learns something
simple and transmits its output to layer two of the network while merging its input into something slightly more difficult and
passing it on to layer three of the network. This process continues as each level of the network draws on the knowledge it
received from the preceding level.

D. Statistical Tools and Econometric Models

This section elaborates on the proper statistical/econometric/financial models, which are being used to forward the study from data
toward inferences. The details of the methodology are given as follows.
1) Convolutional Neural Network (CNN): Deep neural networks that have been specially designed to analyze data with input
forms resembling a 2D matrix are known as convolutional neural networks. Images could be represented by a straightforward
2D matrix. CNN is crucial while using images. It accepts an image as input, assigns weights and biases to various parts and
objects within the image, and then separates them based on significance. The CNN uses filters(sometimes referred to as kernels)
to aid in feature learning and detect abstract ideas such as blurring, edge detection, sharpening, etc., like how the human brain
detects objects in time and space. Because the weights can be reused and there are fewer parameters (2048 to 256), the
architecture fits the picture dataset more effectively.

Fig 1: CNN Process Model

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 1044
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 11 Issue II Feb 2023- Available at www.ijraset.com

Fig 2: Max Pooling operation

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 1045
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 11 Issue II Feb 2023- Available at www.ijraset.com

Fig 4: CNN Architecture for classifying objects.

In this study, a convolutional neural network (CNN) is employed to convert an RGB image into a visual feature vector. The three
most often used CNN layers are convolution, pooling, and fully connected. Additionally, ReLU f(x) = max (0, x), a nonlinear active
function, is used. ReLU is faster than the common equation f(x) = tanh (x). The use of a dropout layer prevents overfitting. The
dropout sets the output of each hidden neuron to zero with a probability of 0.5. The "dropped out" neurons are a part neither of the
backpropagation nor the forward pass.
Due to the millions of parameters that both the CNN and the RNN include, there are specific convergence concerns when they are
merged. For instance, Vinyals et al. found that fixing the convolutional layer's parameters to those trained from ImageNet is optimal.
The only CNN parameters that are learned from caption instances are the RNN parameters and the non-convolution layer
parameters.

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 1046
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 11 Issue II Feb 2023- Available at www.ijraset.com

Table 2: Architectures of CNN:

YEA CNN DEVELOP FEATURES IMPORTANCE NO. OF NO. OF
R ED BY LAYER PARAMETE
S RS
1998 LeNet Yann LeCun 1. Average pooling layer 1. Character 7 layers 60 thousand
with subsampling. Recognition.
2. Activation of the tanh. 2. Classify handwritten
3. MLP is used as the numbers on banks and
final classifier. other financial
4. Sparse layer institutions.
connections will simplify
calculations.
2012 AlexNet Geoffrey 1. ReLU Activation 1. Object detection 8 layers 60 million
Hinton, Ilya function. task.
Sutskever, 2. Batch size is 128.
Alex 3. Ensembling models to
KriZhevsky achieve the greatest
outcomes.
2014 GoogleNe Google 1.1x1 convolution. 1. Image classification 27 layers 4 million
t 2. Inception module. 2. Object recognition
3. Auxiliary Classifier 3. Quantization
for training.
2014 VGG Net Zisserman, 1. Has 2 networks i.e., 1. Large-scale Image 16 layers 138 million
Simonyan VGG-16, VGG-19 Recognition 19 layers
2015 ResNet Kaiming He 1. The skip Connection 1. efficient backbone 34 layers 25 million
technique is used model
2. Residual mapping
2020 Xception Francois 1. Depth Wise separable 1. Image recognition 71 layers 22 million
Chollet Convolutions
2. Takes the tenets of
Inception for a logical
conclusion.

2) VGG16: A convolution neural network (CNN) architecture called VGG16 was utilized to win the 2014 ILSVR (ImageNet)
competition. It is regarded as having one of the best vision model architectures to date. The distinctive feature of VGG16 is that
it prioritized having convolution layers of 3x3 filters with a stride 1 and always utilized the same padding and max pool layer of
2x2 filters with a stride 2. Throughout the entire architecture, convolution and max pool layers are arranged in the same manner.
Two FC (completely connected layers) are present at the very end, followed by a softmax for output. The 16 in VGG16
indicates that there are 16 weighted layers. This network has over 138 million parameters, making it a sizable network.

Fig 5: Architecture of VGG16

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 1047
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 11 Issue II Feb 2023- Available at www.ijraset.com

Fig 6: Process Diagram of VGG16

Fig 7: Flow Chart of VGG16

IV. RESULTS AND DISCUSSION

A. Results of the Training Model
Performance Metrics used: Accuracy
Accuracy is defined as the ratio of the number of accurately predicted image classes to the total number of images. It is the most
straightforward performance metric. However, accuracy is only valid when the class distribution is symmetric, or when there are
nearly equal numbers of images (or observations) in each class. We also plot the Confusion Matrix (4*4) to check how well our
model performs in every class.

Table 3: Training values comparison

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 1048
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 11 Issue II Feb 2023- Available at www.ijraset.com

Table 3 displayed the accuracy obtained after running some set of epochs at each interval. It is observed that as the set of epochs
running increases from interval to interval, the accuracy is increased. Finally, after running 48 epochs at a time it is noted that the
accuracy obtained is 97%.

Fig 8: Accuracy at 15/15 epochs

Fig 9: Graph showing training loss vs Accuracy

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 1049
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 11 Issue II Feb 2023- Available at www.ijraset.com

B. Observation
1) Finally, the learning rate plot demonstrates how the learning rate oscillates between the MIN_LR and MAX_LR values in our
CLR callback:

Fig 10: CLR graph

Cyclical Learning Rate: We trained our neural networks using Keras and Cyclical Learning Rates (CLR). Cyclical Learning Rates
can significantly cut down on the number of tests needed to fine-tune and identify the best learning rate for your model.

2) Finding the best learning rates for our Network on our natural catastrophe dataset using a Keras Learning Rate Finder. With the
Keras deep learning framework, we will use the dataset to train a model for identifying natural disasters.

Fig 11: Learning rate finder

If you look at the plot, you can see that our model immediately picks up speed and learns around 1e-6.
Our loss keeps decreasing until it reaches about 1e-4, at which point it starts to increase once more, indicating overfitting.
Hence, the range of our ideal learning rate is 1e-6 to 1e-4.

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 1050
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 11 Issue II Feb 2023- Available at www.ijraset.com

C. Output

Fig 12: Output displaying over Video

Above fig.12 shows, the output displayed over a video. After the successful completion of training, we uploaded a video as input
and checked its activity. As a result, we obtained good accuracy in detecting video as Flood.

V. ACKNOWLEDGMENT
We would like to thank our guide “prof. Dr. G. Murugan” for helping throughout our project. In addition, special thanks to our
college management (Vardhaman College of Engineering) for standing back for us in the successful completion of the project.

REFERENCES
[1] In Proceedings of Journal of arXiv ISSN: 2008.09188v1, August 2021, Ethan Weber, Nuria Marzo, Dim P. Papadopoulos, Aritro Biswas, Agata Lapedriza,
Ferda Ofli, Muhammad Imran, and Antonio Torralba published "Detecting natural disasters, damage, and occurrences in the wild."
[2] Multimodal classification of crisis occurrences in social media by M. Abavisani, L. Wu, S. Hu, J. Tetreault, and A. Jaimes. In: Computer Vision and Pattern
Recognition Conference Proceedings, IEEE/CVF (2020).
[3] Deep landscape features for enhancing vector-borne disease prediction, N. Abdur Rehman, U. Saif, and R. Chunara. Workshops from the IEEE Symposium on
Computer Vision and Pattern Recognition (2019).
[4] A real-time decision support system for earthquake crisis management, called Ears (earthquake alert and report system), was developed by M. Avvenuti, S.
Cresci, A. Marchetti, C. Meletti, and M. Tesconi. pp. 1749–1758 in SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD).
ACM (2014). (2014).
[5] Inundation modeling in areas with a lack of data: Ben-Haim, Z., Anisimov, V., Yonas, A., Gulshan, V., Shafi, Y., Hoyer, S., Nevo, S. NeurIPS Workshop on
Artificial Intelligence for Disaster Response and Humanitarian Aid (2019).
[6] The multimedia satellite challenge at Medieval 2017: Emergency response for flooding occurrences. Bischke, B., Helber, P., Schulze, C., Venkat, S., Dengel,
A., Borth, D. MediaEval 2017 Workshop Proceedings, pp. 1–3. (2017).
[7] Daly, S., Thom, J.: Exploiting and categorizing social media image posts to assess fires. Pages 1–14 in 13th International Conference on Information Systems
for Crisis Response and Management (ISCRAM) (2016).
[8] Fernandez UAV-based urban structural damage assessment utilizing object-based picture processing and semantic reasoning. Galarreta, J., Kerle, N., Gerke, M.
Earth System Science and Natural Hazards, 15(6), 1087-1101 (2015)
[9] https://ieeexplore.ieee.org/document/7237228
[10] https://www.itm-conferences.org/articles/itmconf/pdf/2022/04/itmconf_icacc2022_03010.pdf

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 1051