TechSenary Journal of Engineering & Research

TSJER, Volume2, Issue 1

A Paper on Forest Fire Prediction Using

Convolution Neural Network

Aadira Pillai, Aishwarya Gaikwad, Dipali Dande, Prof. Deepali Dhadwad
Department of Computer Engineering, Indira College of Engineering and Management, Parandwadi, Pune

Abstract: In order to detect fire, a forest fire image/video recognition method based on convolutional neural
networks is proposed. There are two main types One is based on conventional image processing technology and
the other is based on convolutional neural network technology. The erstwhile is easy to lead to false detection
because of blindness and randomness in the stage of feature selection, while for the hindmost the unprocessed
convolutional neural network is applied directly. Hence, in the proposed work we build a deep learning model
based on the YOLOv3 algorithm that processes a video frame-by-frame to detect such oddity in real-time and
generate an alert for the concerned authorities. The final model has a detection rate of 45 frames per second and
also has been benchmarked on datasets like IMFDB, UGR, and FireNet with accuracies of 89.3
percent,82.6percent and 86.5percent respectively. Experimental results convinces the goal of the proposed model
and also shows a fast detection rate and also shows that the CNN method based on adaptive pooling method has
better performance and has higher recognition rate.

Keywords: Forest fire recognition, deep learning, convolutional neural network, Yolov3, darknet.

1.Introduction

With the rapid expansion of digital camera technology and image processing technology, the
flame detection method based on the computer vision system has gradually replaced the
conventional method and has become an important trend. Due to the compound background
and large space of the forest, certain difficulties are brought to the forest fire identification
process, especially in the feature selection process. Applying the convolutional neural network
(CNN) technology to image/video recognition can avoid the blindness and randomness to a
large extent in the feature extraction process, and theoretically extract deeper features, which
can greatly improve the accuracy. CNN technology has been applied by many researchers.
Wang5 used the logistic regression and support vector machine to classify the images in the
output classification layer of the CNN. The experiments show that the accuracy of the forest

tsjer.techsenary.com 55 | Page
 TechSenary Journal of Engineering & Research
TSJER, Volume2, Issue 1

fire detection algorithm based on deep learning has excelled. Chen et al.6 used the block
detection method to pre-process the video fast on forest fire images, which greatly reduced the
running time of the whole system. This method uses a graphics processing unit (GPU) to
accelerate the texture analysis and improves the real-time execution of the system. The
experimental results show that the CNN-based forest fire flame recognition method has notable
advantages compared with the traditional flame detection method.

Tested on real video, the proposed system achieves better performance than other
related traditional video forest fire detection methods. However, in these most of the original
images were directly input to the CNN to accomplish the training which makes the network
generalization performance very poor. Due to the above factors, traditional image processing
techniques and CNN methods are combined in this system. Firstly, the flame area of fire is
segmented, and then the CNN is used to complete the recognition of fire flames.

2. Literature Survey

Forests are the most valuable and necessary resource and protect earth’s ecological balance,
forest products are essential in our lives in direct and indirect manner. But wildfires can cause
crucial damage to the ecosystem. Wildfires burn acres of land and destroy everything in its way
in mere minutes. The effects of wildfire’s are several and wide-ranging. it causes a hugely
notable impact. Naturally caused wildfires can be predicted using factors like temperature,
humidity, soil moisture, pressure and many more. In this the prediction of forest fires by
machine learning using some operational monitoring over a region and coming up against
changes in climate using different sensors are advocated. The Wildfire Prediction System
monitors and records change in climatic parameters and predicts the severity of forest fire based
on real-time data, avoiding the massive loss due to forest fires but its computational speed is
less.[3] To prevent forest fires, predictions need to be made to find out areas of land that have
the potential to burn based on meteorological conditions acquired from the sensor, it is expected
to lessen the spread of fire before the fire roll out. Meteorological parameters used in the study
for prediction of areas of land that will be affected by forest fires are temperature, wind,
humidity, and rainfall. The method used in this is a neural network with an Extreme Learning
[5]
Machines (ELM) training model. To improve the performance of the ELM method, in this
study several tests will be carried out so that the resulting predictions are the best.

tsjer.techsenary.com 56 | Page
 TechSenary Journal of Engineering & Research
TSJER, Volume2, Issue 1

3.Proposed System
According to the principle of object detection algorithms, the flow of image/video fire detection
algorithms based on convolutional neural networks is designed. Majority of the research is
focused on traditional features extraction methods for detection. The major issues with such
methods are their time-consuming process and also generate a high number of false alarms. To
cope with such issues, we extensively studied and explored deep learning architectures for early
flame detection. We investigated numerous CNNs to improve the flame detection accuracy and
minimize the false warnings rate and also minimize the computational time and memory.
The CNN detection has functions of region proposal, feature extraction and
classifications. CNN takes an image/video as input and outputs region proposals by
convolution, pooling. Then the region-based object detection CNN decides the presence or
absence of fire in proposal regions through convolutional layers, pooling layers or fully-
connected layers, etc. Different from the other neural networks CNN uses connection weight
and weighted sums, the convolutional layer uses image transform filters called convolution
kernels to generate feature maps of original images/videos. The convolutional layer is a
collection of convolution kernels. The convolution kernel slides on the image/video and
computes a new pixel by a weighted sum of the pixels it floats over to generate a feature map.

4.Methodology

Our model is trained using 3000 images of fire from the UGR FireNet dataset. The dataset
contains images of fire in a variety of different angles, positions and orientations. The dataset
is also annotated in () format which was converted into the format suitable for YOLO (). 500
images containing guns will be also used which were downloaded from Google and were
annotated by using Label Img a graphical image annotation tool for testing gun detection, we
are using the UGR handgun testing dataset and the IMFDB dataset. The negative images in the
dataset contain images of objects with shapes similar to a gun like hairdryers, drills, etc. For
testing our model’s performance on images containing fire we are using the images in the
FireNet Dataset. We have also made a custom dataset of 23 images downloaded from Google
which contains images of fire in different intensity and different places. Our dataset is called
Fire Gun dataset. The final model has a detection rate of 45 frames per second and also has

tsjer.techsenary.com 57 | Page
 TechSenary Journal of Engineering & Research
TSJER, Volume2, Issue 1

been benchmark on datasets like IMFDB, UGR, and FireNet with accuracies of 89.3 percent
82.6percent and 86.5percent when tested with these respective datasets.

5. System Architecture Diagram

The architecture of the proposed system is shown in figure 1. It has four modules.

Figure 1: System Architecture

6. Algorithm Strategy

The YOLOv3 algorithm Object detection is a task in computer vision that involves identifying
the presence, location, and type of one or more objects in a given photograph. It is a challenging
problem that involves building up methods for object recognition (e.g., where are they), object
localization (e.g., what are their extent), and object classification (e.g., what are they). In recent
years deep learning techniques are achieving great results for object detection, such as on
standard datasets and in computer vision competitions. Notable is the “You Only Look Once,”
or YOLO, family of Convolutional Neural Networks that achieve great results with a single
end-to-end model that can perform detection in real-time.
YOLO is a Convolutional Neural Network for doing object detection. CNN’s are
classifier-based systems that can process input images as structured arrays of data and identify
patterns between them. YOLO has the major advantage of being much faster compared to other
networks and still maintains accuracy.

tsjer.techsenary.com 58 | Page
 TechSenary Journal of Engineering & Research
TSJER, Volume2, Issue 1

Figure 2: Algorithmic Strategy

7.Mathematical Model

Identify the Users,

U = {u1, u2, u3, ...}
Where ‘U’ is main set of Users
like u1, u2, u3, …
D be the set of Data.
D = {D1, D2, D3, ...}
Input video
Q = {V 1, V 2, V 3, ...}
Where ‘V’ is the main set of Video v1, v2, v3, ...
SYS = {DX, DF, AP, BG}
DX = It Data Extractor which extract the data from the dataset
DF = find fire images using YOLOV3 in the database.
AP = Filter the results of YOLOV3 using Algorithm.
BG = It generate the different frame
Identify the processes as P.
P = {P1, P2, P3, P4}
P1 = {e1, e2} where,
e1 = i | i, database designing from the dataset
e2 = j | j, show all clicks through data from the database
P2 = {e1, e2, e3, e4} where,
e1 = i | i, Take the video from the user t
e2 = i | i, Search query using yolov3

tsjer.techsenary.com 59 | Page
 TechSenary Journal of Engineering & Research
TSJER, Volume2, Issue 1

e3 = j | j, Filter the results of yolov3 using Apriori

e4 = j | j, Generate the frame
Graph G = {E, V
were,
V = {v1, v2, v3, ...} be set of vertices
E = {(v1, v2), (v2, v3)} be set of edges
P3 = {e1, e2} where,
e1 = i | i, Find out the fire and gun images in video

Algorithm: Object Detection using YOLOv3 in Python

Step 1: Download the models. We download the models using the script file getModels.sh
from the command line.
Step 2: Initialize the parameters.
Step 3: Load the model and classes.
Step 4: Read the input.
Step 5: Process each frame.
Step 6: The names of output layers
Step 7: Post-processing the network’s output
Step 8: Draw the predicted box.

It allows the model to look at the complete image at test time so its predictions are informed
by the global context in the image. YOLO “rank” the regions based on their similarities to
predefined classes. High-scoring regions are noted as positive detection of whatever class they
most closely identify with. For example, in a live feed of traffic, YOLO can be used to detect
different types of vehicles depending on which region of the video scores highly in comparison
to predefined classes of vehicles.

Suppose we assume that a specific user wants to monitor the forest. He/She will have to do
the following
1. Login
2. IF invalid credentials then the further access is denied or else the access is granted.

tsjer.techsenary.com 60 | Page
 TechSenary Journal of Engineering & Research
TSJER, Volume2, Issue 1

3. Then the input is also given in the form of live video in the actual scenario. THEN
video is continuously monitored to detect the fire in the forest.
4. IF fire is detected then an alert pops up notifying the same else the surveillance is
continued.
5. Logout

8.Snapshots

Figure 3.1 Home Screen

When a user first accesses the system, he/she visits the home page where the basic
information of the system is present.

Figure:3.2 Admin Login

tsjer.techsenary.com 61 | Page
 TechSenary Journal of Engineering & Research
TSJER, Volume2, Issue 1

The user needs to insert the valid credentials to get access to the system. But if the
credentials are not valid the access is denied to proceed further. So as to avoid any anomalies.

Figure 3.3 Fire Detection Using Bouncing Boxes

The image here shows how the system detects fire using the frames and then represents
it in the form of the bouncing box which has the highest confidence.

Figure 3.4 Opening the video file for detection

These snapshots show the working of the system once the input is given in the form of
stored video/live video for the prediction of fire.

9.Comparison

The graph represents the performance of Proposed System and the Existing System

tsjer.techsenary.com 62 | Page
 TechSenary Journal of Engineering & Research
TSJER, Volume2, Issue 1

Figure 4: Graphical Result

Existing Proposed
System System
Existing System 5.1
Proposed System 3.2
Table 1.1: Result Table

The graph represents the performance of Existing System and Proposed System in
milliseconds.
Formula for Performance:

CPU Time= I*CPI*T

Here,
I= No of instructions
(instruction count: 109 instructions)
CPI= Average Cycle per instructions.
T= clock cycle time (before, after)
(Clock cycle time can be computed from the clock rate to be 0.5×10-9 seconds.)
The measurements are made by setting a timestamp before before the method calls of adaptor
methods and afterwards after. The difference between before and after is taken for the runtime
of the method. For the time stamps the global time is taken.

tsjer.techsenary.com 63 | Page
 TechSenary Journal of Engineering & Research
TSJER, Volume2, Issue 1

In the existing system (or a program task) takes 1 Hundred instructions to execute on a
processor running at 1.9 GHz. Suppose also that 50% of the instructions execute in 5 clock
cycles, 30% execute in 6 clock cycles, and 20% execute in 7 clock cycles. We have the
instruction count: 109 instructions. The clock time can be computed quickly from the clock rate
to be 0.5×10-9 seconds. So, we only need to to compute clocks per instruction as an effective
value:
Value Frequency Product

5 0.5 2.5

6 0.3 1.2

7 0.2 1.4

CPI = 5.1

Table 1.2 CPI for Existing system

Whereas at the time in the proposed system 1 Hundred instructions to execute on a processor
running at 4 GHz. Suppose also that 50% of the instructions execute in 4 clock cycles, 25%
execute in clock cycles.
Value Frequency Product

4 0.5 2

3 0.25 0.75

2 0.25 0.5

CPI = 3.2

Table 1.3 CPI for Proposed system

For calculation of performance of both the systems we compare the CPU time required for
both the systems for producing the desired output. That is calculated by performance ratio

Formula of Performance ratio :

(CPU Time/CPU Time2)*(Ghz1/Ghz2)

tsjer.techsenary.com 64 | Page
 TechSenary Journal of Engineering & Research
TSJER, Volume2, Issue 1

So, in our case performance ratio will be as follows ((5.1/3.2)*(1.9/4.0)) =0.757.

Which is a 0.75% improvement.
The graph shows that the existing system requires 5.1ms and the proposed system
requires 3.2ms to produce the desired output. That is, it requires less computational time and
is more efficient than the already existing system.

10.Conclusion

A real-time frame-based efficient fire detection deep learning model that will have been presented with
a high accuracy metric. The detections per frame are appropriate for real-time monitoring and can be
deployed on any GPU based system. It is a cost-effective fire detection tool using CNN. The main focus
of this is to discuss decision parameters, attributes, and features used for predicting the Fire. The
proposed model has a detection rate of 45 frames per second. The proposed system is more
effective and accurate with its results. Its compilation time is also less as compared to the existing
system compilation is completed in 3.2ms whereas the existing system needs 5.1ms. The datasets are
considered in so many existing techniques that we have discussed using Sensor which is not very cost
effective and requires maintenance after every regular period of time.

REFERENCES

[1] Studying the Factors Affecting the Risk of Forest Fire Occurrence and Applying Neural Networks for Prediction
Nizar Hamadeh, Bassam Daya, Alaa Hilal, Pierre Chauvet SAI Intelligent Systems Conference 2015 November 10-11, 2015
IEEE Xplore ISBN: 978-1-4673-7606-8

[2] A forest fire prediction system based on rechargeable wireless sensor networks Gao Demin1,2, Lin Haifeng2, Jiang
Anna2, Wu Guoxin1 2014 IEEE Xplore (2019) ISBN: 978-1-4799-4734-8

[3] IoT Sensor and Deep Neural Network based Wildfire Prediction System Mrs Vidya Zope, Tarun Dadlani, Ashutosh
Matai, Pranjali Tembhurnikar, Richa Kalani Proceedings of the International Conference on Intelligent Computing and Control
Systems (ICICCS 2020) IEEE Xplore ISBN: 978-1-7281-4876-2

[4] Multi-UAV Oxyrrhis Marina-Inspired Search and Dynamic Formation Control for Forest Firefighting K. Harikumar,
J. Senthilnath, and Suresh Sundaram1545-5955 2018 IEEE Xplore

[5] Prediction of Forest Fire using Neural Network based on Extreme Learning Machines (ELM) Mochammad Anshori,
Mukhammad Wildan Alauddin, Wayan Firdaus Mahmudy, Farhanna Mar’I 2019 IEEE Xplore ISBN:978-1-7281-3880-0

tsjer.techsenary.com 65 | Page