Journal of Theoretical and Applied Information Technology

15th October 2022. Vol.100. No 19

© 2022 Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

USING IOT AND ML FOR FOREST FIRE DETECTION,

MONITORING, AND PREDICTION:
A LITERATURE REVIEW
MOUNIR GRARI1, MIMOUN YANDOUZI2, IDRISS IDRISSI1, MOHAMMED BOUKABOUS1,
OMAR MOUSSAOUI1, MOSTAFA AZIZI1, AND MIMOUN MOUSSAOUI1
1
MATSI Lab., ESTO, Mohammed First University, Oujda, Morocco
2
LARI Lab., FSO, Mohammed First University, Oujda, Morocco

E-mail: [email protected]

ABSTRACT

Forests are large areas gathering trees and other plants. They are so important for protecting the environment;
they filter air and water, provide food and shelter for animals, and help regulate the climate. Wildfires are
one of major hazards of global warming; they destroy forests and speed up the deforestation phenomenon.
Other wildfires are also caused by human errors in wilderness environments. Dry vegetation fuels a wildfire's
rapid ignition and spread. It is difficult to extinguish flames even with the best efforts of forest firefighters.
Smoke and air pollution from wildfires may harm human health and ruin property. Forest fires are difficult
to detect at time or to anticipate it, because they spread rapidly. Early-warning systems that they are more
accurate are really needed. These systems could be implemented with IoT (Internet of Things), machine
learning (ML), or deep learning (DL). In this paper, we focus on this direction of research and we examine
literature proposals utilizing IoT and DL to detect wildfires and their spread via a comprehensive evaluation
and comparison of existing works.
Keywords: Forest Fire, Wildfire, IoT, Machine Learning (ML), Deep Learning (DL).

1. INTRODUCTION fires are frequently sparked in wild areas by lightning

or human neglect. With conducive environmental
The term ‘forest’ refers to a large area of land, conditions, a wildfire can ignite quickly and broadly
densely forested with trees and other plants. For over large areas, fueled by dry vegetation. Forest
many reasons, forest ecosystems are critical. firefighters often deploy all their efforts to contain
Millions of plants and animals call forests their and extinguish wildfires, it takes them days to stop
homes. Additionally, they assist in climate it, and some of them die for this end. Wildfire smoke
regulation, filter air and water, and provide a variety and air pollution can cause a real health risk, and the
of valuable resources. Despite their value, forests flames can destroy or damage homes and businesses.
really face numerous threats, like destruction, In order to safeguard homes and communities
wildfires, or deforestation. Numerous forests are against wildfire dangers, it is critical to have a fire
eliminated to make room for agriculture or to extract prevention policy and disaster response program [2].
valuable resources such as timber and minerals. The task of detecting and predicting wildfires is
Additionally, global warming poses a significant not obvious, as they can ignite anywhere and spread
threat to forests. Because of fires, pests, and diseases, rapidly. However, with the help of technological
forests are increasingly threatened as the Earth's solutions involving IoT and artificial intelligence, it
climate changes. It is critical to safeguard forests that is possible to develop early-warning systems with
are real homes of animals and plants. By adopting relevant accuracy. Using key sensors to measure
environmentally friendly choices, we can all changes in temperature, humidity, and wind speed,
contribute to forest preservation [1]. as well as to detect the presence of smoke and fire, is
While forest fires spontaneously occur in a variety a relevant way to accomplish this task, according to
of ecosystems, they could be extremely destructive our opinion. The collected data from these measures
to people, properties, and natural resources. These could be archived, and next used to develop ML

5445
 Journal of Theoretical and Applied Information Technology
15th October 2022. Vol.100. No 19
© 2022 Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

models capable of forecasting the likelihood of a fire monitored. For instance, people could use their
igniting. We can develop models that predict how a phones to turn on the lights in their homes or adjust
given fire will progress under various conditions by the refrigerator's temperatures [6].
analyzing data from previous fires. Built fire datasets The IoT is = rapidly gaining popularity. Numerous
will be a valuable mean to assist firefighters and companies are developing products that leverage IoT
other emergency responders facing efficiently technologies. IoT has a plethora of potential
wildfires. Additionally, it may aid in identifying applications. Home automation, health care, and city
areas at high risk of wildfires, allowing for the management are just a few examples of the IoT's
implementation of preventative measures [3]. most popular applications. For instance, home
By conducting a systematic review of the automation is gaining popularity as people become
literature, we examine the possibility of using IoT more interested in remotely controlling their homes
and deep learning to detect and predict wildfires. [7].
This paper is structured as follows. The second IoT in reality is inspired from a WSN and internet,
section covers the terminological aspect of this topic but with key improvements. IoT is a network of
and gives the background of used technologies. The physical devices, vehicles, home appliances, and
third section discusses the adopted methodology of other objects that are embedded with electronics,
our literature review. Before concluding, the fourth software, sensors, and connected to the internet, and
section makes a synthesis of our findings. can communicate between each other. Compared to
WSN, IoT is used not only for monitoring and data
2. BACKGROUND collection, but also for control.
2.2.1 IoT protocols.
2.1 Wireless Sensor Networks (WSN) IoT protocols can be divided into two categories:
A WSN is a network of autonomous sensors that communication protocols and application protocols.
can monitor physical or environmental conditions Communication protocols define how devices
like temperature, and transmit their data to a central connect to and communicate with each other. These
location. This latter can be either a standalone protocols can be based on different technologies,
centralized unit or a cluster head (a node within the such as Wi-Fi, Bluetooth, Zigbee, and LoRa. Each
WSN that has the responsibility of collecting and has its own strengths and weaknesses, and the best
sending sensor data to the centralized unit). Wireless protocol for a given application will depend on the
sensor networks are composed of nodes, which are specific requirements [8].
often battery-operated, autonomous, and self-  Wi-Fi: is a widely used protocol for internet
contained with sensors, a radio transceiver, and an connectivity, and it can also be used for IoT
onboard signal-processing unit. Data collected by the applications. It has a good range and data
sensor nodes are processed, then transmitted throughput and is widely supported by devices
wirelessly to the centralized unit using multi-hop and infrastructure. However, it can be power-
wireless links [4]. hungry and is not well suited for applications
WSNs have many potential applications in a that require very low data rates or long battery
variety of fields, such as military, industrial, life.
commercial, and home applications. The main  Bluetooth: is another popular protocol for IoT,
benefits of WSNs appear in their use of monitoring and it has the advantage of being low-power and
remotely physical and environmental conditions supporting very low data rates. It is often used
over a wide area (without the need for physical wires for short-range applications where power
or cables) under reasonable budgets. WSNs are often consumption is a concern, such as in wearables.
more flexible and easier to install than wired However, its range is limited, and it may not be
networks. In addition, WSNs are often more robust suitable for some types of applications.
and reliable than wired networks due to their  Zigbee: is a low-power protocol that is often
distribution and redundancy [5]. used for IoT applications that require long
battery life. It has a good range and supports
2.2 Internet of Things (IoT) mesh networking, which can be useful for
IoT is a term that refers to a future in which certain types of applications. However, it has
everyday objects are connected to the internet and relatively low data throughput and may not be
capable of exchanging data. These objects can range suitable for applications that require high data
from watches to refrigerators. The concept behind rates.
the IoT is that by connecting these devices to the  LoRa: is a long-range, low-power protocol that
internet, they can be remotely controlled and is gaining popularity for IoT applications. It has

5446
 Journal of Theoretical and Applied Information Technology
15th October 2022. Vol.100. No 19
© 2022 Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

an excellent range and can support very low data capabilities, making it ideal for complex IoT
rates, making it ideal for applications that applications.
require long battery life. However, it is not as  IBM Watson IoT: IBM’s platform offers a
widely supported as some other protocols, and robust set of features for building cognitive IoT
may not be suitable for all applications. applications.
While the Zigbee and Bluetooth protocols are  Microsoft Azure IoT: Microsoft’s cloud
useful for connecting devices wirelessly, they are not platform provides a comprehensive set of
well suited for applications that require long-range services and tools for deploying IoT
or low power consumption. The LoRa protocol is applications.
designed specifically for low-power and long-range  Thingsboard: is an open-source IoT platform
applications. LoRa allows devices to communicate that allows to connect devices and applications
over distances of up to 15km, making it ideal for IoT to the cloud and to other devices. It provides
applications that require long-range connectivity [9]. access to data and commands from the devices
Application protocols define how data is and allows to send commands to the devices.
exchanged between devices and how applications  ThingSpeak: is an IoT analytics platform that
can access and use that data. Many different allows to collect, visualize, and analyze data
application protocols can be used for IoT devices, from devices. It also allows to send commands
depending on the specific needs of the devices and to devices.
the network [10]. Some common protocols that are  BOLT Cloud: A highly integrated IoT platform
used for IoT applications include MQTT, CoAP, for developers that enables users to rapidly and
HTTP, XMPP, and AMQP [11]: simply construct IoT projects and products.
 MQTT: a popular protocol for handling data These are just some examples of the many
from IoT devices, as it is lightweight and can be different IoT cloud platforms available today. Each
used over unreliable networks. platform has its own unique benefits and features, so
 CoAP: a protocol designed specifically for it is important to choose the one that best suits the
resource-constrained IoT devices, as it is very specific needs.
lightweight and efficient.
 HTTP: a common protocol that is used for many 2.3 Machine Learning (ML)
different types of networking applications, Machines can learn in a variety of ways. Some
including IoT. machines employ artificial intelligence, while others
 XMPP: a protocol that is commonly used for employ machine learning, a subfield of artificial
instant messaging applications, but can also be intelligence concerned with the design and
used for IoT applications. development of algorithms capable of learning from
 AMQP: a protocol that is designed for reliable and forecasting data. Machine learning algorithms
message queuing, and is often used in IoT can be used to discover patterns in data and make
applications that require high reliability. future predictions [15].
2.2.2 IoT cloud platforms Machine learning is a technique for teaching
IoT cloud platforms provide a comprehensive and computers to learn from data without explicitly
centralized way to manage the data and connectivity programming them to do so. It entails exposing
of IoT devices. These platforms enable enterprises to computers to large data sets and then allowing them
remotely monitor and manage IoT devices and to discover patterns and relationships in the data on
connect them to various applications and data their own. This can be accomplished through the use
sources. IoT cloud platforms also offer features such of a variety of techniques, such as artificial neural
as device management, security, data management, networks, decision trees, and genetic algorithms
and analytics [12]. [16]. Machine learning algorithms can be used for
There are many different types of IoT cloud many different tasks, including [17]:
platforms available, each with its own unique  Regression: Used to predict a continuous value,
benefits and features. Some of the most popular IoT such as a price, a probability, or fire in our case
cloud platforms include [13], [14]: study. Common regression algorithms include
 AWS IoT: Amazon’s cloud platform offers a linear regression, logistic regression, and
comprehensive set of features for managing and support vector regression.
deploying IoT applications.  Classification: Used to predict a discrete value,
 Google Cloud IoT: Google’s platform provides such as a class label (e.g., "Fire" or "No Fire").
powerful data processing and analysis Common classification algorithms include

5447
 Journal of Theoretical and Applied Information Technology
15th October 2022. Vol.100. No 19
© 2022 Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

ANN, decision trees, k-nearest neighbors, and learning is how difficult it can be to learn how to use
support vector machines. it effectively. However, with the appropriate tools
 Clustering: Used to find groups of similar and resources, it is possible to develop models that
examples. Common clustering algorithms outperform humans across a broad range of tasks
include hierarchical clustering and k-means [20].
clustering. Machine learning is limited in several ways when
 Dimensionality reduction: Used to reduce the compared to deep learning. Machine learning
number of features in a data set. Common algorithms are less adept at dealing with data noise,
dimensionality reduction algorithms include which means they are more susceptible to being
principal component analysis and kernel duped by inaccurate or incomplete data. Besides,
principal component analysis. machine learning algorithms are less effective at
Machine learning is a rapidly expanding field that generalizing from training data to new data sets,
is being applied in a variety of industries, including which means they are less effective at making
finance, healthcare, manufacturing, and predictions in situations where the data has never
transportation. Additionally, it is used in a variety of been seen before [21].
research fields, including natural language There are various deep learning algorithms; among
processing and computer vision. the most popular are Convolutional Neural
Artificial Neural Network (ANN) Networks (CNNs) [22], Recurrent Neural Networks
ANN is a computational model inspired by the (RNNs) [23], and Generative Adversarial Networks
structure and functionality of the biological neural (GANs) [24].
networks. Structurally, an ANN is composed of a set 2.4.1 Convolutional Neural Networks (CNN)
of interconnected processing nodes, or neurons, CNNs, or Convolutional Neural Networks, are a
where each neuron is a mathematical function that type of neural network that is optimized for image
computes a weighted linear combination of its input processing. They are similar to other types of neural
values. Functionally, an ANN performs a mapping networks, but they have several features that make
of input values to output values [18]. them particularly effective at image processing [25].
The interconnections between neurons are called CNNs are made up of multiple layers (see Figure 1),
synaptic weights, and these weights can be adjusted, each of which is dedicated to a specific task. The first
or learned, to produce the desired mapping. ANNs layer is the input layer, which is responsible for
are capable of learning to approximate any mapping receiving the image data. The following layer is the
function, given enough data and computational convolutional layer, which concatenates the input
resources. There are different types of neural data. Following this layer is the pooling layer, which
networks, including feed-forward networks, pools the input data. The following layer is the fully
recurrent networks, and convolutional networks connected layer, which is responsible for activating
[18]. the input data. Finally, the output layer is responsible
for generating the final output [26].
2.4 Deep Learning (DL) CNNs are more effective at processing image data
Deep learning is a subset of machine learning that than other types of neural networks because they are
enables computers to discover patterns and insights purpose-built for this purpose. The convolutional
in data. It is a highly effective technique for layer extracts feature from the input data, while the
developing sophisticated models capable of learning pooling layer combines these features into a smaller
to recognize objects, textures, and patterns in data. set of parameters. This increases the network's
Deep learning is particularly well-suited for tasks efficiency and capacity for learning. Additionally,
that humans find difficult, such as object recognition the fully connected layer is capable of learning the
in images and videos or natural language relationships between these features, which enables
comprehension. It has been demonstrated to be the network to produce meaningful outputs [27].
effective across a broad spectrum of applications,
including computer vision, speech recognition, and Input Layer
Hiden Layers
Output Layer

natural language processing [19].

One of the primary benefits of deep learning is that
Output
Input

it enables the training of powerful models using

massive amounts of data. These models can then be
used to perform complex tasks such as natural Pooling layer

Convolutional layer

language comprehension or object recognition in Fully connected

images. One of the primary drawbacks of deep

Figure 1: CNN layers architecture

5448
 Journal of Theoretical and Applied Information Technology
15th October 2022. Vol.100. No 19
© 2022 Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

2.4.2 Recurrent Neural Network (RNN) relatively robust against the vanishing gradient
RNNs are a type of neural network in which node problem, which is a common issue with vanilla
connections form a directed graph along a temporal RNNs.
or sequential path. In other words, the output of each  GRUs are a simplified version of LSTM cells
node in the sequence is fed as input to the next node and are well-suited for tasks that do not require
in the sequence (see Figure 2). This creates an the full power of an LSTM. GRUs are faster to
internal state for the network, allowing it to model train and consume less memory than LSTMs,
temporal dependencies in data – such as natural making them a good choice for applications
language understood by humans. RNNs were first where resources are limited.
developed in the 1980s but have only recently been  Plain RNNs are the simplest type of RNN. They
widely used with the advent of powerful computers are suitable for tasks that do not require long-
and effective training algorithms [27]. term memory and are not as computationally
RNNs are particularly well suited to tasks that demanding as LSTMs or GRUs. However, plain
involve sequences of data, such as time series or text RNNs are more susceptible to the vanishing
data. This is because they can remember information gradient problem.
about what has come before in the sequence, which Plain RNN cell
state
LSTM GRU
x +
is important for understanding the context of the x +
x x
x -1
current data. For example, an RNN could be used to x
predict the next word in a sentence, based on the forget
gate
input
gate
output
gate
reset
gate
update
gate

words that have come before [28]. TanH Sigmoid

Pointwise
x multiplication
Pointwise
+ addition
Vector
concatenation
There are several different applications for RNNs. Figure 3: Plain RNN, LSTM, and GRU blocks
One common use case is time series prediction,
where an RNN is used to predict the next value in a 2.5 Datasets
sequence, based on the previous values. RNNs were A dataset is a structured set of data. It usually
also applied in machine translation, speech contains a set of variables, each with a set of values.
recognition, and image captioning applications. A dataset may also contain other sets of data, such as
Hiden Layers subsets, hierarchies, or networks, they are often used
Input Output
Layer Layer to represent real-world objects, systems, or events
[30].
Datasets can be used to store data in a wide variety
of formats, including text, images, audio, and video.
Output
Input

They can also be used to store data in a more specific

format, such as a spreadsheet or database. Datasets
can be created manually or generated automatically,
they can be created by individuals or organizations,
can be small or large, simple or complex, and can be
Figure 2: Recurrent Neural Networks layers static or dynamic. Datasets are often used in data
While RNNs have several advantages, they also analysis and data mining. They can also be used for
have some limitations. One major limitation is that classification, prediction, and optimization.
RNNs struggle to learn patterns that span long Datasets about forest fires are important for
periods of time. This is because the information that understanding the extent of the problem and for
is passed from one node to the next is forgotten over designing effective mitigation strategies. These
time. LSTM networks are designed to overcome this datasets can help answer questions about the
limitation, as they contain special nodes that can location, size, and severity of forest fires, as well as
remember information for long periods of time (see the factors that contribute to them. Forest fire
Figure 3). datasets are available in a variety of formats,
There are three main types of RNN: long short- including tabular (which includes data such as sensor
term memory (LSTM), gated recurrent unit (GRU), data collected) and image/video (contains images or
and plain RNN. Each has its own pros and cons, and videos of forest fires).
the type that is best for a given situation depends on
the nature of the data and the task at hand [29]. 2.6 Evaluation Metrics
 LSTMs are the most powerful and flexible type In machine learning, evaluation metrics are used
of RNN. They are well-suited for tasks that to evaluate the performance of the machine learning
require long-term memory, such as language models, and the appropriate metric to use depends on
translation and text generation. LSTMs are also the specific type of model and the data. Some

5449
 Journal of Theoretical and Applied Information Technology
15th October 2022. Vol.100. No 19
© 2022 Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

common evaluation metrics for classification models classifier that is very recall-oriented but not very
include accuracy, precision, recall, and f1-score. For precise will have a high recall but low precision. The
regression models, common evaluation metrics ideal classifier would have a high precision and high
include mean absolute error, mean squared error, recall, but this is often not possible.
root mean squared error and R-squared. There is no The trade-off between precision and recall is
single perfect evaluation metric. It is always best to important to consider when designing a classifier. In
use multiple evaluation metrics when assessing a some applications, it is more important to have a high
machine learning model. precision (even at the expense of recall), while in
2.6.1 Accuracy others, it is more important to have a high recall
The accuracy metric is a mathematical formula (even at the expense of precision).
used to determine the accuracy of predictions made The F1-score is a good metric to use when
by a machine learning model. The accuracy metric is comparing the performance of different classifiers,
used to compare the predicted values with the actual as it takes into account both precision and recall.
values. The accuracy metric is used to assess the 2.6.3 MAE, MSE, RMSE, and R2
performance of a machine and deep learning model. The Mean Absolute Error (MAE), Mean Squared
The accuracy metric is a value between 0 and 1, Error (MSE), Root Mean Square Error (RMSE), and
where 1 is the perfect score. The accuracy metric is the coefficient of determination (R2, or R-squared)
used to compare the prediction of a machine learning evaluation metrics are used to assess the
model to the actual values. The accuracy metric is performance of predictive models. Each metric
used to assess the performance of a machine learning measures the difference between the predicted values
model [21]. The accuracy is calculated using the and the actual values [32].
formula (1) [31] presented below: MAE is the simplest of the three metrics to
calculate. Simply take the average of the absolute
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = (1) value of the difference between the predicted values
In the case of forest fires, accuracy would tell us and the actual values. MAE is easy to interpret, as it
how many fires were correctly detected out of the gives the average error in the predictions. It can be
total number of fires. If accuracy is high, then we can calculated using the formula (5):
be confident that the fire detection system is working 𝑀𝐴𝐸 = ∑ 𝑦 −𝑦 (5)
well. Where:
2.6.2 Precision, Recall, and F1-score.  𝑛 = number of data points
In the field of data mining, machine, and deep  𝑦 = observed values
learning, there are three important metrics used to
 𝑦 = predicted values
evaluate the performance of a classifier: precision,
recall, and F1-score.
Precision is the proportion of correct positive MSE is the mean of the squared difference
predictions out of all positive predictions. The recall between the predicted values and the actual values.
is the proportion of correct positive predictions out MSE is more difficult to interpret than MAE, but it
of all actual positive events. The F1-score is the is a more accurate measure of differences. It can be
harmonic mean of precision and recall and is a calculated using the formula (6):
measure of a classifier's accuracy [32]. These metrics 𝑀𝑆𝐸 = ∑ 𝑦 −𝑦 (6)
are calculated using the formulas (2,3, and 4) [21] RMSE is the square root of the MSE. RMSE is
presented below: more difficult to interpret than MAE or MSE, but it
is the most accurate measure of differences. It can be
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = (2) calculated using the formula (7):

𝑅𝑒𝑐𝑎𝑙𝑙 = (3) 𝑅𝑀𝑆𝐸 = ∑ 𝑦 −𝑦 (7)

R2 is a measure of the proportion of the variability
𝐹1 = (4) in the target that is explained by the model. It can be
Where: calculated using the formula (8):
 𝑇𝑃 = True Positive 𝑅2 = 1 − (8)
 𝐹𝑃 = False Positive Where:
 𝐹𝑁 = False Negative  𝑆𝐸 = the sum of squares of the residual
A classifier that is very precise but not very recall- errors
oriented will have a high precision but low recall. A  𝑆𝐸𝑡 = the total sum of the errors

5450
 Journal of Theoretical and Applied Information Technology
15th October 2022. Vol.100. No 19
© 2022 Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

picking up small fires, and they can be affected by
Each of these metrics has its own strengths and clouds. One of the main advantages of using
weaknesses, and there is no one metric that is best satellites for forest fire detection is that they can
for all situations. MAE is relatively easy to interpret cover large areas quickly. This is important because
and is often used when there is a need for a quick and forest fires can spread rapidly, making it difficult for
dirty assessment of a model’s performance. MSE is ground-based systems to keep track of them [1].
more commonly used when there is a need for a more Another advantage of using satellites is that they can
detailed analysis. RMSE is often used when there is provide information about the fire at night when it is
a need to compare the performance of two or more often not possible to see the fire from the ground.
models. R2 is a good metric to use when there is a This is important because many Forest fires start at
need to understand how well the predicted values fit night and can spread very quickly before they are
the actual values. noticed [33].
There are several different types of satellites that
3. FOREST FIRE DETECTION, can be used for Forest fire detection, including
MONITORING, AND PREDICTION passive sensors that detect the heat given off by a
TECHNIQUES fire, and active sensors that use a laser or other
electromagnetic energy to “see” through the smoke.
Forest fires are detected, monitored, or predicted Some satellites are specifically designed for fire
using a variety of approaches, which can be divided detection, while others, such as weather satellites,
into two main categories: imagery-based and can also provide valuable information about a fire.
sensors-based. The first approach by which forest The information provided by satellites can be used
fires are detected is through imaging, this can be by firefighters to plan their attack on a fire, and by
accomplished with fixed cameras, satellites, or fire, managers to monitor the progress of a fire and
drones and it allows authorities to get a bird's eye the spread of smoke. Satellites are an important tool
view of the fire, pinpoint its exact location, and get for Forest fire detection, and their use is likely to
its contour. The second approach that is our review increase in the future as they become more
focus, involves the use of sensors that can detect sophisticated and more widely available [34].
environmental data such as heat, humidity, gases, Airplanes and drones are also used for Forest fire
and so on. These sensors, which can be placed detection. They have the advantage of being able to
strategically throughout the forest, will send an alert get a closer look at a fire, which can help to identify
to authorities if they detect a fire [1] (see Figure 4). it more quickly. However, they can be more
expensive to use, and they can only cover a small
area at a time. One way to do this is by using. By
analyzing collected images and videos, it is possible
to identify potential fire hotspots and track the
progress of the fire. This information can then be
used to warn people in the area and help firefighters
to better target their efforts [35].
There are a number of different ways to detect
forest fires using images and videos. One popular
method is to use of heat detection. This involves
Figure 4: Forest fire detection, monitoring, and looking for areas of high temperature in the image or
prediction; the two main categories video. This can be done using infrared cameras or
other thermal imaging methods.
Thermal imagery is a type of infrared photography
3.1 Imagery-based Detection Techniques that can detect heat sources from a distance. This
One way to detect forest fires is to use imagery. makes it ideal for detecting forest fires, which are
Imagery can come from many different sources, often started by careless campers or hikers who don't
including satellites, airplanes, and drones. Each of realize the risks. In general, the working principle of
these sources has its own advantages and this method is very simple: the thermal cameras will
disadvantages [2]. scan the area and identify the areas with high
Satellites are probably the most commonly used temperatures. These areas are likely to be the sources
source of imagery for forest fire detection. They have of forest fires [36].
the advantage of being able to cover large areas very Compared with other methods, the use of thermal
quickly. However, they can sometimes have trouble imagery has many advantages. First of all, it is very

5451
 Journal of Theoretical and Applied Information Technology
15th October 2022. Vol.100. No 19
© 2022 Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

quick and effective. By using this method, the
firefighters can quickly identify the contours of the
area of the fire and focus their efforts on this area.
Moreover, this method can also help to detect the
potential fires which have not been yet broken out.
This is because the high temperatures usually
indicate the areas where the fires are about to start.
Another advantage of using thermal imagery is that
it can be used in all weather conditions. This is
because the thermal cameras can still detect high Figure 5: Forest fire monitoring and detection using
temperatures even in the heavy fog or the dark night. imagery
This is very important because the other methods,
such as the use of ordinary cameras, will not be 3.2 Sensors-based Detection Techniques
effective in these conditions [37], [38]. A sensor is a device that detects and transforms
Infrared gas detectors are a vital tool in the fight physical quantities into a signal that can be read by
against forest fires. By monitoring the presence of an observer or equipment. There are many different
infrared radiation in the atmosphere, these detectors types of sensors, and they are used in a wide variety
can provide an early warning of fires, allowing of applications. Some of the most common
firefighters to take action to prevent them from applications for sensors are in the detection of forest
spreading [39]. One of the main benefits of using fires [41].
infrared gas detection for fire detection is that it can Sensors are one of the most important tools that
be done from a distance. This means that firefighters we use in order to detect forest fires. By monitoring
can be alerted to a fire before it has had a chance to various environmental factors. One of the best ways
spread. This can be crucial in preventing a disastrous to detect a forest fire early is to use sensors.
wildfire. Another benefit of using infrared gas Temperature, humidity, CO, CO2, smoke, light,
detection is that it can help firefighters to identify the sound, wind speed, soil moisture, GPS, and pressure
source of a fire. This can be very helpful in fighting sensors can all be used to detect a forest fire.
a fire, as it can allow firefighters to target their efforts  Temperature sensors are the most commonly
more effectively [40]. used type of sensor for Forest fire detection.
Another popular method is to look for changes in They can be used to detect hot spots, which are
the color of the smoke. This can be done by areas where the temperature is significantly
analyzing the color of the smoke over time or by higher than the surrounding area. Hot spots can
looking at the smoke plume from different angles. be caused by burning embers or fires.
yet another method is to look for changes in the  Humidity sensors can also be used to detect
brightness of the image. This can be caused by the forest fires. When the air is very dry, it can help
fire itself or by the smoke reflecting sunlight. By to spread a fire. By measuring the humidity,
looking at the brightness over time, it is possible to firefighters can get an idea of how dry the
track the progress of the fire. conditions are and whether or not a fire is likely
Using these camera devices (such as a drone, to spread.
satellite images, …), it is possible to perform this  CO and CO2 sensors can be used to detect the
prediction locally (in the Edge layer) or to send the presence of smoke. Smoke is a major indicator
data to the gateway, where a more capable device (in of a fire. These sensors can help to give
the Fog layer) may do so. Once the fire service firefighters an early warning that a fire is
receives the warning, they are able to make the present.
necessary preparations for the forecast blazes on  Light sensors can also be used to detect fires.
either the Edge or the Fog layer. If a wildfire is When there is a lot of smoke, the light levels will
predicted, a notification will be issued to the local be lower than normal. This can be used to help
fire department (see Figure 5). identify areas where a fire is burning.
All of these methods have their advantages and  Sound sensors can be used to detect the sound
disadvantages. However, by using them together, it of a fire. This can be helpful in determining the
is possible to get a more accurate picture of where a location of a fire.
fire is and how it is progressing. This information can  Wind speed sensors can be used to determine
then be used to help firefighters to better target their the rate at which a fire is spreading. The faster
efforts and to warn people in the area. the wind is blowing, the faster a fire will spread.

5452
 Journal of Theoretical and Applied Information Technology
15th October 2022. Vol.100. No 19
© 2022 Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

 Soil moisture sensors can be used to detect the 4. METHODOLOGY
moisture content of the soil. This information The search strategy used to retrieve papers on
can be used to determine the likelihood of a fire forest fire detection using deep learning and IoT

Figure 6: Forest fire monitoring and detection using deployed sensors

spreading. includes searching for papers in the Scopus database;

 GPS sensors can be used to track the location of which is the leading abstract and citation database of
a fire. This information can be used to help peer-reviewed research literature, making it the most
firefighters to determine the best way to get to popular academic search engines and databases,
the fire. which contains IEEE, Springer, and Elsevier
 Pressure sensors can be used to measure the publications. More than 1.8 billion cited references
amount of pressure in the air. This information dating back to 1970, and more than 84 million
can be used to determine the intensity of a fire. records are indexed in Scopus, making it one of the
 O3 sensors can be used to detect the presence of largest academic search engines in the world [42].
ozone. Ozone is produced when oxygen is
The search terms used were ("forest fire" OR
burned. This can be used to help identify areas
"wildfire") AND ("Machine Learning" OR "ML"
where a fire is present.
OR "DL" OR "Deep Learning") AND (IoT OR
 NH3 sensors can be used to detect the presence
"Internet of Things" OR "Wireless Sensor Network"
of ammonia. Ammonia is produced when
OR "WSN"). Our inclusion criteria for selected
vegetation is burned. This can be used to help
papers consider only those that had been published
identify areas where a fire is present.
in peer-reviewed journals, and deal with forest fire
These are the most used sensors that can be
detection using deep learning and IoT. 41 papers
employed to detect forest fires. Tree-mounted
were retrieved using the search strategy described
sensors that are protected from animals and people
above (see Figure 7). In addition, some of these
while collecting the most data possible from a wide
papers were excluded because they did not meet the
area are ideal for this purpose (temperature,
inclusion criteria or were just first pages of
humidity, CO concentration, and other relative
conference proceedings. The remaining 20 papers
measurements). In order to receive and communicate
were included in the review (see Table 1).
information to and from these sensors’ devices
without interference, data collected by these sensors
is sent to a gateway, which is supposed to be situated
in a high position. Gateways take data and send it to
the cloud, where it may be used for further analysis,
storage, and dashboarding (see Figure 6). By using
these sensors, firefighters can get an early warning
that a fire is present and take action to prevent it from
spreading.

5453
 Journal of Theoretical and Applied Information Technology
15th October 2022. Vol.100. No 19
© 2022 Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

Figure 7: Number of publications per year

5. FINDINGS
The following section analyzes the results of Table
1, 2, and 3. To begin, we will define the comparison
criteria used in Table 1:
 Year: the year in which the paper was published;
 Objective:
o CLA: Classification
o REG: Regression
o OD: Object Detection
o Other objectives
 ML or DL Model: the used machine or deep
learning models.
o ANN: Artificial Neural Network
o CNN: Convolutional Neural Network
o RNN: Recurrent Neural Network
o COELMAE: Compressed Sensing and
Online Extreme Learning Machine
Autoencoder
o LSTM: Long Short-Term Memory
o HNN: Hebbian Neural Network
o AutoML: Automated Machine Learning
o MLR: Multiple Linear Regression
o GRNN: General regression neural network
o LogR: Logistical Regression
o LVQ: Learning Vector Quantization
 EXP/SIM:
o EXP: Experience

5454
 Journal of Theoretical and Applied Information Technology
15th October 2022. Vol.100. No 19
© 2022 Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

EXP / SIM

DL Model

Results
Objective

Dataset
ML OR

Features
Paper

Title
Year

Table 1: Summary of reviewed papers

UCI

EXP / SIM
A smart approach for fire repository X, Y, month, day, FFMC,

DL Model

Results
Objective

Dataset
ML OR

Features
Paper

prediction under uncertain Forest Fires DMC, DC, ISI, temp

Title
2020
Year

[49] CLA SIM ML ACC=72%

conditions using machine Data Set (Temperature), RH
learning [63] (Humidity), wind, rain, area
+Images
UCI Atmospheric
X, Y, month, Data (wind
day, FFMC,
Detection of forest fires using ML: repository
speed, ambient
Sustainable R2=62.21% DMC, DC,Temperature,
ISI, temp
learningPeatland
2015

[51] machine technique: A REG SIM ML: REG

Multi Forest Fires relative Humidity,
RMSE=7.9 (Temperature), RH
2021

[52] Management with IoT

perspective and Data REG EXP - Captured
Data Set
Variate accumulated precipitation),
(Humidity), wind, rain, area
Analytics [63]
REG Ground Data (soil
Real-time identification of Temperature and GWL)
Asmoldering
comparative and flamingof
analysis CO, CO2, Smoke
combustion
LSTM andphasesARIMA in forest
for IoTConcentration,
Data: O3, NH3,Air CO
2016

[43] CLA EXP ANN ACC=82.5% Captured

2021

[53] using
enhanceda wireless sensor
real-time network- REG EXP LSTM
air pollutant - Captured Temperature,
Ground relative
station Data: O3,
based forecasting
levels multi-sensorusing system and
sensor Humidity,
NH3, CO and GPS
Artificial
fusion Neural station
with ground Network data
ForestSensor
Faulty 4.0: Digitalization
Data Detection of in
Re
2021

[62] forest Sensor

using the Internet Using
of - ML, DL - - - Humidity
2017

[56] Wireless Networks Faulty

view SIM ML: LogR ACC=98% Captured Temperature,
Things (IoT)
Logistical Regression
Effective deployment
Advancing of sensors
early forest fire Deploy
in a wireless sensor networks ANN: Light, Temperature, and
2018

[44] detection utilizing smart wireless ment

CLA SIMEXP HNN ACC=89.10% Captured
2021

[59] LVQ - - -
Sound.
using hebbian machine
sensor networks learning
WSN
Low-Latency technique
Energy-Efficient CNN:
Hallasan’s
2020

[45] Cyber-Physical Disaster System CLA EXP MobilNet ACC=91.49% Captured Images
Using Edge Deep Learning Fire V2 [64], Kaggle Temperature, wind, Humidity,
Burned area:
Early forest fire detection system Spread Wildfire fire intensity
Towards mountain fire safety RMSE=2.52 Temperature, Humidity, PM,
and SIM [65], and
2020

[46] using
usingwireless sensor
fire spread network CLA EXP RNN
predictive ML: ACC>99%
MAD=6.42 Captured
X,CO2, CO and Pressure
2020 2022

[60] and deep learning fire Burned (QGI UCI Y, month, day, FFMC,
analytics and mountain AutoML Spread:
area S) repository Temperature,
DMC, DC,Humidity,
ISI, temp Fire
containment
Real in iot street
time adaptive environment
lighting predict RMSE=1653
[55] OD EXP ANN MAD=1229 - Forest Fires
Captured (Temperature),
Detect, RH and
intensity light
system ion Data Set (Humidity), wind, rain, area
Light fault
[63]
Self-Localization
A Machine Learning in Large Scale
Approach
Locali ML: MAE=11 Intel Lab Temperature, Humidity, light,
2020
2022

[57] Wireless
[54] SensorPrediction
to Weather Network inUsing REG SIMSIM ANN - - -
zation MLR RMSE=17.5 dataset [66] and voltage
Machine
Wireless Learning
Sensor Networks
Accurate Location
IoT Sensor and DeepEstimation
Neuralof Temperature, Humidity, Soil
Locali CNN:
2022

[61]
[47] Smart Dustsbased
UsingWildfire
Machine CLA SIM
EXP GRNN ACC=91%
- - -
2020

Network zation Captured moisture, Pressure, Altitude

Learning DenseNet
Prediction System and GPS sensor
Forest fire
Anomaly detection
Detection forsystem
Internet ACC=99.75%
ML: Captured Temperature, Humidity, light
2022

[50] of
using wireless
Things Basedsensor networks
on Compressed CLA EXP
Autoen COEL ACC=86.36%
Pre=93.20% Shuttle intensity level, and CO level
REG (7000r)
2020

[58] and machine learning SIM -

Sensing and Online Extreme coder MAE Rec=92.63% (Harvard)
Learning Machine Autoencoder F1s=92.92%

o SIM: Simulation At first sight, scientific research aimed at forest

 Results: The proposed model's best results, fire detection using ML, and IoT or WSN techniques,
based on the metrics used; is increasing. More than 80% of the papers on the
o ACC: Accuracy Scopus database that address this problem, are
o R-square published over the last three years.
o MAE The majority of the papers studied deal with the
o RMSE problem of forest fire detection by classification
o RECALL [43]–[50]. While the papers [51]–[54] worked on
o F1-Score Regression to predict fire level, [51] determined the
 Dataset: the data used to train and evaluate the fire risk concerning the hydration level of a peat bog,
proposed Machine Learning model; [52] predicted humidity, and [54] estimated the post-
 Features: Temperature and Humidity, Soil fire air pollution level (forecasting three different air
moisture, Pressure… pollution levels such as CO, NH3, and O3) [53]. [55]
implemented an object detection technique in
combination with Fire Sensor. In [56]–[61] the

5455
 Journal of Theoretical and Applied Information Technology
15th October 2022. Vol.100. No 19
© 2022 Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

authors examined complementary aspects of forest the UCI repository (for Forest Fire), the Harvard
fire detection, including the deployment of WSNs, Shuttle, CortezMorais/Corsican Fire,
sensor location, and sensor data transmission errors. Hallasan's/Kaggle Wildfire/UCI, and Intel Lab,
Half of the examined papers (up to ~ 10) [43]– respectively.
[48], [50], [52], [53], [55] were based on real-world Table 2 clearly shows that the most commonly
experiments, others involved simulation, as for [49], used characteristics for detecting forest fires are
[51], [54], [56]–[61]. It is difficult to compare the Temperature, Humidity, CO, and Light. However,
results obtained from the works reviewed since they the papers that used particular characteristics were

Table 2: Summary of reviewed literature based on the used characteristics

Temperature

Soil moisture
Fire detector
Humidity

Pressure
Images

Sound
smoke
Paper

GWL
Wind
Light

CO2

NH3
GPS

PM
CO

O3
% 86% 79% 36% 29% 21% 14% 14% 14% 14% 7% 7% 7% 7% 7% 7% 7% 7%
[43] X X X X X X
[56] X X
[44] X X X
[45] X
[46] X X X X X X
[55] X X X X
[47] X X X X X
[48] X X X
[49] X X X X
[52] X X X X
[53] X X X
[60] X X X
[54] X X X
[50] X X X X

were acquired using several different metrics and trying to explore some specific areas. For the use of
datasets, although practically they are very inspiring. Sound [44], it seems to be interesting, and we will
By examining the models utilized, we can clearly consider it in our future works.
categorize these works into two broad categories: The choice of the communication module is
ML and DL. The following papers [48]–[52], [54], closely related to the targeted switching channel.
[56], [60] worked with regression techniques Apart from the articles that used a complete solution
(Logistical, Multi Linear...). On the other hand, the based on a weather station [48], [52], the most
majority of papers on DL [43]–[47], [53], [55], [57]– commonly used boards are Arduino ones [47], [53]
[59], [61] worked with classification techniques, and their generic NodeMCU [55], and the most used
with the exception of [55] which used image-based communication channel is Wi-Fi, with the exception
object detection. The Dataset is crucial for any of [50]; This latter added the SIM800 Module to be
scientific research using machine learning and IoT to able to use the GSM network. [43], [44]
detect forest fires. Papers [57], [59], [61], [62] made implemented the CC2430 and IRIS XM2110 boards,
no reference to their used dataset. Meanwhile, the both of which integrate ZigBee technology; the latter
following papers [43]–[48], [50], [52], [53], [55], boards of which has been discontinued. [45] used a
[56] collected their own datasets without making Raspberry pi board to capture images. [47] used a
them publicly available. On the other side, papers low-quality LoRa32u4 card from BSFrance that is no
[49], [51], [54], [58], [60] used public datasets from longer commercially available. [48], [52] collected

5456
 Journal of Theoretical and Applied Information Technology
15th October 2022. Vol.100. No 19
© 2022 Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

their own datasets using a weather station, the first addition, the datasets used to this end are mostly
on IBM Cloud and the second on MIMOS Malaysia. synthetic.
[43], [44], [50] performed storage on a single remote The majority of the studied papers use machine
PC using WSN architecture, the first two using learning approaches to address the problem of
ZigBee and the last using GSM and Radio wildfire detection; just few ones are considering
Frequency. [45] store d the captured images in deep learning models. The classification, regression,
Microsoft Azure in an IoT architecture with Wi-Fi as and object detection algorithms were by far the most
a communication channel. [46] used ThingsBoard as popular. According to the findings of our systematic
a Cloud solution in conjunction with WSN literature review, we conclude that scientific works
architecture and LoRA as communication having combined IoT and machine/deep learning,
technology. And [47], [53], [55] have built a Wi-Fi- achieved best performances and constitute a
based IoT architecture with various cloud choices, feasibility proof of the potential and the strength of
including Thingspeak, BOLT Cloud, and Micro soft their approach towards detecting and forecasting
Azure. wildfires.

Table 3: Summary of reviewed literature based on the used boards, sensors, communication channels, WSN/IoT
and cloud storage platforms
Board/module Communicati
Year Paper Sensors WSN/ IoT Cloud Storage
Connexion on
EC805-CO NeMOTO, S-100 ELT,
2016

[43] CC2430 MS5100 OGAM, SHT11 SENSIRON, ZigBee WSN Remote Server
LeadTek GPS 9546
Expansion Connector for Light,
2020 2020 2018

[44] IRIS XM2110 ZigBee WSN Personal Computer

Temperature and Sound

[45] Raspberry Pi Camera WIFI IoT Microsoft Azure

BME280, Nova SDS011, MH-Z14A-

[46] LoRa32u4 Lora WSN ThingsBoard
CO2, ZE07-CO

NodeMCU DHT11, Fire Sensor, Ambient Light

2020

[55] WIFI IoT Thingspeak

Raspberry PI check, LDR Module

Arduino
2020

[47] DHT11, YL-69, BMP280 GPS Sensor WIFI IoT BOLT Cloud
UNO+/Bolt
2020

[48] Weather station - - - IBM Cloud

Weather station
2021

[52] AGROMIST (Piezometers) LORA IoT MIMOS Malaysia

MQ-7 MQ-131 MQ-135

2021

[53] Arduino UNO WIFI IoT Microsoft Azure

Optical Dust Sensors
Arduino PC+ Arduino
2022

[50] nano+nrf24L01 DHT22, LDR, MQ9 GSM/RF WSN nano+nrf24L01+

+ SIM800 SIM800

6. CONCLUSION ACKNOWLEDGMENT
This work is supported by the Mohammed First
In this paper, we conducted a systematic review of University, under the PARA1 Program (Low-cost,
the scientific literature that addressed the challenge real-time Forest Fire Detection System based on
of detecting and predicting forest fires using IoT and Wireless Sensor Networks - SDF-RCSF).
machine/deep learning. The performed data analysis
revealed that the features mostly used to identify and REFERENCES:
detect forest fires are temperature, humidity, CO, and [1] M. Grari, I. Idrissi, M. Boukabous, O. Moussaoui,
light. We also learned that the communication M. Azizi, and M. Moussaoui, “Early Wildfire
channels deployed in this context are mostly based Detection using Machine Learning Model
on one of these protocols WIFI, ZigBee, or GSM. In

5457
 Journal of Theoretical and Applied Information Technology
15th October 2022. Vol.100. No 19
© 2022 Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

Deployed in the Fog / Edge Layers of IoT,” [13] P. P. Ray, “A survey of IoT cloud
Indones. J. Electr. Eng. Comput. Sci., 2022. platforms,” Futur. Comput. Informatics J., vol. 1,
[2] M. Yandouzi et al., “Review on Forest Fires no. 1–2, pp. 35–46, Dec. 2016, doi:
Detection and Prediction using Deep Learning 10.1016/J.FCIJ.2017.02.001.
and Drones,” J. Theor. Appl. Inf. Technol., 2022. [14] A. A. Ismail, H. S. Hamza, and A. M. Kotb,
[3] A. A. A. Alkhatib, “A Review on Forest Fire “Performance Evaluation of Open Source IoT
Detection Techniques:,” Platforms,” 2018 IEEE Glob. Conf. Internet
http://dx.doi.org/10.1155/2014/597368, vol. Things, GCIoT 2018, Jan. 2019, doi:
2014, Mar. 2014, doi: 10.1155/2014/597368. 10.1109/GCIOT.2018.8620130.
[4] C. S. Raghavendra, K. M. Sivalingam, and T. [15] M. Boukabous and M. Azizi, “Crime
Znati, Wireless sensor networks. Springer, 2006. prediction using a hybrid sentiment analysis
[5] J. A. Stankovic, “Wireless sensor networks,” approach based on the bidirectional encoder
Computer (Long. Beach. Calif)., vol. 41, no. 10, representations from transformers,” Indones. J.
pp. 92–95, 2008, doi: 10.1109/MC.2008.441. Electr. Eng. Comput. Sci., vol. 25, no. 2, Feb.
2022, doi: 10.11591/IJEECS.V25.I2.PP.
[6] I. Idrissi, M. Azizi, and O. Moussaoui, “A
Stratified IoT Deep Learning based Intrusion [16] M. Boukabous and M. Azizi, “Review of
Detection System,” in 2022 2nd International Learning-Based Techniques of Sentiment
Conference on Innovative Research in Applied Analysis for Security Purposes,” in Innovations
Science, Engineering and Technology (IRASET), in Smart Cities Applications Volume 4, Springer,
Mar. 2022, pp. 1–8, doi: Cham, 2021, pp. 96–109.
10.1109/IRASET52964.2022.9738045. [17] V. Nasteski, “An overview of the supervised
[7] I. Idrissi, M. Mostafa Azizi, and O. Moussaoui, machine learning methods,” 2017, doi:
“A Lightweight Optimized Deep Learning-based 10.20544/HORIZONS.B.04.1.17.P05.
Host-Intrusion Detection System Deployed on [18] I. Idrissi, M. Azizi, and O. Moussaoui, “IoT
the Edge for IoT,” Int. J. Comput. Digit. Syst., security with Deep Learning-based Intrusion
Jul. 2021. Detection Systems: A systematic literature
[8] S. Al-Sarawi, M. Anbar, K. Alieyan, and M. review,” in 4th International Conference on
Alzubaidi, “Internet of Things (IoT) Intelligent Computing in Data Sciences, ICDS
communication protocols: Review,” ICIT 2017 - 2020, Nov. 2020, pp. 1–10, doi:
8th Int. Conf. Inf. Technol. Proc., pp. 685–690, 10.1109/ICDS50568.2020.9268713.
Oct. 2017, doi: [19] A. Kherraki and R. El Ouazzani, “Deep
10.1109/ICITECH.2017.8079928. convolutional neural networks architecture for an
[9] J. Haxhibeqiri, E. De Poorter, I. Moerman, and J. efficient emergency vehicle classification in real-
Hoebeke, “A Survey of LoRaWAN for IoT: time traffic monitoring,” IAES Int. J. Artif.
From Technology to Application,” Sensors 2018, Intell., vol. 11, no. 1, pp. 110–120, Mar. 2022.
Vol. 18, Page 3995, vol. 18, no. 11, p. 3995, Nov. [20] I. Idrissi, M. Azizi, and O. Moussaoui, “An
2018, doi: 10.3390/S18113995. Unsupervised Generative Adversarial Network
[10] B. H. Çorak, F. Y. Okay, M. Güzel, Ş. Based-Host Intrusion Detection System for IoT
Murt, and S. Ozdemir, “Comparative Analysis of Devices,” Indones. J. Electr. Eng. Comput. Sci.,
IoT Communication Protocols,” 2018 Int. Symp. vol. 25, no. 2, 2022.
Networks, Comput. Commun. ISNCC 2018, Nov. [21] I. Idrissi, M. Azizi, and O. Moussaoui,
2018, doi: 10.1109/ISNCC.2018.8530963. “Accelerating the update of a DL-based IDS for
[11] D. Soni and A. Makwana, “A SURVEY IoT using deep transfer learning,” Indones. J.
ON MQTT: A PROTOCOL OF INTERNET OF Electr. Eng. Comput. Sci., vol. 23, no. 2, pp.
THINGS(IOT) MP-Index View project Analysis 1059–1067, Aug. 2021, doi:
and Survey on String Matching Algorithms for 10.11591/ijeecs.v23.i2.pp1059-1067.
Ontology Matching View project A SURVEY [22] A. Kherraki, M. Maqbool, and R. El Ouazzani,
ON MQTT: A PROTOCOL OF INTERNET OF “Traffic Scene Semantic Segmentation by Using
THINGS(IOT),” 2017. Several Deep Convolutional Neural Networks,”
[12] S. Tayeb, S. Latifi, and Y. Kim, “A survey 2021 3rd IEEE Middle East North Africa
on IoT communication and computation Commun. Conf., pp. 1–6, Dec. 2021, doi:
frameworks: An industrial perspective,” Mar. 10.1109/MENACOMM50742.2021.9678270.
2017, doi: 10.1109/CCWC.2017.7868354.

5458
 Journal of Theoretical and Applied Information Technology
15th October 2022. Vol.100. No 19
© 2022 Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

[23] M. Boukabous and M. Azizi, “A comparative Zerguit, and H. Bouali, “Review on maintenance
study of deep learning based language of photovoltaic systems based on deep learning
representation learning models,” Indones. J. and internet of things,” Indones. J. Electr. Eng.
Electr. Eng. Comput. Sci., vol. 22, no. 2, pp. 424– Comput. Sci., vol. 26, no. 2, May 2022.
432, May 2020, doi: [33] S. J. Goetz, G. J. Fiske, and A. G. Bunn, “Using
10.11591/ijeecs.v22.i2.pp424-432. satellite time-series data sets to analyze fire
[24] M. Berrahal and M. Azizi, “Improvement of disturbance and forest recovery across Canada,”
facial attributes’ estimation using Transfer Remote Sens. Environ., vol. 101, no. 3, pp. 352–
Learning,” in 2022 2nd International Conference 365, Apr. 2006, doi:
on Innovative Research in Applied Science, 10.1016/J.RSE.2006.01.011.
Engineering and Technology (IRASET), Mar. [34] A. Shamsoshoara, F. Afghah, A. Razi, L.
2022, pp. 1–7, doi: Zheng, P. Z. Fulé, and E. Blasch, “Aerial
10.1109/IRASET52964.2022.9737845. imagery pile burn detection using deep learning :
[25] M. Berrahal and M. Azizi, “Review of DL- The FLAME dataset,” vol. 193, 2021.
Based Generation Techniques of Augmented [35] H. Wu, H. Li, A. Shamsoshoara, A. Razi, and F.
Images using Portraits Specification,” in 4th Afghah, “Transfer Learning for Wildfire
International Conference on Intelligent Identification in UAV Imagery,” no. Dl, pp. 1–6,
Computing in Data Sciences, ICDS 2020, Nov. 2020.
2020, pp. 1–8, doi: [36] W. Schroeder, P. Oliva, L. Giglio, and I. A.
10.1109/ICDS50568.2020.9268710. Csiszar, “The New VIIRS 375m active fire
[26] M. Berrahal and M. Azizi, “Augmented Binary detection data product: Algorithm description
Multi-Labeled CNN for Practical Facial and initial assessment,” Remote Sens. Environ.,
Attribute Classification,” Indones. J. Electr. Eng. vol. 143, pp. 85–96, Mar. 2014, doi:
Comput. Sci., vol. 23, no. 2, pp. 973–979, Aug. 10.1016/j.rse.2013.12.008.
2021. [37] Y. Li, A. Vodacek, R. L. Kremens, A. Ononye,
[27] I. Idrissi, M. Boukabous, M. Azizi, O. and C. Tang, “A hybrid contextual approach to
Moussaoui, and H. El Fadili, “Toward a deep wildland fire detection using multispectral
learning-based intrusion detection system for iot imagery,” IEEE Trans. Geosci. Remote Sens.,
against botnet attacks,” IAES Int. J. Artif. Intell., vol. 43, no. 9, pp. 2115–2125, Sep. 2005, doi:
vol. 10, no. 1, pp. 110–120, Mar. 2021, doi: 10.1109/TGRS.2005.853935.
10.11591/ijai.v10.i1.pp110-120. [38] J. A. Pardo, W. G. Aguilar, and T. Toulkeridis,
[28] M. Boukabous and M. Azizi, “Multimodal “Wireless communication system for the
Sentiment Analysis using Audio and Text for transmission of thermal images from a UAV,”
Crime Detection,” in 2022 2nd International 2017 Chil. Conf. Electr. Electron. Eng. Inf.
Conference on Innovative Research in Applied Commun. Technol. CHILECON 2017 - Proc.,
Science, Engineering and Technology (IRASET), vol. 2017-Janua, pp. 1–5, Dec. 2017, doi:
Mar. 2022, pp. 1–5, doi: 10.1109/CHILECON.2017.8229690.
10.1109/IRASET52964.2022.9738175. [39] L. Hua and G. Shao, “The progress of
[29] Y. Hu, A. Huber, J. Anumula, and S.-C. Liu, operational forest fire monitoring with infrared
“Overcoming the vanishing gradient problem in remote sensing,” J. For. Res. 2016 282, vol. 28,
plain recurrent networks,” Jan. 2018, doi: no. 2, pp. 215–229, Dec. 2016, doi:
10.48550/arxiv.1801.06105. 10.1007/S11676-016-0361-8.
[30] A. Paullada, I. D. Raji, E. M. Bender, E. Denton, [40] C. Paton-Walsh et al., “Measurements of trace
and A. Hanna, “Data and its (dis)contents: A gas emissions from Australian forest fires and
survey of dataset development and use in correlations with coincident measurements of
machine learning research,” Patterns, vol. 2, no. aerosol optical depth,” J. Geophys. Res. Atmos.,
11, p. 100336, Nov. 2021, doi: vol. 110, no. D24, pp. 1–6, Dec. 2005, doi:
10.1016/J.PATTER.2021.100336. 10.1029/2005JD006202.
[31] M. Hossin and Sulaiman, “A REVIEW ON [41] P. K. Singh and A. Sharma, “An insight to forest
EVALUATION METRICS FOR DATA fire detection techniques using wireless sensor
CLASSIFICATION EVALUATIONS,” IJDKP networks,” 4th IEEE Int. Conf. Signal Process.
) Int. J. Data Min. Knowl. Manag. Process, vol. Comput. Control. ISPCC 2017, vol. 2017-Janua,
5, no. 2, 2020, doi: 10.5121/ijdkp.2015.5201. pp. 647–653, Sep. 2017, doi:
[32] Y. Hammoudi, I. Idrissi, M. Boukabous, Y. 10.1109/ISPCC.2017.8269757.

5459
 Journal of Theoretical and Applied Information Technology
15th October 2022. Vol.100. No 19
© 2022 Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

[42] Scopus, “Your brilliance, connected,” 2022. . sensor fusion with ground station data,” Cogent
[43] N. M. System, “Real-Time Identification of Eng., vol. 8, no. 1, 2021, doi:
Smoldering and Flaming Combustion Phases in 10.1080/23311916.2021.1936886.
Forest Using a Wireless Sensor,” no. Ccd, 2016, [54] S. S. Patil and B. M. Vidyavathi, “A Machine
doi: 10.3390/s16081228. Learning Approach to Weather Prediction in
[44] P. Pokhrel and H. S. B, Advancing Early Forest Wireless Sensor Networks,” Int. J. Adv. Comput.
Fire Detection Utilizing Smart Wireless Sensor Sci. Appl., vol. 13, no. 1, pp. 254–259, 2022, doi:
Networks. Springer International Publishing, 10.14569/IJACSA.2022.0130131.
2018. [55] S. Garg, S. Ahuja, and S. Randhawa, Real time
[45] Y. S. Patel, S. Banerjee, R. Misra, and S. K. adaptive street lighting system, vol. 1192 CCIS.
Das, “Low-Latency Energy-Efficient Cyber- 2020.
Physical Disaster System Using Edge Deep [56] T. Zhang, “Faulty Sensor Data Detection in
Learning,” in PervasiveHealth: Pervasive Wireless Sensor Networks Using Logistical
Computing Technologies for Healthcare, 2020, Regression,” 2017, doi:
vol. Part F1656, doi: 10.1145/3369740.3372752. 10.1109/ICDCSW.2017.37.
[46] B. Wiame, E. Moussati, M. Omar, and B. [57] A. V. Vidyapeetham, S. Sinha, and A. V.
Mohammed, “Early forest fire detection system Vidyapeetham, “Self - Localization in Large
using wireless sensor network and deep Scale Wireless Sensor Network Using Machine
learning,” Int. J. Adv. Comput. Sci. Appl., vol. 7, Learning,” pp. 1–5, 2020.
no. 2, pp. 469–505, 2021. [58] Y. Yu, X. Wu, and S. Yuan, “Anomaly
[47] V. Zope, T. Dadlani, A. Matai, P. Detection for Internet of Things Based on
Tembhurnikar, and R. Kalani, “IoT Sensor and Compressed Sensing and Online Extreme
Deep Neural Network based Wildfire Prediction Learning Machine Autoencoder,” in Journal of
System,” in Proceedings of the International Physics: Conference Series, 2020, vol. 1544, no.
Conference on Intelligent Computing and 1, doi: 10.1088/1742-6596/1544/1/012027.
Control Systems, ICICCS 2020, 2020, pp. 205– [59] S. M. Kusuma, K. N. Veena, and N. Aparna,
208, doi: 10.1109/ICICCS48265.2020.9120949. “Effective deployment of sensors in a wireless
[48] B. S. Negara, R. Kurniawan, M. Z. A. Nazri, S. sensor networks using hebbian machine learning
N. H. S. Abdullah, R. W. Saputra, and A. technique,” in Proceedings - IEEE 2021
Ismanto, “Riau Forest Fire Prediction using International Conference on Computing,
Supervised Machine Learning,” in Journal of Communication, and Intelligent Systems,
Physics: Conference Series, 2020, vol. 1566, no. ICCCIS 2021, 2021, pp. 268–274, doi:
1, doi: 10.1088/1742-6596/1566/1/012002. 10.1109/ICCCIS51004.2021.9397148.
[49] R. Sharma, S. Rani, and I. Memon, “A smart [60] Imran, N. Iqbal, S. Ahmad, and D. H. Kim,
approach for fire prediction under uncertain “Towards mountain fire safety using fire spread
conditions using machine learning,” Multimed. predictive analytics and mountain fire
Tools Appl., vol. 79, no. 37–38, pp. 28155– containment in iot environment,” Sustain., vol.
28168, 2020, doi: 10.1007/s11042-020-09347-x. 13, no. 5, pp. 1–23, 2021, doi:
[50] U. Dampage, L. Bandaranayake, R. 10.3390/su13052461.
Wanasinghe, and K. Kottahachchi, “Forest fire [61] S. Bashir, O. A. Malik, D. Teck, and C. Lai,
detection system using wireless sensor networks “Accurate Location Estimation of Smart Dusts
and machine learning,” Sci. Rep., pp. 1–11, 2022, Using Machine Learning,” 2022, doi:
doi: 10.1038/s41598-021-03882-9. 10.32604/cmc.2022.024269.
[51] A. Kansal, Y. Singh, N. Kumar, and V. [62] R. Singh, A. Gehlot, S. Vaseem Akram, A.
Mohindru, “Detection of Forest Fires using Kumar Thakur, D. Buddhi, and P. Kumar Das,
Machine Learning Technique : A Perspective,” “Forest 4.0: Digitalization of forest using the
pp. 241–245, 2015. Internet of Things (IoT),” J. King Saud Univ. -
[52] J. T. Liew et al., Sustainable Peatland Comput. Inf. Sci., 2021, doi:
Management with IoT and Data Analytics, vol. 10.1016/j.jksuci.2021.02.009.
629 IFIPAI. 2021. [63] P. Cortez and A. Morais, “A Data Mining
[53] G. Mani and R. Volety, “A comparative Approach to Predict Forest Fires using
analysis of LSTM and ARIMA for enhanced Meteorological Data,” Proc. 13th Port. Conf.
real-time air pollutant levels forecasting using Artif. Intell., pp. 512–523, 2007, [Online].

5460
 Journal of Theoretical and Applied Information Technology
15th October 2022. Vol.100. No 19
© 2022 Little Lion Scientific

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

Available:
http://www.dsi.uminho.pt/~pcortez/fires.pdf.
[64] Imran, S. Ahmad, and D. H. Kim, “A task
orchestration approach for efficient mountain
fire detection based on microservice and
predictive analysis in IoT environment,” J. Intell.
Fuzzy Syst., vol. 40, no. 3, pp. 5681–5696, Mar.
2021, doi: 10.3233/JIFS-201614.
[65] “Forest Fires Data Set|Kaggle,” 2007.
https://www.kaggle.com/elikplim/forest-fires-
data- (accessed May 27, 2022).
[66] “Intel Lab DataSet.”
http://db.csail.mit.edu/labdata/labdata.html
(accessed May 17, 2022).

5461