See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/361739530

REVIEW ON FOREST FIRES DETECTION AND PREDICTION USING DEEP

LEARNING AND DRONES

Article in Journal of Theoretical and Applied Information Technology · June 2022

CITATIONS READS

17 1,967

7 authors, including:

Mimoun Yandouzi Mounir Grari

Université Mohammed Premier Université Mohammed Premier
7 PUBLICATIONS 32 CITATIONS 9 PUBLICATIONS 41 CITATIONS

SEE PROFILE SEE PROFILE

Idriss Idrissi Omar Moussaoui

Université Mohammed Premier Université Mohammed Premier
20 PUBLICATIONS 239 CITATIONS 28 PUBLICATIONS 307 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Computer-Aided Experimentation (CAE) View project

Cloud computing resources optimization View project

All content following this page was uploaded by Idriss Idrissi on 04 July 2022.

The user has requested enhancement of the downloaded file.

 Journal of Theoretical and Applied Information Technology
30th June 2022. Vol.100. No 12
© 2022 Little Lion Scientific

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

REVIEW ON FOREST FIRES DETECTION AND PREDICTION

USING DEEP LEARNING AND DRONES
MIMOUN YANDOUZI1, MOUNIR GRARI2, IDRISS IDRISSI2, OMAR MOUSSAOUI2,
MOSTAFA AZIZI2, KAMAL GHOUMID1, AISSA KERKOUR ELMIAD1
1
LARI Lab., FSO, Mohammed First University, Oujda, Morocco
2
MATSI Lab., ESTO, Mohammed First University, Oujda, Morocco
E-mail: [email protected]

ABSTRACT

Forests everywhere in the world are essential components for protecting the biosphere. They strongly
contribute to the global carbon cycle and sustain a wide variety of plant and animal life forms. In many areas
of the globe, forest fires are one of the major threats to living beings; it leads the ecosystem in jeopardy,
including animals, plants, and even people. Last year, the Mediterranean and North African regions were
devastated by wildfires. The earlier discovery of forest fires is strongly required to save lives and properties.
Forest fires detection or prediction are difficult tasks because wildfires start small and are difficult to see from
a distance, and then can quickly spread to become large and dangerous fires. The combination of drones and
deep learning can be used to detect wildfires using images with high accuracy. The use of drones can help to
identify the location of the fire and its spreading area, while deep learning can be used to identify the
characteristics of the fire. This combination is a key foundation to create a system capable to detect wildfires
more accurately. This paper examines current state-of-the-art published research papers on detecting forest
fires using deep learning and drones.
Keywords: Forest Fires, Wildfire, Deep Learning, Drone, UAV

1. INTRODUCTION fatalities and injuries, as well as extensive property

and ecosystem damage [4].
Forests are one of the most vital ecosystems
on the planet. They are vital to the global carbon Forest fires are common in many parts of
cycle and home to a diverse range of plant and the world and can cause significant environmental
animal species. Humans benefit from forests in a damage. The hot, dry conditions required for a fire to
variety of ways, including timber products, water start can also cause it to spread quickly, destroying
resources, and recreation [1]. Wildfires, trees, homes, and other structures. A fire's smoke and
deforestation, or clear-cutting of forests, can have heat can also be harmful to humans and animals.
serious environmental consequences, including Forest fires can be started by a variety of factors such
biodiversity loss, soil degradation, and increased as lightning, careless campers, or arsonists. Fighting
greenhouse gas emissions. a forest fire often involves a large number of people
and agencies, including firefighters, police, and
Forest fires are a significant issue in many
forestry workers.
parts of the world. They can cause significant
environmental and wildlife damage, as well as Detecting forest fires is an important task
endanger human lives. In fact, in recent years, forest for protecting forests and preventing loss of property
fires have emerged as one of the world's leading and human life. Deep learning and drones have the
causes of natural disasters. This year's forest fires potential to significantly increase the speed and
have ravaged much of the Mediterranean and North accuracy of forest fire detection. The deep learning
Africa, thousands of fires had been reported as of algorithm can be trained to recognize forest fire
2021, consuming nearly hundreds of thousands of characteristics in aerial imagery. Drones can provide
hectares of land [2], from Turkey to Italy to Greece more frequent and accurate images of the forest
to Algeria and Morocco [3], the biggest wildfires in canopy than ground-based imagery can. There are
decades have ravaged these nations, claiming the several reasons why using drones to detect forest
lives of dozens and causing significant economic fires is preferable to satellite imagery (the most
damage. The fires have resulted in numerous known and used aerial imagery approach). For

4565
 Journal of Theoretical and Applied Information Technology
30th June 2022. Vol.100. No 12
© 2022 Little Lion Scientific

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

starters, satellites are not always capable of detecting learning network, for example, could be trained to
smaller fires, and drones can because they can fly recognize the features of a dog, such as its fur, eyes,
lower and collect more accurate data. Second, the and ears. Once trained, a deep learning network can
cost of employing drones is less than that of be used to classify images into various categories. A
employing satellites, and drones can be used for a deep learning network, for example, could be used to
variety of tasks, including fire detection, crop determine whether an image is of a dog or a cat [7].
monitoring, infrastructure inspection, and many
Deep learning for image classification has
others. Finally, drones are more versatile than
the advantage of learning to recognize patterns that
satellites, and they can send images out every day,
are too complex for humans to detect. This means
whereas satellites may only be able to send images
that deep learning networks can frequently
out every few days or weeks. This combination of
outperform traditional image classification methods
deep learning and drones has the potential to
[8].
significantly improve forest fire detection. The
purpose of this paper is to review current state-of- 2.1.2 Segmentation (SEG)
the-art methods for detecting forest fires using deep Deep learning methods have been applied
learning and drones. to the task of image segmentation, which involves
dividing an image into a set of regions that
2. BACKGROUNDS
correspond to different objects or classes of objects.
2.1 Computer Vision (CV) A deep neural network for image segmentation will
Computer vision [5] is the ability of a typically have several layers, each of which is
computer system to interpret and understand digital responsible for gradually refining the image
images. The ability to interpret and understand segmentation [9].
digital images has many practical applications, such
The first layer of a deep neural network
as in the field of automatic inspection and machine
(using CNN in particular) for image segmentation is
vision. In the field of automatic inspection, computer
typically a convolutional layer that learns to detect
vision can be used to inspect the quality of products
image features. These characteristics can be simple,
as they are being manufactured. In machine vision,
such as edges or corners, or complex, such as object
computer vision can be used to "see" the world and
shapes [10]. The convolutional layer produces a set
to guide robots or other machines. There are many
of feature maps, which is then passed on to the next
other applications of computer vision, including
layer. The following layer is typically a pooling
medical diagnosis, video surveillance, and 3D
layer, which reduces the dimensionality of the input
reconstruction. In each of these applications,
by averaging the values within a small patch of the
computer vision can be used to interpret and
input feature maps. This is followed by fully
understand digital images in order to achieve some
connected layers (one layer or more) that learn to
desired goal. In recent years, the application of deep
recognize the objects in the image. The network's
learning in computer vision has achieved great
final layer will generate a set of labels indicating the
success. Compared with traditional machine learning
type of object or objects present in the image [11].
methods, deep learning has the advantage of learning
multiple layers of representations for data, which can Although there are several picture
better capture the complex structure of data and segmentation methods, the most may well be divided
improve the performance of pattern recognition. into three categories: semantic, instance, and
Computer vision has many fields of study, including panoptic:
but not limited to: classification, segmentation, and
a) Semantic segmentation: is a technique for
object detection.
distinguishing items in a picture from the
2.1.1 Classification (CLA) background. This is normally accomplished by
Traditionally, image classification was recognizing and categorizing each object in the
performed by human experts who examined images image using a pre-determined set of labels. A
and determined which category they belonged to. segmentation algorithm, for example, maybe trained
Image classification is now performed by machines, to recognize different sorts of vehicles, such as
which can learn to recognize patterns in images more automobiles, trucks, and buses [9], [12].
accurately than humans can [6].
b) Instance segmentation: is a technique for
Deep learning networks can be trained to identifying and isolating specific objects in an image.
recognize the features of various objects in images This is frequently accomplished by identifying the
when it comes to image classification. A deep pixels that comprise an object and then grouping

4566
 Journal of Theoretical and Applied Information Technology
30th June 2022. Vol.100. No 12
© 2022 Little Lion Scientific

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

them together. A segmentation algorithm, for humans do. Deep learning aims to develop computer
example, could be trained to recognize different parts systems that can learn to recognize patterns and
of the human body, such as the head, torso, and legs insights in data, with the goal of allowing these
[13]. systems to make predictions or decisions in ways
similar to humans. Deep learning algorithms differ
c) Panoptic segmentation: is a technique for
from traditional machine learning algorithms in that
generating a three-dimensional representation of an
they use multiple layers of processing, where each
object from a single image. This is accomplished by
layer is a model that is trained on data, and the output
mapping the surface of the object onto a 3D grid and
of one layer is used as the input for the next. This
then reconstructing the object by interpolating the
approach allows deep learning algorithms to learn
data from the grid cells. This method is frequently
more complex patterns and insights in data than
used to detect objects that are too small or difficult
traditional machine learning algorithms [21].
to detect using other methods [14].
Deep learning has been demonstrated to be
2.1.3 Object Detection (OD)
extremely effective in a variety of applications,
Object detection in computer vision is the
including speech recognition [22], image recognition
task of locating a specific object in an image or video
[23], cybersecurity [24], and natural language
sequence. Object detection can be used to locate a
processing [25]. Some of the most successful deep
single object or a group of objects [15]. Many object
learning applications have been in the field of
detection algorithms, most notably deep learning-
computer vision, where deep learning algorithms
based object detectors, have been developed in
have achieved cutting-edge results in tasks such as
recent years [16]. Deep learning-based object
facial recognition and object recognition [26]. Deep
detectors have achieved cutting-edge performance
learning algorithms have also been shown to be
on a variety of object detection benchmarks. The
effective in the field of machine translation, with
YOLO algorithm, the SSD algorithm, and the Faster
results comparable to human translators [27]. There
R-CNN algorithm are three of the most popular deep
are numerous deep learning algorithms, some of the
learning-based object detectors [17].
most popular are convolutional neural networks
The YOLO algorithm is a fast and efficient (CNNs) [28], recurrent neural networks (RNNs)
object detector that detects objects at different scales [25], and generative adversarial networks (GANs)
[18]. The SSD algorithm is a fast and accurate object [29].
detector capable of detecting objects in real-time
2.2.1 Convolutional Neural Networks (CNN)
[19]. The Faster R-CNN algorithm is a deep
learning-based object detector that outperforms the CNNs [30] are a type of deep learning
competition on a variety of object detection algorithm that mimics the operations of the human
benchmarks [20]. brain. They are made up of a series of interconnected
layers, each of which serves a specific purpose. A
To summarize, classification is the most
convolutional neural network's first layer is typically
basic type of object categorization. Each object in the
a "kernel" layer that performs a series of
image is assigned a category by the algorithm, such
mathematical operations on the input data. This layer
as animal, plant, or person. The categories are fixed
is followed by a series of "feature" layers that extract
and will not change. Classification is a static process
specific data features. The final layer is a "classifier"
that does not take into account the object's location
layer that labels the input data [31].
or size in the image. Classification is simpler than
segmentation, where an image is divided into regions Each layer of neurons in a convolutional
by the algorithm, such as the sky, clouds, and neural network is "connected" to a certain number of
ground. Segmentation is a dynamic process that neighboring layers, and each connection has a
takes the location and size of the object in the image weight. When a neuron in one layer fires, it sends a
into account. And the most difficult type of object series of input signals to the neurons in the layer
categorization is object detection. The algorithm below it. Neighboring neurons then combine those
detects objects in an image, such as people or cars. input signals with their own to generate new signals,
Object detection is a dynamic process that takes into which are then passed on to the next layer of neurons
account the object's location and size in the image. [27]. This process is repeated until the signals reach
the output layer and are interpreted. The weight of
2.2 Deep Learning (DL)
each connection determines its strength. When a
Deep learning is a class of machine learning
neuron fires, it sends a signal to its neighbors, and
that focuses on algorithms that enable computer
the neighbor with the strongest input signal "fires"
systems to learn from data in the same way that

4567
 Journal of Theoretical and Applied Information Technology
30th June 2022. Vol.100. No 12
© 2022 Little Lion Scientific

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

the most. This is referred to as "spiking," and it objects and then detects objects in the video based on
allows the network to learn which connections are their similarity to the objects that have been learned.
important and which can be ignored. Convolutional YOLO has proven to be very accurate and can detect
neural networks have the advantage of being able to and recognize a wide range of objects, including
learn complex patterns and recognize objects in cars, people, and animals. It is also able to detect
images. They can also generalize these patterns, objects in a variety of environments, including
which means they can identify an object not only in outdoors and in closed spaces.
a single image but also in a series of images [32].
The YOLO algorithm is a variant of the
There are several variants of CNN that can CNN algorithm (Convolutional Neural Network), to
be used for object detection, such as R-CNN detect objects in an image, it employs a single neural
(Region-based CNN) and Bi-CNN (Bidirectional network. The detection network and the category
CNN); prediction network are the two main components of
the YOLO algorithm. The detection network detects
R-CNN combines a Region Proposal
objects in an image, and the category prediction
Network (RPN) with a Convolutional Neural
network predicts the object's category. The YOLO
Network (CNN) to detect objects in images. The
algorithm's detection network is made up of a series
RPN generates a set of region proposals, each of
of convolutional layers and a pooling layer. The
which is a rectangular area in the image that may
convolutional layers detect features in the image, and
contain an object. The CNN then classifies each
the pooling layer combines the features detected by
region proposal as containing an object or not [33].
the convolutional layers. A succession of fully
Bi-CNN is similar to R-CNN, but instead of connected layers makes up the category prediction
a single CNN, it uses two CNNs, one for the forward network.
pass and one for the backward pass. This allows the
2.2.3 Mobilenet
Bi-CNN to better learn the relationships between
object proposals and the objects they contain [34]. Mobilenet [35] is a deep learning network
architecture designed specifically for mobile
2.2.2 YOLO
devices. It is intended to provide efficient and
In recent years, there has been a growing accurate deep learning model execution on mobile
interest in developing systems for object detection devices while consuming minimal resources. The
and recognition in videos. One such system is Google Brain team created Mobilenet, which is
YOLO, or “You Only Look Once” [19]. YOLO is a based on the original MobileNet model.
real-time object detection system that can detect and
Mobilenet is suitable for mobile devices
recognize objects in videos with a high degree of
due to its small size and low computational
accuracy. It is fast and efficient and can be run on a
complexity. It also has a small number of parameters,
laptop or mobile device [20].
which reduces its computational requirements even
YOLO works by detecting objects in a further. Mobilenet achieves high accuracy while
video based on their appearance. It uses a deep keeping the model size to a minimum by employing
learning algorithm to learn the features of different a simplified architecture with only five layers.

Hiden Layers
Input Layer
Output Layer
Output
Input

Pooling layer

Convolutional layer
Fully connected
Figure 1: CNN layers architecture

4568
 Journal of Theoretical and Applied Information Technology
30th June 2022. Vol.100. No 12
© 2022 Little Lion Scientific

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

Mobilenet is effective and simple to use. It works network, but each layer in a deep residual network is
with both CNNs and RNNs.The Mobilenet connected to a number of residual layers [40].
architecture has been used to successfully train
The residual layers are connected to the
several deep learning models for mobile devices.
input layer, the hidden layer, and the output layer in
2.2.4 U-Net a deep residual network. The input layer is connected
to the first residual layer, the first residual layer is
The U-Net [36] is a neural network
connected to the second residual layer, and so on.
designed specifically for image recognition tasks.
The residual layers are also connected to each other.
The U-Net has been found to be especially effective
The connection between the first residual layer and
for image classification, semantic segmentation, and
the second residual layer is called the first residual
object detection. The main advantage of the U-net is
connection, the connection between the second
its ability to achieve good performance while
residual layer and the third residual layer is called the
minimizing the number of hidden layers and
second residual connection, and so on.
parameters. As a result, the U-net is relatively simple
to train and efficient in terms of computation. The U- The connections between the residual
net is made up of convolutional layers and pooling layers and the input layer, the hidden layer, and the
layers. Convolutional layers are used to extract output layer are called the input connections, the
features from the input image, whereas pooling hidden connections, and the output connections,
layers are used to reduce the number of parameters respectively.
and improve the network's efficiency. The U-final
2.2.7 AlexNet
net's layer is a fully connected layer that is used to
classify the image. AlexNet [41] is a convolutional neural
network that was designed by Alex Krizhevsky and
The U-net has been found to be especially
trained by himself and his colleagues. AlexNet
effective for image classification, semantic
consists of five convolutional layers and three fully-
segmentation, and object detection.
connected layers, it was originally trained to classify
2.2.5 DenseNet. images into 1000 object categories, such as “Zebra”,
“Horse”, and “Cat”. However, AlexNet has also been
DenseNet [37] is a visualization technique
applied to other tasks such as recognizing letters,
used to improve the accuracy and interpretability of
facial recognition, and automatic speech recognition.
deep neural networks. It was developed in response
to the limitations of the convolutional neural network 2.2.8 FireNET
(CNN) approach, which can lead to over-fitting and
FireNet [42] is a deep learning real-time fire
difficulty in interpreting the results.
detection project, it provides a labeled dataset, pre-
The key difference between DenseNet and trained deep learning models, and inference codes
most other deep learning architectures is that are included in the project. FireNet pre-trained deep
DenseNets connect every layer in a deep network learning models are a low-weight neural network
with every other layer, including the input and output architecture that is well-suited for mobile and
layers. This densely connected network topology is embedded applications and has an excellent
claimed to result in improved learning performance performance in real-time fire detection and
and feature reuse [38]. monitoring. On less powerful, low-cost single-board
computers like the Raspberry Pi 3B, the network
2.2.6 ResNet
operates at a very high frame rate of more than 24
Deep residual networks ResNet [39] have frames per second. The proposed neural network is
achieved great success in image recognition tasks. made up of three convolutional layers and four dense
Several state-of-the-art results on the ImageNet layers (including an output softmax layer).
database have been reported by training a deep
2.3 Drones / Unmanned Aerial Vehicles (UAV)
residual network.
UAVs [43], or drones, have become
To understand the reason for the superior commonplace in the military and law enforcement
performance of the deep residual networks, let's first communities all over the world. Drones used for
take a look at the structure of a deep residual reconnaissance, surveillance, and target acquisition
network. A deep residual network is a deep neural have allowed militaries to operate with greater
network with a large number of layers and a large precision and safety, while also providing troops
number of channels. The structure of a deep residual with greater situational awareness on the
network is similar to the structure of a deep neural battleground. Drones have been used in law

4569
 Journal of Theoretical and Applied Information Technology
30th June 2022. Vol.100. No 12
© 2022 Little Lion Scientific

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

enforcement for search and rescue operations, crime preventing the spread of wildfires as technology
scene investigation, and suspect tracking. advances.
In the internet of things (IoT) industry, 3. METHODOLOGY
drones are growing rapidly. IoT drones are remotely
In our study, we used the search criterion
controlled devices that can be used to collect data or
wildfire OR “wild fire” OR “forest fire”, and we got
perform tasks. They are also becoming more
around 33811 documents as results, indicating a high
sophisticated, with capabilities such as obstacle
level of interest in wildfires and forest fires among
avoidance and automated landing, which makes
academic researchers. Afterward, we narrowed our
them easier to use. Additionally, the development of
search results by including other words to the criteria
5G networks is expected to boost drone use in the
((wildfire OR "wild fire" OR "forest fire") AND
IoT industry, as they will provide the necessary
("deep learning" OR dl) AND (UAV OR drone)) that
bandwidth and speed to support the high-volume
are relevant to the objective of this review, and we
data traffic generated by drones [44]. They have a
obtained around 30 document results. These resulted
variety of uses, including delivery, surveillance, and
papers were published between 2017 and December
agriculture. They are frequently small, light, and
2021. As a consequence of the relevance of wildfire,
agile, which makes them excellent candidates for
we omitted certain findings because they only
difficult or dangerous tasks. Numerous drones are
represented the first few pages of conference
equipped with sensors that enable them to collect
proceedings and not actual articles, and we also
data about their surroundings and use it to optimize
excluded several irrelevant papers since they were
their efficiency or safety.
not related to wildfire, leaving us with 16 papers for
Drones [45] are now being used in this review. All of the information in this table (see
commercial and private sectors, in addition to the Table 1) was obtained from the Scopus database,
military and law enforcement communities. Drones which is the world's largest abstract and indexing
are being used by farmers to survey their crops and database of peer-reviewed literature, and which
identify potential problems such as pests and contains publications as well as conference
diseases. Drones are being used by power companies proceedings, patent records, and websites in the most
to inspect transmission lines and identify damage. important subject fields.
Aerial shots that were previously impossible to
4. FINDINGS
obtain are now possible thanks to the use of drones
by cinematographers. Drones are being used by We analyze below the results of Table 1.
private citizens to record weddings, birthday parties, First, we clarify the comparison criteria employed:
and other special events. Recently, there has been  Year: the year in which the paper was published;
discussion about using drones to assist in the  Deep Learning Model: the used Deep Learning
detection of wildfires in forest areas. This would be models;
a significant benefit because firefighters would be  Dataset: the used data to train and evaluate the
alerted to a potential fire much sooner than is proposed deep learning model;
currently possible.  Best achieved results: the best results for the
Drones have some advantages over proposed model based on the metrics used;
traditional wildfire detection methods. For example,  Type of detection: the used computer vision
can fly over difficult terrain that ground crews cannot techniques; CLA: classification, SEG:
access. They can also collect data faster and more segmentation, and OD: object detection;
efficiently than humans on the ground. Drones can  D/P/A: nature of the treated problem; D: fire
also be used to map the extent of a fire, which is detection; P: fire prediction; A: after fire;
useful for fighting the fire and preventing it from  F/S/ALL: the nature of the treated objects; F:
spreading. However, it can have some potential flame; S: smoke; ALL: a wide range of objects
drawbacks; for example, it may not be able to detect including, but not limited to, flame and smoke.
all fires because some are too small to detect from The majority of the papers in this review
the air. Bad weather conditions, such as fog or rain, focus on either detecting forest fires [46]–[55] or on
can also impede drone operations. detection and curation [56]–[58]. However, papers
[59]–[61] integrate both detection and prevention.
Overall, the potential applications of drones Fire detection are addressed in three major ways:
in wildfire detection are promising. Drones may classification [49], segmentation [46], [47], [53],
become an important tool for firefighters in [58], and object detection [48], [52], [54], [56], [57],
[61].

4570
 Journal of Theoretical and Applied Information Technology
30th June 2022. Vol.100. No 12
© 2022 Little Lion Scientific

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

Table 1: Description of papers studied

DL Model

detection

F/S/ALL
achieved

Type of
Dataset

results

D/P/A
Paper
Year

Best
Recall = 99.21%
Precision = 99.21%
MobileNet v3 MSCOCO+Collected
2021

OD

F/S
[52] Accuracy = 99.57%

D
YOLOv4 images (1844)
Inference time reduction =
75.68%
CLA : Accuracy = 76.23%
FLAME (Fire Luminosity
CNN SEG : Recall =83.88%

SEG
2021

F/S
[53] Airborne-based Machine

D
UNet Precision = 91.99%
learning Evaluation)
F1-score = 87.75%
DenseNet121

ALL
D-A
2021

OD
[56] Resnet152 COCO Dataset MobileNet : Accuracy = 87.5%
MobileNet v2
DenseNet121 UAV Dataset Kaggle
DenseNet :

ALL
D-A
2021

OD
[57] Resnet152 Image from the drone
Accuracy = 93.1%
MobileNet v2 Open-source photos
Author's own

CLA

SEG
D-P
2021

[59] Images captured by drone Accuracy = 81.97%

-
model
R-CNN
Data originated from the Precision = 43.4%

ALL
SEG
2021

[46]

D
ConFoBi project Accuracy = 92.4%
Collected from a
UNet++ forest fire in Andong, the

SEG

D-A
2020

[58] Specificity = 91.77% , 83.11%

-
UNet Republic of Korea, in
April 2020
Precision = 99,38%
DenseNet Images generated using

SEG
2020

F/S
[47] F1-Score = 98,16%

D
CycleGAN CycleGAN
Accuracy = 98.27%
2020

OD

F/S
[48] CNN Images captured by drone *****

Recall = 78% D
2020

OD

[54] YOLOv3 Videos from drone Precision = 84%

D

F
F1-score = 81%
MobileNet v2
CLA-

CNN 2096 images collected Parameters = 2.5M

SEG
2020

F/S

[55]
D

FireNet from the internet Accuracy = 99.3%

AlexNet
Two open-access dataset:
CNN Accuracy (UCM)= 97.84%
CLA

ALL
2020

[49] UCM Dataset

D

RNBFE Accuracy (WHU)= 97.%

WHU-RS Dataset
Precision = 82%
D-P
2019

F/S

[60] YOLOv3 Videos from drone Recall = 79%

-

F1-score = 81%
D-P
2019

OD

F/S

[61] CNN IRIS Dataset *****

Images from Google an

CLA

SEG
2018

F/S

[50] FireNet(DCNN) Baidu Accuracy = 98%

D
-

AInML lab Dataset

YUPENN Dataset
CLA
OD-
2017

F/S

[51] Bi-CNN BUAA Dataset Mean Accuracy = 93%

D

Maryland Dataset

4571
 Journal of Theoretical and Applied Information Technology
30th June 2022. Vol.100. No 12
© 2022 Little Lion Scientific

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

A mixture of classification and natural phenomena, such as sunlight reflecting off of

segmentation was found in the papers [50], [55], trees, these papers such as [46], [47], [57], [60], [61],
[59]. Only [51] used the classification and object [48]–[53], [55], [56] intended to predict not only
detection approaches. Among the papers working on flames but a variety of objects such as smoke, clouds,
object detection, [54] seeks to detect simply the fog and other objects available in the forests.
flame, whereas [48], [51], [52], [61] intend to detect
The results clearly show that object
the fire by the flame or/and the smoke. Reforestation
detection is the best way to solve our problem and
in burnt regions was addressed by [56], [57] in terms
that the DL YOLOv4 model used in the paper [42]
of detecting any form of object in these regions.
provided the best performance. Although the results
Table 2 displays the distribution of papers for almost all of the works are very encouraging, this
by the Deep Learning model as a solution to the is due in most cases to the images chosen, which are
forest fires problem. globally easily classifiable. Furthermore, the angles
of the images used do not correspond to those
The most used technique for forest fire
normally captured by a drone. To account for these
detection is object detection, with various versions of
constraints as well as the geographical and climatic
the YOLO and MobileNet models being the most
peculiarities of the Mediterranean region, especially
frequently adopted.
Morocco's Eastern region, we are working on a
The DL CycleGAN model is not mentioned dataset composed of images obtained from a drone
in Table 2 although it is used in the paper [47] for the (DJI Tello EDU drone) acquired as part of the project
simple reason that it is only exploited for image (Low-cost, real-time Forest Fire Detection System
generation during the data augmentation phase. It's based on Wireless Sensor Net-works - SDF-RCSF).
rare to find a paper that clearly explains how the data
The advantages of deep learning object
used for training, validation, and testing is divided
detection are thus visible in the detection of forest
up.
fires. Some of the benefits of using deep learning
The absence of publicly accessible drone object detection in the detection of forest fires
datasets on wildfires is one of the most common include:
obstacles encountered in the evaluated papers. Due
 Improved performance: Deep learning object
to safety, ethical considerations, and governmental
detection algorithms can learn features from
regulations, collecting comprehensive datasets on
data automatically. This leads to improved
wildfires is challenging. As a consequence, there are
performance because the features learned by the
no publicly accessible datasets that include fire data,
deep learning algorithm are more relevant to the
and several of the evaluated papers relied upon
task at hand.
images collected from the web [50], [52], [55]–[57].
 Improved accuracy: Deep learning detection
Papers [48], [54], [59], [60] experimented using
models can generalize better to previously
either images or videos taken by their own Drones.
unseen data. This means that deep learning
And others [46], [47], [49], [51], [53], [58], [61] are
object detection models can detect objects with
part of project-based research work.
greater accuracy even when the object is not
This review of the literature reveals some explicitly present in the training data.
difficulties, such as distinguishing fire from other
Table 1: Distribution of papers by Deep Learning model
own model
MobileNet

DanseNet

Author's
Bi-CNN

AlexNet
FireNet
R-CNN

ResNet
YOLO
CNN

Unet

SEG [46] [58] [47]

[53] [55]
SEG + CLA [55] [53] [55] [59]
[55] [50]
CLA [49]
CLA + OD [51]
[52] [52]
[48] [56] [56]
OD [54] [56]
[61] [57] [57]
[60] [57]

4572
 Journal of Theoretical and Applied Information Technology
30th June 2022. Vol.100. No 12
© 2022 Little Lion Scientific

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

 Reduced processing time: Deep learning forest, allowing more accurate fire location detection
detection models can learn from data than is possible with satellite imagery.
automatically. Because there is no need to
ACKNOWLEDGEMENTS
engineer features specifically for the task at
This work is supported by the Mohammed First
hand, the processing time is reduced.
University under the PARA1 Program (Low-cost,
 Improved efficiency: Deep learning detection
real-time Forest Fire Detection System based on
models can detect objects with greater accuracy
Wireless Sensor Networks - SDF-RCSF).
even when the object is not explicitly present in
the training data. This is due to the large deep REFERENCES:
learning models' ability to encode a large [1] A. A. A. Alkhatib, “A Review on Forest Fire
amount of information about the data. Detection Techniques:,”
The benefits of deep learning object http://dx.doi.org/10.1155/2014/597368, vol.
detection are thus visible in the detection of forest 2014, Mar. 2014, doi: 10.1155/2014/597368.
fires. Deep learning object detection algorithms can [2] “Copernicus: Mediterranean region evolves
learn features from data automatically. This leads to into wildfire hotspot, while fire intensity
improved performance because the features learned reaches new records in Turkey | Copernicus.”
by the deep learning algorithm are more relevant to https://atmosphere.copernicus.eu/copernicus-
the task at hand. Furthermore, deep learning object mediterranean-region-evolves-wildfire-
detection models generalize better to previously hotspot-while-fire-intensity-reaches-new-
unseen data. This means that deep learning object records (accessed Jan. 01, 2022).
detection models can detect objects with greater [3] “From wildfires to floods, the Mediterranean
accuracy even when the object is not explicitly bears the brunt of catastrophic climate change.”
present in the training data. https://www.rfi.fr/en/international/20210813-
Despite its benefits, using deep learning and from-wildfires-to-floods-the-mediterranean-
drones to detect and predict forest fires has a number bears-the-brunt-of-catastrophic-climate-
of drawbacks. To begin, deep learning algorithms change (accessed Sep. 18, 2021).
need a big quantity of data to train, and this data is [4] “Fires, forests and the future: a crisis raging out
often unavailable during forest fires. Second, deep of control? | WWF.”
learning techniques demand a significant amount of https://wwf.panda.org/discover/our_focus/fore
computer power, which may be costly and sts_practice/forest_publications_news_and_re
challenging to scale. Thirdly, using drones to gather ports/fires_forests/ (accessed Sep. 18, 2021).
data on forest fires may be risky and difficult to [5] R. Szeliski, Computer vision: algorithms and
arrange. Finally, the accuracy of deep learning applications. Springer Science \& Business
system predictions may be difficult to verify in the Media, 2010.
real world, and even minor errors can jeopardize the [6] M. T. Islam, B. M. N. Karim Siddique, S.
forest ecosystem. Rahman, and T. Jabid, “Image Recognition
with Deep Learning,” 2018 Int. Conf. Intell.
5. CONCLUSION Informatics Biomed. Sci. ICIIBMS 2018, pp.
Forest fires can be devastating, burning 106–110, Nov. 2018, doi:
houses, animal habitats, and wood while polluting 10.1109/ICIIBMS.2018.8550021.
the air with potentially toxic pollutants. Fire also [7] L. Perez and J. Wang, “The Effectiveness of
releases carbon dioxide into the environment. To Data Augmentation in Image Classification
avoid the uncontrolled broad spreading of forest using Deep Learning,” Dec. 2017, Accessed:
fires, it is vital to identify wildfires in an earlier stage Jan. 12, 2022. [Online]. Available:
and control their propagation. It is necessary to https://arxiv.org/abs/1712.04621v1.
mobilize appropriate fire apparatus and qualified [8] A. Mikołajczyk and M. Grochowski, “Data
operating people as rapidly as possible to the source augmentation for improving deep learning in
of the fire. image classification problem,” 2018 Int.
Interdiscip. PhD Work. IIPhDW 2018, pp.
In conclusion, this literature review has 117–122, Jun. 2018, doi:
found that deep learning-based classifiers are more 10.1109/IIPHDW.2018.8388338.
accurate than traditional methods, and object [9] A. Kherraki, M. Maqbool, and R. El Ouazzani,
detection was the most commonly used technique for “Traffic Scene Semantic Segmentation by
forest fire detection. We also found that drones can Using Several Deep Convolutional Neural
be used to obtain high-resolution images of the Networks,” 2021 3rd IEEE Middle East North

4573
 Journal of Theoretical and Applied Information Technology
30th June 2022. Vol.100. No 12
© 2022 Little Lion Scientific

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

Africa Commun. Conf., pp. 1–6, Dec. 2021, 2020. [Online]. Available:
doi: https://pjreddie.com/yolo/.
10.1109/MENACOMM50742.2021.9678270. [21] M. Boukabous and M. Azizi, “Review of
[10] S. Minaee, Y. Y. Boykov, F. Porikli, A. J. Learning-Based Techniques of Sentiment
Plaza, N. Kehtarnavaz, and D. Terzopoulos, Analysis for Security Purposes,” in Innovations
“Image Segmentation Using Deep Learning: A in Smart Cities Applications Volume 4,
Survey,” IEEE Trans. Pattern Anal. Mach. Springer, Cham, 2021, pp. 96–109.
Intell., 2021, doi: [22] M. Boukabous and M. Azizi, “Crime
10.1109/TPAMI.2021.3059968. prediction using a hybrid sentiment analysis
[11] S. Ghosh, N. Das, I. Das, and U. Maulik, approach based on the bidirectional encoder
“Understanding Deep Learning Techniques for representations from transformers,” Indones. J.
Image Segmentation,” ACM Comput. Surv., Electr. Eng. Comput. Sci., vol. 25, no. 2, Feb.
vol. 52, no. 4, Aug. 2019, doi: 2022, doi: 10.11591/IJEECS.V25.I2.PP.
10.1145/3329784. [23] M. Berrahal and M. Azizi, “Review of DL-
[12] M. Thoma, “A Survey of Semantic Based Generation Techniques of Augmented
Segmentation,” Feb. 2016, Accessed: Feb. 16, Images using Portraits Specification,” in 4th
2022. [Online]. Available: International Conference on Intelligent
https://arxiv.org/abs/1602.06541v2. Computing in Data Sciences, ICDS 2020, Nov.
[13] L. Yang, N. Xu, and A. Research, “Video 2020, pp. 1–8, doi:
Instance Segmentation.” pp. 5188–5197, 2019, 10.1109/ICDS50568.2020.9268710.
Accessed: Feb. 16, 2022. [Online]. Available: [24] I. Idrissi, M. Boukabous, M. Azizi, O.
https://github.com/. Moussaoui, and H. El Fadili, “Toward a deep
[14] A. Kirillov, K. He, R. Girshick, C. Rother, and learning-based intrusion detection system for
P. Dollar, “Panoptic Segmentation.” pp. 9404– IoT against botnet attacks,” IAES Int. J. Artif.
9413, 2019, Accessed: Feb. 16, 2022. [Online]. Intell., vol. 10, no. 1, pp. 110–120, Mar. 2021,
Available: https://arxiv.org/abs/1801.00868. doi: 10.11591/ijai.v10.i1.pp110-120.
[15] A. R. Pathak, M. Pandey, and S. Rautaray, [25] M. Boukabous and M. Azizi, “A comparative
“Application of Deep Learning for Object study of deep learning based language
Detection,” Procedia Comput. Sci., vol. 132, representation learning models,” Indones. J.
pp. 1706–1717, Jan. 2018, doi: Electr. Eng. Comput. Sci., vol. 22, no. 2, pp.
10.1016/J.PROCS.2018.05.144. 1032–1040, 2021, doi:
[16] X. Zhou, W. Gong, W. Fu, and F. Du, 10.11591/ijeecs.v22.i2.pp1032-1040.
“Application of deep learning in object [26] M. Berrahal and M. Azizi, “Optimal text-to-
detection,” Proc. - 16th IEEE/ACIS Int. Conf. image synthesis model for generating portrait
Comput. Inf. Sci. ICIS 2017, pp. 631–634, Jun. images using generative adversarial network
2017, doi: 10.1109/ICIS.2017.7960069. techniques,” Indones. J. Electr. Eng. Comput.
[17] Z. Q. Zhao, P. Zheng, S. T. Xu, and X. Wu, Sci., vol. 25, no. 2, Feb. 2022, doi:
“Object Detection with Deep Learning: A 10.11591/IJEECS.V25.I2.PP.
Review,” IEEE Trans. Neural Networks Learn. [27] I. Idrissi, M. Mostafa Azizi, and O. Moussaoui,
Syst., vol. 30, no. 11, pp. 3212–3232, Nov. “A Lightweight Optimized Deep Learning-
2019, doi: 10.1109/TNNLS.2018.2876865. based Host-Intrusion Detection System
[18] S. Wang, J. Zhao, N. Ta, X. Zhao, M. Xiao, and Deployed on the Edge for IoT,” Int. J. Comput.
H. Wei, “A real-time deep learning forest fire Digit. Syst., vol. 11, no. 1, pp. 209–216, 2022,
monitoring algorithm based on an improved doi: 10.12785/ijcds/110117.
Pruned + KD model,” J. Real-Time Image [28] I. Idrissi, M. Azizi, and O. Moussaoui,
Process. 2021 186, vol. 18, no. 6, pp. 2319– “Accelerating the update of a DL-based IDS for
2329, May 2021, doi: 10.1007/S11554-021- IoT using deep transfer learning,” Indones. J.
01124-9. Electr. Eng. Comput. Sci., vol. 23, no. 2, pp.
[19] D. Qi, W. Tan, Q. Yao, and J. Liu, 1059–1067, Aug. 2021, doi:
“YOLO5Face: Why Reinventing a Face 10.11591/ijeecs.v23.i2.pp1059-1067.
Detector,” Accessed: Jul. 27, 2021. [Online]. [29] I. Idrissi, M. Azizi, and O. Moussaoui, “An
Available: https://www.github.com/deepcam- Unsupervised Generative Adversarial Network
cn/yolov5-face. Based-Host Intrusion Detection System for IoT
[20] J. Redmon and A. Farhadi, “YOLOv3: An Devices,” Indones. J. Electr. Eng. Comput.
Incremental Improvement.” Accessed: Jul. 26, Sci., vol. 25, no. 2, 2022, [Online]. Available:

4574
 Journal of Theoretical and Applied Information Technology
30th June 2022. Vol.100. No 12
© 2022 Little Lion Scientific

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

http://ijeecs.iaescore.com/index.php/IJEECS/a 13, Feb. 2020, doi: 10.1007/s12652-020-

rticle/view/26735. 01754-0.
[30] Y. Lecun, Y. Bengio, and G. Hinton, “Deep [41] M. Z. Alom et al., “The History Began from
learning,” Nature, vol. 521, no. 7553. Nature AlexNet: A Comprehensive Survey on Deep
Publishing Group, pp. 436–444, May 27, 2015, Learning Approaches,” Mar. 2018.
doi: 10.1038/nature14539. [42] A. Jadon, M. Omama, A. Varshney, M. S.
[31] I. Idrissi, M. Azizi, and O. Moussaoui, “IoT Ansari, and R. Sharma, “FireNet: A
security with Deep Learning-based Intrusion Specialized Lightweight Fire & Smoke
Detection Systems: A systematic literature Detection Model for Real-Time IoT
review,” in 4th International Conference on Applications,” May 2019.
Intelligent Computing in Data Sciences, ICDS [43] G. Chmaj and H. Selvaraj, “Distributed
2020, Nov. 2020, pp. 1–10, doi: Processing Applications for UAV/drones: A
10.1109/ICDS50568.2020.9268713. Survey,” Adv. Intell. Syst. Comput., vol. 1089,
[32] M. Berrahal and M. Azizi, “Augmented Binary pp. 449–454, 2015, doi: 10.1007/978-3-319-
Multi-Labeled CNN for Practical Facial 08422-0_66.
Attribute Classification,” Indones. J. Electr. [44] T. Lagkas, V. Argyriou, S. Bibi, and P.
Eng. Comput. Sci., vol. 23, no. 2, pp. 973–979, Sarigiannidis, “UAV IoT Framework Views
Aug. 2021. and Challenges: Towards Protecting Drones as
[33] R. Girshick, J. Donahue, T. Darrell, and J. ‘Things,’” Sensors 2018, Vol. 18, Page 4015,
Malik, “Region-Based Convolutional vol. 18, no. 11, p. 4015, Nov. 2018, doi:
Networks for Accurate Object Detection and 10.3390/S18114015.
Segmentation,” IEEE Trans. Pattern Anal. [45] P. Cohn, A. Green, M. Langstaff, and M.
Mach. Intell., vol. 38, no. 1, pp. 142–158, Jan. Roller, “Commercial drones are here: The
2016, doi: 10.1109/TPAMI.2015.2437384. future of unmanned aerial systems,” 2017.
[34] R. Cai, X. Zhang, and H. Wang, “Bidirectional [46] D. Bulatov and F. Leidinger, “Instance
Recurrent Convolutional Neural Network for segmentation of deadwood objects in
Relation Classification,” pp. 756–765. combined optical and elevation data using
[35] A. G. Howard et al., “MobileNets: Efficient convolutional neural networks,”
Convolutional Neural Networks for Mobile https://doi.org/10.1117/12.2599837, vol.
Vision Applications,” Apr. 2017. 11863, pp. 299–308, Sep. 2021, doi:
[36] O. Ronneberger, P. Fischer, and T. Brox, “U- 10.1117/12.2599837.
Net: Convolutional Networks for Biomedical [47] M. Park, D. Q. Tran, D. Jung, and S. Park,
Image Segmentation,” Lect. Notes Comput. “Wildfire-Detection Method Using DenseNet
Sci. (including Subser. Lect. Notes Artif. Intell. and CycleGAN Data Augmentation-Based
Lect. Notes Bioinformatics), vol. 9351, pp. Remote Camera Imagery,” Remote Sens. 2020,
234–241, May 2015. Vol. 12, Page 3715, vol. 12, no. 22, p. 3715,
[37] G. Huang, Z. Liu, L. Van Der Maaten, and K. Nov. 2020, doi: 10.3390/RS12223715.
Q. Weinberger, “Densely Connected [48] G. D. Georgiev, G. Hristov, P. Zahariev, and D.
Convolutional Networks,” Proc. - 30th IEEE Kinaneva, “Forest Monitoring System for
Conf. Comput. Vis. Pattern Recognition, Early Fire Detection Based on Convolutional
CVPR 2017, vol. 2017-Janua, pp. 2261–2269, Neural Network and UAV imagery,” 28th Natl.
Aug. 2016, doi: 10.1109/CVPR.2017.243. Conf. with Int. Particip. TELECOM 2020 -
[38] A. Kherraki and R. El Ouazzani, “Deep Proc., pp. 57–60, Oct. 2020, doi:
convolutional neural networks architecture for 10.1109/TELECOM50385.2020.9299566.
an efficient emergency vehicle classification in [49] A. Rajagopal et al., “Fine-Tuned Residual
real-time traffic monitoring,” IAES Int. J. Artif. Network-Based Features with Latent Variable
Intell., vol. 11, no. 1, pp. 110–120, Mar. 2022. Support Vector Machine-Based Optimal Scene
[39] S. T. Krishna and H. K. Kalluri, “Deep learning Classification Model for Unmanned Aerial
and transfer learning approaches for image Vehicles,” IEEE Access, vol. 8, pp. 118396–
classification,” Int. J. Recent Technol. Eng., 118404, 2020, doi:
vol. 7, no. 5, pp. 427–432, 2019. 10.1109/ACCESS.2020.3004233.
[40] H. Chang, D. Zhao, C. H. Wu, L. Li, N. Si, and [50] Y. Zhao, J. Ma, X. Li, and J. Zhang, “Saliency
R. He, “Visualization of spatial matching Detection and Deep Learning-Based Wildfire
features during deep person re-identification,” Identification in UAV Imagery,” Sensors 2018,
J. Ambient Intell. Humaniz. Comput., pp. 1– Vol. 18, Page 712, vol. 18, no. 3, p. 712, Feb.

4575
View publication stats

Journal of Theoretical and Applied Information Technology

30th June 2022. Vol.100. No 12
© 2022 Little Lion Scientific

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

2018, doi: 10.3390/S18030712. doi: 10.1109/ICIAI.2019.8850815.

[51] J. Zheng, C. Xianbin, Z. Baochang, Y. Huang, [61] M. S. Allauddin, G. Sai Kiran, G. S. S. Raj
and Y. Hu, “Bi-heterogeneous Convolutional Kiran, G. Srinivas, G. U. Ratna Mouli, and P.
Neural Network for UAV-based dynamic V. Prasad, “Development of a surveillance
scene classification,” ICNS 2017 - ICNS system for forest fire detection and monitoring
CNS/ATM Challenges UAS Integr., Aug. using drones,” Int. Geosci. Remote Sens.
2017, doi: 10.1109/ICNSURV.2017.8011932. Symp., vol. 2019-July, pp. 9361–9363, 2019,
[52] S. Wang, J. Zhao, N. Ta, X. Zhao, M. Xiao, and doi: 10.1109/IGARSS.2019.8900436.
H. Wei, “ORIGINAL RESEARCH PAPER A
real ‑ time deep learning forest fire monitoring
algorithm based on an improved Pruned + KD
model,” no. 0123456789, 2021, doi:
10.1007/s11554-021-01124-9.
[53] A. Shamsoshoara, F. Afghah, A. Razi, L.
Zheng, P. Z. Fulé, and E. Blasch, “Aerial
imagery pile burn detection using deep
learning: The FLAME dataset,” Comput.
Networks, vol. 193, p. 108001, Jul. 2021, doi:
10.1016/J.COMNET.2021.108001.
[54] Z. Jiao et al., “A YOLOv3-based Learning
Strategy for Real-time UAV-based Forest Fire
Detection,” Proc. 32nd Chinese Control Decis.
Conf. CCDC 2020, pp. 4963–4967, Aug. 2020,
doi: 10.1109/CCDC49329.2020.9163816.
[55] H. Wu, H. Li, A. Shamsoshoara, A. Razi, and
F. Afghah, “Transfer Learning for Wildfire
Identification in UAV Imagery,” 2020 54th
Annu. Conf. Inf. Sci. Syst. CISS 2020, Mar.
2020, doi:
10.1109/CISS48834.2020.1570617429.
[56] G. V. S. Lohit and D. Bisht, “Seed Dispenser
using Drones and Deep Learning Techniques
for Reforestation,” Proc. - 5th Int. Conf.
Comput. Methodol. Commun. ICCMC 2021,
pp. 1275–1283, Apr. 2021, doi:
10.1109/ICCMC51019.2021.9418227.
[57] G. V. S. Lohit, “Reforestation Using Drones
and Deep Learning Techniques,” 2021 7th Int.
Conf. Adv. Comput. Commun. Syst. ICACCS
2021, pp. 847–852, Mar. 2021, doi:
10.1109/ICACCS51430.2021.9442053.
[58] D. Q. Tran, M. Park, D. Jung, and S. Park,
“Damage-Map Estimation Using UAV Images
and Deep Learning Algorithms for Disaster
Management System,” Remote Sens. 2020,
Vol. 12, Page 4169, vol. 12, no. 24, p. 4169,
Dec. 2020, doi: 10.3390/RS12244169.
[59] S. Zheng, W. Wang, Z. Liu, and Z. Wu, “Forest
Farm Fire Drone Monitoring System Based on
Deep Learning and Unmanned Aerial Vehicle
Imagery,” Math. Probl. Eng., vol. 2021, pp. 1–
13, Nov. 2021, doi: 10.1155/2021/3224164.
[60] Z. Jiao et al., “A Deep learning based forest fire
detection approach using uav and yolov3,” 1st
Int. Conf. Ind. Artif. Intell. IAI 2019, Jul. 2019,

4576