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

net/publication/363892041

Analysis of Deep Learning Methods for Early Wildfire Detection Systems: Review

Conference Paper · September 2022

DOI: 10.1109/IICETA54559.2022.9888515

CITATIONS READS
20 353

2 authors:

Ahmed Saleem Mahdi Sawsen Abdulhadi Mahmood

Mustansiriyah University Mustansiriyah University
5 PUBLICATIONS 40 CITATIONS 39 PUBLICATIONS 213 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Ahmed Saleem Mahdi on 08 October 2022.

The user has requested enhancement of the downloaded file.

 5th International Conference on Engineering Technology and its Applications 2022- (5thIICETA2022)

Analysis of Deep Learning Methods

for Early Wildfire Detection Systems: Review
2022 5th International Conference on Engineering Technology and its Applications (IICETA) | 978-1-6654-7215-9/22/$31.00 ©2022 IEEE | DOI: 10.1109/IICETA54559.2022.9888515

1st Ahmed S. Mahdi 2nd Sawsen A. Mahmood

Computer Science Department, College of Education Computer Science Department, College of Education
Mustansiriyah University Mustansiriyah University
Baghdad, Iraq Baghdad, Iraq
[email protected] [email protected]

Abstract— A wildfire is an uncontrollable fire that arises outside large amounts of data and identifying complex patterns and
of a specific focus, damages property, and poses a threat to relationships[1].
human life and health. Thousands of fires are started every year
for a variety of reasons, including dry seasons, thunderstorms, In this paper, variant types of fire detection datasets
and volcanic activity. However, in recent years, the human according to their ensemble way have been surveyed, and
factor has emerged as the main reason for irreparable forest used devices for capturing such as (Satellites, UAVs and
fires. In order to address the problem of wildfire detection, deep Geodata) in addition to sample categories (images and
learning model-based wildfire detection and recognition have videos). A comparison study was performed on state-of-the-
been attracted by researchers. Deep learning models using art techniques based on the detection accuracy of afire event
Convolutional Neural Network (CNN) are essentially required in images and videos. This paper focused on the deep learning
several dataset samples to train the network with high accuracy. models-based researches which adopted the common CNN
Meanwhile, the common wildfire datasets are small, repetitious architecture, such as U-Net [2][3], AlexNet [4][5], GoogleNet
and collected as images or videos samples from the Internet and [6], and YOLO (You Only Look Once) [7][8][9] in addition
experiments and there's no standard dataset to adopt in the to the researches that used R-CNN [10] and ANN [11] to build
training mode and performance evaluation task. This paper the Wildfire detection system.
presents a literature review of fire detection methods based on
Computer Vision systems. It’s focused on recent machine The rest part of this paper is arranged as follows; the
learning models used in addition to the datasets required to related works are stated in section II. The main challenges
construct future research projects for the wildfire detection faced bywildfire detection system design and
system. Further, a comparison study is performed to highlight implementation were discussed in section III. A comparison
the strength and weaknesses key points of current methods study was listed and described in section IV. finally, this
including detection accuracy. paper is concluding
in section V.
Keywords— Wildfire detection system, Wildfire Datasets, CNN,
ANN II. LITERATURE SURVEY
I. INTRODUCTION Since the 1990s, artificial intelligence methods AI have
been adopted in Wildfire Detection Methods, with early
Wildfires, whether caused by climate or by mankind, are
applications including expert systems and neural networks.
among the world's most dangerous and devastating disasters.
Since then, the area has advanced significantly in parallel with
And it's difficult to predict, extinguish, and cause significant
the widespread adoption of machine learning (ML)
financial losses, which adds to their complexity. Even though
technologies in the environmental sciences [12].
fire detection devices are extensively used in indoor
environments, they often require a long time burning to A. Machine Learning Methods: Machine learning is a set of
produce a considerable volume of Fire, which then triggers strategies for detecting patterns in data and utilizing the
the alarm. Furthermore, these devices cannot be used on a discovered patterns to predict data or another outcome
large-scale outdoor, such as in a wild area or a forest. Vision- that interest, Wildfire detection techniques divide into two
based fire detection systems, on the other hand, gather images generations; traditional machine learning-based methods
from cameras and detect fire instantly, making them ideal for and deep learning-based methods.
early fire detection. This type of device is very inexpensive.
Many studies have been conducted to monitor, detect, and 1) Traditional Methods:
prevent Forest Fires utilizing various AI methods and Decision trees: Multiple variable analyses are
strategies, for example, Machine Learning, Big Data, and performed using decision trees method. These methods
Remote Sensing. enable the prediction, explanation, description, and
Deep learning is a branch of machine learning that focuses classification of the output. Decision tree methods are used
on large-scale neural networks. Machine learning succeeds at for instances classification by sorting them down the tree from
tasks that humans struggle with, such as quickly synthesizing root to a leaf node. As a result, a decision tree is created, with
each branch representing a different decision scenario and its

978-1-6654-7215-9/22/$31.00 ©2022 IEEE 271

Authorized licensed use limited to: Universiti Teknikal Malaysia Melaka-UTEM. Downloaded on September 28,2022 at 10:05:59 UTC from IEEE Xplore. Restrictions apply.
 5th International Conference on Engineering Technology and its Applications 2022- (5thIICETA2022)
outcome. The decision tree-based learning method has been difficult to understand or express by a human. Other studies
used successfully by expert systems to capture knowledge for mentioned in TABLE I use the different architecture of CNN.
decision making. The focus is on methods that predict high
fire risks, trying to allow fire prevention steps to be taken, at
the very least, limiting their consequences. For prediction and TABLE I. CNN METHOD STUDIES WITH DATASETS
decision-making, two data mining techniques are used: Back-
propagation and decision tree forward neural networks [13]. Reference
Wildfire Datasets
no.
Support Vector Machines (SVM): is a classifier that [17] CCTV surveillance systems
identifies the hyperplane (decision border) in an N- [18] Corsican Fire Database
dimensional space that separates the boundary of each class, [19] Generative Adversarial Network
for data in n dimensions. SVM determines the best hyperplane [20] GIS database
by maximizing the distance between each class's nearest point [21] GIS database
and the decision boundary[12]. A new approach was proposed [22] GIS database
Ignition Raster Datasets IRDs & Raster Datasets
in [13] using the SVM approach for Wildfire detection with [5]
VRDs
both static and dynamic characteristics. This work has been [23] Images
put to the test with a large number of real video clips, and the [24] Images
results have proven its effectiveness. However, there are very [25] Optical Remote Sensing
weak static and dynamic fire characteristics for fires with tiny [2] Satellite imagery
areas or fire regions obscured in smoke, and hence the [26] Satellite imagery
detection accuracy is still low. Other approaches to this [27] UAV Images
problem will be tried in the future, such as segmenting smoke [28] UAV Images
first and considering both fire and smoke simultaneously. [29] UAV Videos
[30] UAV-based remote sensing technologies
2) Deep Learning Methods:
B. Wildfire Detection Datasets
Disastrous wildfires have become more frequent and
intense in recent years, exacerbated by changes in the climate 1) CCTV surveillance systems:
and long-term droughts. For fire-prevention plans, deep The first dataset type is the CCTV surveillance systems,
learning wildfire susceptibility mapping has proven to be which have always been in fixed places monitoring a lot of
effective, making it a vital tool in wildfire prevention. things like houses, streets and maybe forests, Recent
However, deep learning models have been used to predict developments enabled vision-based systems by integrated
wildfire susceptibility in the past, many researchers of processing for detecting fire when using convolutional neural
Wildfire Detection used deep learning [14]. networks for surveillance. However, such algorithms often
Artificial Neural Network (ANN): Artificial Neural require more processing memory and time, limiting their
Networks are a part of artificial intelligence that was created application in surveillance networks [6].
to mimic human reasoning and intellect. ANN can recognize 2) The Corsican Fire Database [31]:
patterns, organize data, and learn in the same way that the
brain does [15]. Authors in [11] an innovative approach to Is the second one, The Project Feux database [18], which
identifying Wildfire combining colour and multi-colour space is held by the University of Corsica (France) / Laboratory of
local binary patterns of both fire and smoke signatures using Science and Environment, contained 8229 Images in total
a single ANN to contribute to the previous research. (1135 Fire images and 7094 No Fire images); this dataset
obtained a validation accuracy of
Convolutional Neural Network (CNN): Recent over 97 %.
advancements in embedded processing have allowed vision-
based systems to detect fire during surveillance by employing 3) Kaggle Dataset [32]:
convolutional neural networks (CNNs) [16]. The study [6] This dataset contains approximately 500 general fire &
presents a low-cost fire detection CNN architecture for smoke images that were used to train the YOLOv2 detector in
surveillance videos compared to computational cost networks the experiments. These images were selected by Kaggle
like AlexNet, the model is built on the GoogleNet architecture Website from a wide range of real fire and smoke catastrophic
because of its moderate computational complexity and fit for events, study [7] used 400 images and labels in MATLAB,
the job. To balance efficiency and accuracy, the model is fine- and get an accuracy of about 93%.
tuned due to the nature of the problem and fire data.
Experiments using benchmark fire datasets show that the 4) General Advanced Networks (GAN):
framework proposed is effective and that it is suitable for fire This method is a new technique for computer vision
detection in closed-circuit television (CCTV) surveillance applications that have emerged in recent years. GAN
systems when compared to state-of-the-art methods. Research produces high-quality samples by combining the concepts of
[2] combined multiple CNNs structures to achieve higher zero-sum games with adversarial training and can learn and
performance in comparison to original circumstances sets, express features more effectively than traditional machine
and the circumstances where CNNs deviate from original learning algorithms. It's a deep learning-based learning model
algorithms can be advantageous, allowing for more that can generate massive amounts of data intelligently and is
complicated relations and rules, but at the cost of being used to intelligently generate massive volumes of data and has

272

Authorized licensed use limited to: Universiti Teknikal Malaysia Melaka-UTEM. Downloaded on September 28,2022 at 10:05:59 UTC from IEEE Xplore. Restrictions apply.
 5th International Conference on Engineering Technology and its Applications 2022- (5thIICETA2022)
been widely employed in fields including photos, text, speech, the model using historical forest fire locations as inputs and
and language. A Generator (G) and a Discriminator (D) are ignition factors as outputs. As a result, one of the most
the two main components of GAN. In [19] they employ important tasks is to collect high-quality historical forest fire
(GAN) to generate high-quality forest fire data because the data. A forest fire database was used in this study, which
effectiveness of deep learning algorithms is highly dependent contains 2530 historical fire locations from 2007 to 2016.
on the number of labelled examples, limiting their usage in These fire sites were generated from the Ministry of
contexts with limited sample numbers, the validation Agriculture and Rural Development of Vietnam's fire
accuracy is about 97.6 %. database, the last paper that used this dataset type is [22] the
input to the network was chosen from 15 components
5) Image-Based Dataset: collected from 6 different datasets, which resulted in 29
Fire recognition algorithms can be divided into two quantitative features. It compared it to three standard
categories. The first is based on classic image processing, methods: SegNet, multilayer perceptron, and logistic
while the second is based on convolutional neural networks. regression. In 4 of the 6 quantitative criteria studied,
Many researchers have used CNN technology to recognize AllConvNet surpasses the other three baseline approaches.
fire images. To compensate for the lack of training data, one
of the datasets created synthetic smoke images by inserting 7) Satellite-Based Dataset:
real or simulative smoke into a forest background [10], Satellite images accumulated through the many years of
another paper [23] used CNN technology to improve image fire monitoring operations on wildfire activities throughout
recognition by avoiding blindness and unpredictability in the the world, many studies used this dataset, The first study used
feature extraction process and potentially extracting deeper combines Remote Sensing, Big Data, and Data Mining
features, which can considerably improve flame image algorithms (ANN and SVM) to handle satellite photos
recognition accuracy, the maximum accuracy that reached gathered across broad areas [38] and use the information
using this technique is about 90.7%. Authors suggest a deep predicting the occurrence of wildfires and preventing
CNN improve the fire detection rate, so they employ transfer disasters, with results indicating a high prediction accuracy of
learning to neural network training and a window-based 98.32 %, in [39] high-resolution aerial photos are presented,
analysis technique, the other dataset contains 4000 images For the neural network training and testing, two datasets of
collected from the internet (50% Fire and 50% No Fire), and satellite RGB-pictures with varying spatial resolutions were
the results show that the accuracy is about 93.36% [8]. The created, containing 1457 and 393 high-resolution images,
study [9] used a combination of several fire image dataset to respectively. The third study [2] offers a new large-scale
analyze the study: BowFire[33], FD-dataset[34], Forestry active fire identification dataset, with +150,000 image
Images[35], VisiFire[36]. patches taken from Landsat-8 images of wildfires in several
locations. The dataset was divided into two parts, with the
6) Geo Data: first containing spectral images in 10-band with related
This section describes Geo datasets, one of the datasets outputs generated by three well-known handcrafted active fire
provided by [20] includes a huge number of wildfires in detection algorithms, and the second containing manually
California's Eastern Sierra, an area that is heavily impacted by annotated masks. And also, presents research on how
devastating wildfires every year. This data was created handcrafted algorithms used various CNNs architectures to
entirely using the data pipeline, and it is ideally suited for use be approximated these, as well as how combined
as a training and testing set for learning models or other automatically segmented patches training models provide
analyses of correlation between the nature of a given fire better results than the original methods, achieving 87.2 %
event and environmental variables present during that event's precision and 92.4 % recall with the most effective
duration. One of the dataset's most remarkable features is that combination.
the samples obtained from each wildfire's date-time relative
data tensors contain no unexpected randomness, even though 8) UAV Imagery:
both their random sample from the tensor and subsequent An unmanned aerial vehicle (UAV) can deliver high-
shuffling into training, validation and testing sets can be resolution geo-referenced images. The first paper employed
seeded at random, another study [21] produced 21 rasters deep CNN as the LM, one of the significant contributions was
monthly for the period 2013–2019 (20 corresponding to the to provide a benchmark for fire detection. There seems to be
variables and 1 to the fire class raster), the monthly multi- no uniform dataset in previous literature, to the authors’
band rasters were created by stacking these images. For each knowledge, consequently, comparing different fire detection
of these fire-presence multi-band rasters. The goal of the algorithms is difficult. As a result, they collected the fire
study was to use similar samples in both classes to make the sequences from prior studies as well as from their dataset and
model classification task difficult, requiring them to search created annotations [4]. The dataset in [40] used contained
for specific patterns and characteristics that could explain flame & smoke images of various colours and sizes taken
wildfires. As a result, after extracting a class one (fire) from various ranges. They reported an 83 % detection rate and
sample, a class zero (no-fire) sample was randomly selected a frame rate of 6.5 frames per second on a DJI Manifold
from a nearby location, only if the central pixel had not been onboard processor. Another study was conducted using video
sampled earlier was a fresh class zero sample extracted, and recordings of wildfires made by UAVs. There are 6600
if there were no pixels in this new sample's fire class raster positive photos and 15600 negatives images[44]. The dataset
categorized as one, and in the study [37] the wildfire in the study [27] contains 2100 forest fire images taken by e
modelling with Deep-NC utilizes supervised learning, with six-rotor UAV (DJI900) with a SONYA7 camera. Images

273

Authorized licensed use limited to: Universiti Teknikal Malaysia Melaka-UTEM. Downloaded on September 28,2022 at 10:05:59 UTC from IEEE Xplore. Restrictions apply.
 5th International Conference on Engineering Technology and its Applications 2022- (5thIICETA2022)
from a genuine Wildfire were used in the training data sets for burning of heaped trash in an Arizona pine forest. Video
the two groups of studies. The test photos are images of recordings and thermal heatmaps obtained by infrared
Wildfire gathered from the internet. This study presents a cameras are included in the dataset.
Wildfire Dataset obtained by drones during a controlled
TABLE II. SUMMARY OF WILDFIRE DETECTION METHODS: DATASET, TECHNIQUES AND ACCURACY
References Data Accuracy Precision Recall F1-Score
Dataset Type Techniques Year
No. Type % % % %
CCTV [6] CNN - GoogleNet architecture 2018 Videos 94.4 82 98 89
Corsican Fire [18] CNN - Dataset Augmentation 2020 Images 97.0 - - -
Images &
Kaggle [7] CNN - YOLOv2 2021 93.0 97 95.4 97
Videos
GAN then Histogram of Oriented
Generative Adversarial
[19] Gradients (HOG) then CNN) and 2020 Samples 97.6 - - -
Network (GAN)
SVM
Combination of Conventional
[23] Image Processing Techniques and 2019 Image 90.7 - - -
CNN
Collected Image
[8] CNN / MobileNet-V2 2020 Images 93.36 - - -
[9] CNN / Yolov5 2021 Images 94.7 70.5 75.6 -
[41] ResNet v2 2022 Images 99 100 98 99
CNN / Stochastic Spatial
[20] 2020 Geodata 97.5 97.2 95.5 96.5
Sampling
GIS dataset [21] CNN 2021 Geodata 95.3 90 87 88.77
[37] Deep Neural Computing 2021 Geodata 89.4 - - -
[22] All Conv Net 2021 Geodata 78.57 - - -
Artificial Neural Network and
[42] 2019 Nasa 98 97 98 97
SVM
Satellite-based dataset CNN / U-Net neural network
[39] 2020 Images 87.2 - - -
with the ResNet34
[2] CNN / U-Net 2021 Images 92.2 87 92.4 89.7
[4] CNN 2016 Images 97 - - -
Deep Convolutional Neural
[40] 2018 Images 98 - - -
Network (DCNN)
UAV Imagery Local Binary Pattern (LBP) and
[27] 2019 Images 99.81 - - -
SVM
[3] ANN 2021 Images 76.23 98 87 87
[43] DCNN 2021 Images 99.74 100 98 99

274

Authorized licensed use limited to: Universiti Teknikal Malaysia Melaka-UTEM. Downloaded on September 28,2022 at 10:05:59 UTC from IEEE Xplore. Restrictions apply.
 5th International Conference on Engineering Technology and its Applications 2022- (5thIICETA2022)
The collected dataset is annotated and labelled to help • UAV is a very reliable way to detect Wildfire’s event
other researchers in applying their fire detection methods [3]. due to its mobility and covering wide-scale forests
The last study used a normal fire detection benchmark dataset mounted camera and GPS devices.
because there is no wildfire detection official dataset [30].
• The Modern deep learning methods give high accuracy
III. COMPARISON STUDY and low-cost techniques compared with the traditional
method; However, it has some drawbacks as
In this paper, a comparison study is conducted to specify
demonstrated below:
the most efficient parameters that affected mainly wildfire
detection methods. According to the reviewed approaches o The most important issue in modern deep learning
listed and indicated in TABLE II, specific issues related to methods is to find a dataset for forest fire detection
higher performance evaluation of various methods were because there is no official benchmark dataset related
reached. The most accurate result is [27], which employs an to the subject of Wildfire Detection or datasets are
Unmanned Aerial Vehicle (UAV) image-based forest fire often not available, and it is sometimes not enough
detection by using the local binary pattern (LBP) feature due to The researcher can't start a forest fire to have
extraction and SVM classifier for detection purposes. An a good dataset, Extinguishing the fire is more
integration of LBP features extraction and SVM classifier important than photographing it, it's not possible to
leads to obtaining higher detection accuracy. Deep learning photograph the fire in high resolution in the presence
networks-based wildfire detection methods as stated in of smoke, forests are far from places of residence and
[4][18][19][20][40][43] respectively were achieved and photography.
specified second higher detection accuracy with an average
accuracy of around 97.6%, On the other hand, the worst o The training process of CNN is required high
accuracy result is founded in research [3] which offers an computational power and high computer specifics
aerial image of fires dataset along with fire detection Hardware.
methods, and developed an Artificial Neural Network (ANN) o The other difficulties encountered by the modern
method for wildfire detection. The demonstration of the detection methods are due to fire event and its related
accuracy results based on dataset type is depicted in Fig. 1. feature extraction process.
98 ACKNOWLEDGEMENTS
97 This research was supported by Mustansiriyah University
96 hosted by the Department of Computer Science in the college
95 of education, Baghdad, Iraq.
Accuracy %

94
REFERENCES
93
[1] Aristana Scourtas, “WILDFIRE DETECTION,” sagecontinuum.
92 https://sagecontinuum.org/science/wildfire-detection/ (accessed Dec.
91 26, 2021).
90 [2] G. H. de Almeida Pereira, A. M. Fusioka, B. T. Nassu, and R. Minetto,
“Active fire detection in Landsat-8 imagery: A large-scale dataset and
89 a deep-learning study,” ISPRS Journal of Photogrammetry and
Remote Sensing, vol. 178, pp. 171–186, 2021, doi:
10.1016/j.isprsjprs.2021.06.002.
[3] A. Shamsoshoara, F. Afghah, A. Razi, L. Zheng, P. Z. Fulé, and E.
Blasch, “Aerial imagery pile burn detection using deep learning: The
Datasets FLAME dataset,” Computer Networks, vol. 193, no. 2, p. 108001,
2021, doi: 10.1016/j.comnet.2021.108001.
[4] Q. Zhang, J. Xu, L. Xu, and H. Guo, “Deep Convolutional Neural
Fig. 1. The average accuracy of wildfire detection rate based on dataset Networks for Forest Fire Detection,” International Forum on
types. Management, Education and Information Technology Application,
pp. 568–575, 2016, doi: 10.2991/ifmeita-16.2016.105.
IV. CONCLUSION [5] G. Zhang, M. Wang, and K. Liu, “Forest Fire Susceptibility Modeling
Using a Convolutional Neural Network for Yunnan Province of
This paper presented the recent works as a literature China,” International Journal of Disaster Risk Science, vol. 10, no. 3,
review on fire detection methods including; the related pp. 386–403, 2019, doi: 10.1007/s13753-019-00233-1.
studies based on traditional methods, deep learning methods [6] K. Muhammad, J. Ahmad, I. Mehmood, S. Rho, and S. W. Baik,
based on computer vision in addition to the most important “Convolutional Neural Networks Based Fire Detection in
Surveillance Videos,” IEEE Access, vol. 6, no. June, pp. 18174–
dataset used for training and testing systems. In this literature 18183, 2018, doi: 10.1109/ACCESS.2018.2812835.
review, a comparison study was conducted between the state- [7] S. Saponara, A. Elhanashi, and A. Gagliardi, “Real-time video
of-the-art methods of wildfire detection to present the best fire/smoke detection based on CNN in antifire surveillance systems,”
approaches for future works. The following conclusions have Journal of Real-Time Image Processing, vol. 18, no. 3, pp. 889–900,
been reached: 2021, doi: 10.1007/s11554-020-01044-0.
[8] H. Pan, D. Badawi, and A. E. Cetin, “Computationally efficient
wildfire detection method using a deep convolutional network pruned
via fourier analysis,” Sensors (Switzerland), vol. 20, no. 10, 2020, doi:
10.3390/s20102891.

275

Authorized licensed use limited to: Universiti Teknikal Malaysia Melaka-UTEM. Downloaded on September 28,2022 at 10:05:59 UTC from IEEE Xplore. Restrictions apply.
 5th International Conference on Engineering Technology and its Applications 2022- (5thIICETA2022)
[9] R. Xu, H. Lin, K. Lu, L. Cao, and Y. Liu, “A forest fire detection [29] S. S. Moumgiakmas, G. G. Samatas, and G. A. Papakostas,
system based on ensemble learning,” Forests, vol. 12, no. 2, pp. 1–17, “Computer vision for fire detection on uavs—from software to
2021, doi: 10.3390/f12020217. hardware,” Future Internet, vol. 13, no. 8, 2021, doi:
[10] Q. X. Zhang, G. H. Lin, Y. M. Zhang, G. Xu, and J. J. Wang, 10.3390/fi13080200.
“Wildland Forest Fire Smoke Detection Based on Faster R-CNN [30] A. Bouguettaya, H. Zarzour, A. M. Taberkit, and A. Kechida, “A
using Synthetic Smoke Images,” Procedia Engineering, vol. 211, pp. Review on Early Wildfire Detection from Unmanned Aerial Vehicles
441–446, 2018, doi: 10.1016/j.proeng.2017.12.034. using Deep Learning-Based Computer Vision Algorithms,” Signal
[11] F. M. A. Hossain, Y. M. Zhang, and M. A. Tonima, “Forest fire flame Processing, vol. 190, 2021, p. 108309, doi:
and smoke detection from uav-captured images using fire-specific 10.1016/j.sigpro.2021.108309.
color features and multi-color space local binary pattern,” Journal of [31] “Corsican Fire Database,” 2017. http://cfdb.univ-corse.fr/
Unmanned Vehicle Systems, vol. 8, no. 4, pp. 285–309, 2020, doi: [32] A. Upadhyay, “kaggle Dataset,” 2019.
10.1139/juvs-2020-0009. [33] W. D. de Oliveira, “Bowfire dataset,” 2019.
[12] P. Jain, S. C. P. Coogan, S. G. Subramanian, M. Crowley, S. Taylor, www.bitbucket.org/gbdi/bowfire-dataset/downloads
and M. D. Flannigan, “A review of machine learning applications in [34] “FD Dataset.” www.nnmtl.cn/EFDNet
wildfire science and management,” Environmental Reviews, vol. 28, [35] “Forestry Images,” 2018. www.forestryimages.org
no. 4, pp. 478–505, 2020, doi: 10.1139/er-2020-0019. [36] “VisiFire Dataset.” http://signal.ee.bilkent.edu.tr/VisiFire/
[13] A. Karouni, B. Daya, and P. Chauvet, “Applying decision tree [37] H. Van Le, D. Anh, C. Trung, P. Quoc, and V. H. Thi, “Ecological
algorithm and neural networks to predict forest fires in Lebanon,” J Informatics A new approach of deep neural computing for spatial
Theor Appl Inf Technol, vol. 63, no. 2, pp. 282–291, 2014. prediction of wildfire danger at tropical climate areas,” vol. 63, no.
[14] M. Tonini, M. D’andrea, G. Biondi, S. D. Esposti, A. Trucchia, and June 2020, 2021.
P. Fiorucci, “A machine learning-based approach for wildfire [38] Ouladsayadyounes, “Wildfires,” 2018.
susceptibility mapping. The case study of the liguria region in italy,” https://github.com/ouladsayadyounes/Wildfires.
Geosciences (Switzerland), vol. 10, no. 3, 2020, doi: [39] V. Khryashchev and R. Larionov, “Wildfire Segmentation on Satellite
10.3390/geosciences10030105. Images using Deep Learning,” Moscow Workshop on Electronic and
[15] K. Angayarkkani and N. Radhakrishnan, “An Intelligent System For Networking Technologies, MWENT 2020 - Proceedings, 2020, pp. 1–
Effective Forest Fire Detection Using Spatial Data,” vol. 7, no. 1, pp. 5, doi: 10.1109/MWENT47943.2020.9067475.
202–208, 2010. [40] Y. Zhao, J. Ma, X. Li, and J. Zhang, “Saliency detection and deep
[16] S. A. Mahmood, A. M. Abid, and W. A. K. Naser, “Contextual learning-based wildfire identification in uav imagery,” Sensors
Anomaly Detection Based Video Surveillance System,” in 2021 11th (Switzerland), vol. 18, no. 3, 2018, doi: 10.3390/s18030712.
IEEE International Conference on Control System, Computing and [41] R. K. Mohammed, “A real-time forest fire and smoke detection
Engineering (ICCSCE), 2021, pp. 120–125. doi: system using deep learning,” Int. J. Nonlinear Anal. Appl, vol. 13,
10.1109/ICCSCE52189.2021.9530859. 2022, pp. 2008–6822, doi: 10.22075/ijnaa.2022.5899.
[17] A. Dhall, A. Dhasade, A. Nalwade, M. R. Mohan, and V. Kulkarni, [42] Y. O. Sayad, H. Mousannif, and H. Al Moatassime, “Predictive
“A survey on systematic approaches in managing forest fires,” modeling of wildfires: A new dataset and machine learning
Applied Geography, vol. 121, no. November 2018, p. 102266, 2020, approach,” Fire Safety Journal, vol. 104, no. January, , 2019 pp. 130–
doi: 10.1016/j.apgeog.2020.102266. 146, doi: 10.1016/j.firesaf.2019.01.006.
[18] M. S. Khan, “Application of Convolutional Neural Networks For [43] A. Bouguettaya, H. Zarzour, A. M. Taberkit, and A. Kechida, “A
Wild Fire Detection,” IEEE SoutheastCon, 2020. Review on Early Wildfire Detection from Unmanned Aerial Vehicles
[19] Z. Liu, K. Zhang, C. Wang, and S. Huang, “Research on the using Deep Learning-Based Computer Vision Algorithms,” Signal
identification method for the forest fire based on deep learning,” Optik Processing, vol. 190, p. 108309, 2021, doi:
(Stuttg), vol. 223, no. February, p. 165491, 2020, doi: 10.1016/j.sigpro.2021.108309.
10.1016/j.ijleo.2020.165491. [44] D. Alexandrov, E. Pertseva, I. Berman, I. Pantiukhin, and A.
[20] M. E. Green, K. Kaiser, and N. Shenton, “Modeling Wildfire Kapitonov, “Analysis of machine learning methods for wildfire
Perimeter Evolution using Deep Neural Networks,” 2020. security monitoring with an unmanned aerial vehicles,” Conference of
[21] A. Bjånes, R. de La Fuente, and P. Mena, “A deep learning ensemble Open Innovation Association, FRUCT, vol. 2019-April, 2019, pp. 3–
model for wildfire susceptibility mapping,” Ecological Informatics, 9, doi: 10.23919/FRUCT.2019.8711917
vol. 65, no. May, p. 101397, 2021, doi:
10.1016/j.ecoinf.2021.101397.
[22] J. R. Bergado, C. Persello, K. Reinke, and A. Stein, “Predicting
wildfire burns from big geodata using deep learning,” Safety Science,
vol. 140, no. April, p. 105276, 2021, doi: 10.1016/j.ssci.2021.105276.
[23] Y. Wang, L. Dang, and J. Ren, “Forest fire image recognition based
on convolutional neural network,” Journal of Algorithms and
Computational Technology, vol. 13, 2019, doi:
10.1177/1748302619887689.
[24] K. Srinivas and M. Dua, “Fog Computing and Deep CNN Based
Efficient Approach to Early Forest Fire Detection with Unmanned
Aerial Vehicles,” Lecture Notes in Networks and Systems, vol. 98 ,
2020, pp. 646–652, doi: 10.1007/978-3-030-33846-6_69.
[25] P. Barmpoutis, P. Papaioannou, K. Dimitropoulos, and N.
Grammalidis, “A review on early forest fire detection systems using
optical remote sensing,” Sensors (Switzerland), vol. 20, no. 22 , 2020,
pp. 1–26, doi: 10.3390/s20226442.
[26] S. Treneska, “Wildfire detection from UAV collected images using
transfer learning,” 2021.
[27] Y. Chen et al., “UAV image-based forest fire detection approach
using convolutional neural network,” Proceedings of the 14th IEEE
Conference on Industrial Electronics and Applications, ICIEA 2019,
2019, pp. 2118–2123, doi: 10.1109/ICIEA.2019.8833958.
[28] H. Wu, H. Li, A. Shamsoshoara, A. Razi, and F. Afghah, “Transfer
Learning for Wildfire Identification in UAV Imagery,” no. Dl, 2020,
pp. 1–6.

276

Authorized licensed use limited to: Universiti Teknikal Malaysia Melaka-UTEM. Downloaded on September 28,2022 at 10:05:59 UTC from IEEE Xplore. Restrictions apply.
View publication stats