Grenze International Journal of Engineering and Technology, Jan Issue

Forest Fire Detection using AI

Arnav kandhari1, Priyanshu Mohan Ginkal2 and K.Kalaiselvi3
1-2
Department of Networking and communications ,SRM Institute Of Science and Technology, Kattankulathur 603203,
Chennai, Tamil Nadu, India
Email: ak8062@[Link], ginkalpriyanshu@[Link]
3
Faculty of Engineering and technology, Department of Networking and communications ,SRM Institute Of Science and
Technology, Kattankulathur-603203, Chennai, Tamil Nadu, India
Email: mkkalai1981@[Link]

Abstract— Forests are crucial resources for human existence and social progress because they
safeguard the earth's natural balance. However, forest fires occur regularly as a result of
unmanaged human activities and erratic environmental conditions. These fires are the most
destructive natural disasters to forest resources and the human environment. Forest fires have
become more common in this scenario as a result of climate change, human activity, and other
things. Forest fire detection and monitoring has become a global problem for forest fire
prevention organisations. Currently, forest fire detection methods mostly include vigils,
observation from watch towers, and, more recently, satellite monitoring. Although observation
from watch towers is simple and feasible, there are various obstacles. First and foremost, this
strategy requires Keywords — Wireless sensor networks(WSN), Computer vision, aerial
vehicles(UAV), YOLO, Internet of things(IOT), Moderate Resolution Imaging
Spectroradiometer(MODIS) , mAP (mean Average Precision) Many financial and material
resources, as well as a skilled labour force. Second, there are numerous issues with fire
protection manpower, such as inattention, absenteeism from the post, a lack of real-time
monitoring capability, and limited area coverage. A variety of parameters limit the spectrum of
use of satellite detection systems, reducing their efficiency in forest fire detection. A satellite
monitoring system, for example, has a long scanning period and a low resolution of saturated
pixel dots of photographs. Another issue is that cloud layers might obscure photos during the
scanning process, and real-time quantitative quantification of fire parameters is extremely
difficult to perform. Given the limitations of traditional monitoring, we propose a cloud-based
wireless sensor network technology and Explain how it can be used as a monitoring system. This
system can monitor real-time related parameters such as temperature and relative humidity
and promptly transfer the data to the monitoring center's computer. The computer will assess
and manage the collected data. In comparison to standard baroscopic data and fundamental
forest resource data, the system can conduct an immediate assessment of a potential fire threat.
The analytical results will then be provided to the relevant department as the policy-making
foundation on which the department will decide whether to battle fires or prevent fires.

Index Terms— YOLO, Fire, Precision, MODIS

I. INTRODUCTION
The specter of forest fires looms as a relentless and pervasive environmental challenge, casting a long shadowover

Grenze ID: [Link].10.1.315

© Grenze Scientific Society, 2024
ecosystems, biodiversity, and human settlements. These infernos, once ignited, rapidly engulf vast swaths of
forests, leaving a trail of ecological devastation, displacing wildlife, and threatening the very fabric of human
communities. The need for efficient forest fire detection and response mechanisms has never been more
[Link] methods of forest fire surveillance, predominantly reliant on human observation, are beset
by inherent limitations. Their coverage is bound by geographical constraints, rendering vast and remote
forestedareas inaccessible to timely scrutiny. Furthermore, human surveillance is prone to delays, increasing the
risk of catastrophic fire [Link] project focuses on the development of a forest firedetection system that
utilizes YOLO, a state-of-the-art object detection algorithm. The system is designed to process real- time video
footage from a camera and identifythe presence of smoke or flames in forests or other woodedareas. The main
objective of the project is to develop an accurate and reliable forest fire detection system that can identify the
presence of smoke or flames in real-time. The system should be able to detect forest fires in different weather
conditions, lighting conditions, and forest types.
The purpose of this project is to provide a solution for detecting forest fires in real-time that can be implemented
in surveillance systems or drones. The forest fire detectionsystem can help mitigate the impact of forest fires by
enabling quick responses from firefighting teams. Additionally, the project aims to showcase the capabilitiesof
the YOLO algorithm in the field of object detection andhighlight its potential for other real-world applications
suchas wildlife monitoring and search and rescue operations. Inthe pages that follow, we embark on a voyage of
exploration, innovation, and implementation—a journey that seeks to harness the extraordinary potential of AI-
driven image analysis for the paramount task of forest fire detection. In doing so, we strive to redefine the way
we safeguard our forests, protect our precious ecosystems, andensure the safety and security of the communities
that coexist within these vital natural landscapes. Through our project on forest fire detection using AI, we
endeavor to contribute to a future where the threat of forest fires is met with swifter, more accurate, and more
proactive responses,ultimately fostering a harmonious coexistence between nature and civilization.Yolo5 is a
state-of-the- art object detection algorithm that hasdemonstrated high accuracyin detecting objects in real-time,
including smoke and flames in forests. The Yolo5 algorithm is optimized for speed, allowing for real-time
detection of forest fires in real-time. This can provide timelyalerts to firefighting teams and enable quick
response to prevent the spread of fires. Yolo5 can run efficiently on a variety of hardware, including drones and
low-power embedded devices. This can make it a cost-effective solution for forest fire detection without
requiring specialized equipment. Yolo5 can be trained on a wide range of data, allowing it to adapt to different
weather conditions, lighting conditions, and forest types. This can make it a versatile tool for forest fire detection
in various settings. Yolo5 can be integrated with other systems, such as weather sensors and satellite imagery, to
provide acomprehensive solution for forest fire monitoring and prevention. Overall, the proposed system using
Yolo5 has the potential to offer accurate, real- time forest fire detectionwith minimal hardware requirements,
while also beingadaptable and flexible to different weather and forest conditions.

II LITERATURE SURVEY
Shixiao Wu and colleagues' research, which focuses on real-time detection, early fire detection, and reducing
false alarms, addresses significant issues in forest fire detection. The authors use traditional object detection
techniques such as SSD, Faster R-CNN, and YOLO variations (tiny-yolo-voc, tiny-yolo-voc1, yolo- voc.2.0, and
YOLOv3). SSD exhibits enhanced real-time capabilities, increased detection accuracy, and early fire detection
proficiency.
[1]In this research, Feifei Xie et al. offer an original method for using unmanned aerial vehicles (UAVs) to detect
forest fires from the air. Small fire points, smoke interference, and complicated backgrounds present problems
for traditional deep learning algorithms in UAV aerial photography situations. The research suggests a strategy
using an improved Faster RCNN algorithm with transfer learning to address these problems.. This novel
approach holds promise for effective forest fire detection in UAV aerial photography scenarios, offering
significant potential in real-world forest fire prevention and mitigation efforts. [2].The work by E. Menaka et al.
covers the crucial problem of forest fires in tropical locations, especially in nations like India where a sizable
amount of the forested areas is prone to annual fires, having detrimental ecological, economic, and social effects.
This method provides faster, reproducible identification of sparse and high-dry regions over large areas with
higher accuracy. [3]
The research by Xiwen Chen et al. discusses the shortcomings of current forest monitoring techniques in giving
accurate and timely information about wildfires, particularly in their early phases. They emphasize the potential
of drone systems as useful tools for early detection and assessment of wildland fires, particularly in remote and
inhospitable forested areas, due to their 3D mobility, low flight altitude, and rapid deployment capabilities. The
2858
research also highlights the lack of comprehensive aerial datasets as a result of aircraft restrictions during
managed

2859
burns and wildfires. To close this gap, the authors present a multi-modal UAV dataset that includes dual-feed
side- by-side recordings of a prescribed fire in Northern Arizona that include both RGB and thermal pictures.
Additionally, they suggest an enhanced deep learning-based approach for recognizing fire and smoke pixels.
with significantly improved accuracy compared to single-channel video feeds. This dataset, enriched with
contextual information, offers a valuable resource for the development of data-driven fire detection and modeling
techniques, promising advancements in wildfire management and intervention strategies. [4]
Georgi Dimitrov Georgiev et al.'s paper addresses the critical issue of forest fires and the imperative need for
early and accurate detection to mitigate environmental degradation. The paper introduces an autonomous early
fire detection system designed for reliability and minimal human interaction. Central to this system is an object
detection method based on a convolutional neural network, detailed in the paper's main section. Unlike
traditional lookout towers and satellite monitoring, the system utilizes live video feeds from unmanned aerial
vehicles (UAVs) patrolling high-risk areas, providing a broader field of view. To enhance fire probability
predictions, both optical and thermal cameras onboard the UAV are employed. The authors have developed a
web-based platform using Node- RED to present real-time acquired data and notify relevant stakeholders. This
comprehensive approach offers a promising solution to timely forest fire detection and intervention, reducing the
environmental impact of these devastating events. [5]
The important contrast between different types of forest fires is discussed in Alexander A. Khamukhin et al.'s
research, which focuses in particular on dangerous crown fires with their quick spread rates. In order to create
efficient firefighting techniques, early detection is [Link] innovative approach has the potential to
significantly enhance early detection and response strategies for forest fires, contributing to improved
firefighting efforts and reduced environmental damage. [6]
The paper by C. Nithesh et al. emphasizes the significant contribution of wildlife and forests to our ecosystem,
which is increasingly threatened by climate change-induced higher temperatures and the consequent disastrous
forest fires. In addition to being a direct threat, these fires also cause resource depletion. This research offers an
innovative approach that takes use of recent developments in UAV navigation and image processing. It improves
the coordination amongst numerous drones by utilizing the synergy of these fields and a recently created
algorithm. This novel strategy offers major advancements in the early identification and suppression of forest
fires, providing a means to mitigate the damage caused by these increasingly frequent and severe wildfires. [7]
The study by Xiaojun Bai et al. offers a comprehensive system that combines classification and object detection
models for precise forest fire recognition. It introduces a revolutionary method for early forest fire warning. The
method starts with VGG network optimization and transfer learning implementation to train smoke and flame
identification models. The YOLO network for fire detection is introduced to address the issue of spotting small
flames in the early stages of a fire, including enhancements to boost feature extraction and multi-scale feature
fusion. The findings of categorization and detection are combined in the final stage, which uses a decision tree
method for joint decision-making in fire warning.[8]
In order to improve the initial active fire identification method, Liu Shixing et al.'s research goes into the theory
and techniques of using MODIS (Moderate Resolution Imaging Spectroradiometer) data. The paper uses high-
quality MODIS data as a learning tool and proposes an improved technique that uses variance between-class and
a smoke plume mask. Notably, the brightness temperature threshold for possible fire pixels has been raised to
305K, allowing hot fire spots and cool fire spots to be distinguished from the background using the variation
between-class in thermal infrared spectral channels. Furthermore, by employing a smoke plume mask, the system
successfully differentiates low- temperature smolder locations. [9]
The study by Cang Naimeng et al. proposes a novel detection approach and its particular implementation to
overcome the difficulties caused by background interference components in typical forest fire smoke
monitoring videos, such as clouds and skylines. The first step of the process entails gathering early forest fire
smoke source detection footage by multi-rotor aircraft close-range aerial [Link] largest linked area of
these suspected smoke regions is identified as a suspected smoke area if a localized reduction in high-frequency
energy is discovered.. Simulation experiments validate the method's efficacy, demonstrating accurate smoke
detection in video frame images [10]
Research by Wenjie Wang et al. addresses the crucial problem of correctly identifying forest fires in their early
phases in light of the significant harm they may cause to both human communities and the natural environment.
Early warning system innovation was required due to the limitations of traditional fire warning technologies'
sensitivity and accuracy. The study examines several fire detection techniques with a particular emphasis on
using deep learning technology to detect forest fires. The study, in particular, increases calculation efficiency
through strategies for data and feature augmentation. The result of this research is the creation of a lightweight
real-time
2860
fire detection technology that combines extensive experimentation with the training of a deep learning YOLO
model. The results demonstrate the method's great sensitivity and accuracy, especially in flame datasets, making
it a promising advancement in forest fire early warning systems. [11]

III. METHODOLOGY
The methodology for a project is crucial for ensuring its success. It defines the approach, steps, and procedures to
be followed to achieve the project's goals. In this project, the methodology consists of several stages, including
data collection, pre-processing, feature extraction, model training, and testing. Each stage is essential and
contributes to the overall success of the project .The overall work is done in several steps. Figure 1.1 given below
the steps of our workflow.

Figure 1.1 Architecture diagram for Data preprocessing

The methodology used in this project ensures that the model developed is accurate and reliable. By following a
systematic approach, the project team can identify and address any issues or challenges encountered in the
process. Additionally, the methodology used can be replicated for future projects, enabling the development of
efficient and effective solutions.
A. DATASET COLLECTION & PRE-PROCESSING
1) Dataset Collection
The success of any object detection system depends heavily on the quality and quantity of the dataset used for
training. In this project, we collected a dataset of images from various sources, such as satellite imagery and
ground- based sensors, to capture the diverse range of forest fire scenarios. The dataset was collected with great
care and attention to detail. Only high-quality images were chosen for the dataset. To ensure that the images
were suitable for use in the training process, they were checked for factors such as resolution, color balance, and
cloud cover. Only those images that met the desired quality criteria were added to the dataset. In total, we
collected over 1000 images of varying sizes, providing the system with sufficient data to train on and be able to
detect forest fires with a high level of accuracy shown in the Figure 1.2 a

Figure 1.2 a Data collection

2) Dataset Pre-Processing
In our forest fire detection project using YOLO5, we observed that the images in our dataset were of different
sizes, resolutions, and color schemes. To ensure the accuracy of our system, we performed pre-processing of our

2861
dataset. One of the pre-processing steps that we took was to resize all the images to a specific dimension of
640x640 pixels. This was done to ensure that all images had the same size and dimensions, which is important
for accurate feature extraction. Overall, the pre-processing step plays a crucial role in ensuring the accuracy and
efficiency of our forest fire detection system. By standardizing the dataset through resizing and color
normalization, we were able to improve the quality of the dataset and enhance the performance of our object
detection model. shown in the Figure 1.2 b

Figure 1.2 b Data preprocessing

B) TRAINING & VALIDATION

The dataset was split into three parts for training, validation, and testing purposes. Specifically, 70% of the
dataset was used for training the model, 20% for validation, and the remaining 10% for testing. This partitioning
approach helped ensure that the model was not overfitting or underfitting the data in shown in Figure 1.3

Figure 1.3 Training and Validation

During the training process, different hyperparameters were used to fine-tune the model to achieve the best
possible results. The training process was also monitored closely to check for convergence and prevent
overfitting. In addition to the hyperparameters, the model's performance was evaluated using various metrics,
including accuracy, precision, recall, and F1-score. These metrics helped to determine the model's effectiveness
in correctly detecting Forest Fires. Overall, the training phase was a crucial step in the development of the forest
fire detection system. It ensured that the model was optimized to achieve high accuracy, which was necessary for
the successful deployment of the system in real-world scenarios.

III. RESULTS AND DISCUSSIONS

A) Training Results
Confusion Matrix -the test set, indicating that it is capable of accurately detecting fires in real-world images. The
mAP score is a useful measure of the overall performance of our system, taking into account both precision and
recall.
The mAP metric is widely used in the field of computer vision and serves as a key performance indicator for
object detection systems. It provides a standardized way to compare the performance of different models and can
guide the development of more accurate and efficient object detection systems for forest fire detection shown in
the figure 1.4
A confusion matrix is a useful tool for evaluating the performance of a classification model, including forest fire
detection. It is a table that shows the number of correctly and incorrectly classified instances of each class in the

2862
model's predictions. The rows of the matrix represent the actual class labels, while the columns represent the
predicted class [Link] the context of forest fire detection, a confusion matrix would allow us to evaluate the
performance of our model in identifying areas of forest fire. We can determine the number of true positive (TP),
true negative (TN), false positive (FP), and false negative (FN) instances in our predictions.

Figure 1.4 Efficiency

For instance, a TP would represent an image where a forest fire was correctly detected, while a TN would
represent an image where there was no forest fire and the model correctly predicted this. On the other hand, an
FP would represent an image where there was no forest fire, but the model incorrectly predicted that there was,
while an FN would represent an image where there was a forest fire, but the model failed to detect it shown in
Figure 1.5
B) Loss Curves

Figure 1.5 loss

we can identify areas where our model needs improvement and take steps to enhance its performance. We can
also use it to adjust our model's threshold and optimize it for specific use cases. Overall, a confusion matrix is an
essential tool for evaluating the performance of our forest fire detection system.
C) Mean Average Precision
mAP (mean Average Precision) is a widely used evaluation metric in the field of computer vision, particularly in
object detection tasks. It is a measure of the average precision of a model at various levels of recall.
In the context of our forest fire detection project, we used mAP to evaluate the performance of our object
detection system. The system achieved an mAP of 71% on In our forest fire detection project, we utilized the
RoboFlow platform for image annotation and dataset generation, and the YOLOv5 algorithm for object
detection. The YOLOv5 algorithm uses three types of losses: box loss, class loss, and object loss.
The box loss measures the error in the predicted bounding box coordinates of the fire, while the class loss
measures the error in the predicted class probabilities of the fire. The object loss measures the error in the
confidence score of the fire being present in the bounding box.
The box loss, class loss, and object loss curves provide valuable insights into how well the model is learning
during training. The box loss curve shows how well the model is able to predict the coordinates of the bounding
box for the fire. The class loss curve shows how well the model is able to predict the correct class (fire or non-
fire) for each object. he object loss curve shows how well the model is able to predict the confidence score for
the presence of a fire. Analysing these curves allows us to identify any issues with the model's performance and
2863
make necessary

2864
adjustments to improve its accuracy. By fine-tuning the YOLOv5 model and optimizing its hyperparameters to
minimize the box loss, class loss, and object loss, we were able to achieve good results in detecting forest fires
shown in figure 1.6

Figure 1.6 Mean Average Precision

D) PR Curve
Validation is a critical part of the training process as it helps to evaluate the performance of the model during the
training phase. In this project, we used 15% of our dataset for validation. The validation set is used to tune the
hyperparameters of the model and prevent overfitting.
During the validation process, the model's performance was evaluated using various metrics such as precision,
recall, and F1-score. These metrics helped us to assess the model's accuracy and ensure that it could generalize
well to new data.
Moreover, we also used visualization techniques to understand the model's behavior and identify any potential
issues. For example, we plotted the training and validation loss curves to identify any overfitting or underfitting.
Additionally, we visualized the model's predictions on sample images from the validation set to ensure that it
was detecting forest fires accurately.
In conclusion, the validation phase is a crucial step in the training process as it helps to fine-tune the model's
hyperparameters, evaluate its performance, and ensure that it can generalize well to new data. shown in figure
1.7

Figure 1.7 Precision curve

E) F1 Curve
Precision-Recall (PR) curve is a plot of precision (y- axis) and recall (x-axis) at different classification
thresholds, which shows the trade-off between precision and recall for different threshold values. In the context
of forest fire detection, the PR curve is a useful evaluation metric for object detection models because it provides
a more complete picture of model performance than just using accuracy or F1-score.
The PR curve for this project shows how well the model can detect forest fires in the validation dataset. A high
precision means that the model has a low false positive rate and only identifies actual forest fires, while a high
recall means that the model has a low false negative rate and can identify most of the actual forest fires in the
dataset. The ideal model would have a high precision and high recall, resulting in a PR curve that hugs the upper
right corner of the plot.
The PR curve is particularly useful when the dataset is imbalanced, as is often the case in object detection tasks.
In this project, there may be many non-fire images and only a few fire images, so the PR curve can give a more
accurate representation of how well the model is performing for the minority class (forest fires) rather than just
looking at overall accuracy or F1-score.

2865
Precision-Recall (PR) curve is a plot of precision (y- axis) and recall (x-axis) at different classification
thresholds, which shows the trade-off between precision and recall for different threshold values. In the context
of forest fire detection, the PR curve is a useful evaluation metric for object detection models because it provides
a more complete picture of model performance than just using accuracy or F1-score.
The PR curve for this project shows how well the model can detect forest fires in the validation dataset. A high
precision means that the model has a low false positive rate and only identifies actual forest fires, while a high
recall means that the model has a low false negative rate and can identify most of the actual forest fires in the
dataset. The ideal model would have a high precision and high recall, resulting in a PR curve that hugs the upper
right corner of the plot.
The PR curve is particularly useful when the dataset is imbalanced, as is often the case in object detection tasks.
In this project, there may be many non-fire images and only a few fire images, so the PR curve can give a more
accurate representation of how well the model is performing for the minority class (forest fires) rather than just
looking at overall accuracy or F1-score.
G) P Curve
Precision-Confidence Curve is a visual representation that can be used to evaluate the performance of forest fire
detection models. It plots the precision values against the confidence scores for a range of detection thresholds.

Figure 1.8 Precision-Confidence Curve

In the context of this project, the Precision-Confidence Curve can be used to analyse the model's performance on
a per-class basis. This curve will help us to determine the confidence threshold at which the precision is the
highest for detecting forest fires. It will also help us to evaluate the model's overall precision at different levels of
confidence.
The curve will enable us to analyse the trade-off between precision and confidence, which is crucial in forest fire
detection applications. We can use this curve to set the optimal detection threshold for the model based on the
desired level of precision, which is essential for reducing the false positives while ensuring that the actual forest
fires are detected.
Overall, the Precision-Confidence Curve is an important evaluation metric that provides insights into the model's
performance at different levels of confidence. It helps us to understand the model's strengths and weaknesses and
enables us to fine-tune it for better performance in detecting forest fires.

Figure 1.9 recall-confidence curve

H) R Curve
The recall-confidence curve is a visual representation of an object detection model's performance in terms of its
ability to correctly identify forest fires in images at different confidence levels. It plots the recall on the y-axis
and the confidence threshold on the x-axis. By analyzing the curve, we can identify the confidence threshold at
which
2866
the model achieves the highest recall, indicating the optimal trade-off between precision and recall. The recall-
confidence curve can also help identify areas where the model is underperforming and can be used to improve
the model's accuracy in future iterations. It is a useful evaluation metric that provides insights into the model's
performance at different levels of confidence and helps understand the model's strengths and weaknesses,
enabling us to fine-tune it for better performance Figure 1.9

IV CONCLUSION
In conclusion, the creation of a system for detecting forest fires that makes use of picture recognition technology
marks an important step in our capacity to lessen the destructive effects of wildfires. The goal of this project was
to use computer vision and machine learning to build a reliable, effective system that could recognize forest fires
in real-time.
Through the use of image recognition algorithms, their improvement, training on a variety of datasets, and
integration with security cameras or drones, the study's goals were met. The outcome is a technology that can
identify forest fires quickly and precisely, giving it a major advantage in the fight against these terrible natural
[Link] growing threat of forest fires around the world, which is what inspired our research, is still quite
strong. Wildfires are becoming more frequent and severe, necessitating proactive and creative solutions. In the
face of such difficulties, conventional fire detection techniques are ineffective, emphasizing the significance of
adopting technology [Link] project's scope includes both the creation of an image recognition-based
forest fire detection system and the flexibility of that [Link] various forest types and environmental conditions.
Its value is increased by the possibility of further uses, such as monitoring other natural disasters or supporting
efforts to conserve species.
As we proceed, it is crucial to understand the wider effects of implementing automated fire detection systems.
Collaborations with relevant stakeholders, such as forestry departments and environmental agencies, are essential
for the successful application of this technology. Additionally, ethical issues relating to privacy and data security
must be addressed.
To put it simply, this research represents a significant improvement in our capacity to safeguard our forests,
wildlife, and communities against the destructive power of forest fires. It highlights how technology can improve
our ability to respond to environmental disasters and gives us hope for a future that is safer and more
[Link]

REFERENCES
[1] S. Wu and L. Zhang, "Using Popular Object Detection Methods for Real Time Forest Fire Detection," 2018 11th
International Symposium on Computational Intelligence and Design (ISCID), Hangzhou, China, 2018, pp. 280-284, doi:
10.1109/ISCID.2018.00070.
[2] F. Xie and Z. Huang, "Aerial Forest Fire Detection based on Transfer Learning and Improved Faster RCNN," 2023
IEEE 3rd International Conference on Information Technology, Big Data and Artificial Intelligence (ICIBA),
Chongqing, China, 2023, pp. 1132-1136, doi: 10.1109/ICIBA56860.2023.10165603.
[3] E. Menaka, S. S. Kumar and P. Parameshwari, "Assessment of sparse forest and fire detection using threshold watershed
algorithm," IET Chennai 3rd International on Sustainable Energy and Intelligent Systems (SEISCON 2012),
Tiruchengode, 2012, pp. 1-6, doi: 10.1049/cp.2012.2196.
[4] H. Dang-Ngoc and H. Nguyen-Trung, "Aerial Forest Fire Surveillance - Evaluation of Forest Fire Detection Model
using Aerial Videos," 2019 International Conference on Advanced Technologies for Communications (ATC), Hanoi,
Vietnam, 2019, pp. 142-148, doi: 10.1109/ATC.2019.8924547.
[5] X. Chen et al., "Wildland Fire Detection and Monitoring Using a Drone-Collected RGB/IR Image Dataset," in IEEE
Access, vol. 10, pp. 121301-121317, 2022, doi: 10.1109/ACCESS.2022.3222805.
[6] D. Georgiev, G. Hristov, P. Zahariev and D. Kinaneva, "Forest Monitoring System for Early Fire Detection Based on
Convolutional Neural Network and UAV imagery," 2020 28th National Conference with International Participation
(TELECOM), Sofia, Bulgaria, 2020, pp. 57-60, doi: 10.1109/TELECOM50385.2020.9299566.
[7] A. Khamukhin and S. Bertoldo, "Spectral analysis of forest fire noise for early detection using wireless sensor
networks," 2016 International Siberian Conference on Control and Communications (SIBCON), Moscow, Russia, 2016,
pp.
1-4, doi: 10.1109/SIBCON.2016.7491654.
[8] C. Nithesh et al., "RSA based Forest fire spread detection using Drones and
[9] Image Processing," 2022 IEEE 3rd Global Conference for Advancement in Technology (GCAT), Bangalore, India, 2022,
pp. 1-4, doi: 10.1109/GCAT55367.2022.9972001. X. Bai and Z. Wang, "Research on Forest Fire Detection Technology
Based on Deep Learning," 2021 International Conference on Computer Network, Electronic and Automation
(ICCNEA), Xi'an, China, 2021, pp. 85-90, doi: 10.1109/ICCNEA53019.2021.00029.

2867
[10] L. Shixing, Z. Yongming, S. Weiguo and X. Xia, "An Enhanced Algorithm for Forest Fire Detection Based on MODIS
Data," 2010 International Conference on Optoelectronics and Image Processing, Haikou, China, 2010, pp. 200- 203, doi:
10.1109/ICOIP.2010.332.
[11] C. Naimeng, Y. Wanjun and W. Xiaoyu, "Smoke detection for early forest fire in aerial photography based on GMM
background and wavelet energy," 2021 IEEE International Conference on Power Electronics, Computer Applications
(ICPECA), Shenyang, China, 2021, pp. 763-765, doi: 10.1109/ICPECA51329.2021.9362647.
[12] W. Wang, Q. Huang, H. Liu, Y. Jia and Q. Chen, "Forest Fire Detection Method Based on Deep Learning," 2022
International Conference on Cyber-Physical Social Intelligence (ICCSI), Nanjing, China, 2022, pp. 23-28, doi:
10.1109/ICCSI55536.2022.9970702.
[13] Li Jie and Xiao Jiang, "Forest fire detection based on video multi-feature fusion," 2009 2nd IEEE International
Conference on Computer Science and Information Technology, Beijing, China,2009, pp. 19-22, doi:
10.1109/ICCSIT.2009.5234862.

2868