EdgeFireSmoke: A Novel Lightweight CNN Model for Real‑Time

Video Fire Smoke Detection

V. C. Ranganayaki, Javid J. and Jaideep Reddy M.

Department of Computer Science and Engineering, St. Joseph’s Institute of Technology, Chennai, Tamil Nadu, India

Keywords: Lightweight CNN, Real‑Time Detection, Edge Computing, Fire, Smoke Detection, Scalable Solution.

Abstract: Edge Fire Smoke is the lightweight CNN model optimized toward real-time fire and smoke detection in video

streams and especially toward edge computing. The system responds to the surging demand for effective

solutions in fire prevention, industrial safety, urban fire monitoring, and forest fire management. Unlike the

traditional solutions of reliance on centralized processing, Edge Fire Smoke exploits its lightweight

architecture to function easily on devices such as surveillance cameras and drones, as well as IoT devices,

without letting the latency on the fire and smoke pattern detection reduce. Besides, the model has also been

trained in large, heterogeneous datasets ensuring robust performance against changing environmental

conditions. It comes with adjustable sensitivity levels, so it can be configured to the specific application and

operational requirement. Real- time alerting mechanisms are integrated so that users or alarms can be notified

right away upon detection. Comprehensive logging capabilities enable recording of detection events for

further analysis or audits. A user- friendly interface makes it possible to monitor and configure a system with

minimal technical complexity, thereby making the technology available to users without much technical

know-how. Edge Fire Smoke is cost- effective, scalable, and dependable proactive fire management. The

deployment of this technology in edge environments reduces dependence on cloud infrastructure, thereby

lowering costs while improving response times. The new system plays a great role in safeguarding lives,

infrastructural facilities, and the environment against any fire risks.

1 INTRODUCTION

The critical use of real-time video streams in the field

of fire and smoke detection with an urban, forest, or

industrial context is their ability to track fires or

smoke at any possible early stage of development.

This saves lives, guards structures, and prevents

further ecological damage. More often than not,

traditional fire detection schemes depend on cloud

processing; however, this sometimes introduces

latency, incurs additional operational costs, and

consumes vast amounts of computation. It is an

innovative, lightweight CNN particularly designed to

overcome these shortcomings of real-time fire and

smoke detection. The concept of edge computing in

Edge Fire Smoke is utilized to reduce dependence on

cloud infrastructure, thus allowing the detection to be

fast and efficient right on the device, for example,

surveillance cameras, drones, and IoT sensors.

Therefore, a decentralized approach means that the

fire and smoke detection can be done in a highly

delayed manner in environments that have limited

computational power. The most important advantage

of EdgeFireSmoke is its optimized architecture,

making it efficient on resource- constrained edge

devices. This makes it suitable for application in

various applications, from smart cities to the

industrial facility and forest management, among

other wildlife monitorings. It has been

heterogeneously trained on high datasets, therefore

robustly reliable across very different environments.

Whether it is a forest wildfire, an industrial chemical

fire, or a city urban building fire, the EdgeFireSmoke

is designed to detect with accuracy patterns of fires

and smoke at an incredible speed. Additionally,

EdgeFireSmoke features customizable sensitivity

settings because it can be made very sensitive to

specific environmental settings or operational

requirements. This flexibility makes the system

highly adaptable to be deployed in a variety of

scenarios without sacrificing performance. The

system also has real-time alerting that notifies users

immediately upon detection of fire or smoke, thus

enhancing response times and helping prevent

130

Ranganayaki, V. C., J., J. and M., J. R.

EdgeFireSmoke: A Novel Lightweight CNN Model for Real-Time Video Fire Smoke Detection.

DOI: 10.5220/0013909200004919

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 4, pages

130-137

ISBN: 978-989-758-777-1

catastrophic outcomes. The model's user-friendly

interface allows even those without technical

expertise to easily configure and monitor the system.

Apart from this, the logging feature allows wide

detection and audit events along with further events

analysis. Altogether, EdgeFireSmoke is scalable,

dependable, and economical for the solution of

proactive fire management. It promotes the safety

issue, disaster prevention, protection of the

environment, lower operating cost, and response

speed through its capabilities to equip an edge device

with live real-time detection of fires and smoke.

EdgeFireSmoke stands out as a pioneering solution in

light of the increasing demand for efficient, real-time

fire and smoke detection in various critical

environments. It ensures fast and reliable

performance even on resource-constrained devices

like IoT sensors, drones, and cameras, owing to its

lightweight CNN model.

2 LITERATURE SURVEY

The study of Krizhevsky et al. (2012) "ImageNet

Classification with Deep Convolutional Neural

Networks" initiated deep Convolutional Neural

Networks, which have been responsible for the

breakthrough in the image classification and,

ultimately, the basis for fire and smoke detection in

real time. AlexNet contains many convolutional

layers for the purpose of automatic learning of raw

image hierarchical features and, consequently, has

delivered better accuracy than traditional machine

learning approaches. For example, their work with a

large dataset of images such as ImageNet proved that

deep CNNs can indeed be quite powerful for real-

time video analysis, especially in safety-critical

applications such as fire detection.

Redmon et al. (2016) brought out a highly

influential paper titled "You Only Look Once:

Unified, Real-Time Object Detection." In that

research, they developed the framework YOLO and

revolutionized object detection through converting

the problem into one regression. In contrast, in

traditional systems, the images pass through several

stages of scan. YOLO evaluates the whole image in

just one forward pass, which improves the speed and

efficacy. This efficiency works much in real-time fire

and smoke detection, as alert is vital at the time of

such incidences.

Simonyan and Zisserman (2015) in the paper

"Very Deep Convolutional Networks for Large-Scale

Image Recognition" introduced VGGNet, a deep

CNN architecture that is known for its simplicity and

depth. This architecture achieved high accuracy in the

recognition of images, as it uses small convolutional

filters with 3x3 dimensions and up to 19 layers that

allow it to capture very fine details in images. Such is

important because, in detecting fire and smoke, very

early signs of such events appear as very faint smoke

or small flames that most existing systems fail to

detect.

The very first proposal in their pioneering paper

"Fire Detection in Video Sequences Using a Generic

Color Model" for the possible use of color and motion

features in fire detection from video sequences was

given by Celik et al. in 2007. The authors were able

to develop a model that, based on the reddish-yellow

color of flames, was able to identify fires and

incorporated motion detection to discern dynamic

flames from static sources of light, thereby preventing

false positives. This work was among the first to use

video-based analysis for fire detection and has been

the basis of modern systems that use machine

learning and CNNs.

Toreyin et al. (2005) proposed a paper "Computer

Vision Based Method for Real-Time Fire and Flame

Detection," where the method discovers the unique

spatiotemporal properties of flames, especially flicker

frequency, that can be used to isolate flames from

other moving entities in a scene. It is one of the first

works that utilize temporal patterns during fire

detection since flames exhibit some motion

characteristics that may be observable in video

frames. In modern CNN-based systems in

EdgeFireSmoke, Toreyin's design does not use colour

or intensity thresholds solely; it uses the dynamical

behaviour of fire.

In the paper "MobileNetV2: Inverted Residuals

and Linear Bottlenecks," Sandler et al. in 2018

proposed a lightweight CNN architecture optimized

for mobile and edge computing on limited

computational resources. The MobileNetV2 model

based on inverted residuals and linear bottlenecks has

scaled down the parameters drastically, therefore

reducing the computation cost as well. It is a good

choice for real-time fire and smoke detection in edge

devices such as surveillance cameras and IoT sensors

mainly because they work in resource-constrained

environments, and such efficiency can be allowed

without accuracy loss.

Iandola et al. (2016), in the paper "SqueezeNet:

AlexNet-Level Accuracy with 50x Fewer

Parameters and <0.5MB Model Size," presented an

extremely efficient CNN architecture, significantly

cutting down memory usage with great performance.

SqueezeNet uses 1x1 convolutions and strategic

downsampling to achieve AlexNet performance

EdgeFireSmoke: A Novel Lightweight CNN Model for Real-Time Video Fire Smoke Detection

131

while having fewer parameters; thus, it is great for use

on fire and smoke detection on edge devices which

have fewer resources. With a low-memory design, the

EdgeFireSmoke project could make real-time

deployment possible on resource-constrained devices

like drones and cameras for speedy detection and

response to emergencies caused by fire. Kim et al.

(2016), in their paper "Real-Time Fire Detection

Based on Image Processing," proposed a method that

utilized color segmentation and motion analysis for

the detection of fire in real-time. This decreases false

positives due to static light sources because of the

identification of reddish-yellow tints typical of flames

as well as motion detection. This method is best used

in dynamic environments such as industrial sites,

forests, or even cities. For an EdgeFireSmoke, for

example, adding color and motion features to the

CNN model will be great for improvements in

detection.

Ma et al. (2020), implemented lightweight

architecture of MobileNet to deploy fire detection in

the edge device with the algorithm of "Efficient

Video Fire Detection Algorithm Using MobileNet".

It has proven the ability of doing real-time fire

detection at surveillance cameras, or IoT sensors that

run in a low power condition at the very edge of any

location including harsh environment. It ensures

running the fire detection systems across all

environments as computational resources do not have

to be reduced without sacrificing precision. This will

ensure immediate, cloud- agnostic detection and

response for EdgeFireSmoke in any applications

requiring real-time fire management.

Yuan et al. (2020), in their survey "Survey on

Deep Learning-Based Fire and Smoke Detection

Techniques," surveyed various deep learning

approaches wherein it is identified that challenges in

fire and smoke detection exist because of the varied

conditions of visual aspects and the requirement of

real-time processing. Techniques like GANs have

been applied to generate synthetic data. Hybrid

systems use a combination of CNN-based detection

along with sensor data for improvement in accuracy.

give for the conference of the paper.

3 PROBLEM STATEMENT

Rising instances of security threats in public domains

such as schools, airports, and transportation stations

necessitate immediate real-time detection solutions.

Manual monitoring methods have limitations with

traditional surveillance technologies since it takes

some considerable time before a threat can be

detected; therefore, response time increases along

with risks. Most available automated detection

technologies are cost-prohibitive and call for

hardware that may not be found in many public

environments. This project proposes a low-cost,

scalable, and precise weapon detection system using

OpenCV, an open-source library for computer vision,

and Django, the Python-based web framework. Since

Django has the features that help in easy deployment,

user-friendly interfaces, and real-time notifications, it

pairs perfectly with OpenCV, which is known for fast

video and image processing. It would combine pre-

trained deep learning models such as YOLO (You

Only Look Once), and even SSD (Single Shot

MultiBox Detector), distinguish weapons from other

non-threatening objects of interest in real-time video

stream with great accuracy. With OpenCV and

Django, the system ensures that those processes

involved in detection, alerting, and data logging run

inside a user- friendly interface, suitable for use on

standard hardware. It simply means that using related

advanced, affordable technologies to augment public

safety is a great idea, which enables security personnel

with real-time actionable insight. Such an architecture

would be practical, efficient, and easily deployable.

4 METHODOLOGY

4.1 System Overview

A general overview of the system is as presented by

the block diagram in the figure 1 system as follows:

Figure 1: Architecture diagram of weapon detection

The EdgeFireSmoke system makes use of the real-

time fire and smoke detection ability via a lightweight

CNN optimized to run on devices with small

computational powers. The detection is achieved

using video streams sourced from surveillance

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

132

cameras, drones, or IoT sensors in highly dynamic

environments. It therefore integrates the use of pre-

trained CNN models on large heterogenous datasets

that can very efficiently recognize the patterns of fire

and smoke with very high accuracies. The model is

optimized for low-latency processing to make sure

that fire detection is in real- time. That way, it reduces

dependency on cloud-based servers to make the

system work standalone in remote or resource-

constrained settings.

In terms of detection, EdgeFireSmoke operates as

follows: every frame from a video is processed with

the aid of a pre-trained CNN that extracts features of

the image such as color, texture, and patterns in

movement. These features then lead to the

identification of the presence of fire or smoke through

color segmentation, motion analysis, and

spatiotemporal feature extraction. It possesses real-

time alerting mechanisms which automatically alert

users to the case of detection with logging facility to

analyze events and audit them.

4.2 Data Collection Module

Data Collection Module of the EdgeFireSmoke is

crucial to the acquisition of the quality diversification

of gathered data to properly train a CNN for an

accurate real-time detection of fire and smoke. In this

module, the system gathers video streams from

diverse sources including but not limited to, the

surveillance cameras, drones, as well as IoT sensors

that spread throughout the following diverse

environments: urban and industrial sites, and forests.

It has several visual conditions that involve capturing

various types of fires and smoke, lighting variations,

changes in weather, and environmental settings to

make the data more robust and generalizable. The

module captures video frames at a high resolution,

thus saving the fine-grained details of fire and smoke

patterns. It also possesses many features to identify

whether it is fire, smoke, or any other object with

some degree of accuracy, such as color of flame

(reddish-yellow colors), characteristics of motion

(flickering and movement), and environmental

conditions (fog, reflections, and lighting conditions).

Information from sensors, for example, temperature

or gas concentration measurements, can also be

included in order to supplement video information to

add a layer of verification and eliminate false

positives. Data once acquired undergoes

preprocessing. Normalizing video frames and image

resizing are two standard preprocessing techniques

into a CNN model. Others include data augmentation

in a way that increases diversity as well as simulate

different environmental conditions. The dataset is

then further divided into training, validation, and

testing subsets, so the model is tested on multiple

scenarios, thereby making the model more robust and

generalized. To summarize, the Data Collection

Module ensures that the EdgeFireSmoke system is

trained in a variety of conditions; that is, it provides

real-time fire and smoke detection across different

environments in a reliable and accurate manner.

4.3 Preprocessing Module

Pre-processing module of the EdgeFireSmoke system

is very essential because it pre-processes the video

data fed into CNN. It is standardized noise-free and

ready for proper training, hence ensuring that real-

time inference is feasible. The first step among all the

preprocessing steps is to undergo image resizing,

wherein video frames are transformed into a certain

dimension, so they can be presented to the CNN for

further operations; therefore, hence reducing

computation time. Normalization: This makes pixel

values fall within the range 0 to 1 or -1 to 1, which

offers a similar standard for diverse input data and

results in a better convergent rate of the network on

training. Data Augmentation: Increase data varieties

and robustness from changed conditions by using

augmentation technique such as rotation, flips, color

change, and cropped images. This simulates different

environmental conditions such as lighting changes,

vantage or partial occlusions. Noise reduction is also

provided. And this module eliminates all the noise

created by extraneous visual information that

degrades performance in a model, for example, clutter

in the background, unwanted reflections. Motion

analysis has also been included for identifying static

features and dynamic features with focus on motion

patterns characteristic of fire and smoke. The

preprocessed data is then labeled in supervised

learning, whether the fire or smoke is there, so that

the system learns to distinguish the elements within

real-time video.

4.4 Model Selection Module

The Model Selection Module of the EdgeFireSmoke

system is supposed to determine which one of these

CNN architectures will be used for the purpose of

real- time fire and smoke detection under specific

conditions. Given the limits of an edge device, it has

to operate under computational and power

constraints, making the selection of some of the

available lightweight yet accurate CNN models that

provide the best accuracy with a lower computational

EdgeFireSmoke: A Novel Lightweight CNN Model for Real-Time Video Fire Smoke Detection

133

cost. These purposes require models like

MobileNetV2, SqueezeNet, and VGGNet due to their

proven efficiency in managing the image

classification tasks along with a resource-friendly

model. Module trade- offs for detection accuracy,

size, and inference time help with performance

metrics. Model comparison based on comparative

evaluation involves further considering how each

architecture may cope with changing environmental

conditions: light variations, smoke density, and flame

characteristics. It also ensures real-time processing

needs. The selected model will then work on video

frames as quickly as possible without causing any

delays. After the selection, the model is fine-tuned on

the preprocessed data with the objective of improving

performance in specific fire and smoke detection

tasks.

4.5 Fire Detection and Classification

This is a smoke detection and visual elements

classification component of the EdgeFireSmoke

system, detecting and classifying visual elements

within real- time video streams in real time. It applies

a pre-trained CNN that analyzes video frames based

on some characteristics unique to smoke: color,

texture, and movement patterns. Smoke generally

appears as a partly misty or translucent cloud that

assumes irregular shapes and has diffused edges. It

will, therefore, differ in intensity and color,

depending on the fuel source and environmental

factors. It is this subtle differentiation that the CNN

model becomes capable of as it learns from the diverse

nature of the dataset it is fed on. The diversity of such

smoke in the dataset with varying illumination,

weather, and environmental factors would be

different. First, at detection, it separates the areas that

might be potential smoke by using the color

segmentation technique to identify whether they are

shades of gray, white, or black colors that are most

associated with smoke. Motion analysis further

differentiates the smoke from static objects because of

their dynamic and swirling motion. Once the possible

smoke areas have been identified, the CNN will

classify them based on learned patterns to determine

whether they really show smoke or are false alarms

like fog or vapor.

4.6 Alert System Integration

The Alert System Integration of the EdgeFireSmoke

system has been designed to send alerts to users or

automation systems as soon as possible when fire or

smoke is detected. When the model identifies and

classifies a possible fire or smoke in a video stream,

an alert system calls a series of operations to take

place in response so as to act quickly. The system uses

the detection results in real-time and generates

notifications through various channels, such as SMS,

emails, or direct communication with the central

monitoring systems. Other than basic alerts, the

system may integrate with other IoT devices, such as

sprinklers, alarms, or fire suppression systems, and

may enable an automated response to detected

threats. It has the characteristic of configurability in

terms of alert sensitivity so that it will not lead to false

positives and hence give reliable detection.

Information regarding the type of danger, whether it

is fire or smoke, location, and time related to alerts is

of utmost importance to the emergency responders.

4.7 Backend System

The backend system should be strong enough to

integrate real-time fire and smoke detection using the

YOLO-based model; therefore, the computational

power needed for the deployment of deep learning

models and large-scale data can be provided through

cloud infrastructure like AWS or Google Cloud.

Apache Kafka or AWS Kinesis may be used to

continuously ingest real-time video streams, which

could have minimal latency, with which to manage

real time video streams. Tools, for instance,

TensorFlow Serving, or TensorRT will implement

the YOLO model at inference time for enhanced

performance optimization. Additionally, TensorFlow

Lite may be useful for optimizing models for Edge

Devices that have low resource. The video data will

be stored in either Amazon S3 or Google Cloud

Storage, while metadata and other events will be

taken care of by MongoDB. The communication

between edge devices and the backend will be

handled by restful API or AWS API Gateway,

ensuring alert and data transmission fluidly. The

scalability, real-time processing, and fire and smoke

detection will be supported in the backend

architecture on the very resource-constrained device.

4.8 System Testing

The testing of the EdgeFireSmoke project includes

functional, performance, and reliability testing of the

entire fire and smoke detection system. This testing

will validate all the real-world operating capabilities

of the YOLO-based detection model, preprocessing,

classification modules, and the alert system. It also

includes the validation of the system by varied

datasets and simulating changing environmental

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

134

conditions, like light, weather, and density of smoke

to check accuracy and robustness. Real-time testing

for the edge devices, such as cameras and drones, is

under way to test detection and alerting almost for

free from latency. Integration tests are performed to

ensure seamless communication with the edge device,

with the backend infrastructure, and through the alert

system. A stress test is conducted, determining how

well the system behaves with its data loads and high

video frame rates. Such tests are then used as feed

back for further refining the performance of models

and responsiveness of systems.

4.9 Deployment

The EdgeFireSmoke system is sending the trained

YOLO-based detection models to edge devices,

including surveillance cameras, drones, or IoT

sensors for real-time execution. Coupled with

lightweight frameworks, such as TensorFlow Lite or

ONNX Runtime, that enable execution on the

processing resource-less devices, this system makes

central monitoring and alert management the role of

cloud services, like AWS IoT Core or Google Cloud

IoT. It encompasses the establishment of the alerting

system for instant alerts via SMS, email, or app-based

alerts. The system is fully tested post-deployment to

ensure stability in live environments. Scalability is

achieved due to containerization tools such as Docker

whereby the system can be very quickly replicated

across different locations, thereby offering better

coverage fire and smoke detection. All this involves

installation on the edge devices, whereby compact

models like MobileNet or YOLO exist for real-time

detection of events. The device is pre-configured to

send alerts in an autonomous form through the

integrated communication protocols that support

SMS, email, or through mobile apps. Deployment

also includes calibration of the system for fitment into

various environmental setups, thus avoiding minimal

cases of false alarms.

5 RESULT AND OUTPUT

The EdgeFireSmoke system provides accurate,

efficient real-time detection of fire and smoke based

on rigorous testing conditions. Equipped with a

YOLO-based model and advanced preprocessing, the

system obtained great detection accuracy on fire and

smoke patterns while producing minimal false

positives and negatives. It worked well under adverse

conditions such as low-light settings and in changing

weather,

and even partially obscured, it gave

Figure 2: Complete detection of a fire.

and adaptability. There is the real-time video frames

that are annotated with marks to indicate there's fire

or smoke to help fasten the review and subsequent

response. In addition, alerts are received immediately

over SMS, email, and through an application

notification that nobody is left unaware. Because of

its capabilities, its effectiveness in environments with

resources constraint without a cloud was a guarantee.

In this, it is the best in high-risk places and

inaccessible locations. This performance was

maintained at high video frame rates and large data

sets thus indicating scalability and responsiveness.

Full logging and report generation with all the above

features ensures very good insight into post-event

analysis and audits. Thus, the results ensure the

system is a reliable and proactive solution for fire and

smoke emergencies, offering a scalable, low-cost

approach for industrial, residential, and natural

environments. It is through quick detection and

alerting that EdgeFireSmoke crucially goes along in

preventing and mitigating fire- related disasters.

Figure 2 shows the complete detection of a fire.

6 PERFORMANCE ANALYSIS

In reflecting the efficiency and accuracy in fire and

smoke detection performance analysis, it shows under

any scenario, as it underwent rigorous testing of its

heterogeneous dataset on extreme conditions ranging

EdgeFireSmoke: A Novel Lightweight CNN Model for Real-Time Video Fire Smoke Detection

135

from industrial sites to forest environments, urban

and household scenarios. The model YOLO-based

demonstrated precision and recall values close to

perfect even at the conditions of low illumination,

smokes at occlusion, and dynamism in weather

conditions. In experiments, such a low-latency high-

performance framework could directly analyze

multiple input sources of high frame rates onto these

miniature edge devices constituting IoT sensors and

surveillance cameras. A huge amount of data flow

went through the system while still managing to

provide the desired scale of performance.

Resource optimization leads to a lightweight

model with efficient preprocess without losing any

precision in case of detection for its ease-of-

deployment with constrained- resource devices. With

comparative benchmarks, EdgeFireSmoke detected

more compared to the traditional detection system, in

terms of detection rate, response time, and

adaptability of varying environmental conditions. In

addition, real-time alerting systems as well as the

logging mechanism increased its usability; it brought

actionable insights into the hands of the user. The

given results confirm that EdgeFireSmoke is reliable,

efficient, and robust for fire and smoke detection, that

makes it a really valuable tool for safeguarding life

and assets in critical safety applications.

7 CONCLUSION AND FUTURE

WORKS

The EdgeFireSmoke has been well defined by the

advancement that fire and smoke detection by real-

time means are possible through light versions of

CNN models by virtue of edge computing. The strong

capabilities of such high accuracy with responses in

real time, along with adaptability into any setting,

make for a highly reliable solution for this risk to be

mitigated by fire. This system can maximize the use

of resources and is useful when integrated with edge

devices, for example, drones, sensors, and cameras in

supporting fast detection and alerting while being

light on resources. Full logging and analytics

facilities make this system extremely valuable for the

assessment after the event, garnering precious

insights. Future work: The system would improve

further by incorporating multimodal data, for

example heat, gas, or acoustic sensors in enhancing

detection accuracy while suppressing false alarms. It

is in the more complex situation that advanced deep

learning methods like transformer-based models are

used to enhance the performance of the system.

Predictive analytics can be included for risk detection

of fire occurrence based on patterns within

environmental data. The system called

EdgeFireSmoke will serve as a bedrock for future

innovations and challenges in fire safety technologies

as it responds to growing concerns through innovative

solutions.

REFERENCES

A. Boukerche and E. Nakamura, "Localization systems for

wireless sensor networks", IEEE Wireless

Communications, vol. 14, no. 6, pp. 6-12, 2007. doi:

10.1109/mwc.2007.4407221.

D. S. Thomas, D. T. Butry, S. W. Gilbert, D. H. Webb, and

J. F. Fung, "The Costs and Losses of Wildfires,"

National Institute of Standard and Technology (NIST),

Tech. Rep., 2017.

F. Khelifi, A. Bradai, K. Singh, and M. Atri, "Localization

and Energy-Efficient Data Routing for Unmanned

Aerial Vehicles: Fuzzy-Logic-Based Approach," IEEE

Communications Magazine, vol. 56, no. 4, pp. 129-133,

2018. doi: 10.1109/mcom.2018.1700453.

K. Hoover, L. A. Hanson, "Wildfire Statistics 2019 | Centre

for Research on the Epidemiology of Disasters,"

Federal Assistance for Wildfire Response and

Recovery, 2019. Fas.org. (2019).

[Online] Available at: https://fas.org/sgp/crs/misc/IF10

244.pdf [Accessed 26- May-2019].

M. Y. Arafat and S. Moh, "Location-Aided Delay Tolerant

Routing Protocol in UAV Networks for Post- Disaster

Operation," IEEE Access, vol. 6, pp. 59891- 59906,

2018. doi: 10.1109/access.2018.2875739.

M. Y. Arafat and S. Moh, "Routing Protocols for

Unmanned Aerial Vehicle Networks: A Survey," IEEE

Access, vol. 7, pp. 99694 - 99720, 2019. doi:

10.1109/ACCESS.2019.2930813.

M. Y. Arafat, S. Poudel and S. Moh, "Medium Access

Control Protocols for Flying Ad Hoc Networks: A

Review," IEEE Sensors Journal, pp. 1-1, 2020. doi:

10.1109/jsen.2020.3034600.

M. Y. Arafat, M. Habib and S. Moh, "Routing Protocols for

UAV-Aided Wireless Sensor Networks," Applied

Sciences, vol. 10, no. 12, p. 4077, 2020. doi:

10.3390/app10124077.

N. H. Motlagh, T. Taleb, and O. Arouk, "Low-Altitude

Unmanned Aerial Vehicles-Based Internet of Things

Services: Comprehensive Survey and Future

Perspectives," IEEE Internet of Things Journal, vol. 3,

no. 6, pp. 899922, 2016. doi: 10.1109/jiot.2016.26121

19.

Q. Guo, Y. Zhang, J. Lloret, B. Kantarci and W. Seah, "A

Localization Method Avoiding Flip Ambiguities for

micro-UAVs with Bounded Distance Measurement

Errors," IEEE Transactions on Mobile Computing, pp.

1-1, 2018. doi: 10.1109/tmc.2018.2865462

S. Yang, D. Enqing, L. Wei and P. Xue, "An iterative

method of processing node flip ambiguity in wireless

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

136

sensor networks node localization," in Proceedings of

International Conference on Information Networking

(ICOIN), pp. 92 97, 2016. doi: 10.1109/icoin.2016.742

7094.

Z. Lin, H. Liu and M. Wotton, "Kalman Filter-Based Large-

Scale Wildfire Monitoring With a System of UAVs,"

IEEE Transactions on Industrial Electronics, vol. 66,

no. 1, pp. 606-615, 2019.

doi: 10.1109/tie.2018.2823658.

EdgeFireSmoke: A Novel Lightweight CNN Model for Real-Time Video Fire Smoke Detection

137