[Link].

com/scientificreports

OPEN Trust-driven approach to enhance

early forest fire detection using
machine learning
Tayyab Khan1, Karan Singh2, Bhoopesh Singh Bhati1, Khaleel Ahmad3, Amal Al-Rasheed4,
Masresha Getahun5 & Ben Othman Soufiene6
Forest fires pose significant threats to both natural ecosystems and human communities due to their
unpredictable nature and capacity for widespread destruction. Identifying and mitigating fires in the
trunk, ground, and canopy of forests is crucial for reducing their adverse effects on the ecosystem and
climate. The detrimental impacts of forest fires, such as the exacerbation of the greenhouse effect,
the hastening of global warming, and the modification of climatic patterns, underscore the urgent
necessity for the creation of efficient detection systems. This study presents a real-time Universal Trust
Model (UTM) that is specifically designed for the early forest fires detection (FFD) using an intelligent
wireless sensor network and machine learning approaches. Our method seeks to reduce fire detection
time and improve the reliability of the detection process. This is achieved by employing environmental
indicators and moisture levels to swiftly identify fires. The intelligent WSN functions by partitioning
the forest into suitable clusters and intelligently positioning sensor nodes to guarantee extensive
coverage and effective data transmission to the sink. The proposed UTM system’s core component is
the computation of trust ratings for every sensor node. These ratings consider communication, energy,
and data trust factors to evaluate the reliability of the data being delivered. This integrated trust
model enhances the robustness and accuracy of fire detection, especially under difficult environmental
conditions. Furthermore, a machine learning regression model is deployed at the base station to
augment the precision of fire detection. This is accomplished by examining essential attributes such
as temperature, humidity, and CO2 concentrations. We have conducted thorough experiments using
actual datasets consisting of 7200 samples to confirm the efficacy of our proposed UTM in detecting
forest fires at an early stage. The results suggest that our system obtains a high rate of data processing
and a reduced time delay in comparison to existing systems. This renders it a promising solution
for the imperative need to promptly detect and prevent forest fires. Our technique combines trust
mechanisms with machine learning algorithms to create a very advanced forest fire detection system.

Keywords Trust, WSN, Forest fire detection, Sensor, Machine learning

A wireless sensor network (WSN) is comprised of energy-efficient sensor devices (SDs) that detect ecological
factors such as temperature, barometric pressure, moisture, light intensity, as well as chemical parameters like
CO, CO2, and NO2. The purpose is to derive valuable insights from the collected data. The collected data is
transmitted to the central station for additional analysis1. The SDs rely on battery and solar panels as their main
and subordinate power sources, respectively. WSNs are extensively employed in the surveillance and tracking of
targets in distant locations, utilizing a microcontroller, transceiver module, and power components2. Forest fires
are a worldwide issue that inflicts significant harm on human lives, as well as plants and animals, particularly
in the Mediterranean region3. Figure 1 depicts the causes, effects, prerequisites, and remedies for forest fires.
Forest fire detection (FFD) utilizing intelligent sensors is a subject of particular interest that has seen significant
progress in recent years. Forests have a substantial impact on the planetary, biological, and environmental
systems due to their influence on the vast quantity of greenhouse gases, absorption of atmospheric carbon,

1IndianInstitute of Information Technology Sonepat, Khewra, Haryana, India. 2Jawaharlal Nehru University, New
Delhi 110067, India. 3Department of Computer Science & Information Technology, Maulana Azad National Urdu
University, Hyderabad, India. 4Department of Information Systems, College of Computer and Information Sciences,
Princess Nourah bint Abdulrahman University, P.O. Box 84428, Riyadh 11671, Saudi Arabia. 5Department of
Computer Science and Information Technology, College of Engineering and Technology, Kebri Dehar University,
Kebri Dehar, Ethiopia. 6PRINCE Laboratory Research, ISITcom, University of Sousse, Hammam Sousse, Tunisia.
email: masreshaggetahun@[Link]

Scientific Reports | (2025) 15:14480 | [Link] 1

[Link]/scientificreports/

Fig. 1. Forest fire- reasons, consequence, requirements and solutions.

and mitigation of soil erosion4. Wildland fires are mostly ignited by lightning strikes in open, uninhabited,
and isolated regions, often due to delays in promptly detecting and extinguishing them. Wildland fires have
significant impacts on animals, human lives, material assets, and can disrupt weather patterns on a global scale.
Furthermore, the projected yearly decline in several nations is on the rise as a result of forest (wildland) fires. The
conventional FFD techniques5,6,3,7,8 are subject to various constraints, including the lack of reliability in human
observations and the limited spatial and temporal resolution of satellite pictures. In addition, they failed to take
into account the specifications of early FFD systems, as depicted in Fig. 1. Therefore, employing a trust-based
forest fire detection (FFD) system that utilizes an intelligent WSN appears to be an advantageous approach
for mitigating the impact on wildlife, preventing property damage, and safeguarding the surrounding weather
during the initial stages of a flame, before it escalates into a catastrophic event. This approach also increases
the likelihood of achieving more accurate and timely FFD, as evidenced by previous research9–11. In this study,
we partition the forest region into the most efficient clusters and install different types of intelligent sensor
nodes (ISNs or IoT sensors) capable of detecting various environmental parameters, including thermal level,
atmospheric humidity, illumination intensity, and carbon monoxide (CO) concentration. The ISNs, designed in
a spherical shape, has the ability to perceive their environment, collect and analyze significant data, and transmit
it to the group heads (GHs). The GHs thereafter transmit the aggregated data to the base station (BS). If the BS
is situated at a significant distance from the GH, denoted as x, the GH (x) will transfer its data to another GH
(y), which will subsequently relay the data to the BS. Each GH is outfitted with a lithium-ion battery, solar panel,
Arduino Nano, and nRF24L01 module. The BS, which is powered by household electricity, combines a PC,
Arduino Nano, nRF24L01, and SIM800L module. The trust model being given evaluates the trustworthiness of
each ISN by calculating a trust score and categorizing it as highly trusted, valid, or flawed, depending on how it
compares to an optimal trust threshold. The suggested machine learning method takes the combined monitored
data as input and assesses it to improve the accuracy of detection. When a forest fire is detected, a quick alarm
is promptly sent by text message to the chosen officers’ cellphones, urging them to take fast action. Data was
collected from several locations and weather conditions, encompassing both fire and non-fire scenarios, for the
purpose of training and assessing the model. A total of 7000 data samples were collected in succession, with
each sample accurately measuring temperature, relative humidity, light intensity, and CO levels at specific time
intervals. Among them, 80% were chosen at random to train the model, while the remaining 20% were allocated
for testing.

Motivation
Forests are essential ecological and economic assets that sustain life on Earth. Forests enhance the pleasing
appearance of the environment, govern regional weather patterns, manage precipitation, and offer benefits such
as clean water and air, valuable timber, medicinal plants, habitats for wildlife, diverse land-based ecosystems,
and stable soil3. Forest loss and deforestation have negative effects on natural ecosystems and biodiversity, and
they also have significant impacts on the livelihoods of 1.6 billion people. Forest degradation can be caused by
several events, including avalanches, fire, drought, smoldering, and flooding. Among these, fire is a significant
uncontrolled natural disaster that has a big impact on the forest ecosystem and environmental regulations4.
Forest fires are capricious and deadly calamities that result in extensive devastation globally, impacting finances,
biology, and the environment5. Detecting forest fires in hard and dynamic conditions is a difficult and complex
task globally. Early and precise fire detection systems are in great demand to prevent catastrophic incidents

Scientific Reports | (2025) 15:14480 | [Link] 2

[Link]/scientificreports/

such as fire accidents and forest fire disasters, in order to safeguard human lives and property6. WSNs are highly
effective monitoring systems that can accurately identify the start of forest fires. This is achieved by strategically
deploying a range of intelligent sensors in appropriate locations, ensuring optimal coverage across the targeted
regions3. Positioning the ISN optimally can significantly enhance the detection system and increase the quality
of the data12. The intelligent WSN analyzes the collected data and promptly notifies the appropriate personnel
upon detecting the occurrence of a fire10.

Our contribution
To overcome the shortcomings of the existing3,5–8,13–19 FFD methods, we develop a robust and universal trust
model (UTM) to improve collaboration among ISNs that consists of subsequent unique features.

1. Assign unique ID and location to each type of ISNs to exclusively identify it. Deploy the ISNs with optimal
distance in an optimal clustered architecture.
2. Employ a logical time window to record features (temperature, humidity, etc.) and trust scores. It forgets
older records and focuses on the most recent recorded information to achieve higher accuracy.
3. Incorporation of a proficient trust evaluation strategy to isolate faulty ISNs. Moreover, UTM incorporates
threshold ratio and dynamic trust functions for communication trust, data trust, and peer recommendation
trust to provide reliable information. Threshold (temperature, trust, etc.) values can be set according to the
weather requirements that make UTM superior to the traditional FFD approach. Furthermore the dynamic
trust function employs a natural abnormal factor to provide a space for natural calamities.
4. Extract an adequate number of features (light intensity level, temperature, humidity, CO level, etc.) from
the real data set collected using various types of ISNs. 80% of the obtained data was randomly allocated as
training data for the model, while the remaining 20% was designated as test data. We analyze the behavior of
forest fires from the recorded real data.
5. UTM employs real data set observed in different climatic zone along with xBD dataset20, FD-dataset21, For-
estryImages22, VisiFire23, BowFire24 to evaluate the performance of proposed algorithms.

The experimental outcomes obtained after several successive experiments exhibit that the developed FFD system
is highly suitable for real-time early FFD. Finally, this paper will provide a new direction to researchers working
in the sensor-enabled FFD. The rest of the research paper is organized as follows. The subsequent sections of
the research article are structured in the following way. The literature pertaining to current papers is examined
in Section II. Section III introduces the proposed algorithm and its evaluation. In Section IV, we present a
collection of experimental findings that demonstrate the practicality and resilience of the suggested approach.
Section V provides the conclusions and outlines further work.

Related work
This section reviews some recent existing forest fire detection (FFD) systems. Some of the recent proposals for
WSNs to detect danger spots and fires are listed below. Table 1 shows the comparative analysis of current forest
fire detection methods found in the literature review.
Dampage et al.13 present a scientific report on “Forest fire detection system using WSN and ML” to detect
forest fire at early stages. Different kinds of SDs (DHT22, LDR, MQ9) with adequate battery power are placed
on the trees at an appropriate height (one meter) in the clustered architecture. Experimental results prove that
the suggested model can detect forest fire at early stages but it does not consider attack scenarios. Moreover,

Previous methods Observation Advantage Limitation

Reduce the fire detection time, easy Inefficient due to different geographic location and unreliability of human
Fire watch tower4,25 Manned observation towers
observation of forest (wildland) observations, costly, prone to lightning strikes, long detection delay
Vision-based Manually position and install Continuous sampling, cheap and easy
Night images and bad weather images problems, high cost, moderate efficiency,
systems (camera each camera in the most to install, detect fire immediately, high
long detection delay
surveillance)4,25 optimal location. accuracy26
Low resolution, significant delay in fire scanning, quality of satellite photos
Expensive satellite are used for Coverage of large areas, low faulty alarms can be influenced by weather conditions, not feasible to install satellites to
Satellite images27
forest fire detection repetition continuously monitor all forests, cannot forecast forest Fires, high cost, very
long detection delay, moderate Fire localising accuracy
Fire pixel classification,
False positives due to complicated lighting conditions (day and night, artificial
Image processing extract fire pixel, focus on Real-time monitoring, fast and effective
lights, light reflection, and shadows), wrongly detect fire-colored ordinary
technique3,14 various features (shape, color, to detect fire, less cost26
objects
texture) of fire
Low cost, reliable28, easily detect
Resource constraint, Coverage of large areas is difficult, substantial quantity of
Wireless sensor Different type of sensors are chemical properties13, accurate early
SNs must be densely distributed across a forest, resulting in a significant cost for
network (WSN)6 deployed detection is possible, efficiency and low
achieving comprehensive coverage.
power consumption17
low cost, flexible devices, easily fly in Expensive, Limited Flight Time, problem in High Wind Conditions, regular
Unmanned Aerial Wildfire surveillance and
remote (critical) areas to detect fire, offer service is required for optimum use, Invasion of privacy, Low operation speed,
Vehicles (UAVs)15,29 suppression using drones
a high spatiotemporal resolution limited battery life
Good dataset required, difficult to design efficient algorithm, feature
Artificial Extract features with deeper Reduce false positives, better learning
identification, extraction and selection problem, problem of correctly prediction
intelligence5 semantic information13 than human, early fire detection ability30
a large number of unseen patterns from a few known ones, high complexity15,31

Table 1. Comparative analysis of existing forest fire detection methods.

Scientific Reports | (2025) 15:14480 | [Link] 3

[Link]/scientificreports/

it does not compute the communication overhead and complexity of the suggested scheme. Mahmoud et
al.14 suggested a multi-stage video-based FFD system using Image Processing and Support Vector Machine
(SVM). Experimental results on MATLAB (R2017a) exhibit a good detection rate and low false alarms.
Special wavelet analysis and SVM are employed to differentiate and classify the real fire and non-fire objects
respectively. However, this approach is suitable for some good scenarios since various day-night scenarios are
not considered. Kinaneva et al.15 discussed early FFD using Unmanned Aerial Vehicles (UAVs, drones) and
artificial intelligence. Two types of drones (fixed-wing and a rotary-wing) furnished with thermal cameras are
employed to capture the image and video data. It triggers an alarm if any UAVs detect fire ignitions. The main
problems with UAVs are limited flight time, weather conditions, costly and low operation speed, etc. Xu et
al.16 investigated an Ensemble Learning-based real-time FFD system in diverse scenarios. The author employs
deep learning techniques to learn and extract forest fire features. Fire detection is carried out by integrating
Yolov5 and EfficientDet learners. Then, EfficientNet learner is launched to lead the detection process to diminish
false positives. Benzekri et al.17 proposed an early FFD system using clustered-WSN and deep learning with
LoRa communication signal modulation. If the suggested deep learning model predicts fire occurrence then
UAVs are launched to measure temperature values and sensor reading in a mobile way. The UAVs send the
collected information to the base station for necessary action. Vikram et al.7 designed an energy-efficient FFD
technique and localization technique by employing SVM. The entire forest area is divided into different zones
to propose a semi-supervised classification model. Moreover, a greedy forwarding technique is used to forward
the packets from the highly active zone to the base station with good throughput and less congestion. In8, “the
author proposed a real-time and embedded implementation of a fire and smoke detection technique that can
reuse standard video cameras of surveillance systems. The target of this work is to develop smart IoT devices for
fire/ smoke detection in indoor and outdoor environments. The algorithm is simple to design and can be trained
fast. The proposed solution achieved promising results for accuracy in comparison to the state-of-the-art. In
terms of false positive, it reduced the background mistakes in non-fire/smoke videos. Indeed, the proposed
method performed very well, even with videos containing challenging features such as sun and clouds. The
proposed detection solution stands for its low-latency and real-time performance when compared to the other
regional-based detectors. Wu et al.18 proposed a GNN based on the similarity of multi-view forest fire images,
which successfully estimates the similarity between images. Then, a dynamic feature method was proposed to
segment the fire areas from images. This method simplifies the complex preprocessing of images, and effectively
extracts the key features of images, enhancing the robustness of forest fire recognition. In addition, a self-made
forest fire image dataset was contributed. Kizilkaya et al.19 designed an effective and energy-efficient framework
is proposed for the detection of forest fires. Unlike existing studies, multimedia sensors employing deep learning
models are used alongside scalar sensors, and detection is performed using the fusion of information at different
levels of the system architecture. In addition, the efficiency of communication is improved by the introduction
of edge computing for decision-making. The proposed framework uses the efficiency of scalar sensors and the
accuracy of multimedia sensors. Through the establishment of a hierarchical infrastructure, the new framework
balances energy efficiency and detection accuracy and provides a sustainable emergency surveillance” system.
In study31, active fire detection was performed on a medium spatial resolution dataset (Landsat-8 imagery)
where the extent of active fires was very low. A deep CNN (Fire-Net) was proposed in this work to detect active
forest fires in various regions. Specifically, the USA and Australia regions were used to train the network whereas
testing was done for Africa, Brazil, Ukraine, Australia (the parts was not involved in training) regions. The
results depicted a high transferability of the proposed method. Then, the Fire-Net was compared with another
state-of-the-art deep network, i.e., MSR-U-Net and other common ML algorithms. The results for active fire
detection were qualitatively and quantitatively assessed. The performance evaluations showed there was a
trade-off between active fire and non-active fires detection. A well-maintained IoT and Drone-based Forest
Fire Detection and Counteraction System is presented in29 using a subnet mechanism. For the detection of fire,
different types of sensors are deployed in the subnets that can communicate to the gateways, alarms are used to
warn humans and protect animals, and drones are utilized to combat the fire. The subnets are connected to the
control room through gateways, which transmit collected sensor information to the control room. The control
room is responsible for instructing drones for visualization and fire counteraction. The sequence of the system
is presented through a sequence diagram and graph theory is used to represent the model, and then formal
specifications of the model are developed using” VDM-SL.

Problem description
Forest fires have extensive detrimental ecological, economic, and societal consequences, such as contributing to
global warming, depleting precious resources, and reducing biodiversity, and a raise in the percentage of carbon
dioxide in the atmosphere, etc. The main reasons for forest fires are natural causes (lightning, high atmospheric
temperatures, low humidity, etc.) as well as manmade causes (electric spark, naked flame, smoking or barbeque
parties, etc.). The key parameters that initiate a forest fire are air, heat, and fuel available in the forest area.
Unluckily, a forest fire is frequently observed in remote areas only when fire covers a large forest area, making
it challenging to control within a short period which results in loss of natural resources and fuel sources31. Due
to the non-rigid nature and dynamic shape of fires, early detection of forest fire by monitoring the potential
risk areas using WSN can significantly reduce the potential damage, forest fire handling cost, and reaction time.
Although there exist numerous research contributions for FFD3,5–8,13–18,29–31, but they are unable to fulfill the
requirements (Fig. 1) of early FFD systems and lacking in insider attack detection. Moreover, existing research
work did not consider any attack scenarios. Hence no trust-based literature review for FFD is available till
now. SDs are susceptible to various internal and external attacks1. Internal attacks such as on-off attacks, black
hole attacks2, etc. try to misguide the SD to destroy the network performance9. The sensor-based FFD system
must defend against such internal attacks to obtain genuine information since internal attacks alter /drop the

Scientific Reports | (2025) 15:14480 | [Link] 4

[Link]/scientificreports/

information to reduce the network throughput. Additionally, ISNs are lack-temper resistant hardware so in-
depth attention is required to protect the ISNs from wild animals and climate conditions. Moreover, existing
FFD systems3,5–8,13–18,29–31 don’t incorporate the trust factor before any decision-making since trust-enabled
systems improve the lifespan of the network by providing accurate information. Furthermore, existing FFD
systems ignore energy requirements and energy trust factors before decision-making that resulting in a fast
energy drain of SDs since multimedia data transmission is a costly operation. Existing FFD systems do not focus
on minimization of visual data transmission that requires additional power. Finally, an excellent dataset with
high-quality features needs to extract and integrate well for better detection performance. Hence, efficient low-
cost, accurate, robust, and trust-based early forest fire detection is required to resolve the aforesaid problems in
such dynamic environments using available moisture contents.

Proposed method
In this research work, we assign the unique pair < ID, Location > to each GPS-enabled ISN and organize the ISNs
into an optimal number of clusters. These ISNs monitor forests and sense various environmental parameters
such as temperature, humidity, smoke, windy speed, level of toxic gases, light intensity etc. The ISNs are deployed
in such a way that sensors can cover large forest areas without overlap coverage situations. Moreover, due care
is given to the forest density during the deployment process since some barriers (such as trees) block the sensor
range. The base station (sink) is placed near to the group members due to the resource-constrained nature
of ISNs. We incorporate a novel factor known as trust score at the sensor level to assess the reliability of the
monitored data since ISNs can be damaged/attacked by some adversaries in a hostile environment. Moreover,
trust factors determine and diminish the effect of distance on the data integrity during data exchange between
two communicating ISNs. If any ISN is found malicious at any stage, it will be removed from the network to
get an accurate result. Within a group, each group member (GM) senses the environmental parameters, stores
them into a logical sliding window1, and sends the observed data along with the location to the group head
(GH). The GH maintains the record of data packets sent by GMs, processes them, and further sends them to
the base station. If any packet contains abnormal events such as a spark of fires, smoke, etc. then it is processed
and forwarded on high priority. At the base station, a ML algorithm will test the monitored data and generate
an alert message if fire accidents are detected. Moreover, if fire intensity and density are high then nearby ISNs
themselves generate an alert message provided that ISN is a trusted ISN. The Fig. 2 shows the complete flowchart
of proposed FFD system.
The trained data set consists of day and night scenarios, Peaceful fire, wildfire, low and high fire density levels,
downpour scenarios, road presence, fire intensity and volume, small-width fire propagating over a large distance,
fire hazard, etc. to minimize the false alarm rates. The distances between two ISNs are computed using Eq. (1)
where SR denotes the sensing range of each ISN.
D = 2 ∗ [SR ∗ sin(60◦ )](1)
−

Using the Eq. (1), the distance between two ISNs and GHs are 8.66 m and 88.66 m respectively. The sensor
coverage distance function is determined by the forest density (d) at the sensor’s location (f dj ), the forest density
at the specific point (f di ), and the sensor’s maximum coverage distance (SR). Thereby, the coverage sensor
(f d
j + fi
d
distance is [SR∗( 1 − 2∗fmax
d )]. We compute the threshold ratio ( RTh ) using Eq. (2) for each monitored
feature by performing various trials in different climatic zones. In Eq. (2), X may be any feature such as
Temperature, humidity, light intensity etc.
average value of the logical sliding window for feature (X)
RTh = (2)
latest value of the feature (X)

If threshold ratio is greater than threshold value of feature defined in our FFD system then it might be occurrence
of a fire.

Optimal number of clusters

The primary purpose for calculating the most optimal number of clusters WSN is to minimize the overall energy
required for data collection, in comparison to alternative clustering patterns within the WSN. Additionally, it
prolongs the lifespan of the WSN. If the number of clusters is not optimized, the overall energy consumption
of the network per round increases exponentially. The optimal result cannot be achieved when the clusters are
generated randomly. Clusters and network life span are closely linked so need to determine the optimal number
of clusters for a network. In WSNs, increasing the number of clusters helps distribute the workload evenly among
GHs, resulting in shorter communication distances between an ISN and its corresponding GH. Consequently,
the total energy usage is correspondingly diminished. Conversely, when the number of clusters is increased, the
communication line between a SD and the BS will involve additional GH, leading to a greater overall energy
usage. Thus, determining the ideal number of clusters is a crucial aspect for WSNs. We examine a network
consisting of Ns ISNs that are randomly placed in a M X M m2 forest area. This network is then partitioned
into Kopt (optimal value of clusters), which is the ideal number of clusters. Each cluster in WSN consists of
N
Kopt
SDs, with one SD designated as the ISN and the remaining ( Kopt N
-1) ISNs serving as members of
that cluster. As the BS is at the boundary consequently, free spaces (fs) as well as multipath (mp) losses are
considered. The Kopt are obtained using Eq. (3) as follows

Scientific Reports | (2025) 15:14480 | [Link] 5

[Link]/scientificreports/

Fig. 2. Flow chart of proposed work.

Scientific Reports | (2025) 15:14480 | [Link] 6

[Link]/scientificreports/

√
Ns ∗ ϵf s M
kopt = ∗√ (3)
2π ∗ Eavg ϵmp ∗ D2 + ϵf s

Where D represents the distance from the BS to the GH, and Ns is the total number of SDs divided among the
clusters. The Eq. (3) considers the average energy available per ISN (Eavg ) to reflect a realistic energy constraint
on ISNs which balances communication efficiency with ISN lifetime. On the other hand, it models the trade-off
between intra-cluster and inter-cluster communication very well through explicit modeling both free-spaceϵf s )
and multipath (ϵmp ​) energy. This modification enables the Eq. (3) to be broadly scalable across a diverse range of
network densities, forest coverages, and energy profiles. This improvement does benefit the network lifespan by
decreasing energy consumption yet allowing efficient data transmission. It also makes the optimization of cluster
design in WSNs much more generalized and thus can be adaptable to situations with a high level of complexity
and heterogeneity, such as those with limited energy or dynamic environments.

Trust assessment model

In this subsection, we discuss the trust evaluation model to assess the reliability of ISNs. The trust score of ISN
depends upon its activities (packet forwarding behavior, processing, ISN location, and ISN energy) monitored by
its neighboring ISNs in ∆ t time. The trust score is updated after every ∆ t time and stored into the forwarding
table maintained by ISNs. A malicious ISN always tries to drop/alter packet to damage the network. We compute
communication trust (direct and indirect), data trust, and energy trust to isolate selfish ISNs from the network.
Communication trust reflects the reliability that ISNs interact in the data transmission and message exchange.
Data trust evaluates the accuracy and reliability of the data shared by ISNs. The energy trust calculation typically
involves monitoring the energy levels of ISNs and adjusting trust scores accordingly. If a ISN’s energy drops
below a certain threshold, its trust score can be significantly punished, reflecting that low energy ISNs are either
unreliable or may not work as expected. Since sink ISN and GHs are trustworthy and powerful ISNs, we only
compute the GH to GM trust and ignore the base station to GH trust evaluation because it will increase the
unnecessary load on the networking which will reduce the power level of ISNs. A logical timing window1 is
employed to record the cooperative and non-cooperative interactions among ISNs. It records recent interaction
history and removes older information to make accurate decisions about the ISN’s reliability. Each ISN takes
0.05 joule to send and receive a data packet. Figure 3 shows the flowchart for trust evaluation.
Each GH stores configuration information of each GM present in its cluster in the form of a matrix < ID,
location, Energy > as shown in Eq. (4).
 
ID1, (X1 , Y1 ) E1
ID2, (X2 , Y2 ) E2 
M= 
... ... ... (4)
IDk−1, (Xk−1 , Yk−1 ) Ek−1

The GM (say x) to GM (say y) direct communication trust ( Tx,y C

(∆ t)) and data trust ( DTx,yD
(∆ t)) is
computed using Eqs. (5 and 6) as follows. The symbol S and U are amount of cooperative and non-cooperative
interactions between ISN (x) and ISN (y) at time (∆ t) . The parameter U reflect how the suggested system
handles data inconsistencies or false data injections. A higher value of U penalizes ISNs severely for data
discrepancies, discouraging malicious behavior. The symbol q is natural abnormal factor (NAF) that determine
the number of non-cooperative interactions due to natural calamities such as earthquake etc. In the suggested
research work, the symbol φ denotes the penalty component, whereas θ represents the threshold value for the
trust score. The parameter q reflects the extent to which negative interactions affect the trust score, while ϕ
describes the punishment for behaving uncooperatively. Higher value of these parameters are meant for higher
susceptibility of the network to misbehaviors such as packet loss or malicious attacks. For networks with high
packet loss rates, increasing the value of q would increase the penalty to poor communications. Tmax is the
highest trust value utilized in the research study. During the experimental investigation, we select 10 as the
maximum value for Tmax . Algorithm 1 calculates and updates the trust values inside a cluster. algorithm 2
provide the pseudo-code for data transmission between ISNs and BS.
[ ( ) ]
Sx,y (∆t) + 1 Sx,y (∆t) 1
C
Tx,y (∆t) = Tmax × ∗ ∗ √ (5)
(Sx,y (∆t) + q ∗ Ux,y (∆t)) + 2 Sx,y (∆t) + 1 2
φ ∗ Ux,y (∆t)
 ( D (∆t)+1
)
( ) Ux,y

 D
Sx,y(∆t) + 1 D (∆t)+q∗U D (∆t)+2
Sx,y 
(∆t) = Tmax ×
D x,y
DTx,y  (6)
(Sx,y
D
(∆t) + q ∗ Ux,y
D
(∆t) + 2)

The trust is constrained by a maximum value (T_max), preventing it from increasing without limit. The
( )
Sx,y (∆ t)+1
ratio represents the Beta likelihood ratio, indicating the probability of success
(Sx,y (∆ t)+q∗Ux,y (∆ t))+2
relative to the probability of failure, based on the recorded successes and failures for ISN (x) and ISN (y). When
Sx,y (∆ t) significantly exceeds Ux,y (∆ t), the ratio will approximate 1, demonstrating a high level of trust.
If Ux,y (∆ t) significantly exceeds Sx,y (∆ t), the ratio will diminish, indicating a lack of trust. Thus, the Beta
distribution’s correlation in Eq. (5) illustrates how the trust score is affected by both successes and failures,

Scientific Reports | (2025) 15:14480 | [Link] 7

[Link]/scientificreports/

Fig. 3. Flowchart for dynamic trust evaluation.

encouraging the model to adjust the level of trust accordingly. The term denotes the ratio of success probability to
failure probability, modified by constants that are frequently employed to maintain calculation smoothness and
avert divide-by-zero errors. This association with the Beta distribution underscores that the trust computation
is adaptively modified according to recorded successes and failures, similar to way in which a Beta distribution
revises its parameters when further communication exchanges is gathered.
The Eqs. (5 and 6) incentivizes positive interaction and penalizes negative behavior depending on the q and
ϕ parameters, which indicate the effectiveness of proposed model to managing untrustworthy activity. Data
trust can similarly be quantified using Eq. (6) that prioritizes consistency and accuracy in the data supplied by
ISNs, while energy trust can be computed by a linear or non-linear function contingent upon energy thresholds.
The above equations represents a balance between positive interactions and untrustworthy interactions. Positive
interactions enhance the trust score, whereas the score diminishes with an increase in unsuccessful interactions,
particularly if they are malevolent or erroneous. The penalty factor q and the severity factor ϕ enable precise
regulation of the impact of failures on confidence. The normalization process, achieved by multiplying Tmax in
Eq. (5), guarantees that the trust score is confined within an acceptable range (limited by Tmax) and prevents any
S (∆ t)
single factor from disproportionately influencing the trust computation. The second term Sx,yx,y (∆ t)+1 is vital
because successful interactions enhance the trust score; nevertheless, a large number of successful interactions
should not result in unreasonably increased trust values. Normalizing the success rate mitigates the excessive

Scientific Reports | (2025) 15:14480 | [Link] 8

[Link]/scientificreports/

1
inflation of the trust score. The third part √
2
introduces a severity factor to punish untrustworthy
ϕ *U x,y (∆ t)
behavior more effectively that moderates the penalty for untrustworthy interactions. As the frequency of
unsuccessful interactions escalates, the punishment intensifies. The factor ϕ modulates the penalty according
to the severity of untrustworthy behaves that deserve punishment. The square root guarantees that although
the penalty increases with Ux,y (∆ t) it does so in a manner that is not excessively severe. This segments of the
Eq. (5) evaluates the seriousness of failures and finds their impact on the trust score. Increased failure rates,
particularly those of a malevolent nature (such as blackhole attacks or selective forwarding), incur a more
severe penalty. A practical case would clarify the model’s functionality in real-world contexts. For instance, in
an industrial WSN application, consider that ISNs A, B, and C participate in communication. ISN A possesses a
record of trustworthy contact with ISN B, yet exhibits low energy, whereas ISN C frequently transmits erroneous
data yet maintains high energy levels. According to the weighted trust considerations, ISN A may be deemed
reliable for communication but penalized for energy consumption, whereas ISN C may face penalties for data
trustworthiness but receive rewards for energy efficiency. The ultimate trust score will represent the combination
of all of these factors, with the weighting modified according to the application’s priorities (e.g., communication
prioritized over energy or vice versa).

Algorithm 1. Trust Evaluation Algorithm

Algorithm 1 states that if the cooperative contacts between ISN(x) and ISN(y) are larger than zero and
there are no unsuccessful interactions, then the proposed system will assign the maximum trust value Tmax
. Furthermore, if the successful interactions between ISN(x) and ISN(y) are nonexistent and the unsuccessful
interactions are more than zero, the recommended approach will assign a trust value of zero to SNs. In addition,
if the number of favorable interactions between ISN(x) and ISN(y) is more than zero, and the number of negative
contacts is also greater than zero, then the suggested system assigns trust levels based on the calculations from
Eqs. (5 and 6). In the event that there are no successful or unsuccessful contacts, the suggested system utilizes
suggestions from peers to improve the accuracy of the proposed technique in a hostile environment. When
evaluating trust in peer recommendations, we only take into account trusted neighbors who have a direct
connection. This helps to minimize communication burden and enhance the accuracy of the Unified confidence
Model (UTM). The peer recommendation trust estimation enhances the resilience of trust value as it mitigates
the potential compromise of trust value by malicious ISNs. The peer recommendation trust is determined by
utilizing Eq. (7).
[( ∑ z )]
T
j=1 x,j
× Tj,y
P Rx,y (∆ t) = (7)
|z|

Where z represents the collection of SDs that are directly trusted. In this context, we exclude ISNs with
Tx,y (∆ t)(< Tmax
2 , which
) means we eliminate malicious ISNs in order to get an adaptable trust value. Final
trust value fx,y
T
(∆ t) is computed by simply aggregating Eqs. (5, 6 and 7) as follows using Eq. (8). the sum of
weights in Eq. (8) is 1. The status of a ISN is determined using Eq. (9) as follows.

Scientific Reports | (2025) 15:14480 | [Link] 9

[Link]/scientificreports/

x,y (∆ t) + w2∗ DTx,y (∆ t) + w3∗ P Rx,y (∆ t)

w1∗ T C D
T
fx,y (∆ t) = (8)
3
   Tmax   
Tmax +1
 
 2
; 2  highly trusted ISN 
 malicious ISN
S fx,y (∆t) =
T
  (0; θ)   (9)
 Tmax +1
θ;  legitimate ISN 
2

The significance of communication, data, and energy trust factors is dependent upon the context and the function
of the ISN within the IWSN. In an IWSN, when ISNs are required to deliver vital and highly precise data, data
trust may be prioritized over energy trust. In contrast, in situations where the network’s primary objective is to
optimize the operational longevity of ISNs, such as in remote monitoring or sensor networks, energy trust may
assume greater importance. The trade-off process can be represented as a weighted sum, in which each trust
element is multiplied by a weight that signifies its significance in the specific circumstance. The weights can
be dynamically adjusted according to the real-time performance of ISNs or the changing requirements of the
network. For example, if a ISN demonstrates persistently poor communication performance yet remains energy-
efficient, the system may recalibrate the weight assigned to communication trust to diminish its impact on the
overall score, while augmenting the weight of energy trust. The trade-offs of trust factors can be optimized using
empirical data or simulations that replicate various network situations. The weights allocated to each aspect
must be dictated by the network’s specific objectives, such as emphasizing dependability over energy efficiency.
An illustrative scenario could be an emergency response network where the dependability of communication
is crucial, therefore the communication trust component may prevail, however in a long-term monitoring
application, energy trust could be the determining factor. The symbol θ is “application-dependent trust threshold
whose value depends on application requirements. The energy trust of an ISN x at (∆ t) time ( ET x (∆ t)) is
computed with the help of a predefined threshold value Eth which is 20% of the total energy. If the current
energy level (CEL) of an ISN (let’s say x) is greater than 50% of total energy, then ISN (x) is highly trusted. If
20% of total energy < CELx ≤ 50% of total energy, then ISN (x) is trusted, else it is non-trusted. Let us,
we define the number of ISNs in WSN is Ns and the number of clusters (groups/cluster heads) are g then size (n)
of each cluster is Ns /g. We divide the total communication overhead into intra-cluster communication overhead
and inter-cluster communication overhead. In intra-cluster trust evaluation, ISN x sends and receive one GH
feedback request to interact with ISN y i.e. total communication overhead of 2 request packets. In the worst
case, if ISN x wants to interact with all (n − 2) ISNs then total communication overhead is 2 ∗ (n − 2) request
packets. If all GMs (except GH) wants to interact with each other then maximum communication overhead is
2 (n − 2)(n − 1). During intra-cluster feedback trust calculation, GH sends r requests to only direct trusted
members and receive r response where ( r ≤ n − 1). The total communication overhead due to feedbacks by
GH at GM level is 2 ∗ r request and response packets. Thus total communication overhead in intra-cluster trust
computation Cintra = 2 (n − 2)(n − 1) + 2r. In intra-cluster trust evaluation, GH (i) sends and receive
one BS feedback request to interact with GH (j) i.e. total communication overhead of 2 request packets. Thus
total communication overhead for g groups in inter-cluster trust computation” Cinter = 2g. Hence, in the most
unfavorable scenario, the highest level of communication overhead (CM ax ) for the suggested scheme (UTM)
can be expressed as:” CMax = g ∗ Cinter + Cinter = g ∗ (2 (n − 2) (n − 1) + 2r ) + 2g.

Scientific Reports | (2025) 15:14480 | [Link] 10

[Link]/scientificreports/

Algorithm 2. Pseudo-code for data transmission between ISNs and base stations

Fire detection algorithm (FDA)

This subsection discussed the fire detection algorithm (FDA) in forest areas. Algorithm 3 is an efficient fire
detection algorithm (FDA) that takes various environmental features as input and provides a correct outcome
in terms of fire ignitions. According to algorithm 3, if sensed parameter value exceeds the predefined threshold
value then it indicates the occurrence of a forest fire. In this scenario, an alarm will sound so that all the officers
will come to know about the forest fire at an earlier stage for necessary action. Sometimes malicious ISNs don’t
provide correct information and neither provides genuine feedback about neighbor ISN intentionally then in
such cases, the proposed algorithm computes the final trust score of ISN and removes the malicious ISNs. The
data sensed and processed by good ISNs are forward to the BS through GHs where a ML algorithm is applied
to the data for testing the model. If any fire ignitions are found during the testing process then Initiate an alert
notification, indicating the occurrence of a fire. The detailed fire detection algorithm (FDA) (algorithm 3) is
given below as follows.

Scientific Reports | (2025) 15:14480 | [Link] 11

[Link]/scientificreports/

Algorithm 3. Fire detection algorithm (FDA)

Results and discussion

This section discusses “the effectiveness of the proposed model (UTM) by conducting a series of experiments
at different times on real recorded data in Sanjay Van Delhi, Hauz Khas Village Delhi, and Jawaharlal Nehru
University Campus Delhi. Publicly available datasets were also analyzed in this study. The data can be found
here: FD-dataset21, ForestryImages22, VisiFire23, BowFire24. The forest fire data set of 7000 images (875 non-
forest fire images) is prepared in the best possible way by observing day-night scenario, morning, evening, and
sunshine scenario, normal and artificial fire (controlled) scenario, road presence scenario, highly dense and low
dense area, fire outbreak, fire spread etc during different months ( June to August) of the year 2021. Some fire
images are collected by crawling online forest fire images. Moreover, xBD dataset20 is one of the public datasets
of high resolution labeled satellite images. The dataset on natural disaster images is updated by the Massachusetts
Institute of Technology (MIT). It covers 22,068 images of 19 different disasters. The image resolution is
1,024 × 1,024. This paper also uses the images containing forest fires in xBD dataset. The operation system is
Windows 10 (version 21H1); the graphics processing unit (GPU) is GeForce GTX 1080Ti; the central processing
unit (CPU) is Intel Core i710500U with a RAM of 32GB and a hard disk of 2 T B . Our model was compared
with recent models7,8, using the learning rate of 0.01 and batch size of 64. All input images were adjusted to the
size of 256 × [Link] environmental conditions (temperature, humidity, etc) are measured on every 7 seconds,
and observed data were sent to the BS every 30 seconds. The performance of UTM is evaluated using various
metrics such as false positives, false negatives, precision, recall, F- Score, accuracy shown in Table 2. A false
positive indicates erroneously that a fire has occurred, whereas a false negative fails in identifying when a fire is
really” occurring. Table 3 provides the list of parameters and smart sensors used in the research work.
In our UTM-based fire detection system, ISNs are employed with careful placement to maximize coverage
and energy-efficient data acquisition. The forest region is divided into clusters and ISNs are placed in every cluster
to monitor different environmental factors like temperature, humidity, intensity of light, carbon monoxide, and
concentration of smoke. Each ISN is given a distinct (ID, location) pair from GPS for efficient tracking and

Parameter Formula
F alse positive (F P )
False positive rate (FPR) F alse positive (F P )+T rue positive (T P )
∗ 100
F alse negative (F N )
False negative rate (FNR) F alse negative (F N )+T rue positive (T P )
∗ 100
T rue positive (T P )
Precision F alse positive (F P )+T rue positive (T P )

T rue positive (T P )
Recall F alse negative (F N )+T rue positive (T P )

2 ∗ P recision
∗Recall
F- score
P recision
+Recall
T rue positive (T P )+ T rue negative (T N )
Accuracy Number of positives +Number of negatives

Table 2. Formulas used in the work.

Scientific Reports | (2025) 15:14480 | [Link] 12

[Link]/scientificreports/

DHT22 Sensor To monitor temperature and relative humidity

LDR Sensor To monitor light intensity
MQ9 Sensor To monitor CO level
PIR Sensor Observe the locomotion of human people within the forest
IR Flame Sensor and MQ-2 sensor Fire and smoke sensor
DHTII Soil moisture sensor
Microcontroller Device Arduino nano board
Transceiver nrf24l01 Module (used for data transmission)
Primary Power Source Rechargeable Lithium Iron Cells (2 Cells In Series, 2 Cells In Parallel)
Secondary Power Supply Solar Panel (It Gives 5 W Power And 12 V Voltage Output)
Number of ISNs 200
Duty Cycle 0.33%
Toal SDs 200
Maximum number of packets 1400
Maximum sensing range (SR) of SD 5m
Maximum sensing range (SR) of cluster head 50 m
Distance between two SDs 8.66 m
Distance between two cluster heads 86.66 m
Number of data samples collected 7200
Training data 80%
Test data 20%
Humidity (0 to 100 ppm), temperature (0 to 590 °C), Light intensity
(0 to 10000 lx), carbon dioxide (500 to 5000 ppm), smoke, and
Features/parameters
flame presence, soil moisture content, barometric pressure, carbon
monoxide, and nitrogen dioxide level
Threshold value According to weather conditions (Day, night, summer winters)
Maximum trust value (Tmax) 10
Trust threshold ( θ ) 5
Natural abnormal factor (q) 1

Table 3. List of parameters and sensors used in the research work.

Fig. 4. Forest images (fire and non-fire) in our dataset, including ground fire, canopy fire and trunk fire.

coordination. Optimal number of clusters is calculated using proposed Eq. (3) of this research paper, maintaining
intra-cluster and inter-cluster communication for energy conservation. ISN distance is 8.66 m and GH distance
is 88.66 m for effective data relay to the base station. Each ISN is equipped with rechargeable lithium-ion battery
with solar panel backup and sends data using the nRF24L01 module. Arduino Nano, nRF24L01, and SIM800L
are used by the base station for data processing and alarm generation.
Figure 4 shows the some forest images of our data set that include various types of fire including ground
fire, canopy fire, and trunk fire. Figure 5 shows the detection of forest fire using the proposed model (UTM)
and highlights their locations. We randomly give some forest images to the ML algorithm and it outputs the fire
occurrence status by testing the input images with trained data set. Table 4 shows the effect of logical window size

Scientific Reports | (2025) 15:14480 | [Link] 13

[Link]/scientificreports/

Fig. 5. Forest fire detection and highlighting there location.

Window size Number of experiments Fires detected False positives False negatives
10 25 17 3 5
20 25 19 2 4
30 25 20 2 3
40 25 25 0 0
50 25 25 0 0

Table 4. Effect of window size.

Nitrogen Barometric Carbon

Humidity Light Intensity Carbon dioxide dioxide level Soil moisture pressure monoxide
Temperature (°C) (ppm) (lux) level (ppm) (ppb) Smoke (%) (mb) (ppm) Output
10 90 150 600 50 200 30 1000 50 No Fire
15 40 200 700 60 100 20 1001 100 No Fire
20 60 250 900 70 250 40 1007 150 No Fire
25 10 300 1100 100 50 60 1003 200 No Fire
30 15 75 1500 150 125 70 1009 250 No Fire
35 65 96 2000 180 300 50 1005 130 No Fire
45 45 400 1100 200 170 35 1006 270 No Fire
70 8 600 3000 230 350 25 1012 300 Fire
90 35 700 3500 250 225 45 1008 350 Fire
120 45 1000 4000 280 400 40 1010 400 Fire
150 60 2000 1200 300 350 30 1015 423 Fire
200 75 3000 1400 325 450 40 1020 450 Fire
300 80 5000 3600 340 500 50 1022 500 Fire

Table 5. Experimental output for different situations.

on fire detection accuracy. We performed several trials with different window sizes and observed that window
size in the range of 40 to 50 reduces the false alarms.
We also perform various experiments using the environmental parameters observed by the smart sensors
in the forest area. Table 5 shows a detailed experimental analysis with the outcome generated by the proposed
scheme. We give a set of parameter values to the fire detection algorithm as input and get correct output with low
latency. The observed findings presented in this research substantiate the significance and efficacy of UTM in any
environmental condition because we have tried to collect all possible scenarios of a forest in our dataset. Table 6

Scientific Reports | (2025) 15:14480 | [Link] 14

[Link]/scientificreports/

Method Accuracy (%) Recall (%) F- score (%) Precision (%)

UTM 98.5 98 98.05 98.12
7 96.58 97 97 97
EEFFL8 93.73 95.5 94.75 94

Table 6. Performance comparison of the UTM vs. other methodologies EEFFL7,8.

Method False-positive (%) False-negative (%) Accuracy (%)

UTM 3.1 2.3 98
7 3.4 2.9 96.82
EEFFL8 6.2 5.5 93.73

Table 7. Performance of the UTM vs. EEFFL7,8.

Fig. 6. Effectiveness of UTM against malicious activity.

provides the performance comparison of the UTM vs. other methodologies EEFFL7,8 in terms of accuracy (%),
recall (%), F-score (%), precision (%). The performance metrics are evaluated using the formulas given in Table 2.
Moreover, we evaluated the percentage of false positive and false negative with accuracy (%) as given in Table 7.
We have analyzed the effectiveness of the proposed model in the presence of malicious ISNs using Fig. 6. We
insert 30% malicious ISNs into the network and observe that 99% malicious ISNs are identified; 98% malicious
ISNs are mitigated in the network of 20 ISNs. Moreover, we achieved 98.5% accuracy in the detection and
mitigation of malicious behavior under the network load of 20 ISNs.
Similarly, we observed the performance by increasing the network load in the presence of malicious ISNs.
Finally, we observed the performance in the presence of 200 ISNs (140 good ISNs, 60 malicious ISNs) and found
that 90% malicious ISNs are identified 84% malicious ISNs are mitigated with 87% accuracy. The primary factor
contributing to this exceptional success is the utilization of an effective trust evaluation model in this research
work. The proposed trust model identifies the malicious behaviors by continuously observing the ISNs and
mitigates them. Furthermore, our trust model improves accuracy because it employs a multi-trust approach.
The multi-trust approach considers communication patterns among ISNs, data exchanged among ISNs, and
residual energy. Hence it is not possible to misguide our proposed trust model in a open environment. Even, the
suggested trust model considers the effects of natural calamities using q. for simplification, we choose q value

Scientific Reports | (2025) 15:14480 | [Link] 15

[Link]/scientificreports/

Fig. 7. ISN placement height (m) from the ground vs. average delay (s).

as 1 during simulation. Figure 7 shows the ISN placement height (m) vs. variation in average delay (sec). After
conducting several experiments, the good sensing range and ISN mounting height is determined as 5 m and 1 m
respectively.
Moreover, Fig. 7 depicts that ISN placement height greater than 1 m leads to great delay. Hence 1 m mounting
height is determined as optimal height for ISN placement. The average delay at 1 m height is 30 s, at 1.5 m height
is 50 s and at 2 m, the average delay is 123 s. Figure 8 shows the experimental outcome obtained by maintaining
distance (m) between fire ignition location (FIL) and ISN and the time required for the ISN to sense the threat
(sec). It is observed that when the distance between FIL and ISN is 8 m then the time required for the ISN
to sense the threat is 5 s, 3 s, and 2 s in EEFFL7,8, and UTM respectively. Furthermore, when we increase the
distance between FIL and ISN then the time required for the ISN to sense the threat increases rapidly. When
the distance between FIL and ISN is 20 m then the time required for the ISN to sense the threat is 18 s, 13 s, and
11 s in EEFFL7,8, and UTM respectively. Furthermore, when the distance between FIL and ISN is 40 m then time
required for the ISN to sense the threat is 40 s, 30 s, and 25 s in EEFFL7,8, and UTM respectively. Figure 9 depicts
the average fire detection time (sec) with the rise in the number of clusters (groups) in normal and abnormal
environmental conditions (flooding, ecological imbalance, downpour, etc.). Each cluster contains an optimal
number of ISNs. When there are 2 clusters, the average time it takes to notice a fire is 3 s in normal environmental
conditions and 7 s in abnormal environmental conditions. In addition, when there are 4 clusters, the average
time it takes to discover a fire is 9 s in normal environmental conditions and 17 s in abnormal environmental
conditions. Similarly, when the number of clusters is increased to 10, the average time it takes to detect a fire is
25 s in normal environmental settings and 38 s in severe environmental conditions. Finally, when the number
of clusters is increased to 18, the average time it takes to notice a fire is 50 s in normal environmental settings
and 80 s in abnormal environmental conditions. Therefore, it may be inferred that atypical circumstances lead
to a longer average duration for detecting fires. Our technique, UTM, is capable of successfully detecting fires in
atypical circumstances, but with a slightly longer response time.
Figure 10 shows the percentage of fires correctly identified with the rise in ISNs in the forest area. It is
observed that when the numbers of ISNs are 20 then the percentage of fires correctly identified is 89%, 93%, and
98% in EEFFL7,8, and UTM respectively. Furthermore, when we double the ISNs (i.e. 40 ISNs) then percentage of
fires correctly identified 85%, 90% and 97% in EEFFL7,8 and UTM respectively. Again, when we double the ISNs
(i.e. 80ISNs) then percentage of fires correctly identified 80%, 85% and 92% in EEFFL7,8 and UTM respectively.
Similarly, we test the accuracy with 160 ISNs and observed that the percentage of fires correctly identified is 63%,
75%, and 82% in EEFFL7,8, and UTM respectively. In the end, we place all 200 ISNs in the forest area and found
that the percentage of fires correctly identified was 57%, 70%, and 78% in EEFFL7,8, and UTM respectively. In
each case, our scheme (UTM) correctly identifies the highest percentage of fires in the forest as well as a hostile
environment.

Scientific Reports | (2025) 15:14480 | [Link] 16

[Link]/scientificreports/

Fig. 8. Average distance (m) between fire ignition location (FIL) and sensor node vs. Time required for the
sensor node to sense the threat (sec).

Fig. 9. Number of clusters vs. Average fire detection time (sec).

Scientific Reports | (2025) 15:14480 | [Link] 17

[Link]/scientificreports/

Fig. 10. Number of ISNs vs. percentage of fires correctly identified.

Fig. 11. Number of packets vs. throughput.

To check the lifespan of intelligent WSN, we compute the throughput as depicted in Fig. 11. Throughput
is defined as the “number of packets successfully received at the BS”. In our experiment, when the number of
packets in the network is 200 then the number of packets reached successfully at the destination ISN (BS) are
195,189, 200 in EEFFL7,8, and UTM respectively. After some time, when the number of packets increased to 500
then the number of packets reached successfully at the destination ISN (BS) are 498,456, 499 in EEFFL7,8, and
UTM respectively. Furthermore, when the number of packets reached to 1000 in the WSN then the number of
packets reached successfully at the destination ISN (BS) are 990,900, 998 in EEFFL7,8, and UTM respectively.
From Fig. 11, It is clear that UTM achieves the highest throughput in the normal environment in the absence of
malicious ISNs. However, if we inject some malicious ISNs into the network then throughput will reduce. Energy
is one of the most important parameters for the survival of WSNs. We have analyzed the energy consumption
(in KJ) with the rise in the number of packets in the network as shown in Fig. 12. it is observed that when the
number of packets in the network is 500 then energy consumption is 0.007 KJ,0.006KJ,0.005KJ in EEFFL7,8,

Scientific Reports | (2025) 15:14480 | [Link] 18

[Link]/scientificreports/

Fig. 12. Number of packets vs. energy consumed (KJ).

and UTM respectively. After some time, when number of packets increased to 800 then energy consumption is
0.0168KJ, 0.017KJ, 0.015KJ in EEFFL7,8 and UTM respectively. Furthermore, when number of packets reached
to 1100 in the WSN then energy consumption is 0.045KJ, 0.05KJ, 0.03KJ in EEFFL7,8 and UTM respectively.
Additionally, when number of packets increased to 1400 in the WSN then energy consumption is 0.07KJ, 0.08KJ,
0.05KJ in EEFFL7,8 and UTM respectively.
We have also analyzed the remaining energy level percentage with the rise in the distance among ISNs using
Fig. 13. It is observed from Fig. 13 that when the distance among ISNs is 8 m then the remaining energy is
88%, 90%, and 95% in EEFFL7,8, and UTM respectively. When the distance among ISNs is increased to 10 m
then the remaining energy is 82%, 88%, and 92% in EEFFL7,8, and UTM respectively. Furthermore, When the
distance among ISNs is increased to 20 m then the remaining energy is 56%, 60%, and 65% in EEFFL7,8, and
UTM respectively. We increase the distance among ISNs up to 40 m and found that the remaining energy is
5%, 7%, and 10% in EEFFL7,8, and UTM respectively. Form the above discussion; it is clear that the remaining
energy level in UTM is better than EEFFL7,8 at every distance. Another conclusion can be drawn that the optimal
distance among ISNs is approximately 8 m.

Conclusion and future work

This research study examines the application of the trust model and ML in early forest fire detection (FFD)
systems through the use of an intelligent WSN. The research study presents a ML-based trust-aware early forest
fire detection (FFD) system (UTM) that utilizes an intelligent WSN. This system provides precise and reliable
results even in the presence of varying temperature, humidity, smoke, light intensity, and other factors. A high-
quality dataset consisting of 7200 data samples is created by carefully studying real photos, publicly available
datasets, and forestry datasets. Out of these samples, 80% are designated as training data, while the remaining
20% are allocated for testing purposes. The many types of deployed ISNs detect environmental characteristics,
such as temperature and humidity, and transmit the data to the group leader. The group leader analyzes
the obtained data and transmits it to the central station, where the ML algorithm is employed to detect fire
outbreaks. If any instances of fire ignition are detected at any point, an alert will be activated to notify forest
officers. We utilize a highly effective trust evaluation technique to assess the dependability of each ISN. If any ISN
is determined to be malicious at any point, it is expelled from the network and cannot continue to participate.
Only reliable data is transmitted to the central station to facilitate effective decision-making. The UTM system is
a forest fire detection (FFD) system that operates at two levels. The first level of detection occurs at the individual
ISN level, while the second level of detection takes place at the base station using a ML algorithm. Experimental
data are collected in both normal and abnormal environments to analyze the identification and mitigation of
malicious behavior. The experiments measure several factors such as accuracy, node placement height, latency,
false negative rate (FNR), false positive rate (FPR), precision, recall, F-score, fire detection time, throughput,
and residual energy. The testing results consistently demonstrate the exceptional performance of our suggested

Scientific Reports | (2025) 15:14480 | [Link] 19

[Link]/scientificreports/

Fig. 13. Average distance (m) between ISNs vs. remaining energy level (%).

approach, Universal Terrain Model (UTM), across various climatic circumstances. Although the UTM greatly
improves early-stage detection of forest fires using intelligent WSNs and ML technique, we recognize that some
aspects may affect its efficiency. One such significant limitation is sensor placement and density dependence—
sparse placement can leave a gap in coverage, causing delayed detection or false negatives. Second, the system
relies on accurate sensor readings, but harsh environmental factors like heavy rains, heavy winds, or thick fog
may impair sensor integrity, generating more false positives. Energy constraints of sensor nodes can affect trust
computation, which may penalize low-energy nodes even if providing accurate information, which may degrade
overall detection accuracy. Last but not least, although the trust model tries to identify malicious nodes, it may
be vulnerable to coordinated attacks, e.g., collusion among existing malicious nodes that adjust trust ratings to
avoid detection. The machine learning module, although efficient, uses pre-trained models, which may need
to be re-trained occasionally to accommodate new fire scenarios, which is hard to do in real-time applications.
Lastly, network congestion in real-world large-scale deployments may affect data transmission efficiency,
resulting in higher latency in fire alerts. To address these limitations, future work will address optimizing sensor
placement strategies, utilizing adaptive learning models, and edge computing for real-time computation. We will
also investigate federated learning for dynamic model adaptation, adaptive clustering algorithms for low-latency
communication, and hybrid energy-harvesting mechanisms to prolong sensor lifespan. We believe by outlining
these potential limitations in our manuscript, we will give a more complete picture and further support our
contribution.

Data availability
the dataset was uploaded to the opensource website ([Link] The data
can be found here: FD-dataset21, ForestryImages22, VisiFire23, BowFire24.

Received: 9 August 2024; Accepted: 16 April 2025

References
1. Muhammad, K., Khan, S., Elhoseny, M., Ahmed, S. H. & Baik, S. W. Efficient fire detection for uncertain surveillance environment.
IEEE Trans. Ind. Inform. 15(5), 3113–3122 (2019).
2. Nemade, B. et al. IoT-based automated system for water-related disease prediction. Sci. Rep. 14, 29483. ​ht​ ​t​p​s​:​/​/​d​o​i.​ ​o​r​g​/​1​0​.​1​03​ ​8​/​s​4​
1​5​9​8​-0​ ​2​4​-​7​9​9​8​9​-​6​​​​ (2024).
3. Almeida, J. S., Huang, C., Nogueira, F. G., Bhatia, S. & de Albuquerque, V. H. Edgefiresmoke: A novel lightweight cnn model for
real-time video fire–smoke detection. IEEE Trans. Ind. Inform. 18(11), 7889–7898 (2022).
4. Neji, W., Othman, S. B. & Sakli, H. T-LEACH: Threshold sensitive low energy adaptive clustering hierarchy for wireless sensor
networks. In 2020 20th International Conference on Sciences and Techniques of Automatic Control and Computer Engineering (STA),
Monastir, Tunisia, 336–342, [Link] (2020).

Scientific Reports | (2025) 15:14480 | [Link] 20

[Link]/scientificreports/

5. Harjoko, A., Dharmawan, A., Adhinata, F. D., Kosala, G. & Jo, K. H. Real-time forest fire detection framework based on artificial
intelligence using color probability model and motion feature analysis. Fire 5(1), 23 (2022).
6. Bouakkaz, F., Ali, W. & Derdour, M. Forest fire detection using wireless multimedia sensor networks and image compression.
Instrum. Mes. Métrol. 20, 57–63 (2021).
7. Vikram, R., Sinha, D., De, D. & Das, A. K. EEFFL: energy efficient data forwarding for forest fire detection using localization
technique in wireless sensor network. Wireless Netw. 26(7), 5177–5205 (2020).
8. Saponara, S., Elhanashi, A. & Gagliardi, A. Real-time video fire/smoke detection based on CNN in antifire surveillance systems. J.
Real-Time Image Proc. 18(3), 889–900 (2021).
9. Anilkumar, P. et al. An adaptive DeepLabv3+ for semantic segmentation of aerial images using improved golden eagle optimization
algorithm. IEEE Access 11, 106688–106705. [Link] (2023).
10. Fabi, A. K. & Thampi, S. M. A trust management framework using forest fire model to propagate emergency messages in the
Internet of Vehicles (IoV). Veh. Commun. 33, 100404 (2022).
11. Li, C. et al. Fast forest fire detection and segmentation application for UAV-assisted mobile edge computing system. IEEE Internet
Things J. 11, 26690 (2023).
12. John, J., Harikumar, K., Senthilnath, J. & Sundaram, S. An efficient approach with dynamic multiswarm of UAVs for forest
firefighting. IEEE Trans. Syst. Man Cybernet. Syst. 58, 2860 (2024).
13. Dampage, U., Bandaranayake, L., Wanasinghe, R., Kottahachchi, K. & Jayasanka, B. Forest fire detection system using wireless
sensor networks and machine learning. Sci. Rep. 12(1), 1–11 (2022).
14. Mahmoud, M. A. I. & Ren, H. Forest fire detection and identification using image processing and SVM. J. Inf. Process. Syst. 15(1),
159–168 (2019).
15. Kinaneva, D., Hristov, G., Raychev, J., & Zahariev, P. Early forest fire detection using drones and artificial intelligence. In 2019 42nd
International Convention on Information and Communication Technology, Electronics and Microelectronics (MIPRO) 1060–1065.
(IEEE, 2019).
16. Xu, R., Lin, H., Lu, K., Cao, L. & Liu, Y. A forest fire detection system based on ensemble learning. Forests 12(2), 217 (2021).
17. Benzekri, W., El Moussati, A., Moussaoui, O. & Berrajaa, M. ‘Early forest fire detection system using wireless sensor network and
deep learning. Int. J. Adv. Comput. Sci. Appl 11(5), 496–503 (2020).
18. Wu, D. et al. Forest fire recognition based on feature extraction from multi-view images. Traitement du Signal 38(3), 775 (2021).
19. Kizilkaya, B., Ever, E., Yatbaz, H. Y. & Yazici, A. An effective forest fire detection framework using heterogeneous wireless
multimedia sensor networks. ACM Trans. Multimed. Comput. Commun. Appl. (TOMM) 18(2), 1–21 (2022).
20. Bai, Y. et al. Pyramid pooling module-based semi-Siamese network: A benchmark model for assessing building damage from xBD
satellite imagery datasets. Remote Sens. 12(24), 4055 (2020).
21. FD-Dataset. Availabe online: [Link] (Accessed on 01 Januray 2021).
22. ForestryImages. Availabe online: ​h​t​t​p​s​:​​/​/​w​w​w​.f​​ ​o​r​e​s​t​​r​y​i​m​a​g​​e​s​.​o​r​​g​/​b​r​o​w​​s​e​/​s​ub ​ ​​t​h​u​m​b​.​​c​fm
​ ​?​s​u​b​=​7​4​0 (Accessed on 01 Januray 2021).
23. VisiFire. Availabe online: [Link] (Accessed on 01 Januray 2021).
24. BoWFire Dataset. Availabe online: ​ht​ ​t​p​s​:​​​/​​/​b​i​t​b​u​c​k​e​​t​.​or​ ​​g​/​g​​b​​d​i​/​b​​o​w​f​i​​r​​e​-​da​ ​t​a​​​s​e​t​/​d​​o​w​n​l​o​a​d​s​/ (Accessed on 01 Januray 2021).
25. Alkhatib, A. A. A review on forest fire detection techniques. Int. J. Distrib. Sens. Netw. 10(3), 597368 (2014).
26. Yaacob, N., Najib, M. S. M., Tajudin, N., Yusof, A. L. & Kassim, M. Image processing based forest fire detection using infrared
camera. J. Phys. Conf. Ser. 1768(1), 012014 (2021).
27. Chowdary, V., Gupta, M. K. & Singh, R. A Review on forest fire detection techniques: A decadal perspective. Networks 4, 12 (2018).
28. Bayo, A., Antolín, D., Medrano, N., Calvo, B. & Celma, S. Early detection and monitoring of forest fire with a wireless sensor
network system. Procedia Eng. 5, 248 (2010).
29. Tehseen, A., Zafar, N. A., Ali, T., Jameel, F. & Alkhammash, E. H. Formal modeling of IoT and drone-based forest fire detection
and counteraction system. Electronics 11(1), 128 (2022).
30. Zheng, X., Chen, F., Lou, L., Cheng, P. & Huang, Y. Real-time detection of full-scale forest fire smoke based on deep convolution
neural network. Remote Sens. 14(3), 536 (2022).
31. Seydi, S. T., Saeidi, V., Kalantar, B., Ueda, N. & Halin, A. A. Fire-Net: A deep learning framework for active forest fire detection. J.
Sensors 2022, 1 (2022).

Acknowledgements
This research was financially supported by Princess Nourah bint Abdulrahman University Researchers Support-
ing Project number (PNURSP2025R235), Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia.

Author contributions
All authors contributed equally to the conceptualization, formal analysis, investigation, methodology, and writ-
ing and editing of the original draft. All authors have read and agreed to the published version of the manuscript.

Funding
This research was financially supported by Princess Nourah bint Abdulrahman University Researchers Support-
ing Project number (PNURSP2025R235), Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia.

Declarations

Competing interests
The authors declare no competing interests.

Additional information
Correspondence and requests for materials should be addressed to M.G.
Reprints and permissions information is available at [Link]/reprints.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.

Scientific Reports | (2025) 15:14480 | [Link] 21

[Link]/scientificreports/

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives

4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in
any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide
a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have
permission under this licence to share adapted material derived from this article or parts of it. The images or
other third party material in this article are included in the article’s Creative Commons licence, unless indicated
otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence
and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to
obtain permission directly from the copyright holder. To view a copy of this licence, visit ​h​t​t​p​:​/​/​c​r​e​a​ti​ ​v​e​c​o​m​m​o​
n​s.​ ​o​r​g​/​l​i​c​e​ns​ ​e​s​/​b​y​-​n​c​-​n​d​/​4​.​0​/​​​​.​​

© The Author(s) 2025

Scientific Reports | (2025) 15:14480 | [Link] 22