Emedia:ISSN1529-7306 Emedia/Online Inc.
Emedia: ISSN 1529-7306 Vol.39_1, 1-23
Copyright © 2026 The Authors. Published by Emedia/Online Inc. This work is licensed under a
Creative Commons Attribution 4.0 International License, which permits unrestricted use,
distribution, and reproduction in any medium, provided the original work is properly cited.
Fine-Scale Forest Fire Monitoring and Early Warning via
IoT and Remote Sensing
Junbo Liu1*,YufengTai1,Saipeng Zhang2,Changlong Gao2,Feilong Yi1,Jiashuai
Li2,Jianchi Yu3
1Jilin Electric Power Research Institute Co., Ltd.,China
2State Grid Jilin Electric Power Company Limited Electric Power Research Institute,China
3Tonghua Power Supply Company,State Grid Jilin Electric Power Co.,Ld.,China
*Corresponding Author:Junbo Liu
Email:junboliu87@163.com
Abstract:Forest fires pose a significant threat to natural habitats, leading to profound
changes in the biological, chemical, and physical properties of forests. Accurate and timely
prediction of fire occurrence is essential for proactive forest management and disaster
mitigation. However, existing monitoring systems lack the precision and timeliness required
for effective early fire detection and risk mitigation. This study proposes a hybrid wildfire
monitoring and early alert framework that combines IoT ground sensors with remote sensing
data to forecast forest fire risk, enabling proactive early warnings for forest protection and
disaster mitigation. By collecting IoT sensors and satellite imagery, the data are refined
through preprocessing steps that involve filtering out noise and intelligently estimating
missing values, as well as satellite image correction. Followed by multi-modal feature fusion
through locality-sensitive hashing and rank Gauss normalisation to create a uniform feature
set. From the fused data, a convolutional neural network-based analytical model enhanced
with triplet loss is applied to predict fire ignition probability and classify risk levels (high
risk/ no risk). The findings indicate that the proposed model achieved an accuracy of 97.20%,
a precision of 96.98%, a recall of 97.71%, and an F1-score of 97.34%, significantly
outperforming existing architectures. These results validate the framework as a scalable,
data-driven early warning solution for fine-scale wildfire prediction and sustainable forest
management.
Keywords: Forest Fire Monitoring, Early Warning System, IoT Sensors, Remote Sensing,
Convolutional Neural Network, and Disaster Mitigation
1. Introduction
Forest fires are unchecked or mismanaged fires that take place in grasslands and other
natural terrain, which in most cases have considerable ecological, economic and social
impacts [1]. Such fires have the potential to destroy vegetation, degrade biodiversity and
release a lot of carbon dioxide in the air [2], and are a major source of threat to human safety.
Wildfires are a long-standing phenomenon in the world, and their prevalence is rising amidst
climate change, prolonged droughts, hot temperatures and other human activities [3]. A
review of literature within 20-30 years revealed that wildfire events have increased
significantly, especially in places like North America, Australia [4]. In the Mediterranean
region, dry seasons and lightning sparks are among the factors contributing to an increase in
fire incidences. History and satellite-based measurements have revealed that millions of
Junbo Liu et al.
2
hectares of forests are hit every year, and some areas are known to have repeated seasonal
fires for decades [5], [6]. The ecological and economic effects of such fires are far-reaching,
including depletion of vegetation cover, degradation of soil, smoke emissions, and other
effects that pose a great danger to both wildlife and human beings [7]. In the past, monitoring
relied on the capabilities of ground patrols and watchtowers. However, recent technological
advances have enabled the integration of remote sensing (satellites and aerial drones),
allowing for continuous surveillance [8]. These monitoring systems, combined with current
computational methods, can produce approximate estimates of fire-prone regions, monitor
the progress of current fires, and estimate possible outbreaks [9], thereby enabling effective
preparations and resistance.
Artificial Intelligence (AI) has become a disruptive technology in the field of environmental
surveillance and disaster management in recent years [10]. These calculation procedures
enable machines to perform functions that previously required human intelligence. Deep
learning (DL) is a subdiscipline of AI to determine intricate characteristics of large and
heterogeneous data [11]. In the last ten years, the technologies have achieved considerable
momentum, and multiple studies have shown that the technologies can process satellite
images [12], Internet of Things (IoT) sensor data [13] and meteorological data with accuracy
and speed never before possible. DL combined with IoT devices enable continuous tracking
and offers a stream of data in constant motion inside forested areas [14]. On the same note,
remote sensing platforms, such as satellites and drones, provide high-resolution space and
time imagery that can identify vegetation conditions, thermal signals, and the presence of
fire patches and hotspots [15]. With the aid of the integration of IoT sensor measurements
and satellite-derived data, AI models can sufficiently acquire both time-dependent and
space-dependent trends, which will allow making strong predictions with probability and
risk classification [16]. The benefits of applying this to enhance proactive and timely
warnings, and the safeguarding of large forested regions through proper and precise
monitoring [17]. Therefore, the integration of AI has transformed traditional wildfire
response in forest fire control methods to enhance situational awareness and enable
proactive decision-making in real-time, thereby minimising ecological and socio-economic
impacts.
1.1. Problem Statement
The benefits of forest fire monitoring and AI-based prediction are evident, but existing
methods still have several unresolved issues. To begin with, a majority of satellite-based fire
detection techniques have low temporal resolution [18], thus making it slow to detect small-
scale ignition sources and limiting their ability to provide fine-scale early warnings.
Moreover, the system based on IoT sensors can also be continuous, but due to the noise in
data, it is usually restricted and is unable to assess the fire risk correctly [19]. In addition,
the current models have a poor system of integrating multimodal data, a factor that makes
the usefulness of the data streams' complementary information underutilised, thus lowering
the strength of the predictions [20]. Also, many AI applications for wildfire risk prediction
struggle with skewed data and poor feature representations [21], which limits their ability to
generalise effectively across various forest regimes and changing climate dynamics. Despite
the numerous approaches investigated, these outstanding challenges underscore the need for
a fine-scale framework that enables more precise monitoring of forest fires and early
warning systems to mitigate losses and damages from disasters. Thus, the proposed
methodology has been developed to fill these gaps.
Fine-Scale Forest Fire Monitoring and Early Warning via IoT and Remote Sensing
3
The major innovations introduced by this framework are listed as follows.
Develops a hybrid system of forest fire monitoring that combines the use of IoT sensors and
remote sensing data to predict forest fire risk and make proactive early warnings to protect
forests.
Processes of IoT data to handle missing values, satellite image processing, atmospheric
correction and index retrievals to provide high-quality inputs.
Utilise Locality-Sensitive Hashing (LSH) to combine multi-modal features of IoT sensors
and remote sensing data to generate a unified set of features in predicting fire risks.
Determines the probability of forest fires based on a CNN-based model with triplet loss as
the optimal predictor of fire ignition potential and risk categorisation (high risk/ no risk).
The paper’s layout and flow are organised as follows: Section 2 provides an overview of
current wildfire monitoring approaches reported in the literature. Section 3 provides system
methodology. The analysis and discussion of the results of the experiment are presented in
Section 4. Lastly, Section 5 concludes the research and presents possible avenues of future
research.
2. Literature Review
Forest fire monitoring has been a primary focus in most countries, with researchers such as
Zheng et al. [22], Pang et al. [23], and Lin et al. [24] developed an AI-based prediction model
to improve early detection and risk assessment. They proposed methods utilising
convolutional neural networks (CNN) to extract informative features from forest blaze
imagery, thereby capturing discrimination. Their proposed 15-layer CNN model detects the
presence of high-risk fire zones by combining multi-source information, fire hotspots and
past fire records (2003-2016) in China, utilising multiple AI methods. Similarly, LSTM
architectures are integrated with GIS and remote sensing data to enhance predictive
capabilities, and their correlation is assessed using multicollinearity testing to produce
dynamic forest fire risk maps. Additional development was witnessed in work by Lai et al.
[25] and Fitriany et al. [26], who used a sparse autoencoder-based DL with the balancing of
data to predict fire. Based on Portuguese data in the Montesinho Natural Park, the prediction
error decreased by 19.3 points, indicating improved future success in wildfire management.
Meanwhile, in Indonesia, crowdsourced social media information was used in fire detection,
with Twitter messages regarding forest fires in Riau, Sumatra (2014-2019), being processed.
The results suggest that AI models, combined with unusual data sources like social media,
could enhance national fire management systems. Nevertheless, these studies remained
limited to multimodal integration against various forest conditions.
In parallel, researchers have investigated IoT-based sensor networks for the management of
forest fires. A system that identifies wildfire occurrences using distributed IoT sensors was
presented by Alkhatib et al. [27] and Lertsinsrubtavee et al.[28] to monitor the ecological
conditions and detect fire hazards. This system has its ability to predict natural fire outbreaks
hours in advance. It achieved this by continuously monitoring environmental conditions to
enable the quantification of fire risk. They identified three active fire periods and three non-
fire periods by selecting two for model training and the remaining for validation. A J48
decision-tree algorithm was applied and achieved an accuracy of 72% in distinguishing fire
from non-fire events. Also, Mahaveerakannan et al. [29] and Haque et al. [30] integrated IoT
data collection and AI-driven prediction to enhance the functions of early warning. They
Junbo Liu et al.
4
focused on object detection models with Efficient-Det being applied in detecting fire-related
events. The use of a Calculus Optimisation was introduced to optimise the training of DL.
Also, the incorporation of virtual sensors in the wireless sensor networks increased
predictive ability in intricate fire situations caused by thunder or lightning. The effectiveness
of using a hierarchical perceptron network to classify and predict early warning and fire
scenarios using an experimental setting was proven. However, the issues of scalability and
adjustability were not tackled.
Traditional wildfire prediction approaches had largely relied on satellite-based remote
sensing for monitoring fire activity. Huot et al. [31] and Tian et al. [32] built a multivariate
database of historical wildfires in the United States as a cumulation of 10 years of satellite
images. This dataset is a fusion of both fire data and explanatory factors that provide an
input with rich features to the ML models. The neural networks trained on this data
demonstrated significant potential for predicting wildfire propagation. A different study,
created by Avetisyan et al. [33] developed mountain fire in Muli County (March 2020) was
analysed using a combination of meteorological, combustible, terrain, and human activity
variables. The researchers identified the key factors influencing wildfire occurrence and
assessed vegetation cover before and after two major fires in Sichuan Province using a
random forest model. This enabled managers to correlate fire intensity with vegetation
destruction, informing a more effective mitigation strategy. More inputs were recorded by
Asadollah et al. [34] and Wang et al. [35], who utilised NASA satellite data and an ML
classifier to simulate fire behaviour between 2000 and 2021. CMIP6 EC-Earth3-SSP245
climate data were used to develop the future fire projections (2030-2050). In addition, the
classification was performed spatio-temporally, eliminating false hotspots of fire that were
periodic or recurring sources of heat, and thus generating a dataset. This method was used
in Hunan Province, and the mistakes in the detection of hotspots were minimised.
Nevertheless, there were still difficulties when it came to the generalisation of different
ecological areas.
Several studies have examined AI approaches for wildfire identification and forecasting.
However, current models frequently fail to provide useful multi-modal data fusion, are
unscaleable and cannot merge ground-level IoT data with satellite images to create accurate
early warning messages. To overcome these weaknesses, the proposed work forms a hybrid
framework that has been improved to facilitate proper forest fire monitoring and early
warning.
3. Materials and Methodology
The architecture seeks to create a system for wildfire surveillance and proactive alerts by
utilising a hybrid approach that integrates IoT ground sensors with satellite remote sensing
data. Through the use of multi-source data acquisition, the framework further monitors the
environment with the help of IoT sensors, and monitors the macro-level forest and fire
indicators of the forest using high-resolution satellite images. The collected data undergo
intelligent preprocessing to ensure they are cleaned and guaranteed. Thereafter, a multi-
modal feature fusion layer would be used to integrate spatial information of remote sensing
and time trends of IoT sensors using Locality-Sensitive Hashing (LSH) to align the results
of feature integration correctly and Rank Gauss Normalisation to balance features
appropriately.
The combined dataset is subsequently processed using an AI-based predictive model that
employs a CNN with a triplet loss function to extract high-level information from the
Fine-Scale Forest Fire Monitoring and Early Warning via IoT and Remote Sensing
5
satellite images. Meanwhile, IoT trends are incorporated to enhance temporal perception and
ensure robust, representative learning by comparing similar and dissimilar environmental
patterns. While transformer-based and hybrid models are known for their ability to capture
long-range dependencies and temporal dynamics, the proposed framework uses a CNN with
triplet loss due to its computational efficiency, strong spatial feature extraction, and
suitability for fine-scale fusion of heterogeneous data. The temporal patterns from IoT
sensors are encoded as structured temporal grids and embedded alongside spatial features
from remote sensing imagery. This allows the CNN to learn joint spatio-temporal
representations without the overhead of transformer self-attention mechanisms, which are
often data-hungry and less interpretable in low-resource, real-time forest monitoring settings.
Moreover, the triplet loss enhances discriminative learning across fire/no-fire classes,
compensating for temporal nuances by enforcing similarity-preserving embeddings. This
enables advance warnings to protect the forest and prevent disasters. The system’s
operational workflow is visualised in Figure 1.
Figure 1: Proposed IoT-Remote Sensing Framework Architecture of Forest
Fire Monitoring and Early Warning System.
3.1. Multi-Source Data Acquisition
3.1.1 IoT Sensors
Junbo Liu et al.
6
The ground-level data comprises the IoT sensor data collected by the forest fire detection
management system. These sensors are also strategically mounted in forest areas to monitor
important environmental parameters (temperature, relative humidity, smoke concentration
(CO/CO₂) and wind speed). The data include time and geo-tagging, providing temporal and
spatial data on microclimatic changes, enabling the identification of early fire ignition
evidence. This IoT information is developed through the deployment of low-power, wireless
sensor nodes in forest areas that capture and transfer data in real-time, through
communication protocols that tracking of the area, even where connectivity is low.
3.1.2 Remote Sensing Data (Aerial Level)
The satellite imagery is obtained to complement the ground measurements in areas of
wildfires. These datasets will include high-resolution images of Earth that are taken by Earth
observation satellites with multispectral and thermal sensors. The satellite images are
acquired periodically to ensure temporal continuity, but can be used with drone observations
to provide a better spatial resolution. The images show thermal deviations, vegetation
droughtiness and burn markings to provide macro-level data on forest conditions that are
difficult to detect with ground sensors. The data is interpreted to contain spectral indices,
including normalised burn ratio and other vegetation indices, that illustrate potentially fire-
prone regions or regions that are already burnt.
3.2. Intelligent Preprocessing
Raw data received by IoT and remote sensors is typically noisy and inconsistent,
significantly undermining the performance of predictive models. Such problems are typical
in the current methodology, as they are caused by the weaknesses in sensor precision. Thus,
the framework will include a preprocessing step to optimise the quality and usability of the
data to overcome these issues.
3.2.1. IoT Data Cleaning
Noise Reduction
In this context, Kalman filtering is used to minimise noise in measurements from IoT sensors.
The Kalman filter is an optimal recursive estimator that is employed to estimate the current
condition of a system by leveraging past observations with the new information obtained by
the sensor, in effect separating actual environmental changes and random noise. When
continuous streams of IoT data are filtered using this filter, a clean input signal is generated.
Missing Data Handling
In this case, Multiple Imputation by Chained Equations (MICE) will be used to address
missing or incomplete data in the IoT sensor data. MICE operate through the sequential
prediction of each of the missing values by a set of regression models in which a given
variable with missing data is represented as a function of the other observed variables. In
this process, initial approximate guesses of the missing entries are produced and then refined
through repeated chaining steps, resulting in multiple possible imputed datasets. The average
Fine-Scale Forest Fire Monitoring and Early Warning via IoT and Remote Sensing
7
of these different imputations is then computed as the final value of each missing entry,
producing the statistical reliability required and maintaining the underlying data distribution.
3.2.2. Satellite Image Processing
Atmospheric Correction
Under this, Radiative Transfer Models (RTMs) are used to preprocess satellite images,
eliminating cloud cover and even atmospheric distortion. RTMs model radiation propagation
in the Earth's atmosphere and include scattering, absorption and reflection of radiation by
gases, aerosols and clouds. With the modelling of such interactions, the RTM can estimate
and remove the atmospheric contributions of the observed satellite signal and, in effect,
isolate the reflectance by the surface of the forest. When there are cloud-affected pixels, the
model detects anomalies in spectral signatures caused by cloud presence and uses spatial
and temporal interpolation with RTM outputs to correct them, ensuring the reflected surface
properties.
Index Extraction
In this case, the satellite images are processed using the Normalised Burn Ratio (NBR) to
determine the burnt or fire-prone regions in the forest. The NBR is based on the NIR and
SWIR bands of satellite imagery by taking advantage of the fact that healthy vegetation
strongly reflects the NIR band but absorbs the SWIR radiances. On the other hand, burnt
areas are more reflective of SWIR compared to NIR.
The NBR generates a contrast map by computing the ratio (NIR SWIRNIR SWIR)
for each pixel, indicating areas affected by or at risk of fire. In this flow, the index is used to
correct small residual atmospheric and cloud terms by emphasising the spectral differences
in vegetation health, but not atmospheric artefacts. The resultant NBR maps provide both
quantitative indications and precise locations of forest conditions.
3.3. Multi-Modal Feature Fusion
In this context, combining data from multi-source sensors that contain IoT and remote sensor
information represents the environment's imagery, identifies localised environmental
variations, and reveals more general trends that contribute to forest fire risk. Nevertheless,
the problem of heterogeneity and high dimensionality of IoT sensors data produces time-
series vectors of continuous measurements of temperature, humidity, smoke and wind,
whereas satellite images create high-dimensional space vectors (spectral indices and thermal
anomalies). Locality-Sensitive Hashing (LSH) is applied to combine these modalities in a
framework. It is a probabilistic dimensionality reduction algorithm that is used to condense
complex, high-dimensional data into a simpler, low-dimensional representation, retaining
the similarity relationships between them.
Given a set of IoT feature vectors (X) and satellite feature vectors (Y), LSH applies a family
of hash functions (ℋ) to project these vectors into a shared hash space, as expressed in
equations (1-4).
(1)
Junbo Liu et al.
8
(2)
Each hash function hi is defined as:
(3)
(4)
Where ai indicates a random vector ℝd drawn from a p - stable distribution (Gaussian for
Euclidean distance), bi is a random offset, and r is the bucket width controlling the
quantisation resolution. The product aiv captures the projection of the high-dimensional
vector along a random direction, and the floor operation discretises the projection into hash
bins. By applying L such functions in parallel, each feature vector is mapped to a hash code
H defined in equation 5.
(5)
This mapping ensures that vectors that are close in the original space (representing similar
environmental conditions or fire risk patterns) are highly likely to collide in the same hash
bucket, effectively preserving locality. The fused feature set is then constructed by aligning
IoT and satellite vectors that share common hash codes:
fused
bucket (6)
In equation (6), ℱfused represents the final multi-modal feature matrix used for downstream
AI-based prediction. In this formulation, xi and yj are paired if they collide in the same hash
bucket B, allowing the framework to efficiently align heterogeneous data streams without
exhaustive pairwise comparisons, which would be computationally prohibitive. By using
LSH in this manner, the framework achieves similarity-preserving fusion of IoT and remote
sensing features, ensuring that both localised sensor trends and broader spatial patterns are
jointly considered. This results in enhancing the system’s ability to identify high-risk regions
for proactive forest fire warnings.
3.3.1 Uniform Scaling Function
Following the alignment of features, the fused data often contains heterogeneous
distributions due to differences in measurement scales, units, and variability across sensor
types and spectral indices. To ensure that all features contribute equally to the subsequent
AI-based prediction model and to stabilise training convergence, the framework applies
Rank Gauss Normalisation, which transforms arbitrary feature distributions into a standard
Gaussian form. This approach handles skewed and heavy-tailed data in environmental and
remote sensing measurements.
Given a feature vector ℱfused f1f2fi of length i, the Rank Gauss proceeds by first
computing the rank of each element within the vector, as shown in equation (7).
Fine-Scale Forest Fire Monitoring and Early Warning via IoT and Remote Sensing
9
(7)
Where ri represents the sorted position of fi within f, with ri1 for the smallest value and
rin for the largest. The rank is then rescaled into a uniform quantile in the interval 01 :
(8)
In equation (8), ui denotes the empirical cumulative probability associated with fi, and the
subtraction of 0.5 ensures centring within each quantile. Finally, the quantile is mapped onto
a standard Gaussian distribution by applying the inverse of the normal distribution’s
cumulative function, as shown in equation (9).
norm (9)
Where Φ1 is the inverse of the standard normal function, producing a normalised value
fi
norm with zero mean and unit variance, and approximately Gaussian distribution. In this
formulation, fi
norm is now directly comparable across all features, allowing the AI model to
learn patterns without bias from differences in scale or distribution shape.
By applying Rank Gauss Normalisation, the framework achieves uniform scaling across all
sensor and satellite features, thereby mitigating the influence of extreme values, enhancing
numerical stability, and improving the model's robustness and generalisation. This step
involves multi-modal fusion, ensuring that heterogeneous environmental and spectral
information contribute equally to the fire risk assessment.
3.4. Fire Risk Prediction
To enable fine-grained and reliable prediction of wildfire occurrence, a CNN integrated with
a Triplet Loss learning strategy is employed. The CNN is responsible for hierarchical feature
extraction from both remote sensing imagery and IoT sensor trends, whereas the Triplet Loss
ensures that the learned embeddings preserve similarity between related samples (fire–fire,
no fire-no fire) while maximising the separation between dissimilar instances (fire vs. no
fire). This approach guarantees a discriminative representation learning and robust
generalisation of heterogeneous data inputs. Figure 2 indicates the detailed CNN
architecture for forest fire prediction.
To ensure reproducibility and transparency of the CNN–Triplet Loss model training for
wildfire risk prediction, the following hyperparameters were employed: the model was
trained using the Adam optimiser with a learning rate of 0.0003, a batch size of 64, and for
120 epochs with early stopping based on validation loss stagnation over 10 epochs. A triplet
loss margin of 0.4 was used to enforce discriminative separation between fire and no-fire
embeddings, while a dropout rate of 0.3 was applied after the fully connected layer to
mitigate overfitting and enhance generalisation. These parameters were selected via grid
search to balance predictive accuracy across heterogeneous IoT and remote-sensing inputs.
Junbo Liu et al.
1
0
Figure 2: CNN Architecture for IoT and Remote Sensing-Based Wildfire Prediction
The CNN architecture processes an input tensor X ℝHWC, where HW, and C denote
the height, width, and channel depth of the input (multi-spectral satellite imagery combined
with sensor-derived temporal grids). Each convolutional operation can be mathematically
represented as in equation (10).
(10)
Where Fijk is the activation at the spatial location ij in the k-th feature map, Wmnck are
the learnable convolutional filter weights of spatial extent M N bk is the bias, and σ is the
non-linear activation function (ReLU). This operation progressively extracts local spatial
and temporal features relevant to wildfire dynamics.
The resulting feature maps are pooled to ensure translation invariance and dimensionality
reduction. A max pooling operation is expressed as in equation (11).
(11)
Fine-Scale Forest Fire Monitoring and Early Warning via IoT and Remote Sensing
11
Where Pijk is the pooled activation, uv are the feature map, and Ωij denotes the local
neighbourhood window centred at ij . This reduces redundancy while retaining the
strongest activation responses, which are indicative of wildfire patterns.
Once the deep features are extracted, the CNN projects the input sample x into an embedding
space using a fully connected layer, as shown in equation (12).
(12)
Where fθ represents the CNN parameterised by weights θ , and z is the d - dimensional
embedding vector ℝd. This learn an embedding where samples from the same class ("fire")
are close, and those from different classes ("fire" vs. "no fire") are far apart.
To enforce this discriminative property, this work uses the Triplet Loss. For an anchor
sample xa, a positive sample xp (same class), and a negative sample xn (different class), the
embeddings are computed as zafθxa , zpfθxp , and znfθxn . Thus, the Triplet
Loss (ℒtriplet ) is then defined as in equation (13).
triplet
(13)
Where α 0 is the margin parameter. This ensures that the anchor-positive distance is at
least α smaller than the anchor-negative distance, thereby pulling together similar samples
while pushing apart dissimilar ones. In practice, the optimisation process leverages the
gradient of the Triplet Loss with respect to the embeddings, guiding the CNN to refine its
filters and parameters. The gradient with respect to the anchor embedding () is given by
equation (14).
triplet
(14)
This indicates that the network simultaneously reduces the distance to the positive and
increases the distance to the negative, thereby sharpening class boundaries. To further
stabilise training, a SoftMax classification head is added to the embedding space, predicting
the probability of fire ignition. The SoftMax probability (P) for the class c 01 (no fire
vs. fire) is expressed as in equation (15).
(15)
Where wc and bc are the class-specific weights and bias. This transforms the embedding into
probabilistic predictions interpretable as ignition likelihood. The cross-entropy loss ℒCE is
then used to guide the classifier, as shown in equation (16).
(16)
Junbo Liu et al.
1
2
Where yc is the ground-truth one-hot label. Combining this with the Triplet Loss, the final
joint objective function becomes ℒfinal , defined as in equation (17).
final triplet (17)
Here, λ1 and λ2 are weighting coefficients that balance representation learning and
classification accuracy. This ensures discriminative embeddings and reliable probabilistic
classification. This process produces a system capable of predicting the probability of
wildfire ignition within the range of 0 to 1, expressed as a fire/no fire classification.
Algorithm 1: Pseudocode for CNN with Triplet Loss for fire risk prediction
Input: X (satellite images & IoT sensor data)
Output: Classification result P: {High Risk / No Risk}
Begin
For each input sample x in X do
Convolution () with activation
Max Pooling ()
Fully Connected ()
End For
Select triplets ()
Compute embeddings:
Cross-Entropy Loss:
Final Joint Loss:
Update parameters:
For unseen input :
Generate embedding
Compute fire ignition probability
If threshold Predict "High Risk (Wildfire)"
Else → Predict "No Risk"
End If
Fine-Scale Forest Fire Monitoring and Early Warning via IoT and Remote Sensing
13
End For
End
4. Validation and Results
To confirm the suggested model, tests were conducted using IoT sensor data and satellite
imagery. The entire implementation was done in Python on a 12th Gen Intel(R) Core (TM)
i5-12400 (2.50 GHz) with 8 GB RAM (7.75 GB usable). The AI model was trained on a
fused dataset comprising both IoT sensor streams and satellite images from forest areas with
historical fire incidents. A total of 18,240 samples were obtained across different ecological
zones to ensure generalisability. The dataset was first divided into three parts using stratified
sampling: 70% for training, 15% for validation, and 15% for testing, to maintain class
balance. After that, the model was evaluated using a 5-fold cross-validation protocol with a
fixed random seed (42) to ensure reproducibility in the final reporting of accuracy, precision,
recall and F1-score. This confirms the system's resilience and its fire-early warning
capability in real-world scenarios.
4.1 Performance Analysis
This part presents evidence on the efficiency of the AI-based hybrid forest fire monitoring
model through a wide range of experiments and comparative performance analysis. The
performance of the AI architecture, including the integration of a CNN with triplet-loss
optimisation, showed high learning efficiency, as it extracted small-scale spatial features
from the integrated data. The findings support the validity and effectiveness of the suggested
AI model as a way to issue early, fine-scale, and accurate warnings about forest fires and,
therefore, contribute to swift response and efficient management of disasters.
To evaluate the predictive performance of the CNN-Triplet Loss model, a inclusive set of
standard classification metrics was employed. These included (equations 18-23):
(18)
It measures the overall proportion of correctly classified instances.
(19)
It quantifies the proportion of predicted fire events that were actual fires.
(20)
It measures the model's ability to correctly identify actual fire events.
Precision Recall
Precision Recall (21)
It represents the harmonic mean of precision and recall to provide a balanced assessment.
Junbo Liu et al.
1
4
(22)
It indicates the proportion of non-fire events incorrectly classified as fire.
(23)
It represents the proportion of actual fire events that the model misses.
In this context, TPTNFP, and FN represent the numbers of true positives, true negatives,
false positives, and false negatives, respectively. For every fold in the k-fold cross-validation,
all metrics were computed and their mean values were reported to provide a statistically
robust and unbiased evaluation of the model's performance.
Figure 3: Probabilistic Satellite Contrast Map for Wildfire Classification
Figure 3 presents a comparative visualisation of satellite image samples. It portrays the no
wildfire samples in urban and suburban areas with low ignition probabilities (0.01-0.05),
and highlights wildfire samples with dense vegetation cover and thermal anomalies that have
high ignition probabilities (0.95-0.98). This serves as a qualitative measure to demonstrate
the model's discriminatory ability in distinctive fire-prone and safe areas. It validates the
model's capacity to learn meaningful representations from heterogeneous data and classify
fire risk by contrasting spatial features with corresponding probability scores. Thus, aiding
early warning and disaster mitigation.
Fine-Scale Forest Fire Monitoring and Early Warning via IoT and Remote Sensing
15
Figure 4: Performance Curve of CNN–Triplet Loss Model
The dynamics of training and validation of the proposed AI model are shown in Figure 4
during the 60 epochs. This shows an accuracy progression in which training accuracy
increases gradually and reaches a plateau at 0.98. Whereas, validation accuracy varies at
97.20, indicating high generalisation. Also, the loss trajectory with training loss smoothly
declining to 0.032 and validation loss stabilising around 0.054, confirms effective
convergence and minimal overfitting. This validates the model’s learning stability and
robustness in predicting wildfire risk from fused IoT and satellite data for real-time early
warning systems.
Figure 5: Classification Performance for Wildfire Risk Prediction
Figure 5 provides the classification metric, which summarises the performance of the
proposed CNN-Triplet Loss model in terms of several metrics. The metrics evaluate the
model's capability to detect high-risk wildfire areas compared to safe areas, utilising fused
IoT and satellite data. The model had an Accuracy of 0.9720, a Precision of 0.9698, a Recall
Junbo Liu et al.
1
6
of 0.9771 and an F1-score of 0.9734, which is reliable in both detection and classification.
Therefore, this model justifies its balanced performance by proving that it effectively
minimises false positives and negatives, enabling forest fire management systems to issue
timely and accurate early warnings.
Figure 6: Error Distribution for Wildfire Classification
Figure 6 depicts the distribution of the errors, and the figure illustrates the mistakes made
by the DL model in terms of false positive rate against false negative rate (FPR-FNR). These
factors determine the consistency of early warning systems, where a false positive can result
in unwarranted alerts, and a false negative can lead to undetected fire outbreaks. The model
had an FPR of 0.0337 and an FNR of 0.0229, indicating a high bias towards true fire events,
but with a low false alarm rate. In this way, it proves the accuracy of the model in
discriminating the situation of wildfire to support its applicability in proactive forest fire
monitoring and risk reduction.
Fine-Scale Forest Fire Monitoring and Early Warning via IoT and Remote Sensing
17
Figure 7: Precision–Recall Curve for Wildfire Classification Robustness
Figure 7 shows the precision-recall (PR) curve of wildfire classification, which determines
the discriminatory performance of the AI model during imbalanced data conditions. The
curve shows the accuracy against the recall at different classification thresholds, which
reveals a trade-off between the ability to detect an occurrence of wildfires correctly and the
reduction of the false alarms. The model achieves a high average precision score of 99.4,
indicating perfect departure between fire and no-fire classes. This confirms the CNN
model’s reliability in prioritising true wildfire detections while maintaining minimal false
positives for disaster mitigation.
Figure 8: ROC Curve for Wildfire Detection Sensitivity
Junbo Liu et al.
1
8
Figure 8 presents the ROC curve for wildfire detection, which evaluates the classification
performance of the CNN–Triplet Loss model by plotting the TPR against the FPR across
varying decision thresholds. The curve indicates strong discriminative capability, with the
model achieving an Area Under the Curve (AUC) of 99.4, signifying near-perfect
classification. This assesses the model’s sensitivity and specificity, confirming its ability to
accurately detect wildfire events while minimising false alarms, an indispensable trait for
reliable early warning systems in forest fire management.
Figure 9: Confusion Matrix for Wildfire Classification Precision
Figure 9 shows the heatmap of the confusion matrix of wildfire classification, which
measures the percentage of the prediction of the wildfire classes (actual and predicted) by
the DL model. This is shown in the matrix of 513 True Positives (wildfire correctly
predicted), 459 True Negatives (no wildfire correctly predicted), 16 False Positives (no
wildfire misclassified as wildfire) and 12 False Negatives (wildfire missed). This will
measure the stability of the model in practice by highlighting detection success and error
margins. The low false rates and high true classification numbers verify the quality and
accuracy of the model in differentiating the wildfire events and hence can be used in
proactive forest fire early warning systems.
4.2. Comparative Analysis
The section will compare the functionality of the DL Model with the existing fire detection
model to demonstrate that it is superior. It also showcases the use of multi-modal data fusion
and Triplet Loss optimisation as superior relative to conventional single-source and image-
only methods. In this comparison, the work demonstrates the efficiency and early-warning
capacity of the proposed system in real-time forest fire monitoring.
Table 1: Comparative Performance Analysis of Forest Fire Classification Models
Fine-Scale Forest Fire Monitoring and Early Warning via IoT and Remote Sensing
19
Model
Accuracy
(%)
Precision
(%)
Recall (%)
F1 Score
(%)
VGG 19 [36]
95
95.7
94.2
94.96
InceptionV3 [37]
92.25
84.5
84.5
84.5
YOLO [38]
-
93.91
88.87
88.35
Proposed CNN–
Triplet Loss
97.20
96.98
97.71
97.34
The results of the proposed CNN-Triplet Loss hybrid model are compared with the state-of-
the-art models, including VGG19, InceptionV3 and YOLO, on the forest fire classification
task in Table 1. The proposed model had the highest accuracy, with a total of 97.20, which
is higher in all the main indicators of precision, recall, and F1-score. Such an increase in
performance is primarily associated with the incorporation of multi-modal IoT and remote
sensing data, as well as the application of triplet loss optimisation, which has contributed to
improved feature discrimination and lower false detection. The model, therefore,
demonstrates greater strength and early warning capabilities in real-time monitoring of
forest fires and threat prediction.
4.3. Discussion
As the experimental outcomes of the suggested hybrid forest fire monitoring and early
warning system indicate, combining data from IoT-based ground sensors with satellite-based
remote sensing imagery significantly increases the accuracy and timeliness of fire hazard
predictions. The accuracy (97.20%), precision (96.98%), recall (97.71%) and F1-score
(97.34%) of the DL model allow inferring that the system is capable of learning spatial as
well as temporal relationships between forest fire ignition and particular characteristics.
Contextually, these findings indicate that ground-level IoT sensors provide fine-scale
microclimatic data, whereas remote sensing imagery is an indicator of macro-scale spatial
patterns. This combination enabled the CNN architecture to identify the complicated
interdependencies. In comparison to the works, the current article presents a DL-based
fusion paradigm and data-driven early warning models that can provide real-time predictions
of risks.
These discoveries, in a more generalised manner, add to the new field of AI-driven
environmental intelligence. This work, through sensor networks combined with remote
sensing analytics, illustrates how interdisciplinary integration can defeat the conventional
constraints of temporal resolution and spatial coverage of fire detection. The framework not
only enhances the technical edge of forest fire prediction but also provides a model that can
be reproduced in multi-modal early warning systems for other environmental areas, e.g.,
flood prediction or air quality monitoring. All these innovations have created a new standard
in forest fire risk analytics, proving that smart integration of heterogeneous sources of data
Junbo Liu et al.
2
0
can lead to significant improvements in the accuracy of predictions, false alarms, and timely
and active mitigation of disasters.
5. Conclusion
In this study, a hybrid fine-scale forest fire monitoring and early warning framework was
successfully developed by integrating IoT-based ground sensor data with satellite-derived
remote sensing imagery for accurate and timely forest fire risk prediction. An end-to-end
pipeline was used that featured a multi-modal feature fusion layer to combine heterogeneous
data into one analytical form. The integrated data was processed with a CNN trained with
triplet loss, which enables the creation of strong distinctions between areas at risk of
wildfires and those not at risk. Experimental testing showed the best performance with an
accuracy of 97.20%, a precision of 96.98%, a recall of 97.71%, and an F1-score of 97.34%,
which is much higher than benchmark models. These findings confirm the strength of the
model, its convergence stability and its ability to generalise its results to a wide range of
forests. Fine-grained and real-time fire prediction with a low false alarm rate (FPR: 0.0337,
FNR: 0.0229) can be achieved by integrating IoT and remote sensing data, providing a
consistent early-warning tool for active control of forest works and disaster prevention.
Overall, this work presents a brilliant solution that can turn the traditional fire monitoring
frameworks into dynamic early warning systems that would make forest ecosystems more
resilient to the risks of wildfires caused by climate change. Nevertheless, the use of IoT and
satellite data fusion is mainly used in this work, which can be a limitation in situations when
the data acquisition is compromised by sensor malfunctioning, cloud cover, or signal
interference, which breaks the continuity of real-time monitoring.
5.1. Future work
This framework should be extended in the future to incorporate Unmanned Aerial Vehicle
(UAV) data, meteorological variables, and social sensing inputs to make it more robust and
situation-aware. The addition of edge intelligence and the federated learning paradigm could
also enable on-device processing and collaboration with distributed forest stations in a
privacy-preserving manner. This creates a multi-modal environmental watch, which plays
an important part in the pre-emptive disaster prevention, ecological resilience, and
advancing sustainable forestry practices to adapt to worsening climate conditions
Data Availability
The datasets were obtained publicly in this study. The data on IoT were collected from the
Forest Fire Detection Management System published on GitHub, and the data on remote
sensing were taken from the Wildfire Prediction Dataset published on Kaggle.
IoT data: Forest Fire Detection Management System – GitHub
Remote sensing data: https://www.kaggle.com/datasets/abdelghaniaaba/wildfire-
prediction-dataset
Fine-Scale Forest Fire Monitoring and Early Warning via IoT and Remote Sensing
21
Acknowledgment
This paper is supported by
Jilin Electric Power Research Institute Co., Ltd. Science and Technology Project，“
Research and Application of Refined Monitoring and Early Warning Technology for Wildf
ires in Jilin Province”（Number: KY-GS-25-01-04）.
References
[1] P. Barmpoutis, P. Papaioannou, K. Dimitropoulos, and N. Grammalidis, “A Review on
Early Forest Fire Detection Systems Using Optical Remote Sensing,” Sensors, vol. 20,
no. 22, p. 6442, Jan. 2020, doi: 10.3390/s20226442.
[2] F. Carta, C. Zidda, M. Putzu, D. Loru, M. Anedda, and D. Giusto, “Advancements in
Forest Fire Prevention: A Comprehensive Survey,” Sensors, vol. 23, no. 14, p. 6635, Jan.
2023, doi: 10.3390/s23146635.
[3] J. S. Littell, D. L. Peterson, K. L. Riley, Y. Liu, and C. H. Luce, “A review of the
relationships between drought and forest fire in the United States,” Glob. Change Biol.,
vol. 22, no. 7, pp. 2353–2369, 2016, doi: 10.1111/gcb.13275.
[4] M. W. Jones et al., “Global and Regional Trends and Drivers of Fire Under Climate
Change,” Rev. Geophys., vol. 60, no. 3, p. e2020RG000726, 2022, doi:
10.1029/2020RG000726.
[5] E. Chuvieco, I. Aguado, J. Salas, M. García, M. Yebra, and P. Oliva, “Satellite Remote
Sensing Contributions to Wildland Fire Science and Management,” Curr. For. Rep., vol.
6, no. 2, pp. 81–96, June 2020, doi: 10.1007/s40725-020-00116-5.
[6] L. Hua and G. Shao, “The progress of operational forest fire monitoring with infrared
remote sensing,” J. For. Res., vol. 28, no. 2, pp. 215–229, Mar. 2017, doi:
10.1007/s11676-016-0361-8.
[7] S. Ivanova, A. Prosekov, and A. Kaledin, “A Survey on Monitoring of Wild Animals
during Fires Using Drones,” Fire, vol. 5, no. 3, p. 60, June 2022, doi:
10.3390/fire5030060.
[8] S. Sudhakar, V. Vijayakumar, C. Sathiya Kumar, V. Priya, L. Ravi, and V.
Subramaniyaswamy, “Unmanned Aerial Vehicle (UAV) based Forest Fire Detection and
monitoring for reducing false alarms in forest-fires,” Comput. Commun., vol. 149, pp.
1–16, Jan. 2020, doi: 10.1016/j.comcom.2019.10.007.
[9] J. J. Roldán-Gómez, E. González-Gironda, and A. Barrientos, “A Survey on Robotic
Technologies for Forest Firefighting: Applying Drone Swarms to Improve Firefighters’
Efficiency and Safety,” Appl. Sci., vol. 11, no. 1, p. 363, Jan. 2021, doi:
10.3390/app11010363.
[10] F. Abid, “A Survey of Machine Learning Algorithms Based Forest Fires Prediction
and Detection Systems,” Fire Technol., vol. 57, no. 2, pp. 559–590, Mar. 2021, doi:
10.1007/s10694-020-01056-z.
[11] C. Mambile, S. Kaijage, and J. Leo, “Application of Deep Learning in Forest Fire
Prediction: A Systematic Review,” IEEE Access, vol. 12, pp. 190554–190581, 2024, doi:
10.1109/ACCESS.2024.3515215.
[12] R. Alkhatib, W. Sahwan, A. Alkhatieb, and B. Schütt, “A Brief Review of Machine
Learning Algorithms in Forest Fires Science,” Appl. Sci., vol. 13, no. 14, p. 8275, Jan.
2023, doi: 10.3390/app13148275.
Junbo Liu et al.
2
2
[13] F. Cui, “Deployment and integration of smart sensors with IoT devices detecting fire
disasters in huge forest environment,” Comput. Commun., vol. 150, pp. 818–827, Jan.
2020, doi: 10.1016/j.comcom.2019.11.051.
[14] R. Sahal, S. H. Alsamhi, J. G. Breslin, and M. I. Ali, “Industry 4.0 towards Forestry
4.0: Fire Detection Use Case,” Sensors, vol. 21, no. 3, p. 694, Jan. 2021, doi:
10.3390/s21030694.
[15] N. Guimarães, L. Pádua, P. Marques, N. Silva, E. Peres, and J. J. Sousa, “Forestry
Remote Sensing from Unmanned Aerial Vehicles: A Review Focusing on the Data,
Processing and Potentialities,” Remote Sens., vol. 12, no. 6, p. 1046, Jan. 2020, doi:
10.3390/rs12061046.
[16] C. C. Chan, S. A. Alvi, X. Zhou, S. Durrani, N. Wilson, and M. Yebra, “A Survey on
IoT Ground Sensing Systems for Early Wildfire Detection: Technologies, Challenges,
and Opportunities,” IEEE Access, vol. 12, pp. 172785–172819, 2024, doi:
10.1109/ACCESS.2024.3501336.
[17] S. Giannakidou et al., “Leveraging the power of internet of things and artificial
intelligence in forest fire prevention, detection, and restoration: A comprehensive survey,”
Internet Things, vol. 26, p. 101171, July 2024, doi: 10.1016/j.iot.2024.101171.
[18] M. Naderpour, H. M. Rizeei, and F. Ramezani, “Forest Fire Risk Prediction: A Spatial
Deep Neural Network-Based Framework,” Remote Sens., vol. 13, no. 13, p. 2513, Jan.
2021, doi: 10.3390/rs13132513.
[19] I. Novkovic et al., “GIS-Based Forest Fire Susceptibility Zonation with IoT Sensor
Network Support, Case Study—Nature Park Golija, Serbia,” Sensors, vol. 21, no. 19, p.
6520, Jan. 2021, doi: 10.3390/s21196520.
[20] Y. J. Chew, S. Y. Ooi, Y. H. Pang, and K.-S. Wong, “A Review of Forest Fire
Combating Efforts, Challenges and Future Directions in Peninsular Malaysia, Sabah,
and Sarawak,” Forests, vol. 13, no. 9, p. 1405, Sept. 2022, doi: 10.3390/f13091405.
[21] S. Kondylatos et al., “Wildfire Danger Prediction and Understanding With Deep
Learning,” Geophys. Res. Lett., vol. 49, no. 17, p. e2022GL099368, 2022, doi:
10.1029/2022GL099368.
[22] S. Zheng, P. Gao, W. Wang, and X. Zou, “A Highly Accurate Forest Fire Prediction
Model Based on an Improved Dynamic Convolutional Neural Network,” Appl. Sci., vol.
12, no. 13, p. 6721, Jan. 2022, doi: 10.3390/app12136721.
[23] Y. Pang et al., “Forest Fire Occurrence Prediction in China Based on Machine
Learning Methods,” Remote Sens., vol. 14, no. 21, p. 5546, Jan. 2022, doi:
10.3390/rs14215546.
[24] X. Lin, Z. Li, W. Chen, X. Sun, and D. Gao, “Forest Fire Prediction Based on Long-
and Short-Term Time-Series Network,” Forests, vol. 14, no. 4, p. 778, Apr. 2023, doi:
10.3390/f14040778.
[25] C. Lai, S. Zeng, W. Guo, X. Liu, Y. Li, and B. Liao, “Forest Fire Prediction with
Imbalanced Data Using a Deep Neural Network Method,” Forests, vol. 13, no. 7, p. 1129,
July 2022, doi: 10.3390/f13071129.
[26] A. A. Fitriany, P. J. Flatau, K. Khoirunurrofik, and N. F. Riama, “Assessment on the
Use of Meteorological and Social Media Information for Forest Fire Detection and
Prediction in Riau, Indonesia,” Sustainability, vol. 13, no. 20, p. 11188, Jan. 2021, doi:
10.3390/su132011188.
[27] A. A. A. Alkhatib and K. M. Jaber, “FDPA internet of things system for forest fire
detection, prediction and behaviour analysis,” IET Wirel. Sens. Syst., vol. 14, no. 3, pp.
56–71, 2024, doi: 10.1049/wss2.12076.
Fine-Scale Forest Fire Monitoring and Early Warning via IoT and Remote Sensing
23
[28] A. Lertsinsrubtavee, T. Kanabkaew, and S. Raksakietisak, “Detection of forest fires
and pollutant plume dispersion using IoT air quality sensors,” Environ. Pollut., vol. 338,
p. 122701, Dec. 2023, doi: 10.1016/j.envpol.2023.122701.
[29] Mahaveerakannan R, C. Anitha, Aby K Thomas, S. Rajan, T. Muthukumar, and G.
Govinda Rajulu, “An IoT based forest fire detection system using integration of cat
swarm with LSTM model,” Comput. Commun., vol. 211, pp. 37–45, Nov. 2023, doi:
10.1016/j.comcom.2023.08.020.
[30] A. Haque and H. Soliman, “Smart Wireless Sensor Networks with Virtual Sensors for
Forest Fire Evolution Prediction Using Machine Learning,” Electronics, vol. 14, no. 2,
p. 223, Jan. 2025, doi: 10.3390/electronics14020223.
[31] F. Huot, R. L. Hu, N. Goyal, T. Sankar, M. Ihme, and Y.-F. Chen, “Next Day Wildfire
Spread: A Machine Learning Dataset to Predict Wildfire Spreading From Remote-
Sensing Data,” IEEE Trans. Geosci. Remote Sens., vol. 60, pp. 1–13, 2022, doi:
10.1109/TGRS.2022.3192974.
[32] Y. Tian, Z. Wu, M. Li, B. Wang, and X. Zhang, “Forest Fire Spread Monitoring and
Vegetation Dynamics Detection Based on Multi-Source Remote Sensing Images,”
Remote Sens., vol. 14, no. 18, p. 4431, Jan. 2022, doi: 10.3390/rs14184431.
[33] D. Avetisyan, E. Velizarova, and L. Filchev, “Post-Fire Forest Vegetation State
Monitoring through Satellite Remote Sensing and In Situ Data,” Remote Sens., vol. 14,
no. 24, p. 6266, Jan. 2022, doi: 10.3390/rs14246266.
[34] S. B. H. S. Asadollah, A. Sharafati, and D. Motta, “Satellite-based ensemble
intelligent approach for predicting forest fire: a case of the Hyrcanian forest in Iran,”
Environ. Sci. Pollut. Res., vol. 31, no. 15, pp. 22830–22846, Mar. 2024, doi:
10.1007/s11356-024-32615-4.
[35] H. Wang, G. Zhang, Z. Yang, H. Xu, F. Liu, and S. Xie, “Satellite Remote Sensing
False Forest Fire Hotspot Excavating Based on Time-Series Features,” Remote Sens.,
vol. 16, no. 13, p. 2488, Jan. 2024, doi: 10.3390/rs16132488.
[36] A. Khan, B. Hassan, S. Khan, R. Ahmed, and A. Abuassba, “DeepFire: A Novel
Dataset and Deep Transfer Learning Benchmark for Forest Fire Detection,” Mob. Inf.
Syst., vol. 2022, no. 1, p. 5358359, 2022, doi: 10.1155/2022/5358359.
[37] V. E. Sathishkumar, J. Cho, M. Subramanian, and O. S. Naren, “Forest fire and smoke
detection using deep learning-based learning without forgetting,” Fire Ecol., vol. 19, no.
1, p. 9, Feb. 2023, doi: 10.1186/s42408-022-00165-0.
[38] T. Luan, S. Zhou, L. Liu, and W. Pan, “Tiny-Object Detection Based on Optimized
YOLO-CSQ for Accurate Drone Detection in Wildfire Scenarios,” Drones, vol. 8, no. 9,
p. 454, Sept. 2024, doi: 10.3390/drones8090454.
