Edge AI System for Real-Time and Explainable Forest Fire Detection
Using Compressed Deep Learning Models
Sidi Ahmed Mahmoudi
a
, Maxime Gloesener, Mohamed Benkedadra
b
and Jean-S
´
ebastien Lerat
ILIA Lab, Faculty of Engineering, University of Mons, Mons, Belgium
{sidi.mahmoudi, maxime.gloesener, mohamed.benkedadra, jean-s
´
Keywords:
Forest Fire Detection, Deep Learning, CNN Networks, Vision Transformers, Edge AI, XAI.
Abstract:
Forests are vital natural resources but are highly vulnerable to disasters, both natural (e.g., lightining strikes)
and human induced. Early and automated detection of forest fire and smoke is critical for mitigating damages.
The main challenge of this kind of application is to provide accurate, explainable, real-time and lightweight
solutions that can be easily deployable by and for users like firefighters. This paper presents an embedded
and explainable artificial intelligence “Edge AI” system, for real-time forest fire, and smoke detection, using
compressed Deep Learning (DL) models. Our model compression approach allowed to provide lightweight
models for Edge AI deployment. Experimental evaluation on a preprocessed dataset composed of 1500 images
demonstrated a test accuracy of 98% with a lightweight model running in real-time on a Jetson Xavier Edge
AI resource. The compression methods preserved the same accuracy, while accelerating computation (3× to
18× speedup), reducing memory consumption ( 3.8× to 10.6×), and reducing energy consumption (3.5× to
6.3×).
1 INTRODUCTION
Forests are crucial in combating climate change, ab-
sorbing 2.6 billion tons of carbon annually. (Madeira,
2023). However, in recent years, the increasing fre-
quency and intensity of forest fires have posed a sig-
nificant threat to these vital ecosystems, as well as to
human lives and property. This underscores the ur-
gent need for advanced technologies capable of de-
tecting and mitigating these disasters in real-time. To
address this critical challenge, this paper introduces a
novel Embedded Edge AI system for real-time for-
est fire and smoke detection, enabling continuous
monitoring and early warning. Leveraging cutting-
edge techniques from the field of deep learning, we
combine the strengths of Convolutional Neural Net-
works (CNNs) and Vision Transformers to create a
highly accurate, efficient and explainable model. De-
spite its powerful capabilities, the model is optimized
and compressed to operate within the resource con-
straints of edge devices such as the Jetson Nano, Jet-
son Xavier, Jetson Orin, and Jetson Orin Nano
1
en-
suring its applicability in remote and challenging en-
a
https://orcid.org/0000-0002-1530-9524
b
https://orcid.org/0000-0002-2590-8344
1
Jetson Systems: https://www.nvidia.com/fr-
fr/autonomous-machines/embedded-systems/
vironments. The integration of artificial intelligence
(AI) with edge computing has unlocked new possibil-
ities for environmental monitoring and disaster pre-
vention. Our system allows for the detection, lo-
calization and identification of forest fires and forest
smoke. As a consequence, it can trigger accurate and
timely alerts, enabling rapid response efforts that can
save lives and reduce the ecological and economic im-
pact of these disasters. By harnessing the power of
Embedded AI at the edge, we present a scalable, cost-
effective, and robust solution for real-time forest fire
detection.
Despite the promising results of Deep Neural Net-
works (DNNs) for fire detection in the literature, no
existing solution achieves an optimal balance between
accuracy, explainability, computation time, and mem-
ory/energy efficiency. Our contribution addresses this
gap through three innovations :
• Innovative development approach: best DNN
models in terms of accuracy and explainability ;
• Innovative compression approach: compress
DNN models to significantly accelerate inference
while reducing memory and energy consumption.
Our innovation lies in delivering the best trade-off
across all these metrics within a single model.
• Innovative library “pytorch bench”: comprehen-
Mahmoudi, S. A., Gloesener, M., Benkedadra, M. and Lerat, J.-S.
Edge AI System for Real-Time and Explainable Forest Fire Detection Using Compressed Deep Learning Models.
DOI: 10.5220/0013382500003912
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 20th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2025) - Volume 3: VISAPP, pages
847-854
ISBN: 978-989-758-728-3; ISSN: 2184-4321
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.
847

sive benchmarking library for deep neural net-
works, enabling precise evaluation across multi-
ple metrics, including accuracy, loss, computation
time, memory usage, and energy consumption.
2 RELATED WORK
2.1 CV and AI for Fire Detection
Traditional fire and smoke detection relies on low-
cost sensors, such as smoke, fire, and heat detec-
tors. While effective at identifying the presence of
fire, these sensors often lack capabilities for localiza-
tion, incident analysis, and reliable data on the fire’s
origin and progression (Hamzah et al., 2022). Vision-
based systems are proposed as a superior alternative,
offering more precise and detailed information about
the type, source, and severity of fire and smoke. This
enhanced data supports improved resource allocation
and more effective incident response.
Traditional fire detection and classification meth-
ods heavily relied on handcrafted features such as
color, texture, and shape. These approaches required
significant domain knowledge and often struggled to
generalize across different fire scenarios. Recently,
Deep Learning allowed for the automatic extraction of
high-level features, which are difficult to obtain with
traditional techniques.
Convolutional Neural Networks (CNNs) have
demonstrated remarkable performance in fire classi-
fication tasks. (Zhao et al., 2018) proposed a 15-
layer CNN architecture named ’Fire Net,’ inspired by
AlexNet, which works as both a feature extractor and
a classifier. Transfer learning has also emerged as
a valuable technique in fire detection and classifica-
tion, with many methods leveraging this approach.
(Sousa et al., 2020) proposed a method using transfer
learning by fine-tuning the weights of the Inceptionv3
model for fire classification tasks. Similarly, (Govil
et al., 2020) proposed the ’ForestResNet’ method,
which is based on the ResNet50 model. (Park et al.,
2021) explored various pre-trained models, including
VGG16, ResNet50, and DenseNet121, for fire detec-
tion purposes. (Shamsoshoara et al., 2021) employed
the Xception model for fire detection tasks.
Recent advancements have focused on model ef-
ficiency, leading to the development of architectures
tailored for resource-constrained environments. For
example, (Khan and Khan, 2022) and (Wu et al.,
2020) utilized MobileNetV2 for real-time fire detec-
tion, achieving a balance between accuracy and com-
putational efficiency.
2.2 XAI for Machine & Deep Learning
Deep learning models are seen as black boxes where
the provided results cannot produce partial or com-
plete explanations (Stassin et al., 2024) (Mahmoudi
et al., 2022) (Englebert et al., 2023) . In this context,
post-hoc explainability methods can be applied to in-
put images in order to describe the model reasoning.
The post-hoc methods are represented by three cate-
gories: 1) Gradient-based methods (Simonyan et al.,
2014) that track the neuron values propagation from
the network output to the input, allowing to identify
the importance of each input feature; 2) Perturbation-
based methods (Zeiler and Fergus, 2014) (Petsiuk
et al., 2018) that disturb the inputs and analyze the
performance drop of evaluation metrics. A high drop
of accuracy metrics means that the disturbed inputs
are important and vice versa ; 3) CAM methods that
are mainly used to explain CNN networks by the use
of convolutional layers activations in order to produce
saliency maps showing the importance of each pixel
in the model decision (Wang et al., 2020). Otherwise,
although their high accuracies, vision transformers ar-
chitectures (Dosovitskiy et al., 2020) are also hin-
dered by their lack of explainability. A basic approach
for explaining ViTs consists of the usage of attention
weights as explanation. However, this approach re-
mains inadequate for explaining transformer results
since it considers the query and key elements of the
self-attention, but not the value (Pruthi et al., 2020).
2.3 Optimization and Compression
Recently, a significant improvement of embedded AI
Hardware has been noted where GPUs can be inte-
grated in small Edge AI devices. This makes from
these devices a promising solution for the deploy-
ment of machine and deep learning models. Although
the provided power, several limitations (computation
power, RAM memory, storage size, etc.) may be en-
countered when deploying complex neural architec-
tures for processing massive data and more partic-
ularly HD images. Thus, we can propose restruc-
turing and compression of deep neural architectures
since the latter are generally over-parametrized where
several weight and bias values can be retired without
negative influence on models accuracy. We can cite
3 compression techniques: pruning (Blalock et al.,
2020), quantization (Nagel et al., 2021) and knowl-
edge distillation (Gou et al., 2021).
1. Pruning: Identify and remove unnecessary neu-
rons and connections, retaining only the most rel-
evant components within a deep neural network.
While pruning reduces model size, to produce less
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
848

parametrized (sparse) architectures, the resulting
sparse architecture often remain computationally
inefficient as a consequence of the unchanged op-
eration counts. To address this, the block prun-
ing method (Lagunas et al., 2021) produces dense,
pruned architectures that achieve reduced mem-
ory consumption and computation time.
2. Quantization: An approximation process is ap-
plied to neural networks to represent weights
and/or activations with lower precision values.
Hence, minimizing memory usage and computa-
tion time without compromising on model accu-
racy (Nagel et al., 2021).
3. Knowledge Distillation: Consists of training a
small “Student Network” that can learn during the
training from a large “Teacher Network”. The ob-
jective is to have a small, simplified neural net-
work that can make decisions similar to those
of the larger and more complex one (Gou et al.,
2021).
3 MODEL DEVELOPMENT
APPROACH
3.1 DL Models Selection
In this step, transfer learning is applied to pre-
trained classification models initialized with Ima-
geNet weights, to benefit from previously acquired
learning weights for solving our classification prob-
lem, consisting of fire detection using three classes
(fire, smoke, no fire). We utilized deep neural archi-
tectures suited for image classification like convolu-
tional Neural Networks (CNN), Vision Transformers
(Vit) and even models that combine the usage of con-
volutional layers and attention mechanism like Con-
vNeXt (Liu et al., 2022). Vision transformers belong
to the family of transformers that are represented by
encoder–decoder architectures using multihead self-
attention. The Vision transformer (ViT) architectures
are mainly proposed for images classification tasks
(recently for object detection, and image segmenta-
tion as well), where the embeddings vector is di-
rectly used for classification. With the encoder part,
the ViT represents images as sequences and classes
where each input image is divided to a sequence of
patches. Every patch is flattened to a vector contain-
ing all pixel channels. Then, a linear projection is
applied to provide the desired input dimension. The
decoder part exploits the outputs of the encoder and
uses a Multilayer Perceptron architecture for image
classification (Dosovitskiy et al., 2020). This study
selects the most accurate and generalizable architec-
tures based on their performance with ImageNet pre-
trained weights
2
.
• CNN Architectures: VGG family (VGG11,
VGG13, VGG16), ResNet family (ResNet18,
ResNet34, Resnet50, ResNet101, ResNet152)
EfficientNet family(EfficientNet V2 S, Efficient-
Net V2 M, EfficientNet V2 L)
• Vision Transformers: Vit B 16, Vit B 32
• Combined Architectures: ConvNeXt Small,
ConvNeXt Base, ConvNeXt Large.
3.2 Data Preprocessing
To improve the quality and efficacy of our dataset, we
implemented a preprocessing pipeline using the fol-
lowing steps:
1. Data Collection: we have collected several pub-
lic videos that represent different situations re-
lated to fire or/and smoke in forest. From each
video, we have extracted 3 frames per second in
order to generate a large variety of images rep-
resenting normal situations, fire and smoke situa-
tions. At the end of this step, we got 4000 images
(fire: 1500 images, smoke: 1500 images, no fire:
1000 images)
2. Removal of Duplicated Images: this step con-
sists on removing similar images after calculating
the difference between images. Once the similar
images eliminated, we applied data augmentation
in order to increase the size of our dataset from
the selected unique images.
3. Extreme Luminance Filtering: in this phase, we
removed images with extreme luminance values,
as they provide limited useful information for our
specific use case.
4. Blur Detection and Removal: we used a blur de-
tection algorithm to identify and remove severely
blurred images. The top 3% images with the high-
est blur metrics were eliminated from the dataset.
5. Label Verification: an inspection process was
conducted to identify and remove images with in-
correct labels, ensuring the integrity of our ground
truth data. If an image is presenting both fire and
smoke, we propose ”fire” label since the priority
is prevent fire.
6. Near-Duplicate Detection: to address the issue
of near-duplicate images, particularly those origi-
nating from video frame extraction, we employed
2
Models and pre-trained weights.
https://pytorch.org/vision/main/models.html
Edge AI System for Real-Time and Explainable Forest Fire Detection Using Compressed Deep Learning Models
849

Figure 1: Model Development Pipeline from Data Collection and Model Selection to Explainability.
a similarity threshold approach. Images with a
Structural Similarity Index (SSIM) score above
92% were grouped together and from each group
of near-duplicates, only one representative image
was retained, while the others were discarded.
Fig. 2 illustrates some image examples that have
been removed due to their excessive darkness, bright-
ness or blurring. This preprocessing pipeline sig-
Figure 2: A. dark images removal ; b. bright images re-
moval ; c. high blurred images removal.
nificantly improved the quality and diversity of our
dataset. The final size of our dataset is represented by
1500 images divided in a balanced way between the 3
classes (500 images for each class). The last step con-
sists of splitting the data into three (03) subsets: train
(70%), validation (15%) and test (15%). This allowed
to ensure a fair validation during and after the training
process.
3.3 Models Training and
Hyperparameters Optimization
This step focuses on the optimization of hyperparam-
eters for each model. For this aim, we used Optuna
framework (Akiba et al., 2019) that allowed to explore
in an efficient way the hyperparamters space using
grid search algorithms such as as Bayesian optimiza-
tion within Tree-structured Parzen Estimator (TPE).
In our case, the objective function is represented by
the ”loss val” and the researched hyperparameters are
represented by the values of: optimizer (gradient de-
scent method, batch size, learning rate,etc.), regular-
ization (L1, L2, batch-normalization, droupout, etc.)
and data augmentation. Notice that we used an early
stopping process in order to stop the training at the
best moment and avoid overfitting.
3.4 Model Explainability
We have applied an explainability method among
those proposed in Section 2.2. For this aim, we se-
lected Grad-CAM (Selvaraju et al., 2020) technique
due to its high efficiency and also its compatibility
with CNN, Vit and combined (CNN-Vit) architec-
tures. This explainability, applied within the inference
phase (post-hoc), allowed to identify the responsible
pixels of each model decision and then detect even-
tual bias in the model or/and in the training datasets
(Fig. 1). As shown in this figure, we can note that
the model focuses on the good fire are for detection
which means that our model is not biased.
3.5 Evaluation Metrics
Before analyzing the computational, memory, and en-
ergy performance of our models, we first evaluate
their accuracy using the following metrics: accuracy,
loss and confusion matrix. Note that the accuracy and
loss metrics are calculated for the training, validation,
and test datasets.
4 MODEL COMPRESSION
APPROACH
In order to deploy these models in production, we use
Edge AI resources. For that, we propose and study
model compression for producing light weight accu-
rate models.
4.1 Edge AI System
Figure 3: Edge Surveillance System.
We propose to use a system composed of (Fig.3):
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
850

• HD Camera: that can be connected with USB
port and collect 2 images in real-time ;
• Jetson Nvidia Device: including an embedded
GPU allowing to apply model inference on video
frames. Our solution is compatible with several
Jetson devices proposing variable computational
capacities depending on the device price
3
.
• Portable Monitor: to visualize the processed
video showing the detection results and notifica-
tions all evaluation metrics.
4.2 Model Compression
Figure 4: Two Steps Edge Optimized Model Deployment
Pipeline: (1) None Edge Resource based Pruning + Distil-
lation (2) On Edge Resource Quantization and Deployment
In order to propose a compression method that al-
lows to reduce both computation, memory and en-
ergy consumption, we propose to apply an hybrid
method combining the techniques of pruning, quan-
tization and knowledge distillation. Therefore, our
compression algorithm consists of three steps: block
pruning, knowledge distillation and int8 quantization.
• Block Pruning: this method reduces model com-
plexity by eliminating less important blocks from
the neural network, resulting dense and com-
pressed architectures. This approach is more ad-
vantageous than the unstructured method, which
removes less important weights by converting
them to zero values. While the unstructured
method reduces model size, it does not decrease
computational complexity and leads to sparse
neural networks. Notably, block pruning is ap-
plied during the training process to ensure that the
highest possible accuracy is achieved.
• Knowledge Distillation: used with the teacher
model (base model) and the student model
(pruned model) to help the pruned model recover
3
Jetson Nvidia Comparison:
https://developer.nvidia.com/embedded/jetson-modules
its accuracy. The base model guides the learning
process, enabling the pruned model to regain per-
formance despite its reduced complexity.
• Int8 Quantization: allows to store both weights
and activation (static quantization) values using a
lower precision format (int8) instead of the orig-
inal format (float 32). We propose to apply the
quantization after training (on Edge device) and
not during the training (quantization aware train-
ing) in order to identify the format according to
the target Edge resource. Notice that static quan-
tization provided better results than dynamic one
which justify its selection for this work. The
quantization format of int8 was selected empiri-
cally after experimentation within several config-
urations.
4.3 Evaluation Metrics
To ensure a fair and comprehensive evaluation of
our models, we analyze key metrics that account
for accuracy, computation time, memory and stor-
age usage, energy consumption, and model complex-
ity. For this purpose, we developed a library called
”PyTorch Bench” to compute the following metrics:
Number of model parameters (weights & biases),
Model size (in MB), Number of MACs (complexity
evaluation), Maximum memory usage (in MB, Cur-
rent memory usage (in MB), FPS (frames processed
per second), Estimated energy emissions (in gCO
2
),
Total energy consumption (in mWh).
5 EXPERIMENTAL RESULTS
In this section, we evaluate and compare CNN, VIT
and combined architecture within metrics related to
accuracy, loss, explainability, computation perfor-
mance, memory size and energy consumption. Our
experimental results are presented in four subsec-
tions: experimental setup and database, models eval-
uation before compression, models evaluation after
compression, discussion.
5.1 Experimental Setup and Database
Experimentations were conducted using the following
hardware:
• CPU Processor of training server: Intel I9-
13900K, 5.8 GHZ, 24 Cores, 64G RAM DDR5
• GPU Processor of training server: RTX 4090
24GB, 16000 CUDA cores
Edge AI System for Real-Time and Explainable Forest Fire Detection Using Compressed Deep Learning Models
851

• Edge AI device: Jetson Xavier, 512 CUDA cores,
16 GB of memory
In the following parts, the deep learning mod-
els are trained and evaluated using the framework
Pytorch, known by its computation efficiency (Lerat
et al., 2023) compared to other deep learning frame-
works. As shown in Section 3.2, the models are
trained within a 3 classes balanced dataset composed
of 1500 images where 70% are used for training, 15%
for validation and 15% for test.
5.2 Evaluation Before Compression
This subsection is related to present the results of the
model development approach outlined in Section 3
where several steps (DL models selection, data pre-
processing, models training and hyperparameters op-
timization, models explainability and evaluation met-
rics) have been followed to provide the best results
within CNN architectures, Vit and ConvNeXt archi-
tectures. Notice that the compression is not yet ap-
plied in this subection. Table 1 illustrates the obtained
results within the different deep neural architectures,
where different metrics are calculated.
We note a high test accuracy (from 94% to 98%)
achieved by these models, thanks to our data pre-
processing and hyperparameters optimization process
(the latter allowed us to improve models precision
with around 10% of test accuracy) . We also note that
CNN architectures provided the best results (lightly)
compared to vision transformers or even architectures
that combine convolutions and attention mechanisms
such as ConvNeXt, which are more complex and con-
suming in time. This can be interpreted by the nature
of our images and classes where fire and smoke areas
are generally present within connected zones that do
not require necessarily the usage of attention mech-
anisms adapted for taking into account the correla-
tion between different zones (tokens). Notice also
that best results are provided by more complex ar-
chitectures such as EfficientNet V2 L that uses more
parameters and hence more computation time. This
motivates our proposition of developing compressed
deep neural architectures.
5.3 Evaluation after Compression
Based on the previous subsections, we selected the
most accurate model for each DNN category. We then
applied our compression approach, which combines
block pruning, knowledge distillation, and quantiza-
tion (Section 4.2), and evaluated the models on the
Jetson Xavier to simulate real-world scenarios. As
shown in Table 2, the combination of pruning and
knowledge distillation allowed to accelerate and re-
duce the complexity of our different models (CNN,
Vit and ConvNeXt) with the maintain of a high
model accuracy. Notice that pruning has mainly con-
tributed on reducing the model computational com-
plexity while knowledge distillation helped recover
the accuracy lost as a result of the pruning step. Thus,
the combination between the processes of pruning and
knowledge distillation provided excellent balance for
all metrics. On the other hand, quantization applied
on the Edge AI resource allowed to improve signifi-
cantly the computation performance. As a result, our
compression approach enabled the following:
• Reduce the complexity of our models thanks to
the block pruning: with a ratio of 50%
• Accelerate the computation: with a speedup rang-
ing from 3× to 18×
• Reduce the memory consumption: with a factor
ranging from 3.8× to 10.6×
• Reduce the model sizes: with a factor ranging
from 1.8× to 7.6×
• Reduce the energy consumption: with a factor
ranging 3.5× to 6.3×
Notice that we have also analyzed the explainabil-
ity results where the compression has not affected the
calculation of attention heatmap that are always fo-
cusing on the detection classes zones (Fig. 1). We
note that the pruning was applied with a rate of 50%
which means that models complexity is reduced to the
half. We also note that the proposed Edge AI sys-
tem can run during several hours in constrained envi-
ronments since the deployed models are highly com-
pressed and do not cause overrun problems.
5.4 Discussion
Experimental results demonstrated the interest of us-
ing CNN or/and Vision Transformer based neural ar-
chitectures for fire detection, where the accuracy can
reach 98% with a high variated test dataset. We also
note that our steps of data preprocessing and hyper-
parameters optimization allowed to increase the test
accuracy with a range of 10%. The application of
regularization techniques (L1, L2, Dropout, Batch-
normalization and early stopping) allowed to avoid
the problem of overfitting where no variance is noted
between train, validation and test accuracy values. On
the other hand, experimentations allowed to validate
our compression approach since we were able to re-
duce significantly the models consumption in terms
of memory, storage, energy and computation with the
maintain of a high accuracy with a good explanation
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
852

Table 1: Performance comparison of CNN and Transformer-based models for fire detection: before compression.
DNN Model
Category
DNN Model
Family
DNN Model
Name
# Param
(M)
Val
Acc
Val
Loss
Test
Acc
Test
Loss
CNN Networks
VGG
VGG11 132.9 96.44 0.186 95.11 0.092
VGG13 133.0 95.56 0.176 96.00 0.150
VGG16 138.4 96.89 0.200 96.44 0.160
VGG19 143.7 97.33 0.088 97.33 0.095
ResNet
ResNet18 11.7 96.00 0.100 97.78 0.060
ResNet34 21.8 96.00 0.110 96.44 0.083
ResNet50 25.6 96.89 0.090 97.33 0.093
ResNet101 44.5 96.89 0.072 95.56 0.095
ResNet152 60.2 97.33 0.088 95.56 0.100
EfficientNet
EfficientNet V2 S 21.5 96.00 0.074 96.00 0.081
EfficientNet V2 M 54.1 97.78 0.072 96.00 0.110
EfficientNet V2 L 118.5 96.89 0.059 98.22 0.071
Vision Transformers
VIT
Vit B 16 86.6 96.89 0.110 97.33 0.057
Vit B 32 88.2 95.55 0.150 96.44 0.076
Combined
Conv + Attention
Mechanism
ConvNeXt-Small 50.2 96.44 0.074 96.00 0.131
ConvNeXt-Base 88.6 96.88 0.104 94.66 0.122
ConvNeXt-Large 197.8 97.78 0.089 96.00 0.092
Table 2: Performance comparison of various DNN Models (CNNs and Vision Transformers) under different optimization
scenarios (Base, Pruned & Distilled, and Pruned & Distilled & Quantized).
Model
Base
Model
Pruned &
Distilled Model
Pruned & Distilled &
Quantized Model
FPS
Mem
(MB)
Energy
(mWh)
Test
acc
FPS
Mem
(MB)
Energy
(mWh)
Test
acc
FPS
Mem
(MB)
Energy
(mWh)
Test
acc
VGG19 18.2 2128 0.082 97.33 51.2 835 0.063 98.66 326.6 547 0.013 98.66
ResNet18 196.9 195 0.016 97.78 256.9 94 0.014 97.78 606.5 80 0.004 96.88
EfficientNet V2 L 18.3 737 0.148 98.22 20.2 601 0.142 96.44 98.1 255 0.033 96.44
Vit B 16 21.9 611 0.099 97.33 34.6 479 0.089 97.33 102.4 121 0,028 96.88
ConvNeXt-Large 9.2 834 0.322 96.00 17.1 421 0.177 96.66 78.9 223 0.055 96.66
that confirms the absence of model and data biases.
It is also worth noting that, in some cases, the com-
pressed model achieves better accuracy than the orig-
inal model. This is due to the fact that pruning acts
as a form of regularization, enhancing the model’s
performance. We also note that our compression ap-
proach has provided these improvements across the
different DNN architectures (CNN, Vision transform-
ers and combined architectures) which is very promis-
ing for a widespread use of our solution.
6 CONCLUSION
In this paper, we proposed an innovative approach
providing accurate and explainable deep neural ar-
chitectures for forest fire detection. The proposed
approach is based on an efficient preprocessing step
followed by the selection of CNN or/and transformer
based architectures with an optimization of their hy-
perparameters. Once a high accuracy obtained, we
focused on the compression of the models with an hy-
brid approach combining the processes of block prun-
ing, knowledge distillation and quantization. As re-
sult, the generated models were improved in terms
of accuracy and performance (memory, storage, size,
complexity) with the maintain of a good explainabil-
ity of results. These improvements enabled easy de-
ployment of our models on Edge AI devices, allow-
ing real-time processing. As future work, we plan to
apply our compression approach on other DNN ar-
chitectures such as those used for image detection or
segmentation and even those using Large vision mod-
els (Shi et al., 2025). We plan also to accelerate the
training process using parallel and distributed learn-
ing (Lerat and Mahmoudi, 2024).
Edge AI System for Real-Time and Explainable Forest Fire Detection Using Compressed Deep Learning Models
853

REFERENCES
Akiba, T., Sano, S., Yanase, T., Ohta, T., and Koyama, M.
(2019). Optuna: A next-generation hyperparameter
optimization framework. In Proceedings of the 25th
ACM SIGKDD International Conference on Knowl-
edge Discovery and Data Mining.
Blalock, D., Gonzalez Ortiz, J. J., Frankle, J., and Guttag, J.
(2020). What is the state of neural network pruning?
In Dhillon, I., Papailiopoulos, D., and Sze, V., editors,
Proceedings of Machine Learning and Systems, vol-
ume 2, pages 129–146.
Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn,
D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer,
M., Heigold, G., Gelly, S., et al. (2020). An image
is worth 16x16 words: Transformers for image recog-
nition at scale. ICLR 2021 · The Ninth International
Conference on Learning Representations.
Englebert, A., Stassin, S., Nanfack, G., Mahmoudi, S. A.,
Siebert, X., Cornu, O., and De Vleeschouwer, C.
(2023). Explaining through transformer input sam-
pling. In Proceedings of the IEEE/CVF International
Conference on Computer Vision (ICCV) Workshops,
pages 806–815.
Gou, J., Yu, B., Maybank, S. J., and Tao, D. (2021). Knowl-
edge distillation: A survey. International Journal of
Computer Vision, 129(6):1789–1819.
Govil, K., Welch, M. L., Ball, J. T., and Pennypacker, C. R.
(2020). Preliminary results from a wildfire detection
system using deep learning on remote camera images.
Remote Sensing, 12(1).
Hamzah, S. A., Dalimin, M. N., Som, M. M., Zainal, M. S.,
Khairun, Ramli, N., Utomo, W. M., and Yusoff, N. A.
(2022). High accuracy sensor nodes for a peat swamp
forest fire detection using esp32 camera.
Khan, S. and Khan, A. (2022). Ffirenet: Deep learning
based forest fire classification and detection in smart
cities. Symmetry, 14(10).
Lagunas, F., Charlaix, E., Sanh, V., and Rush, A. M. (2021).
Block pruning for faster transformers. Proceedings of
the 2021 Conference on Empirical Methods in Natural
Language Processing, pages 10619–10629.
Lerat, J.-S. and Mahmoudi, S. A. (2024). Scalable deep
learning for industry 4.0: Speedup with distributed
deep learning and environmental sustainability con-
siderations. In Artificial Intelligence and High Per-
formance Computing in the Cloud, pages 182–204,
Cham. Springer Nature Switzerland.
Lerat, J.-S., Mahmoudi, S. A., and Mahmoudi, S. (2023).
Single node deep learning frameworks: Comparative
study and cpu/gpu performance analysis. Concur-
rency and Computation: Practice and Experience,
35(14):e6730.
Liu, Z., Mao, H., Wu, C.-Y., Feichtenhofer, C., Darrell, T.,
and Xie, S. (2022). A convnet for the 2020s. In Pro-
ceedings of the IEEE/CVF conference on computer vi-
sion and pattern recognition, pages 11976–11986.
Madeira, E. M. (2023). The crucial role of forests in com-
bating climate change. Trends in Biotechnology and
Plant Sciences.
Mahmoudi, S. A., Stassin, S., Daho, M. E. H., Lessage, X.,
and Mahmoudi, S. (2022). Explainable Deep Learn-
ing for Covid-19 Detection Using Chest X-ray and CT-
Scan Images, pages 311–336. Springer International
Publishing, Cham.
Nagel, M., Fournarakis, M., Amjad, R. A., Bondarenko, Y.,
Van Baalen, M., and Blankevoort, T. (2021). A white
paper on neural network quantization. arXiv preprint
arXiv:2106.08295.
Park, M., Tran, D. Q., Lee, S., and Park, S. (2021). Mul-
tilabel image classification with deep transfer learn-
ing for decision support on wildfire response. Remote
Sensing, 13(19).
Petsiuk, V., Das, A., and Saenko, K. (2018). Rise: Ran-
domized input sampling for explanation of black-box
models. In BMC.
Pruthi, D., Gupta, M., Dhingra, B., Neubig, G., and Lipton,
Z. C. (2020). Learning to deceive with attention-based
explanations. Proceedings of the 58th Annual Meeting
of the Association for Computational Linguistics.
Selvaraju, R. R., Cogswell, M., Das, A., Vedantam, R.,
Parikh, D., and Batra, D. (2020). Grad-cam: visual
explanations from deep networks via gradient-based
localization. International journal of computer vision,
128:336–359.
Shamsoshoara, A., Afghah, F., Razi, A., Zheng, L., Ful
´
e,
P. Z., and Blasch, E. (2021). Aerial imagery pile
burn detection using deep learning: The flame dataset.
Computer Networks, 193:108001.
Shi, B., Wu, Z., Mao, M., Wang, X., and Darrell, T. (2025).
When do we not need larger vision models? In Euro-
pean Conference on Computer Vision, pages 444–462.
Springer.
Simonyan, K., Vedaldi, A., and Zisserman, A. (2014).
Deep inside convolutional networks: Visualising im-
age classification models and saliency maps. ICLR
2014 · International Conference on Learning Repre-
sentations. Workshop Poster.
Sousa, M. J., Moutinho, A., and Almeida, M. (2020).
Wildfire detection using transfer learning on aug-
mented datasets. Expert Systems with Applications,
142:112975.
Stassin, S., Corduant, V., Mahmoudi, S. A., and Siebert, X.
(2024). Explainability and evaluation of vision trans-
formers: An in-depth experimental study. Electronics,
13(1).
Wang, H., Wang, Z., Du, M., Yang, F., Zhang, Z., Ding, S.,
Mardziel, P., and Hu, X. (2020). Score-cam: Score-
weighted visual explanations for convolutional neural
networks. In CVPR Worshop on TCV.
Wu, H., Li, H., Shamsoshoara, A., Razi, A., and Afghah,
F. (2020). Transfer learning for wildfire identification
in uav imagery. In 2020 54th Annual Conference on
Information Sciences and Systems (CISS), pages 1–6.
Zeiler, M. D. and Fergus, R. (2014). Visualizing and under-
standing convolutional networks. In European confer-
ence on computer vision, pages 818–833. Springer.
Zhao, Y., Ma, J., Li, X., and Zhang, J. (2018). Saliency de-
tection and deep learning-based wildfire identification
in uav imagery. Sensors, 18(3).
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
854