Received 15 March 2024, accepted 27 March 2024, date of publication 1 April 2024, date of current version 17 April 2024.

Digital Object Identifier 10.1109/ACCESS.2024.3383653

FireDetXplainer: Decoding Wildfire Detection

With Transparency and Explainable
AI Insights
SYEDA FIZA RUBAB , ARSLAN ABDUL GHAFFAR , AND GYU SANG CHOI , (Member, IEEE)
Department of Information and Communication Engineering, Yeungnam University, Gyeongsan 38541, Republic of Korea
Corresponding author: Gyu Sang Choi ([email protected])
This work was supported in part by Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by
the Ministry of Education (NRF-2021R1A6A1A03039493) and in part by the 2024 Yeungnam University Research Grant.

ABSTRACT Recent analyses by leading national wildfire and emergency monitoring agencies have
highlighted an alarming trend: the impact of wildfire devastation has escalated to nearly three times
that of a decade ago. To address this challenge, we propose FireDetXplainer (FDX), a robust deep-
learning model that enhances the interpretability often lacking in current solutions. FDX employs an
innovative approach, combining transfer learning and fine-tuning methodologies with the Learning without
Forgetting (LwF) framework. A key aspect of our methodology is the utilization of the pre-trained
MobileNetV3 model, renowned for its efficiency in image classification tasks. Through strategic adaptation
and augmentation, we have achieved an exceptional classification accuracy of 99.91%. The model is
further refined with convolutional blocks and advanced image pre-processing techniques, contributing to
this high level of precision. Leveraging diverse datasets from Kaggle and Mendeley, FireDetXplainer
incorporates Explainable AI (XAI) tools such as Gradient Weighted Class Activation Map (Grad-CAM)
and Local Interpretable Model-Agnostic Explanations (LIME) for comprehensive result interpretation. Our
extensive experimental results demonstrate that FireDetXplainer not only outperforms existing state-of-the-
art models but does so with remarkable accuracy, making it a highly effective solution for interpretable image
classification in wildfire management.

INDEX TERMS Deep learning, explainable AI (XAI), transfer learning, wildfire detection.

I. INTRODUCTION However, the adoption of these advanced AI models in

The escalating severity and frequency of wildfires globally critical areas like wildfire management is often impeded
necessitate advanced detection and management strategies. by their complexity and lack of interpretability. This chal-
Recent studies highlight the role of climate change and lenge is addressed by integrating Explainable AI (XAI)
human activities in exacerbating these events, making timely methods such as SHAP, LIME, and CAM, which aim to
and accurate detection more crucial than ever [1], [2]. provide clearer insights into the decision-making processes
In response, the field of artificial intelligence, particularly of AI models, thereby enhancing their reliability and
deep learning, has seen significant advancements, offering trustworthiness [5], [6].
promising solutions for wildfire detection and analysis. Another essential aspect in the evolution of AI models
The development of efficient convolutional neural networks for environmental applications is the ability to continually
(CNNs), especially architectures like MobileNetV3, has learn and adapt to new data without forgetting previously
revolutionized image classification tasks, which are integral acquired knowledge. Techniques like transfer learning and
to detecting and monitoring wildfires [3], [4]. the Learning without Forgetting (LwF) framework have
become crucial in this regard, facilitating efficient and
comprehensive learning processes [7].
The associate editor coordinating the review of this manuscript and The increasing incidence and impact of wildfires, partic-
approving it for publication was Laxmisha Rai . ularly in Mediterranean countries, underscore the urgency

2024 The Authors. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
52378 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ VOLUME 12, 2024
S. F. Rubab et al.: FDX: Decoding Wildfire Detection With Transparency and XAI Insights

for effective detection and response mechanisms. Studies detection to offer transparent and understandable expla-
have emphasized the role of explainable artificial intelligence nations for its predictions, thereby increasing user trust
(XAI) in enhancing wildfire prediction models, advocating and model interpretability.
for models that are not only accurate but also interpretable [8]. • Robust Testing Across Diverse Datasets: By lever-
Similarly, research by [9] highlights the importance of aging datasets from Kaggle and Mendeley, the model
early fire detection using AI and deep learning techniques, demonstrates unparalleled adaptability and robustness,
stressing the need for models that maintain their effectiveness showcasing its potential for real-world application
on new datasets without compromising on initial training. across various conditions and scenarios.
Reference [10] explores the development of lightweight • Setting a New Benchmark: FDX represents a paradigm
models for wildfire image classification and illustrates the shift in the use of deep learning for environmental
need for efficient and scalable solutions in this domain. applications, combining high accuracy, adaptability, and
Furthermore, the advancement in model-agnostic explainable explainability in a way that sets a new benchmark for
AI for object detection, as discussed in recent studies, aligns future research in the field.
with the growing demand for transparent AI systems in Looking ahead, our research paves the way for significant
various applications, including wildfire detection [11]. advancements in detecting wildfires more accurately and
In a world increasingly reliant on artificial intelligence responsively. Building on the solid foundation laid by the
(AI), the field of explainable AI (XAI) has emerged as a current study, future efforts will focus on refining our
beacon of clarity, aiming to demystify the decisions made model to distinguish fires from similar phenomena more
by AI systems. A systematic review [12] of XAI literature precisely. This will involve integrating a broader spectrum
reveals a growing discourse on the necessity of explanation of data, including varied weather conditions and temperature
in AI, highlighting four key debates central to the ‘black- changes, to train the model to identify actual fires amidst
box’ problem. This research suggests an urgent need for a everyday natural occurrences like clouds or sunsets [14]. The
unified conceptual understanding within XAI, emphasizing ambition is to develop a highly precise detection tool that can
the importance of generating trust through transparency, quickly alert to forest fires, enabling swift action to mitigate
ensuring compliance with regulations, promoting social damage and protect communities [15]. By harnessing cutting-
responsibility, and establishing accountability. The review edge technologies and expanding our dataset to cover a wider
calls for empirical studies that address diverse stakeholder array of wildfire scenarios, we aim to create a comprehensive
needs, aiming to measure XAI’s true efficacy in making AI solution for early wildfire detection and management. This
more comprehensible and trustworthy for all. Reference [13] future work will not only enhance the model’s accuracy
complemented by exploration of convolutional neural net- and reliability but also its applicability across different
works in various industrial technologies, underscoring the environments, setting a new standard in the field of wildfire
versatility and potential of these models in diverse applica- management [16].
tions.
Highlighting the major contributions and novelty, this II. LITERATURE REVIEW
study presents the following key advancements in the field Wildfires present a growing global challenge, and the race to
of wildfire detection and management: detect them early is on. Aiming to improve response times
• Innovative Integration of Techniques: FireDetX- and mitigate damage, a study has harnessed the continuous
plainer (FDX) uniquely combines transfer learning, stream of hyperspectral satellite data. Their deep learning
fine-tuning methodologies, and the Learning without model showcases an impressive capability to detect wildfires
Forgetting (LwF) framework to address catastrophic rapidly and accurately, promising a significant leap forward
forgetting, enhancing the model’s ability to retain in the domain of remote sensing and disaster prevention.
knowledge while learning new information. Recent research [17] leveraged hyperspectral satellite
• Utilization of MobileNetV3: FDX leverages the pre- imagery for early wildfire detection, utilizing deep learning to
trained MobileNetV3 model, known for its efficiency in pinpoint fires at the pixel level. A novel system outperformed
image classification, to achieve exceptional accuracy in baselines, yielding a 94% F1-score and faster detection
wildfire imagery classification. This marks a significant by 1.5 times, despite challenges like night-time fires.
advancement in applying pre-trained models to environ- This method, tested on GOES-16 data, also integrated an
mental challenges. interactive dashboard for detailed regional analysis, marking
• Advanced Model Adaptation: The research intro- progress in the field of remote sensing for fire identification.
duces strategic adaptation and augmentation techniques, Building on the imperative for early detection of forest
including convolutional blocks and advanced image pre- fires, a novel study [18] implemented a deep learning model
processing, to improve the model’s capability in identi- using Himawari-8 satellite data, aimed at reducing detection
fying complex patterns specific to wildfire imagery. latency. This model, leveraging both spatial and temporal
• Integration of Explainable AI (XAI) Tools: FDX data, improved detection times significantly, achieving an
incorporates Explainable AI tools like Grad-CAM and average initial detection time of just 12 minutes. The
LIME, making it one of the first models in wildfire convolutional neural network (CNN) method outshined

VOLUME 12, 2024 52379

 S. F. Rubab et al.: FDX: Decoding Wildfire Detection With Transparency and XAI Insights

traditional random forest (RF) approaches, showcasing its model represents a significant advancement in the domain-
spatial pattern recognition capabilities. It demonstrated an incremental learning field, providing a robust framework
overall accuracy and F1-score of 0.98 and 0.74, respectively, for fire detection in integrated space-air-ground observation
with the added advantage of robust performance under networks.
diverse environmental conditions. The CNN’s swift detection The study [22] examines efficient CNN architectures for
ability marks a notable advancement in the utilization of real-time fire detection, proposing two compact models:
geostationary satellite data for timely forest fire monitoring. NasNet-A-OnFire and ShuffleNetV2-OnFire. These archi-
Tackling the significant environmental threat posed by tectures are optimized for computational efficiency and
forest fires, researchers [19] have developed a deep learning speed, with ShuffleNetV2-OnFire achieving a 2.3× speed
system designed to quickly detect fires and alert authorities. increase for binary classification and a 1.3× speed increase
Utilizing a Convolutional Neural Network (CNN) model, this for superpixel localization, at 40 fps and 18 fps, respectively.
system processes images from a custom dataset, achieving The models demonstrate robust performance, with NasNet-
commendable detection accuracies of 93% in training and A-OnFire and ShuffleNetV2-OnFire both achieving 95%
92% during testing. The system, supported by a Raspberry accuracy for full-frame classification and 97% for superpixel
Pi setup, not only identifies fires but also sends notifications localization. Notably, the ShuffleNetV2-OnFire architecture
through various channels. With its high accuracy and real- is over six times more compact than previous works, with
time alert capabilities, this model presents a practical solution around 0.15 million parameters, while still maintaining high
for early fire detection, potentially limiting the damaging accuracy. This advancement demonstrates the potential for
impact of wildfires on ecosystems. deploying these models on low-powered devices, like the
In the face of escalating forest fires, [20] have developed a Nvidia Xavier-NX, which reached 49 fps, highlighting their
deep learning-based system designed for early detection and suitability for real-world applications.
classification. This system capitalizes on the Inception-v3 In the research [5], the effectiveness of LIME, SHAP,
convolutional neural network’s transfer learning capabilities, and CAM, three explainable AI methods, was put to the
enhanced by Radial Basis Function Networks (RBFNs) with test for image classification. LIME came out on top, closely
Rapid and Accurate Image Super Resolution (RAISR). The matching human decision-making by identifying key image
proposed model outshines previous CNN models with a segments with an Intersection Over Union (IOU) score of
notable accuracy of 97.55% and an F-Score of 93.33%, 0.566. Although CAM wasn’t as precise, it was the fastest,
showcasing its effectiveness in identifying fire and non- processing images in just 0.012 seconds. This comparison
fire images. It also utilizes the water wave optimization highlights the balance between accuracy and speed needed
technique for image feature selection and enhancement. This for real-time AI applications. The study suggests more work
method sets a new precedent for forest fire categorization, is needed to further enhance these methods for broader
promising to aid in developing strategic responses to such uses.
disasters. While utilizing the capabilities of unmanned In a study [8] exploring wildfire detection in South-
aerial vehicles (UAVs), a groundbreaking wildfire detection ern Europe, an XAI framework using a Random Forest
system was crafted, employing deep convolutional neural model was employed to interpret wildfire occurrence. The
networks for heightened accuracy in identifying fires from framework’s accuracy in predicting wildfire-prone regions
aerial photos [21]. Through rigorous testing, GoogLeNet was impressive, with an area under the curve (AUC) of
and a modified VGG13 network demonstrated superior 81.3%. The study’s performance metrics showed an accuracy
performance, with GoogLeNet achieving an impressive of 69.7%, an F1 score of 62.0%, and a high sensitivity
99.0% accuracy. of 78.7%, indicating its effectiveness in detecting wildfire
The modified VGG13, tailored for the specific task of occurrences. Despite these promising results, the precision
fire detection, also showed promising results with a 96.9% was relatively modest at 50.9%, suggesting a balance between
accuracy rate. These networks, by swiftly processing vast correctly predicting wildfires and avoiding false alarms is
quantities of high-resolution UAV imagery, provide a crucial needed. This research underlines the potential of XAI in
tool in the early detection and monitoring of wildfires, aiding forest management and disaster mitigation strategies.
potentially saving lives and reducing economic losses. The FireXplainer model [2] has marked a significant step
The study [1] introduces a Dynamic Equilibrium Network forward in the detection and interpretation of wildfires.
(DENet) to enhance fire detection across varied domains, This innovative approach employs advanced techniques like
specifically targeting data from spaceborne, airborne, and transfer learning and the Learning without Forgetting (LwF)
terrestrial sensors. Tested on the Flame And Smoke Detection method, enhancing the precision of wildfire classification.
Dataset (FASDD) and the FLAME dataset, DENet showcased The usage of the XAI Grad-CAM method also adds a
a remarkable balance in learning, evidenced by a 7.71% layer of transparency, pinpointing the exact image regions
improvement in mean average precision (mAP) over the influencing the model’s decisions. The model’s performance
PODNet model on existing datasets. It maintained strong is impressive, boasting a 99.1% precision and a 99.3% recall,
performance on new datasets, achieving an mAP of 97.55% leading to an overall F1 score of 99.0%. These figures not
on older datasets and 92.20% on newer datasets. This only surpass previous state-of-the-art models but also offer

52380 VOLUME 12, 2024

S. F. Rubab et al.: FDX: Decoding Wildfire Detection With Transparency and XAI Insights

TABLE 1. Overview of recent studies in wildfire detection and classification using deep learning techniques.

an interpretable framework that could significantly aid in formed for other severity levels. On average, the BCE-
wildfire management and response strategies. MSE approach achieved a Root Mean Square Error (RMSE)
This research [9] delves into the crucial task of forest fire of 0.54, outperforming the Dice-MSE’s RMSE of 0.65,
and smoke detection through AI and deep learning. It tackles indicating a more accurate prediction of wildfire-affected
the challenge of identifying such calamities early on, aiding areas.
in prompt disaster management. The study employs Convo- The recent study [10] has revealed the efficacy of a new
lutional Neural Networks (CNNs) and explores the efficiency lightweight wildfire image classification model, LW-FIRE,
of transfer learning on pre-trained models like VGG16, which has been meticulously tested across several datasets.
InceptionV3, and Xception. Remarkably, the Xception model The model has been evaluated against existing state-of-the-
shines with an accuracy of 98.72%. When applying the art methods and has shown promising results. Particularly
Learning without Forgetting (LwF) technique, Xception notable is the model’s performance on the LW-FIRE1504 ×
demonstrates robust performance on both new and original 2 configuration, where it achieved an impressive True
datasets, showing a remarkable adaptability without losing Positive Rate (TPR) of 0.980, a False Positive Rate (FPR)
its initial detection capabilities. The research confirms that of 0.091, and an Accuracy (ACC) of 0.944. Additionally,
these advanced models, with LwF, set a new benchmark in LW-FIRE1001 × 2 excelled with a TPR of 0.966 and an
the field, enhancing the precision of detecting forest fires ACC of 0.951. Moreover, LW-FIRE1504 × 2 outperformed
and contributing significantly to environmental conservation its counterparts with a TPR of 0.976 and an ACC of
efforts. 0.973. These results showcase the model’s robustness and its
In the study [23], the researchers improved the clas- potential as a reliable tool for real-time wildfire detection and
sification of wildfire severity from satellite imagery by classification.
experimenting with different loss functions in deep learning In the innovative study [24] of multi-sensor fire detection,
U-Net models. They found that the Binary Cross-Entropy the proposed Recurrent Trend Predictive Neural Network
(BCE) coupled with Mean Square Error (MSE) loss function (rTPNN) has significantly advanced the field. The rTPNN
produced better results than the conventional Dice-MSE model showcases remarkable accuracy, achieving a 96%
approach. For severity classes 2 and 3, the IoU-SoftIoU success rate in fire detection. Its unique architecture allows it
configuration showed the best outcomes, but it underper- to predict fire trends swiftly, outperforming traditional meth-

VOLUME 12, 2024 52381

 S. F. Rubab et al.: FDX: Decoding Wildfire Detection With Transparency and XAI Insights

ods like Linear Regression and Support Vector Machines. TABLE 2. Distribution of image types across datasets used in the study.
Notably, rTPNN excels in real-time responsiveness, trigger-
ing fire alarms in a mere 11 seconds, substantially faster
than alternative models. Its robust design ensures low false
positive and negative rates, affirming its potential as a reliable
solution for urgent fire detection scenarios.
In the study [25] examining wildfire detection using multi-
including Kaggle and Mendeley [26], ensuring a diverse and
sensor satellite imagery and deep learning, the findings
representative collection of visual data. The dataset is curated
were noteworthy. The Sentinel-2 model excelled in clear
to provide a balanced representation of wildfire scenarios,
conditions, achieving an F1 score of 0.83 with 92% of fire
enhancing the model’s ability to accurately identify and
perimeters detected, while the Sentinel-1 model struggled,
classify fire occurrences in various environmental conditions.
with only a 0.46 F1 score and 26% detection rate. The
This rich dataset forms the foundation of the study’s deep
fusion of Sentinel-2 and Sentinel-3 data yielded the highest
learning approach, enabling robust training and validation
detection accuracy in clear conditions, significantly improv-
of the FireDetXplainer model. As illustrated in Table 2, the
ing results with an F1 score reaching 94%. When cloud
distribution of images across the datasets is balanced, offering
cover was a factor, combining Sentinel-1 and Sentinel-2 data
an equal number of fire and non-fire images from Mendeley,
proved advantageous, showcasing the potential for enhanced
while Kaggle provides a larger proportion of fire images.
detection capabilities in adverse weather conditions. These
outcomes underscore the efficacy of employing deep learning
for analyzing complex satellite data, thereby offering a robust B. FIREDETXPLAINER ARCHITECTURE
tool for early wildfire detection and management. To wrap The FireDetXplainer (FDX) model, an innovative framework
up, the latest researches shows that using deep learning and in wildfire detection, represents a significant stride in
detailed images from satellites is changing how quickly and leveraging cutting-edge machine learning for environmental
accurately we can spot wildfires. These modern methods safety. The architecture of FDX is carefully designed to
are a big step forward, as they can catch fires early and combine advanced neural network capabilities with the
help keep people and the environment safer. Tools like LW- pressing need for accurate and efficient wildfire detection and
FIRE and rTPNN have shown that with the right technology, classification. Central to the FDX model is the integration
we can react to wildfires much faster than before. The work of MobileNetV3 for transfer learning, emphasizing not just
done here is not just about new gadgets; it’s about making on high performance but also on computational efficiency,
a real difference in fighting fires and maybe even saving making it suitable for real-time applications. This section
lives and homes. This is just the beginning, and the hope is delves into the core components of FDX’s architecture,
that these smart systems will be used more widely to tackle highlighting its unique approach in handling the complex task
wildfires everywhere. A comparative overview of recent of wildfire detection. Figure 1 presents a detailed illustration
studies in wildfire detection and classification using different of the FireDetXplainer (FDX) architecture, clearly outlining
methodologies is shown in the Table 1. its components and workflow. This algorithm in Table 3
outlines a machine learning process for wildfire image
III. PROPOSED METHODOLOGY analysis, involving data loading, preprocessing, and splitting.
In this section of FireDetXplainer study, we apply It employs a pre-trained MobileNetV3 model, adjusted for
MobileNetV3 for transfer learning and detail image feature extraction and fine-tuning with wildfire data. The
preprocessing for training efficiency. We also incorporate model is trained, evaluated, and integrated with Explainable
Explainable AI (XAI) techniques, such as Grad-CAM and AI techniques like Grad-CAM and LIME for insightful visu-
LIME, to enhance the model’s transparency and decision- alizations and interpretability, culminating in the aggregation
making clarity. of performance metrics and trained models.

A. DATASETS C. TRANSFER LEARNING WITH MOBILENETV3

The FireDetXplainer study utilizes a comprehensive dataset 1) OPTIMIZING MOBILENETV3 INTEGRATION
to train and evaluate its wildfire detection model. This dataset • Integration: The FireDetXplainer (FDX) model
comprises a total of 3,520 images, with 2,216 categorized integrates MobileNetV3, recognized for its effi-
as ‘fire’ images and 1,304 as ‘no-fire’ images. The dataset ciency and effectiveness in image classification tasks.
contains images captured in clear weather conditions, both MobileNetV3’s design is particularly advantageous
during the day and at night. This selection ensures the for applications where computational resources are
model’s accuracy in accurately differentiating fire and non- limited, making it an ideal choice for real-time wildfire
fire cases with maximum visibility. The data assembled detection.
by the model covers diverse landscapes comprising dense • Adaption: FDX capitalizes on the pre-trained
forests, open fields and even fringes of urban regions. MobileNetV3, which has been trained on extensive
These images are sourced from multiple online platforms, datasets. This allows the model to effectively identify

52382 VOLUME 12, 2024

S. F. Rubab et al.: FDX: Decoding Wildfire Detection With Transparency and XAI Insights

TABLE 3. Algorithm of fireDetXplainer.

and interpret unique patterns in wildfire images, model to specialize in recognizing fire and non-fire images.
leveraging its deep learning capabilities. This step involves a lower learning rate to make small, precise
• Preservation and Adaptability: A key aspect of FDX’s updates to the weights, ensuring the model adapts to the
approach is the preservation of MobileNetV3’s initial specifics of wildfire imagery without losing its pre-learned
layers. These layers are kept frozen to retain their general knowledge.
fundamental image processing skills. Meanwhile, the
latter layers of the model are fine-tuned to focus 3) HYPER-PARAMETER OPTIMIZATION
specifically on the unique characteristics of wildfire In the FireDetXplainer (FDX) model, hyper-parameter opti-
images. This dual strategy ensures that FDX maintains mization is a critical process aimed at refining the model’s
a solid foundation in general image recognition while performance. This optimization includes:
being highly effective in identifying wildfire-specific • Optimization Strategy: The model undergoes a
features. detailed process where key parameters such as learning
rate, batch size, and the number of training epochs
2) FINE TUNING TECHNIQUES are carefully adjusted. These parameters significantly
Fine-tuning in the FireDetXplainer (FDX) architec- influence the model’s learning ability and overall
ture involves selectively training the top layers of the accuracy.
MobileNetV3 model. This process is crucial for adapting • Balancing Act: The goal of this optimization is to find
the model to recognize and classify new data - specifically, the ideal balance that prevents the model from over-
images related to wildfires. The aim of this fine-tuning is to fitting (being too narrowly adjusted to the training data)
strike a balance between general and specific learning: or under-fitting (failing to capture the underlying trends
• Selective Training: The process targets the top layers in the data). Achieving this balance ensures that the
of the model, which are most adaptable to new types of model performs reliably across various scenarios and
data. By refining these layers, FDX becomes attuned to datasets.
the nuances of wildfire images.
• Balanced Learning: This approach ensures that while 4) LEARNING WITHOUT FORGETTING (LWF)
the model remains versatile enough to process a METHODOLOGY
variety of image types, it also becomes proficient in The Learning without Forgetting (LwF) approach in FireDe-
identifying critical features indicative of wildfires. The tXplainer (FDX) focuses on dual objectives:
underlying principle is to retain the model’s broad image • Dual Focus: LwF is crucial for ensuring the model
recognition capabilities while enhancing its ability to retains its original capabilities, even as it learns new
detect specific elements characteristic of wildfires. tasks. In the context of FDX, this means maintaining the
The fine-tuning process of the MobileNetV3 model involves model’s general image processing skills while adapting
adjusting the top layers of the pre-trained network to better to the specific requirements of wildfire detection.
suit wildfire detection tasks. Initially, the model, trained • Retaining Core Knowledge: This methodology is vital
on a large and diverse dataset, possesses general image for preserving the foundational strengths of the model.
recognition capabilities. During fine-tuning, the final layers It allows FDX to keep its initial learned behaviors intact,
are selectively retrained on the wildfire dataset, allowing the ensuring a solid base of knowledge is always present,

VOLUME 12, 2024 52383

 S. F. Rubab et al.: FDX: Decoding Wildfire Detection With Transparency and XAI Insights

that variations in image brightness, contrast, or color do

not hinder the model’s learning and performance.

6) TRAINING PROCESS
The FireDetXplainer (FDX) employs a sequential training
approach to optimize wildfire detection:
• Methodical Learning: The training process is designed
to be systematic, enabling the model to incrementally
learn from the enriched dataset. This approach combines
the pre-learned knowledge from MobileNetV3 with new
patterns specific to wildfire imagery, creating a robust
learning environment.
• Focus on Efficacy and Efficiency: The core objective
of this training phase is to ensure the model is not
just learning effectively but also efficiently. This means
optimizing the use of computational resources while
ensuring the model achieves high accuracy in detecting
wildfires. The process is fine-tuned to strike a balance
between rapid learning and depth of understanding,
making the model proficient in recognizing a wide range
of fire scenarios.
In the development of the FDX model, the transfer learning
and the framework is Learning without Forgetting (LwF)
which is brought to combination with the use of the pre-
existing knowledge to keep the model adaptable to new
data while forgetting what previously has been learned not
forgotten. At first, Transfer Learning is involved with using
MobileNetV3 architecture, which is trained up on large-scale
data set. That way FDX can begin with already well-trained
feature-set, which in turn speeds up the training process and
make results of fire detection tasks better.
In parallel with the usage of LwF technique, there is
also a mechanism that is integrated to tackle with likely
scenario of catastrophic forgetting - the scenario of learning
new information that will make previously learnt knowledge
forgotten. LwF applies the structure where the sub-net is
FIGURE 1. Detailed Architecture of FireDetXplainer.
preserved for the pre-learned tasks but yet the other sub-
nets would be modified by the new data. Such an approach
even as the model evolves to handle the complex task of will guarantee that FDX realizes their efficiency in the
identifying wildfires. already mastered tasks and increases the new knowledge
simultaneously, making a harmonious process where former
5) PREPROCESSING TECHNIQUES and new stones are constantly interacting with one another.
In the FireDetXplainer (FDX) architecture, preprocessing Through reconciliation of transfer learning with LwF, FDX
plays a pivotal role in preparing the data for optimal model thrives due to the robust feature extraction competences
training: of a pre-trained model and it can constantly adapt to a
• Data Augmentation: This involves employing a variety particular wildfire detection problems during its lifespan.
of image manipulation techniques, such as rotations, This integrative scheme is the fundament of the architecture
flips, and color adjustments. By artificially expanding of the model and gives it the capacity of high precision and
the dataset, FDX can better understand and generalize adaptability in fastening fire risk in different environments
from the training data to real-world wildfire scenarios, and weather conditions.
enhancing its detection capabilities.
• Normalization and Standardization: These processes D. MODEL TRAINING
are essential for maintaining consistent image quality In this phase of the research on FireDetXplainer, the training
and format. By standardizing the input data, FDX can process is meticulously outlined in alignment with the data
more effectively learn and recognize patterns, which is provided in the code. The dataset is initially segregated into
critical for accurate wildfire detection. This step ensures distinct subsets for training, validation, and testing, following

52384 VOLUME 12, 2024

S. F. Rubab et al.: FDX: Decoding Wildfire Detection With Transparency and XAI Insights

an 80:10:10 ratio which is a common practice to ensure the TABLE 4. Comparison of the state of art models based on evaluation
metrics.
model is exposed to a variety of data while also being able
to validate and test its predictions accurately. In the splitting
of data in this fashion, a model is tested in a way that is not
only on the data it has already been exposed to but also on
other unseen data of which it has no prior acquaintance. This
would improve its capability to generalize.
Hyper-parameter tuning is a pivotal aspect of training
deep learning models, essential for achieving optimal
performance. In this research, critical parameters such as
the learning rate, batch size, and number of epochs were
meticulously fine-tuned. Specifically, the learning rate was
set to a low value of 0.0001, carefully chosen to ensure
gradual and precise adjustments to the model’s weights it ensures reliable fire detection while minimizing false
during training, thus minimizing the risk of overshooting. The alarms, which is crucial in emergency response scenarios.
batch size, affecting the model’s convergence and its ability Table compares the FireDetXplainer results with state of
to generalize, was optimized at 32, striking a balance between the art model based on evaluation matrices. The model’s
computational efficiency and effective learning. Furthermore, high accuracy rate of 99.91% stands as a testament to its
the model underwent training for 100 epochs, a decision robustness and efficiency in real-world conditions. Such a
driven by the need to provide sufficient iterations for the high accuracy rate is indicative of the model’s comprehensive
network to learn from the entire dataset without falling into learning from the training data and its ability to generalize
the trap of over-fitting. These choices in hyper-parameter this learning to new, unseen data. This level of accuracy is
settings underscore the meticulous approach taken to fine- rarely achieved in wildfire detection models, highlighting the
tune the model’s learning process, where the learning rate advanced capabilities of FDX. Table 4 shows a comparison
influences the magnitude of weight updates, the batch size of FireDetXplainer with state of the art models in terms
impacts the model’s learning granularity, and the epochs of evaluation matrices such as precision, recall, f1-score
determine the depth of exposure to the training data. and accuracy score. Furthermore, the model’s complexity,
An early stopping strategy is employed to prevent over- represented by its 6.5 million parameters, is well-justified
fitting, which is a condition where the model performs well by its performance metrics. This complexity enables the
on the training data but fails to generalize to unseen data. model to capture intricate patterns and nuances in the
This technique monitors the model’s performance on the data, which is essential for accurate wildfire detection.
validation set and halts the training once the performance Despite the large number of parameters, the model remains
ceases to improve, thus saving computational resources and computationally efficient, as evidenced by its relatively
time. low training and validation losses (0.0107 and 0.0084,
The code segment uses TensorFlow and Keras libraries, respectively). These losses indicate that the model is not
which are popular for building and training deep learning over-fitting and is well-tuned to the problem at hand. While
models. The training leverages callbacks like ReduceL- the model’s intricate design is crucial for detecting wildfires
ROnPlateau for dynamic learning rate adjustments, and accurately, this complexity, with its 6.5 million parameters,
EarlyStopping for halting the training at the optimal moment. inherently brings challenges, particularly in computational
After training, the model’s performance is evaluated using resource requirements. Such a sophisticated model may
metrics like accuracy, precision, recall, and F1-score, which demand more memory and processing power during real-time
provide insights into the model’s classification prowess. analysis, posing potential constraints for deployment on less
The training process is powered by Google Colab’s GPUs, powerful systems or in mobile applications where efficiency
which are highly efficient at performing the matrix and vector is paramount. Developing strategies to optimize the model for
operations central to deep learning. Using such hardware faster inference without compromising accuracy is essential
accelerates the training process, allowing for more extensive for broadening its applicability across various platforms.
experimentation with hyper-parameters and larger datasets. The FireDetXplainer (FDX) model’s training showcases
its ability to effectively learn from the dataset without over-
IV. RESULTS fitting, as evidenced by a training loss of 0.0107. Coupled
The FireDetXplainer (FDX) model showcases an impressive with this, the model demonstrates exceptional generalization
performance that sets it apart in the realm of wildfire to unseen data, indicated by a validation loss that significantly
detection. Its precision of 99.88% and recall of 99.97% decreases to 0.0084. This reduction in validation loss is
reflect an exceptional ability to correctly identify both fire critical, reflecting the model’s precision in making accurate
and non-fire scenarios with minimal errors. The F1-score of predictions on new data. This indicates that the model has
99.93% further solidifies its accuracy, indicating a balanced effectively learned general patterns and features from the
precision-recall trade-off. This is particularly significant as training data, which are essential for accurate classification,

VOLUME 12, 2024 52385

 S. F. Rubab et al.: FDX: Decoding Wildfire Detection With Transparency and XAI Insights

without over-fitting. Over-fitting is a common issue in It generates a ‘‘heatmap’’ or a visual explanation that
machine learning where a model becomes too attuned to highlights the regions of the image that were most influential
the training data and loses its ability to generalize to new, in the model’s prediction. This heatmap is created by
unseen data. These low values in loss metrics are indicative capturing the gradients flowing into the final convolutional
of a well-optimized model that is capable of making accurate layer of the network, which indicates the importance of each
predictions. Moreover, the model achieves a noteworthy pixel in the image for the classification task. The result is a
milestone by maintaining a 100% validation accuracy from visual overlay that can be placed on the original image to
an early stage, which is maintained throughout the training show which parts contributed most to the decision, such as
process. This perfect validation accuracy, alongside the identifying specific areas of fire in an image when the task
low validation loss, provides a compelling narrative of is to detect wildfires. This method is particularly useful for
the model’s robustness and its reliable performance across validating the focus of the model and ensuring that it aligns
various conditions. The balanced combination of low training with human interpretation and domain knowledge.
and validation losses, together with impeccable validation In this research, Grad-CAM plays a pivotal role in enhanc-
accuracy, underscores the FDX model’s optimized and effec- ing the interpretability of wildfire detection. By applying
tive approach in wildfire detection, ensuring its applicability Grad-CAM, we can visualize which parts of an image the
and effectiveness in real-world scenarios. neural network focuses on when predicting the presence of
wildfires. This is crucial for confirming the model’s reliability
and ensuring it bases its decisions on relevant features, such
as the actual flames or smoke, rather than on irrelevant
background features.
In the Grad-CAM processed image Figure 3, this technique
highlights the areas in the image that led to the model’s
prediction of fire. The warmer or more intense the coloration
on the heatmap, the more significant that region was for the
model’s decision. This visualization allows for an intuitive
understanding of the model’s functioning, ensuring that the
model is paying attention to the right cues in the image,
such as the bright, fiery areas that typically represent a
wildfire, which are prominently marked in the image. This
not only builds trust in the model’s predictions but also
FIGURE 2. Visualization of Evaluation Metrics of FireDetXplainer Model, provides insights into potential improvements for further
Highlighting Scores in Precision, Recall, F1-Score, and Accuracy. model training.

In summary, the FireDetXplainer model’s exceptional

performance metrics - precision, recall, F1-score, and accu-
racy - combined with its efficient learning as indicated
by its loss metrics, clearly demonstrate its superiority over
existing models. The balance between a large number
of parameters and computational efficiency, along with
the outstanding accuracy, positions the FDX model as a
significant advancement in the field of wildfire detection and
management. As illustrated in Figure 2, the FireDetXplainer FIGURE 3. Visualization of Original wildfire image and Grad-CAM
processed image.
model achieved remarkable performance across all key eval-
uation metrics, with precision, recall, F1-score, and accuracy
score. These results underscore the model’s robustness and B. LIME
reliability in fire detection and explanation tasks. LIME, which stands for Local Interpretable Model-agnostic
Explanations, is a technique used to explain the predictions
V. EXPLAINABLE AI (XAI) of machine learning models. It works by approximating the
To improve the interpretability of our model, we incorporate model locally with an interpretable one, such as a linear
two distinct Explainable AI (XAI) techniques: Gradient- model or decision tree. LIME perturbs the input data and
weighted Class Activation Mapping (Grad-CAM) and Local observes the corresponding changes in the output, allowing it
Interpretable Model-agnostic Explanations (LIME). to identify which features significantly influence the model’s
predictions. Besides these, a LIME-based approach emerged
A. GRAD-CAM and proved to have an powerful degree of interpretability
Grad-CAM is a technique that helps to understand the through diverse fields. More specifically, the outcomes of
decision-making process of convolutional neural networks. AIPs-SnTCN, iAFPsEnC-GA, and cACP-DeepGram have

52386 VOLUME 12, 2024

S. F. Rubab et al.: FDX: Decoding Wildfire Detection With Transparency and XAI Insights

indicated LIME as a model which improved a understanding

of a predictive models of biomedical research.
AIPs-SnTCN employs LIME in the context of predicting
anti-inflammatory peptides, combining word embedding
techniques with a self-normalized temporal convolutional
network to achieve high predictive accuracy and interpretabil-
ity [27].
iAFPs-EnC-GA integrates LIME for antifungal peptide
prediction, utilizing a genetic algorithm-based ensemble clas-
sifier and showing the interpretive contribution of features to
model predictions [28]. FIGURE 5. Visualization of Original wildfire image and LIME processed
cACP-DeepGram applies LIME to anticancer peptide image.
prediction, leveraging deep neural network models and
demonstrating improved interpretability and prediction accu- measure of the model’s accuracy, stands at 99.93% and
racy [29]. 99.91% respectively, indicating a robust performance in
In this research, LIME provides a detailed explanation identifying fire-affected areas.
of individual wildfire predictions made by the model. The model benefits from the Learning without Forgetting
By highlighting the specific features within an image that (LwF) methodology, ensuring that it remains adept at its
contribute to the model’s classification decision, LIME offers original tasks while adapting to new challenges. Through
a granular view of the model’s reasoning process. This Grad-CAM visualizations, we can pinpoint which aspects of
can help researchers and practitioners understand and trust the image most inform the model’s predictions. LIME further
the model’s decisions and identify any biases or areas for elucidates the model’s reasoning by highlighting influential
improvement. features. These XAI methodologies not only provide a deeper
The LIME processed images further elucidate this point. understanding of the model’s inner workings but also pave
For example, in Figure 4, LIME highlights areas of smoke the way for potential enhancements. The discussion also
and regions with a particular color or texture that the model contemplates the model’s training efficiency, reflected in the
associates with fire. It outlines these regions, showing their stability of its learning curves over time. The curves showcase
contribution to the model’s prediction. In Figure 5, LIME the model’s ability to maintain high performance, even with
illustrates how certain patterns in the landscape, such as the introduction of new data during the learning process. This
specific terrain features, are weighted by the model in adaptability is critical for real-time applications in diverse and
its decision-making process. These visual explanations are unpredictable environmental conditions.
instrumental in validating the model’s performance and Lastly, the integration of XAI enhances the user’s trust
ensuring that it makes predictions based on relevant and in the model by offering clear and interpretable insights
accurate features. into its decision-making process. Such transparency is vital
for the deployment of AI in critical sectors like wildfire
management, where understanding the rationale behind
algorithmic decisions can have significant implications.
Figure 6, the accompanying graph illustrates these points,
with the accuracy curve plateauing at high levels, confirming
the model’s reliable performance. The loss curve’s decline
suggests that the model has effectively minimized errors
throughout its training. These visual metrics serve as tangible
proof of the model’s efficacy and the successful application
of its learning strategies.
While the FireDetXplainer (FDX) model demonstrates
FIGURE 4. Visualization of Original wildfire image and LIME processed significant advancements in wildfire detection through high
image. accuracy and interpretability, it is important to acknowledge
its limitations and identify areas for further improvement.
VI. DISCUSSION One notable limitation is the dependency on the quality
In discussing the FireDetXplainer model, we focus on its and diversity of the training data. Despite the model’s
strengths in wildfire detection through deep learning. The robust performance in current test scenarios, its effectiveness
model’s integration of advanced neural network architectures may vary in real-world conditions that were not adequately
like MobileNetV3, coupled with Explainable AI (XAI) tech- represented in the training dataset. This underscores the
niques, allows for a high degree of accuracy and transparency. need for continuous expansion and diversification of the
Our results underscore the model’s capability, with precision dataset to encompass a wider range of wildfire scenarios,
and recall rates both exceeding 99%, presenting a significant including different vegetation types, weather conditions, and
leap from traditional methods. The F1-score, a balanced fire behaviors.
VOLUME 12, 2024 52387
 S. F. Rubab et al.: FDX: Decoding Wildfire Detection With Transparency and XAI Insights

FIGURE 6. Visualization of FireDetXplainer model’s training accuracy and loss convergence over epochs.

VII. ETHICAL CONSIDERATIONS IN AI DEPLOYMENT FOR results, anchored by an impressive accuracy rate of 99.91%,
WILDFIRE MANAGEMENT underscore the efficacy of the model, further enhanced by
In this study, the emphasis will be on using AI to improve Explainable AI techniques like Grad-CAM and LIME, which
the process of wildfire detection and accuracy. Fest covers the clarify the model’s decision-making process. These XAI tools
role of critical side in avoiding any case of bias in training data not only provide transparency but also point towards areas for
for machines to not favour any scenario over others. With a potential refinement. The research concludes with reflections
demand for parity being a major factor in AI system adoption, on future work, suggesting avenues to differentiate between
it is imperative to make the AI algorithms suitable for all fire-related images and other environmental textures more
living areas and climates. effectively. This could potentially involve developing more
It is recommended nowadays for AI decision making nuanced models or incorporating additional sensory data
transparency tools to be built-in. Providing our systems with to reduce false positives and increase reliability in diverse
such transparency is vital for creating trust in AI systems, conditions. The goal remains to advance the model’s practical
specifically in critical applications such as wildfire detection application in real-time wildfire detection and prevention.
that carry many issues for communities and the environment.
The significance of human oversight in AI deployment
IX. FUTURE WORK
is emphasized. Accuracy of AI and fast processing of data
is undoubtedly superior; nevertheless, humans must involve For future works, the actions of the current study would be
with AI making the final decision-making. This approach extended to the wildfire detection model improvement to the
ensures that AI serves as a support tool, not replacing critical point when the model makes a right discrimination between
human decision-making processes. Building a rapport with a fire and similar natural phenomena. This can be accomplished
broad range of stakeholders, including emergency managers, through the use of more data, such as different weather
scientists, and neighborhood harms are inevitable factors variables and temperature changes, to teach the model to
at the AI design and refinement stages. Their insights and distinguish a fire from uneventful things like clouds or sunsets
feedback ensure the AI system is effective, equitable, and thereby ascending the precision of the model. The primary
trustworthy. intent is to design a tool that can precisely detect forest fires
The next step is the consideration of these aspects, which with immediate reaction, therefore addressing the issue early,
main purpose is ensuring that AI will be utilized ethically and and minimizing harm. Such remediation aims to employ the
beneficial in wildfire management. Also the goal is to develop most advanced technologies for wildfire detection, response,
such system, which don’t introduce new challenges. and effective community protection.

VIII. CONCLUSION ACKNOWLEDGMENT

In summarizing the research, we begin with the challenge The authors extend their gratitude to OpenAI’s ChatGPT-4
of accurately detecting wildfires, a problem compounded by for its assistance in refining the sentence structures and
the similar visual characteristics of fire, smoke, and other organizing the article. This work was supported in part
natural elements. The approach of this research relies on by Basic Science Research Program through the National
using transfer learning. It makes use of the MobileNetV3 Research Foundation of Korea (NRF) funded by the Ministry
model, which is carefully adjusted and fine-tuned to improve of Education (NRF-2021R1A6A1A03039493) and in part by
accuracy in identifying different types of images. The the 2024 Yeungnam University Research Grant.

52388 VOLUME 12, 2024

S. F. Rubab et al.: FDX: Decoding Wildfire Detection With Transparency and XAI Insights

REFERENCES [23] S. Monaco, A. Pasini, D. Apiletti, L. Colomba, P. Garza, and E. Baralis,

‘‘Improving wildfire severity classification of deep learning U-Nets from
[1] M. Wang, D. Yu, W. He, P. Yue, and Z. Liang, ‘‘Domain-incremental satellite images,’’ in Proc. IEEE Int. Conf. Big Data (Big Data), Dec. 2020,
learning for fire detection in space-air-ground integrated observation pp. 5786–5788.
network,’’ Int. J. Appl. Earth Observ. Geoinf., vol. 118, Apr. 2023, [24] M. Nakip, C. Güzelís, and O. Yildiz, ‘‘Recurrent trend predictive
Art. no. 103279. neural network for multi-sensor fire detection,’’ IEEE Access, vol. 9,
[2] U. Ejaz, M. A. Khan, H. Park, and H.-C. Kim, ‘‘Firexplainer: An pp. 84204–84216, 2021.
interpretable approach for detection of wildfires,’’ in Proc. Korea Inf. Sci. [25] D. Rashkovetsky, F. Mauracher, M. Langer, and M. Schmitt, ‘‘Wildfire
Soc. Academic Presentation Papers, 2023, pp. 1109–1111. detection from multisensor satellite imagery using deep semantic segmen-
[3] A. Hussain, B. Barua, A. Osman, R. Abozariba, and A. T. Asyhari, tation,’’ IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 14,
‘‘Performance of MobileNetV3 transfer learning on handheld device-based pp. 7001–7016, 2021.
real-time tree species identification,’’ in Proc. 26th Int. Conf. Autom. [26] A. Khan and B. Hassan, ‘‘Dataset for forest fire detection,’’ Mende-
Comput. (ICAC), Sep. 2021, pp. 1–6. ley Data, vol. 1, p. 2020, Aug. 2020. [Online]. Available: https://
[4] S. Qian, C. Ning, and Y. Hu, ‘‘MobileNetV3 for image classification,’’ data.mendeley.com/datasets/gjmr63rz2r/1, doi: 10.17632/gjmr63rz2r.1.
in Proc. IEEE 2nd Int. Conf. Big Data, Artif. Intell. Internet Things Eng. [27] A. Raza, J. Uddin, A. Almuhaimeed, S. Akbar, Q. Zou, and A. Ahmad,
(ICBAIE), Mar. 2021, pp. 490–497. ‘‘AIPs-SnTCN: Predicting anti-inflammatory peptides using fastText and
[5] H. T. T. Nguyen, H. Q. Cao, K. V. T. Nguyen, and N. D. K. Pham, transformer encoder-based hybrid word embedding with self-normalized
‘‘Evaluation of explainable artificial intelligence: Shap, lime, and cam,’’ temporal convolutional networks,’’ J. Chem. Inf. Model., vol. 63, no. 21,
in Proc. FPT AI Conf., 2021, pp. 1–6. pp. 6537–6554, Nov. 2023.
[6] R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and [28] A. Ahmad, S. Akbar, M. Tahir, M. Hayat, and F. Ali, ‘‘IAFPs-EnC-GA:
D. Batra, ‘‘Grad-CAM: Visual explanations from deep networks via Identifying antifungal peptides using sequential and evolutionary descrip-
gradient-based localization,’’ in Proc. IEEE Int. Conf. Comput. Vis. tors based multi-information fusion and ensemble learning approach,’’
(ICCV), Oct. 2017, pp. 618–626. Chemometric Intell. Lab. Syst., vol. 222, Mar. 2022, Art. no. 104516.
[7] F. Zhuang, Z. Qi, K. Duan, D. Xi, Y. Zhu, H. Zhu, H. Xiong, and Q. He, [29] S. Akbar, M. Hayat, M. Tahir, S. Khan, and F. K. Alarfaj, ‘‘CACP-
‘‘A comprehensive survey on transfer learning,’’ Proc. IEEE, vol. 109, DeepGram: Classification of anticancer peptides via deep neural network
no. 1, pp. 43–76, Jan. 2021. and skip-gram-based word embedding model,’’ Artif. Intell. Med., vol. 131,
[8] R. Cilli, M. Elia, M. D’Este, V. Giannico, N. Amoroso, A. Lombardi, Sep. 2022, Art. no. 102349.
E. Pantaleo, A. Monaco, G. Sanesi, S. Tangaro, R. Bellotti, and
R. Lafortezza, ‘‘Explainable artificial intelligence (XAI) detects wildfire
occurrence in the Mediterranean countries of southern Europe,’’ Sci. Rep., SYEDA FIZA RUBAB received the M.S. degree
vol. 12, no. 1, p. 16349, Sep. 2022. in computer science from the Department of
[9] V. E. Sathishkumar, J. Cho, M. Subramanian, and O. S. Naren, ‘‘Forest
Artificial Intelligence, The Islamia University of
fire and smoke detection using deep learning-based learning without
forgetting,’’ Fire Ecology, vol. 19, no. 1, pp. 1–17, Feb. 2023.
Bahawalpur, Pakistan, in 2022. She is currently
[10] A. Akagic and E. Buza, ‘‘LW-FIRE: A lightweight wildfire image pursuing the Ph.D. degree with the Department
classification with a deep convolutional neural network,’’ Appl. Sci., of Information and Communication Engineering,
vol. 12, no. 5, p. 2646, Mar. 2022. College of Engineering, Yeungnam University,
[11] M. Moradi, K. Yan, D. Colwell, M. Samwald, and R. Asgari, ‘‘Model- Gyeongsan, South Korea. Her work focuses on
agnostic explainable artificial intelligence for object detection in image enhancing the understanding and application of
data,’’ 2023, arXiv:2303.17249. AI technologies. As an aspiring academic, she is
[12] J. Gerlings, A. Shollo, and I. Constantiou, ‘‘Reviewing the need for dedicated to contributing to the evolving landscape of artificial intelligence.
explainable artificial intelligence (xAI),’’ 2020, arXiv:2012.01007. Her research interests include data mining, machine learning, artificial
[13] T. Turay and T. Vladimirova, ‘‘Toward performing image classification and
intelligence, explainable AI, and image processing. With a commitment to
object detection with convolutional neural networks in autonomous driving
advancing these fields, she has been actively involved in various research
systems: A survey,’’ IEEE Access, vol. 10, pp. 14076–14119, 2022.
[14] O. Davydko, V. Pavlov, and L. Longo, ‘‘Selecting textural characteristics projects and academic endeavors. She was a recipient of a prestigious
of chest X-rays for pneumonia lesions classification with the integrated scholarship for the master’s and Ph.D. studies.
gradients xai attribution method,’’ in Proc. World Conf. Explainable Artif.
Intell. Cham, Switzerland: Springer, 2023, pp. 671–687. ARSLAN ABDUL GHAFFAR received the M.S.
[15] K. Avazov, A. E. Hyun, A. A. Sami S, A. Khaitov, A. B. Abdusalomov,
degree in computer science from the Department
and Y. I. Cho, ‘‘Forest fire detection and notification method based on AI
of Artificial Intelligence, The Islamia Univer-
and IoT approaches,’’ Future Internet, vol. 15, no. 2, p. 61, Jan. 2023.
[16] C. Yang, P. Zhou, and J. Qi, ‘‘Integrating visual foundation models for sity of Bahawalpur, in 2022. He is currently
enhanced robot manipulation and motion planning: A layered approach,’’ pursuing the Ph.D. degree with the Department
2023, arXiv:2309.11244. of Information and Communication Engineering,
[17] T. C. Phan and T. T. Nguyen, ‘‘Remote sensing meets deep Yeungnam University, South Korea. In addition,
learning: Exploiting spatio-temporal-spectral satellite images for he is actively engaged as a Researcher in this field,
early wildfire detection,’’ Tech. Rep., 2019. [Online]. Available: specializing in data mining, machine learning,
https://infoscience.epfl.ch/record/270339 artificial intelligence, explainable AI, and image
[18] Y. Kang, E. Jang, J. Im, and C. Kwon, ‘‘A deep learning model using processing.
geostationary satellite data for forest fire detection with reduced detection
latency,’’ GISci. Remote Sens., vol. 59, no. 1, pp. 2019–2035, Dec. 2022.
[19] S. Mohnish, K. Akshay, P. Pavithra, and S. Ezhilarasi, ‘‘Deep learning GYU SANG CHOI (Member, IEEE) received
based forest fire detection and alert system,’’ in Proc. Int. Conf. Commun., the Ph.D. degree from the Department of Com-
Comput. Internet Things (IC3IoT), Mar. 2022, pp. 1–5. puter Science and Engineering, The Pennsylva-
[20] M. Prakash, S. Neelakandan, M. Tamilselvi, S. Velmurugan, S. B. Priya, nia State University, University Park, PA, USA,
and E. O. Martinson, ‘‘Deep learning-based wildfire image detection in 2005. From 2006 to 2009, he was a Research
and classification systems for controlling biomass,’’ Int. J. Intell. Syst., Staff Member of the Samsung Advanced Insti-
vol. 2023, pp. 1–18, Sep. 2023.
tute of Technology (SAIT), Samsung Electronics.
[21] W. Lee, S. Kim, Y.-T. Lee, H.-W. Lee, and M. Choi, ‘‘Deep neural networks
for wild fire detection with unmanned aerial vehicle,’’ in Proc. IEEE Int. Since 2009, he has been a Faculty Member with
Conf. Consum. Electron. (ICCE), Jan. 2017, pp. 252–253. the School of Software Convergence, Yeungnam
[22] W. Thomson, N. Bhowmik, and T. P. Breckon, ‘‘Efficient and compact University, South Korea. His research interests
convolutional neural network architectures for non-temporal real-time fire include data mining, natural language processing, and reinforcement
detection,’’ in Proc. 19th IEEE Int. Conf. Mach. Learn. Appl. (ICMLA), learning.
Dec. 2020, pp. 136–141.

VOLUME 12, 2024 52389