Received 26 December 2023, accepted 8 January 2024, date of publication 12 January 2024, date of current version 26 January 2024.

Digital Object Identifier 10.1109/ACCESS.2024.3353688

Underwater Target Detection Using Deep

Learning: Methodologies, Challenges,
Applications, and Future Evolution
ANWAR KHAN 1 , MOSTAFA M. FOUDA 2 , (Senior Member, IEEE),
DINH-THUAN DO 3 , (Senior Member, IEEE), ABDULAZIZ ALMALEH4 ,
ABDULLAH M. ALQAHTANI 5 , AND ATIQ UR RAHMAN 6
1 Department of Electronics, University of Peshawar, Peshawar 25120, Pakistan
2 Department of Electrical and Computer Engineering, Idaho State University, Pocatello, ID 83209, USA
3 School of Engineering, University of Mount Union, Alliance, OH 44601, USA
4 College of Computer Science, Information Systems Department, King Khalid University, Abha 61421, Saudi Arabia
5 School of Computer Science and Information Technology, Jazan University, Jazan 45142, Saudi Arabia
6 Faculty of Computer Information Science, Higher Colleges of Technology, Ras Al Khaimah Campus, Abu Dhabi, United Arab Emirates

Corresponding author: Anwar Khan ([email protected])

ABSTRACT This paper provides a study of the latest target (object) detection algorithms for underwater
wireless sensor networks (UWSNs). To ensure selection of the latest and state-of-the-art algorithms, only
algorithms developed in the last seven years are taken into account that are not entirely addressed by the
existing surveys. These algorithms are classified based on their architecture and methodologies of operation
and their applications are described that are helpful in their selection in a diverse set of applications. The
merits and demerits of the algorithms are also addressed that are helpful to improve their performance in
future investigation. Moreover, a comparative analysis of the described algorithms is also given that further
provides an insight to their selection in various applications and future enhancement. A depiction of the
addressed algorithms in various applications based on publication count over the latest decade (2023-2013)
is also given using the IEEE database that is helpful in knowing their future application trend. Finally, the
challenges associated with the underwater target detection are highlighted and the future research paradigms
are identified. The conducted study is helpful in providing a thorough analysis of the underwater target
detection algorithms, their feasibility in various applications with future challenges and defined strategies
for further investigation.

INDEX TERMS Underwater target detection, deep learning, underwater object detection, YOLO,
convolutional neural networks, ConVNNs.

I. INTRODUCTION detection of crashed ships and boats [8] and target/object

Underwater wireless sensor networks (UWSNs) is one of detection [9], to mention a few. Underwater target/object
the latest realms of research that aims to explore the detection processes signals (image, audio, video, acoustic
underwater environment for a number of applications. These vibrations, radio/optical radiations) and extracts the infor-
applications include Tsunami detection and prediction [1], mation content that provides an ultimate insight to the
military surveillance [2], underwater navigation [3], secure underwater environment. This information is useful in a
communications [4], oil detection [5], fault monitoring number of applications such as object detection and tracking
in underwater cables [6], water quality monitoring [7], by underwater robots [10], [11], objects detections by
radars for military and civilian purposes [12], surveillance
The associate editor coordinating the review of this manuscript and systems [13], [14], [15], [16], ship tracking [17], [18], mine
approving it for publication was Xuebo Zhang . detection [19], [20], [21], water quality and environmental

2024 The Authors. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
12618 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ VOLUME 12, 2024
A. Khan et al.: Underwater Target Detection Using Deep Learning

features selection and classification of signals/objects are

performed in an automated fashion by the deep learning
algorithms.
There exists a number of surveys in literature related
to underwater target detection and identification [30], [31],
[32]. However, the addressed algorithms in these surveys
are not state-of-the-art to address the requirements of the
latest applications. Besides, they lack a thorough and in-depth
insight and comparative analysis of the addressed algorithms.
Keeping in view the importance and scope of the underwater
target detection, challenges, potential applications and its
direct and in-direct effects on the planet earth, this paper
studies the latest underwater target detection algorithms
designed in the last seven years so that only the most
recent algorithms are taken into account. The algorithms are
classified into various categories based on their architecture.
These categories include algorithms based on you only look
once (YOLO) architecture, convolutional neural networks
(ConVNNs) and their various types for varied applications
FIGURE 1. Applications of underwater object detection in various fields.
and hybrid algorithms that combine various techniques
to construct a single target detection architecture. Such a
classification makes algorithms selection convenient for the
appropriate applications. The algorithms are also described in
impacts [22], [23], [24], and toxins in water [25], [26], terms of their operational strategies, merits and demerits that
to mention a few. Figure 1 summarizes the various appli- helps in knowing their working mechanisms and improving
cations of underwater object detection ranging from noise their demerits in future enhancement. In addition, the
detection in underwater channel to waste products, precious operation of the architecture of each class of algorithm
materials, mines, marine animals, radar signals, ships and is described with suitable figures that helps in not only
their paths and position finding, to mention a few. There are a understanding of these algorithms but provides the areas
number of challenges inherently carried by the underwater where the performance of these algorithms can excel.
environment [27]. For instance, the electric and magnetic A comparative analysis of the classified algorithms is
fields of radio waves interact with water and other particles also provided that helps in selection of the specific class
and transfer their energy to them. Due to this, radio energy of algorithms for various applications as well as provides
is lost and, consequently, radio waves are not generally a path for their future improvement. For instance, a bar
preferred for underwater communications instead acoustic chart is provided that shows the calculated accuracy in target
waves are used. However, the speed of acoustic waves is detection of the compared algorithms. This helps not only in
almost five times slower than the speed of radio waves in the use of these algorithms in various applications depending
water due to which underwater communications bear long upon the accuracy requirement but also provides clues for
delay. In addition, the acoustic spectrum is limited (almost the further investigation towards enhanced accuracy. The
to 100 kHz) that further narrows the available bandwidth described mathematical models of the algorithms further
for underwater communications. Consequently, underwater provide an insight to their work, operation and object
communications have low data rates. Besides, underwater detection strategies. Moreover, the use of the classified
nodes have a limited battery lifetime. This restricts the life algorithms in diverse object/target detection applications in
span of UWSNs and demands for smart, efficient, reliable and the latest decade (2023-2013) is graphically depicted that
optimized operation and working strategies. provides an idea of the latest trends of these classes of
The use of deep learning algorithms for underwater algorithms in terms of applications to real world problems.
object detection has recently captured the attention of Finally, the challenges associated with the underwater target
researchers due to a number of advantages over the traditional detection are revealed and future research directions are
detection techniques in terms of accuracy of prediction, specified.
speed, generalization and automatic processing of tasks [28], In summary, this paper provides a study of the underwater
[29]. Figure 2 shows the the basic concept of object detection target detection algorithms developed in the last seven
by deep learning and traditional algorithms. Both types of years by addressing their architectures, operational strategies,
algorithms perform signal operations on the input signals merits, demerits, comparative analysis, target detection
(such as de-noising, filtering, image enhancement, to mention accuracy and applications in various fields in the latest decade
a few). However, the operations of features extractions, (2023-2013). Moreover, the challenges with underwater

VOLUME 12, 2024 12619

 A. Khan et al.: Underwater Target Detection Using Deep Learning

FIGURE 2. Object detection by (a): traditional algorithms and (b): deep learning algorithms. The latter perform features
extraction, selection and classification processes automatically by the deep networks rather than by varied techniques in
the former.

of YOLO, its basic architecture and various versions and

their comparison as well as object detection algorithms based
on YOLO. Section III discusses algorithms using ConVNNs
for underwater target detection, their architecture, operation
and sub-classification into various applications. Section IV
focuses on discussion of hybrid algorithms for underwater
target detection. A comparative analysis of the classified
algorithms is performed in Section V. Finally, Section VI
discusses the challenges in underwater target detection while
Section VII concludes the paper with directions for future
investigation.

II. UNDERWATER TARGET DETECTION USING YOLO

A. THE BASIC ARCHITECTURE OF YOLO SERIES
YOLO considers real time object detection and was first
designed by [33]. It divides an image into an S × S grid with
bounding box regression applied on each cell of the grid with
a confidence score C that measures the probability P(Obj) of
existence of an object in each box and is defined as:
truth
C = P(Obj) × IoUpredicted , (1)

where IoU is the intersection over union operation having

values between 0 and 1 with the latter being the ideal value.
Union represents the total area of the predicted bounding
box and the ground truth while the intersection signifies
the overlapping area of the predicted bounding box and the
ground truth. The conditional class probability of each cell of
the grid given that it has an object is denoted by P(Classi |
Obj) and is mathematically defined as:
truth
P(Classi | Obj) × P(Obj) × IoUpredicted
FIGURE 3. Organization and flow of the manuscript. truth
= P(Classi ) × IoUpredicted . (2)

To ensure accurate object detection, the bounding box and

target detection are highlighted and the possible solution
center of each prediction is corrected by a loss function given
strategies are described for future investigation.
by:
Figure 3 shows the organization and flow of the
2
manuscript. The introduction discusses the importance of s X
A
1Obj
X
underwater object detection, its applications and the contri- Loss = λCoord 2
ij [(bxi −bx̂i )] + [(byi − bŷi )]
2

butions of this paper. Sections II deals with the description i=0 j=0

12620 VOLUME 12, 2024

A. Khan et al.: Underwater Target Detection Using Deep Learning

FIGURE 4. Basic architecture of the YOLO series.

represent the dimensions of the bounding box, the parameters

λCoord and λNoobj are used to give more weight to boxes
with objects and less weight to boxes with no objects and
p(c) represents classification prediction. The parameter 1ij
Obj

has a unity value in case the jth bounding box in the ith cell
predicts an object, otherwise its value is zero. Similarly, if the
predicted object is in the ith cell then 1i is unity else it
Obj

is zero. Figure 4 shows the basic architecture of the YOLO

series. The input images are processed by the ConvNNs for
features extraction. The fully connected layer reduces the
dimensions of the extracted features and obtains a features
map and classifies it. The bounding box regression assigns
attributes to the classified features. Figure 5 shows flow chart
of the basic YOLO architecture in which an input image is
divided into an S × S grid with bounding box regression to
identify the objects. The intersection over union operation is
then applied to find objects in multiple boxes followed by
identifying boxes with high detection content.
In the lines to follow, the algorithms using the YOLO
series for objects detection are considered. Table 1 shows
a comparison among the various versions of the YOLO
architecture with enhancements and added features as the
architecture evolved from the the initial to the latest version.
It shows that the evolution of YOLO started with real-time
FIGURE 5. Flow chart of the basic YOLO architecture. An input image is object detection and with the development and progression,
divided into a grid of S × S cells with each cell processed by bounding it continued to include more features. Each version puts
box regression to detect object. It is followed by the intersection over
union operation for selection of objects in multiple boxes and boxes with emphasis on the overall better, more efficient, reliable and
high detection threshold are identified. accurate detection of objects that the earlier version(s).

B. OBJECT DETECTION USING YOLOV8 AND V7

2
s X
A
A comparison of the various YOLO algorithms is made
1Obj
X
2 2
+λCoord ij [(bwi −bŵi )] +[(bhi − bĥi )] in [34] for synthetic and real world data and concludes
i=0 j=0 that YOLOv5 exhibits the best results on synthetic data
s2
A while YOLOv8 outperforms all the other versions on real
1Obj
XX
2
+ ij (Ci − Ĉi ) datasets. The authors in [35] conclude that YOLOv7 is the
i=0 j=0 best when compared with faster region-based ConVNNs
s X
A 2 (R-ConVNNs), single-shot detector (SSD) and Centernet for
1Noobj
X
+ λNoobj ij (Ci − Ĉi )2 land and aquatic small object detection. The authors in [36]
i=0 j=0 compare YOLOv7 with YOLOv5 series and find it better in
s2 A terms of object detection accuracy and performance in image
1Obj challenging conditions. A loss function is proposed in [37] for
X X
+ i (pi (c) − p̂i (c)2 , (3)
i=0 c∈Classes
YOLOv7 based on the concept of a bag of features to optimize
the error and enhance the accuracy and speed of marine
where A is the number of boxes assigned to each grid, bx and object detection and classification. The concept of supervised
by are the center co-ordinates of each prediction, bw and bh features learning is introduced in [38] based on efficient

VOLUME 12, 2024 12621

 A. Khan et al.: Underwater Target Detection Using Deep Learning

TABLE 1. A comparison of the various YOLO versions.

aggregation network to improve the scaling calculation on detection accuracy of 3-5 %. The authors in [47] make
objects. three changes to the YOLOv5s to improve its performance.
First, they use a multi-head self-attention technique having
C. OBJECT DETECTION USING YOLOv5 contextual information that replaces the convolutional block
The images are first processed by the deep wave net scheme for better features extraction. Second, a hybrid convolutional
in [40] that uses ConVNNs for enhancement and then feeds module is added for reduction in parameters number. Third,
the output to YOLOv5 for object detection. The authors a path aggregation network is used to collect features from the
in [41] first enhance the images using Gridmask method fol- shallow and deep layers. Sonar images are first pre-processed
lowed by adding intersection over union to the non-maximum in [48] to overcome the internal and external noise and then
suppression method to improve detection accuracy when the improved YOLO5 is used for enhanced accuracy of object
the detected objects overlap. The features pyramid network detection, especially in overlapped objects.
(FPN) in the main architecture is modified to detect small The concept in [49] embeds a camera with an autonomous
objects. Drones are used in [42] to utilize YOLOv5s to detect underwater vehicle (AUV) that captures underwater images
submerged objects in water with an effective accuracy and of the target in a swimming pool, which are then detected by
precision. The authors in [43] analyse the vocal behaviour the YOLOv5 and the information is extracted. The training
of mammals by processing their signals using YOLOv5. of the deep learning module is performed by the images
The obtained information, such as central frequency of the processed on the Google-Collab over the cloud and the output
signals, duration and bandwidth are effective in knowing is then processed by the on-board computer of the AUV that
the behaviour of these mammals. The network parameters consists of Raspberry-Pi4 having a coral USB accelerator.
of YOLOv5 are first reduced by using GhostBottleneck In [50], the realization of detection and tracking of the
in [44] and then important features weight is increased by target is performed by the deep sort algorithm. Due to poor
the addition of a convolutional block attention layer to the lightening conditions, the obtained images are also enhanced
final layer of the backbone architecture and the intersection leading to an overall 96% detection accuracy. The concept
over union feature is modified to enhance the accuracy of in [51] optimizes the performance of the original YOLOv5 by
the overlapped objects. The methodology in [45] uses the optimizing its main architecture (CSPDarknet). The features
coordinate attention mechanism and a bidirectional feature are selected and extracted using the cross stage partial
pyramid network to improve the target detection accuracy of (CSPNet), a convolutional layer that utilizes the contextual
YOLOv5 for ship target detection. A comparison of YOLOv5 block streaming (CBS) as its fundamental architecture for the
is made with YOLOv3 for seaweeds detection in [46] and it is recognition of useful information. The elementary layers are
found that the former is faster than the later but with a reduced changed by ConVNNs followed by swin transformer.

12622 VOLUME 12, 2024

A. Khan et al.: Underwater Target Detection Using Deep Learning

TABLE 2. Target Detection using YOLOv8 [34], YOLOv7 [35], [36], [37], [38] and YOLOv5 [39], [40], [41], [42], [43], [44], [45], [46], [47], [48], [49], [50], [51],
[52], [53], [54], [55], [56], [57], [58], [59], [60]. The symbol x represents the unspecified value.

The authors in [52] argue that underwater target tracking architecture of the YOLOv5 is modified in three ways.
requires precision and accuracy and for this purpose the Firstly, the bottleneck count is increased from one to three.

VOLUME 12, 2024 12623

 A. Khan et al.: Underwater Target Detection Using Deep Learning

Secondly, the bottleneck module is replaced by a module with a fusion mechanism that uses attention mechanism for
with coordinated attention so as to enhance the level of multiple features; that learns and obtains the features of a
attention in the image of the target object. Thirdly, a module is number of characteristics utilized for object detection. This
introduced to the backbone of the model for further enhancing approach provides a balance between speedy of detection
the attention level of the image and ignoring the unimportant and accuracy. The framework in [63] also uses YOLOv4 and
features of the image. The approach adopted in [53] first obtains encouraging detection rate. The technique in [64]
detects and then locates the position of underwater objects. studies the YOLOv3 and the deep-sort-multi-target tracking
It modifies the YOLOv5 model for detection time and algorithm for fish detection. The research considers the
accuracy efficiency. First it downsamples the target image and coordinates of a fish to track the path it follows.
extracts the features followed by the fusion to obtain a map of The authors in [65] use the YOLOv3 model that makes
features. Then the image features coordinates and the imaging use of marine pasture biological targets and uses open source
system coordinates determine the position of the target. The images to train the model. It is followed by the testing
method given in [54] first applies the image enhancement phase along with optimizing the tuning parameters of the
on the objects followed by the YOLOv5 with ghost modules learning process. The output of the learning then results
(instead of backbone features extraction) to detect objects in knowing the objects in the images, their locations and
with reduced parameters and computation. Moreover, the the classification. The image detection module is based on
selective kernel convolution is applied for features extraction the Pytorch framework and is trained by the open source
that has better results than the traditional convolution. Finally, SeaCLEF image database until the desired optimization level
the positive and negative sample imbalance is overcome by is achieved. The underwater images have usually low light
the focal loss optimization. The authors in [55] first apply the conditions and contrast and, therefore, they are treated based
various image enhancement algorithms on various detection on Fuzzy contrast and enhanced in [66] using a self-adaptive
techniques and find that the YOLOv5 is the best in terms of technique followed by the application of YOLOv3 for object
objection detection and differentiation from background in detection. The target detection is performed in [67] using the
underwater images. The small target objects are detected in YOLOv3 model in an underwater sonar image and its position
side-scan sonar images in [56] by the modified YOLOv5. The is also identified followed by the recurrent neural networks
re-clustering of the anchor frames of the target is performed in for tracking the path of a dynamic target. The YOLOv3
the first stage using K-means. It is followed by a new layer for model is modified in [68] for real time target detection
capturing shallow features and then an attention mechanism by adding the feature clear and pooling layers to achieve
for extracting deep features. Small samples overfitting is enhanced and effective extraction of features. Moreover, the
overcome by new connections. The authors in [57] use images are processed by the augmentation, enhancement and
the YOLOv5 model to detect two similar objects in an equalization processes for improved accuracy of recognition.
underwater sonar image. The knowledge obtained for training The idea in [69] uses YOLOv3 for the recognition and
images is used to identify similar images and the objects in detection of objects in side-scan sonar images. The image
them. A method for underwater plastic waste is proposed features are extracted using the various maximum bounding
in [58]. It modifies the YOLOv5n by reducing its backbone boxes of high credibility and Darknet53 is used as the
size and the problem of insufficient features is overcome by backbone network for extraction. The algorithm in [70]
modifying the feature pyramid, followed by inserting a loss reduces the detection scale of YOLOv3 by a single decrement
function in the bounding box regression loss of the model. and re-clusters the anchor boxes so as to make them
A comparison of the YOLOv5, YOLOv5-TR and YOLOX appropriate for the considered datasets during the training
is performed in [59] for real-time detection and localization process. This reduces computational complexity and still
of target (a harbor’s wall) and proves that the YOLOX has maintains a certain degree of accuracy in detection of garbage
the best detection rate of 91.3 %. The concept given in [60] in water by robots. The authors in [71] up-sample the
modifies the YOLOv5 by pre-training, clustering, fine-tuning down-sampling rate and add splicing and features fusion
the pre-training and adding features extraction for high-level techniques to YOLOv3 to enhance its performance for small
features for forward-looking sonar images. target detection.

D. TARGET DETECTION USING YOLOv4 AND YOLOv3 E. OBJECT DETECTION USING YOLO
The authors in [61] modify YOLOv4. They add a deep The authors in [72] apply the YOLOX algorithm to recognize
separable convolutional layer to the backbone of the network and detect underwater objects for forward looking SONAR.
with a feature that allows detection of small objects followed The algorithm first extracts the features from the images
by K-clustering of the bounding box of the dataset with followed by obtaining enhanced features using the FPN
improved size of the box according to the clustering. Also, and the detected images are then recognized. A mechanism
a spatial pyramid pooling module is added that increases given in [73] dynamically chooses feature layer channels,
the complexity but also enhances the accuracy. In the last termed as DC block and is combined with YOLOX to make
phase, multi-scale training of the model is performed for YOLOX-DC. A network establishment concept with defined
effective results. The concept in [62] combines YOLOv4 local points in underwater environment is given in [74]

12624 VOLUME 12, 2024

A. Khan et al.: Underwater Target Detection Using Deep Learning

TABLE 3. Target detection using YOLOv4 [61], [62], [63], YOLOv3 [64], [65], [66], [67], [68], [69], [70], [71] and YOLO [72], [73], [74], [75], [76]. The symbol x
represents the unspecified value.

FIGURE 6. Basic architecture of convolutional neural networks. Features are extracted by the
convolutional layer that are further reduced in dimensions by the pooling layer. The fully connected layer
makes a bigger features map by combining the collected features.

that uses the YOLO version for automatic target detection Figure 6 shows the basic architecture of a ConVNN for object
and eases the manual measurement in future trials. The detection. Features are extracted from an object of interest
YOLO algorithm is modified in [75] and transfer learning using the convolutional operation between an image (or any
is adopted to ease the complexity of training and target signal of interest) I of size m × n and a kernel or filter F of
detection. The concept of histogram equalization is used to size L × L as:
deal degradation of image quality. The similarity in structures XL X L
of frames is utilized for enhancing the frame detection rate. O(i, j) = F ∗ I = I (i + k − 1, j + 1 − l)F(k, l),
The authors in [76] use three different datasets to train YOLO k=1 l=1
for detection of fish. The results showed that the model did (4)
not detect fish in datasets that were not used in the training
where O(i, j) is the output of the convolution matrix at the ith
process that advocated the use of diverse datasets during
row and jth column considering a single channel convolution
training.
and the symbol ∗ represents the convolution operation. The
pooling layer removes unnecessary information content from
III. TARGET DETECTION USING ConVNNs the information it receives from the features extraction layer
This section describes the algorithms that make use of the and reduces its dimension that is further processed by the
ConVNNs or their variants for underwater target detection. fully connected layer that combines all the features in a

VOLUME 12, 2024 12625

 A. Khan et al.: Underwater Target Detection Using Deep Learning

external calibration parameters. Moreover, it uses the feature

pyramid ConVNNs for detection of the target in the image.
The various videos of the motion of the target are optimized
for the trajectory calculation as well. The method in [79]
identifies the jellyfish and its density. A camera takes the
real-time picture of jellyfish and processed by the ConVNNs.
The obtained image is enhanced and its edges and their
closure are detected and holes are filled in the gray-scale
and the binary image is obtained. To detect a lobster in the
image in [80], the initial data enhancement is performed
in the pre-processing stage by the generative adversarial
network and increment method followed by the use of the
separable ConVNNs that compress the fully connected layer
and make the model light for computation. The voice calls of
the underwater mammal species are processed for features
extraction by the fractional Fourier transform in [81] as
they behave like modulated pulses with a linear frequency.
The obtained features are then input to the ConVNNs for
voice recognition and, therefore, detect the corresponding
mammals. The authors in [82] use limited training data
for fish detection. It involves various convolutional layers
and residual blocks to detect and segment the target. The
weights of the features of interest are increased. The features
are effectively presented by the residual blocks after the
concatenation of the shallow and deep layers of the model.
The authors in [83] design a model that send signals towards
the moving fish and their reflections are collected and
analyzed using the ConVNNs to detect the moving target.
A trade-off existed between the detection bound and accuracy
FIGURE 7. Flow chart of the the basic ConVNNs. The features are for training the network with real and synthesized data.
extracted from an input signal (image, for instance) through convolution
operation that are further processed by the pooling layer for dimensions
A recurrent neural network model is also given for online
reduction and the fully connected layer to obtain a diverse features map processing with low accuracy. The combination of improved
and classify/detect objects. faster region-based ConVNNs and FPN is performed for
target detection in [84]. The accuracy and speed of detection
are improved by replacing the intersection over union by
single array to construct a bigger features map for information the distance intersection over union. The authors in [85]
classification and object detection. The architecture of the improve the detection accuracy and training speed of the
ConVNNs for object detection involves features extraction faster R-ConVNNs in detection of jellyfish in underwater
from an input image (signal) using convolution operation images. The obtained images are preprocessed to improve
that are further processed by the pooling layer to reduce their brightness and contrast followed by the integration of the
dimensions and fully connected layer to obtain a features restnet50 into the network for effective features extraction.
map and classify and detect objects. These operations are The training speed is improved by using the semi-precision
further elaborated in the flow chart of the ConVNNs depicted floating point method.
in Figure 7. The target detection using ConVNNs is further
classified as given in the lines to follow.
B. TARGET DETECTION FOR SAFETY PROVISION
A. UNDERWATER ANIMALS/MOVING OBJECTS The authors in [86] make use of the time and time-frequency
DETECTION spectra of underwater images and process them by ConVNNs
The scheme designed in [77] uses deep learning with to recognize these objects. In addition, the neural network
improved regression network to detect the target followed parameters such as the pooling rate, learning rate and batch
by prewitt feature enhancement to minimize the features size are studied for optimal performance. To detect divers
loss and their uneven distribution. The binocular vision is and underwater intruders, the authors in [87] first obtain a
used to determine the position of the target (fish swarm) background image of the surrounding using ConVNNs and
and their spatial information are displayed on a radar map. then the underwater moving objects are detected using the
The concept given in [78] studies the underwater camera difference of the current image and predicted image using
imaging with light refraction and the various internal and plan position indication. The ConVNNs are used in [88] to

12626 VOLUME 12, 2024

A. Khan et al.: Underwater Target Detection Using Deep Learning

process the images of the drowning objects and detect them that automatically removes the noise in the real data and is
for rescue robots. They also focus on reducing the cost and then used to train the model.
difficulties in the deployment of the existing rescue robots. The authors in [97] propose an algorithm that blind detects
A deep ConVNNs algorithm is proposed in [89] for the the acoustic signals in underwater communications. First
automatic detection and classification of underwater mines it pre-processes the noise using the generative adversarial
in sonar images captured by the synthetic aperture radar. network that mitigates the noise in the signal and then
The realistic images are synthetically generated to train the the ConVNNs are used to extract the features from the
model. signals and differentiate them from the noise. In addition,
a data transfer model is used to overcome the issue of
C. TARGET DETECTION FOR SONAR IMAGERY
insufficient underwater data for training the respective model.
An approach based on ConVNNs is adopted in [90] that The authors in [98] improve the detection accuracy of the
enhances the features of the target in a sonar image and single-shot multibox detector algorithm, which is based on
weakens the background so that the false detection is ConVNNs. It obtains the position and detail information of
overcome and missed detection is minimized. A ConVNN the object using channel-spatial attention mechanism for high
that has prior training knowledge of the features uses pixel value features to improve detection.
intensity as the features extraction metric to recognize The authors in [99] use the reflection signals of an active
objects/anomalies in the seabed images with small to no acoustic emitter to localize, detect and track moving under-
false detection is proposed in [91]. The authors in [92] water targets with convolutional denoising auto-encoder. The
first identify the target region in a forward-looking sonar concept of faster R-ConVNNs is used in [100] for object
echoscope image during the pre-processing stage using detection that involves the use of the swin transformer as
graph-based manifold ranking and then processed by the deep the backbone of the architecture, a path aggregation network
ConVNNs for the extraction of features and recognition of for fusing the deep and shallow features maps, online hard
the target. The method of transfer learning is used to cope mining and using an improved pooling layer to remove
with the requirement of the availability of sufficient data. quantization error and improve detection.
The optimization of the network is performed by the gradient
descend method. IV. TARGET DETECTION USING HYBRID ALGORITHMS
The authors in [101] combine YOLO, ConVNNs and SSD
for object detection with a comparative analysis to detect
D. TARGET DETECTION FOR MISCELLANEOUS even small objects. It is argued in [102] that the existing
APPLICATIONS underwater object detection algorithms involve frequent
The authors in [93] first apply the method of convolutional human-computer interaction, which is not feasible for the
downsampling to the features of underwater acoustic target automatic operation of the detecting devices. Therefore,
to be recognized, which reduces the cost of computational a long short-term memory (LSTM) deep learning model
processing. It is followed by learning about the local and based algorithm is utilized to extract and classify the features
global features using a varied set of conformer blocks. of the target noise by an underwater glider. It first obtains the
Finally, the splicing of the output of the various conformer data input samples including various noise frequencies and
blocks is carried out and the features of the speech are signals, normalizes them and then inputs them to the LSTM
extracted by the mechanism combining pooling layer and for the extraction and classification of features. The authors
the attention statistics. An algorithm is developed in [94] for in [103] explore the resonant behavior of the low frequency
the detection and modeling of underwater propeller noise sound waves when they are incident on unexploded ordnance.
in acoustic signals. The ConVNNs architecture is used to So two deep learning approaches are modeled to detect
detect and classify the noise for various signal-to-noise the unexplored ordnance in the sonar imagery of synthetic
ratio (SNR) levels. The authors in [95] argue that sonar aperture radar. These algorithms use sequence models to
images have limited availability of data and are not suitable correlate the spatial features in the resonant sound waves.
for deep learning. To overcome this problem, the striation A method of automatic detection of underwater objects
images are used that minimize the need for training data of is given in [104] that uses Fuzzy C-means and K-means
the ConVNNs, the optimization of fuzzy or discontinuous global clustering of the images to get many regions of
fringes is performed and the shared latent sparse scheme is interest followed by local segmentation using the pulse
used to represent the interference fringes. These features are coupled neural network to differentiate the boundary of
then correspondingly used to train the network. A method the target. Extraction of multiple features from the target
proposed in [96] combines the spatial and spectral features area is performed and are input to the nonlinear converter
of the target obtained through 3D ConVNNs that are further to increase the distance of the features and the Fisher
fine-tuned using the depth information, as these features discrimination is used to compute a classification threshold
vary with the water depth. The depth information is also and detect the target. A shallow neural network algorithm is
used to predict the accurate depth of the target. The network proposed in [105] that considers the temporal variations in the
synthetically generates a copy of the actual hyperspectral data amplitude and frequency of the target and clutter signals from

VOLUME 12, 2024 12627

 A. Khan et al.: Underwater Target Detection Using Deep Learning

TABLE 4. ConVNNs for detection of Animals/moving objects [77], [78], [79], [80], [81], [82], [83], [84], [85], safety [86], [87], [88], [89], sonar imagery [90],
[91], [92] and miscellaneous applications [93], [94], [95], [96], [97], [98], [99], [100].

pre-processed spectrographs. The authors in [106] estimate enhanced in [107] using the max-RGB and shades of grey
the monocular depth to restore the image affected by the techniques and then a correlation filter tracking method is
underwater channel properties. This helps later in the target combined with the R-ConVNNs to extract the regions of
detection based on the depth learning. First images are interest and detect objects.

12628 VOLUME 12, 2024

A. Khan et al.: Underwater Target Detection Using Deep Learning

TABLE 5. Target Detection using hybrid deep learning techniques [102], [103], [104], [105], [106].

TABLE 6. Comparative analysis of the classified categories of algorithms.

V. COMPARATIVE ANALYSIS OF THE TARGET DETECTION architecture provides a fast and real-time object detection but
ALGORITHMS struggles with the detection of small objects. The ConVNNs
Based on the description of the classified algorithms, Table 6 and 3D ConVNNs are effective in extraction of features but
shows a comparative analysis of the classified algorithms they have enhanced computational complexity and require
for underwater target detection. It shows that the YOLO intensive training data. The LSTM is capable of retaining

VOLUME 12, 2024 12629

 A. Khan et al.: Underwater Target Detection Using Deep Learning

FIGURE 8. The maximum computed detection percentage of the compared

techniques.

FIGURE 9. The (normalized) publication count for object (target) detection as searched in the IEEE Explore database for a diverse set of applications for
the latest decade (2023-2013).

short and long-term information of the target detection but of the percentage target detection by the compared classes
it suffers from the vanishing gradient problem, where the of algorithms. The plotted values indicate the recorded
weight of the information gradually decreases with the length percentage detection as mentioned by the researchers. The
of information sequences. Figure 8 shows a comparison algorithms of the researchers that have not measured and

12630 VOLUME 12, 2024

A. Khan et al.: Underwater Target Detection Using Deep Learning

mentioned the detection percentage are not included. The plot architecture and their operational mechanisms, merits and
shows that the maximum computed percentage detection of demerits were identified for further future enhancements.
98.65 is achieved by the ConVNNs due to the use of multiple A comparative analysis is also performed for further pro-
features processing layers with enhanced complexity. The viding an insight to the understanding of the classified
second and third highest percentage detection values are algorithms. The applications of the described algorithms for
also achieved by the ConVNNs, which are 98.29 and 97.80, the recent decade (2023-2013) in object detection is graphi-
respectively. cally depicted that provided their scope and importance. The
Figure 9 shows the normalized number of articles (for bet- classified algorithms and techniques are effective and useful
ter presentation) published in the latest decade (2023-2013) in underwater object detection in a diverse set of applications
involving various applications in object detection using the such as underwater military and civil radars, precious
IEEEE Explore database. These applications include, for materials, debris, mines and submarine detection, tracking
instance, defects in bicycles, outdoor smoking, wood pith, the path of autonomous underwater vehicles and position
X-rays, traffic data, security warning, music instruments, calculation of mission robots. They are also beneficial to
railways, microalgae and remote sensing, to mention a few. analyze underwater objects for military and civilian purposes,
The ConVNNs and YOLO techniques exhibit a major boom detect noise over underwater channel for communications,
in application to object detection tasks. water quality monitoring and ensuring underwater explo-
ration. The conducted study is useful to provide a thorough
VI. CHALLENGES IN UNDERWATER TARGET DETECTION analysis of underwater target detection algorithms and their
Keeping in view the challenging underwater conditions, there mutual comparison in terms of methodologies, structure and
are a number of challenges associated with underwater target operation to highlight their effectiveness and robustness. The
detection, as described in the lines to follow [108], [109]. challenges in these algorithms are useful to provide future
• The underwater environment has poor light conditions enhancement paths. The discussion of the merits and demerits
and there is complete darkness beyond a certain depth. of the algorithms provides an insight to differentiate them
This challenges the target detection and identification, for utilization in specific underwater applications. It also
especially when the target is in motion. Due to these provides clues to outline strategies in the development of
conditions, the underwater target resembles with its more robust, sophisticated, efficient and effective algorithms
background that challenges the detection probability. than the existing algorithms.
• The size of the underwater target is generally smaller The following strategies are effective in future investi-
than its surrounding that makes the detection process a gation to cope with the challenges in underwater target
challenging task. detection [30].
• Underwater target are subjected to noise from various • Requirement of a Diverse and Balanced Dataset.
sources; such as thermal, shipping, wave and ambient Deep learning models need to be trained to acquire
environment, due to which the images of the targets are the information patterns hidden in the input datasets so
blurry and distorted. The spectrum of these noise types as to predict the objects in the testing phase. Future
varies with frequency. Therefore, specific frequencies research investigation needs to have thorough, diverse,
filters design is required to cope with it. As a result, data robust, balanced and comprehensive datasets owing to
pre-processing and enhancement become necessary for the diversified zones and regions of the sea environment
underwater target detection. so that object detection is performed at a diverse level.
• When underwater targets are in motion, sophisticated • Deep Transformer for Efficient Processing. The use
and fast response circuitry design is required to capture of deep learning techniques such as transformer [110]
the attributes of the target well before they change with would reduce the computational delay due to its parallel
motion. processing capability unlike the use of the already
• The presence of various occlusions; such as full or prevailing algorithms that struggle with computational
partial covering of the objects by marine life, debris efficiency.
and accumulated waste products, to mention a few, • Transfer Learning for Ease of Training and
challenges the target detection. Detection/Prediction. The transfer learning techniques
• The differentiation and separation of overlapped objects have the capability of training the deep networks on
is specifically challenging as the bounding box approach datasets and then using the information obtaining during
usually counts all the objects in a box as a single object. training to detect/predict similar and related objects
This becomes further challenging when the overlapped without training the deep network again. This avoids the
objects are small or tiny. need for computational rigor and, consequently, results
in time-efficient processing.
VII. CONCLUSION AND FUTURE WORK • Development of Hybrid Detection Techniques. Com-
A survey of the latest and state-of-the-art underwater target bining the advantages and merits of various data
detection algorithms is addressed. These algorithms were processing and object detection techniques could results
classified into various categories depending upon their in a bulk performance enhancement. For instance,

VOLUME 12, 2024 12631

 A. Khan et al.: Underwater Target Detection Using Deep Learning

the techniques for features extraction of convolutional [13] J. Yang, Y. Yao, and D. Yang, ‘‘Particle filter based on Harris hawks
neural networks could be used for multiple bounding optimization algorithm for underwater visual tracking,’’ J. Mar. Sci. Eng.,
vol. 11, no. 7, p. 1456, Jul. 2023.
box regressions within a single box to detect the tiny and [14] M. Vadivel, S. P. Vimal, V. G. Sivakumar, V. V. Baskar, and M. Selvi,
overlapped objects that are challenging to detect with the ‘‘Internet based defence surveillance robot to prevent intruder activities
traditional methods. and auto combat system using SONAR technology,’’ in Proc. Int. Conf.
Innov. Data Commun. Technol. Appl. (ICIDCA), Uttarakhand, India,
• Multiple Signal Processing Techniques. Statistical Mar. 2023, pp. 1055–1059.
signal processing techniques such as entropy, Fourier [15] N. Morozs, P. D. Mitchell, and Y. Zakharov, ‘‘Target detection using
transform and Wavelet transform; to mention a few, underwater acoustic networking,’’ in Proc. OCEANS, Limerick, Ireland,
Jun. 2023, pp. 1–5.
could be utilized to extract only the informative parts
[16] T. Rong, C. Wang, Q. Luo, and Y. Li, ‘‘Underwater shipwreck target
of the signals that could significantly reduce the detection based on audio-visual heterogeneous data,’’ in Proc. OCEANS,
computational cost. Limerick, Ireland, Jun. 2023, pp. 1–6.
• Development of Sophisticated Cameras and Data [17] H. Park, S.-H. Ham, T. Kim, and D. An, ‘‘Object recognition and tracking
in moving videos for maritime autonomous surface ships,’’ J. Mar. Sci.
Processing Circuitries. With the involvement of big Eng., vol. 10, no. 7, p. 841, Jun. 2022.
data and heavy training data requirement by deep learn- [18] H. Li, X. Cui, and S. Chen, ‘‘PolSAR ship detection with optimal
ing techniques, future investigation needs to consider polarimetric rotation domain features and SVM,’’ Remote Sens., vol. 13,
fast, efficient, modular, reliable and adaptive circuitries no. 19, p. 3932, Sep. 2021, doi: 10.3390/rs13193932.
[19] D. Munteanu, D. Moina, C. G. Zamfir, Ş. M. Petrea, D. S. Cristea, and
to detect the changes in the objects orientation, position N. Munteanu, ‘‘Sea mine detection framework using YOLO, SSD and
and status and include them in the actual status of the EfficientDet deep learning models,’’ Sensors, vol. 22, no. 23, p. 9536,
objects before the changes happened. This will lead to Dec. 2022.
[20] S. Hożyń, ‘‘A review of underwater mine detection and classification in
more reliable, accurate and trustworthy detection. sonar imagery,’’ Electronics, vol. 10, no. 23, p. 2943, Nov. 2021.
[21] V. Padmaja, V. Rajendran, and P. Vijayalakshmi, ‘‘Retraction note to:
REFERENCES Study on metal mine detection from underwater sonar images using data
mining and machine learning techniques,’’ J. Ambient Intell. Humanized
[1] L. L. C. Benisto, R. Sukumaran, and M. Saravanan, ‘‘Architecture,
Comput., vol. 14, no. S1, p. 377, Apr. 2023.
localization techniques, routing protocols and challenges for UWNS,’’ in
[22] D. Zhao, B. Tan, H. Zhang, and R. Deng, ‘‘Monitoring marine oil spills
Proc. Int. Conf. Data Sci. Netw. Secur. (ICDSNS), Tiptur, India, Jul. 2023,
in hyperspectral and multispectral remote sensing data by the spectral
pp. 1–7.
gene extraction (SGE) method,’’ Sustainability, vol. 14, no. 20, p. 13696,
[2] N. Nomikos, P. K. Gkonis, P. S. Bithas, and P. Trakadas, ‘‘A survey
Oct. 2022.
on UAV-aided maritime communications: Deployment considerations,
applications, and future challenges,’’ IEEE Open J. Commun. Soc., vol. 4, [23] C. Şerban, C. Maftei, and G. Dobrică, ‘‘Surface water change detection
pp. 56–78, 2023. via water indices and predictive modeling using remote sensing imagery:
A case study of Nuntasi-Tuzla Lake, Romania,’’ Water, vol. 14, no. 4,
[3] E. Coccolo, C. Delea, F. Steinmetz, R. Francescon, A. Signori, C. N. Au,
p. 556, Feb. 2022.
F. Campagnaro, V. Schneider, F. Favaro, J. Oeffner, B.-C. Renner, and
M. Zorzi, ‘‘System architecture and communication infrastructure for [24] V. Assumma, M. Bottero, C. Caprioli, G. Datola, and G. Mondini,
the RoboVaaS project,’’ IEEE J. Ocean. Eng., vol. 48, no. 3, pp. 1–24, ‘‘Evaluation of ecosystem services in mining basins: An application in the
Jul. 2023. Piedmont region (Italy),’’ Sustainability, vol. 14, no. 2, p. 872, Jan. 2022.
[4] Z. A. Zukarnain, O. A. Amodu, C. Wenting, and U. A. Bukar, ‘‘A survey [25] H. I. Farhat, A. Gad, A. Saleh, and S. M. A. El Bakey, ‘‘Risks assessment
of Sybil attack countermeasures in underwater sensor and acoustic of potentially toxic elements’ contamination in the Egyptian red sea
networks,’’ IEEE Access, vol. 11, pp. 64518–64543, 2023. surficial sediments,’’ Land, vol. 11, no. 9, p. 1560, Sep. 2022.
[5] M. Chaudhary, N. Goyal, A. Benslimane, L. K. Awasthi, A. Alwadain, [26] R. Ou, L. Cai, J. Qiu, H. Huang, D. Ou, W. Li, F. Lin, X. He, L. Wang,
and A. Singh, ‘‘Underwater wireless sensor networks: Enabling technolo- and R. Wu, ‘‘Simulation experiment of environmental impact of deep-sea
gies for node deployment and data collection challenges,’’ IEEE Internet mining: Response of phytoplankton community to polymetallic nodules
Things J., vol. 10, no. 4, pp. 3500–3524, Feb. 2023. and sediment enrichment in surface water,’’ Toxics, vol. 10, no. 10, p. 610,
[6] K. S. Kumar and V. Anand, ‘‘Underwater acoustic sensor network: Oct. 2022.
Architecture, challenges, future possibilities in the perspective of IoT,’’ in [27] R. Astya and C. K. Jha, ‘‘UWSN survey: Challenges and application
Proc. 2nd Int. Conf. Electr., Electron., Inf. Commun. Technol. (ICEEICT), issues,’’ in Proc. 4th Int. Conf. Adv. Comput., Commun. Control Netw.
Trichirappalli, India, Apr./Jun. 2023, pp. 1–7. (ICACN), Greater Noida, India, Dec. 2022, pp. 1855–1858.
[7] C. Lin, G. Han, J. Jiang, C. Li, S. B. H. Shah, and Q. Liu, ‘‘Underwater [28] J. R. Prasad, A. Parikh, and H. K. Prasanth, ‘‘Exploration of deep learning
pollution tracking based on software-defined multi-tier edge computing in based underwater image processing techniques,’’ in Proc. 10th Int. Conf.
6G-based underwater wireless networks,’’ IEEE J. Sel. Areas Commun., Comput. for Sustain. Global Develop. (INDIACom), New Delhi, India,
vol. 41, no. 2, pp. 491–503, Feb. 2023. Mar. 2023, pp. 1222–1225.
[8] A. Khan, M. Imran, A. Alharbi, E. M. Mohamed, and M. M. Fouda, [29] T. Yin, X. Jiang, Q. Cen, and Z. Zhang, ‘‘An efficient real-time marine
‘‘Energy harvesting in underwater acoustic wireless sensor networks: object detection algorithm based on convolutional attention,’’ in Proc.
Design, taxonomy, applications, challenges and future directions,’’ IEEE 8th Int. Conf. Intell. Comput. Signal Process. (ICSP), Xi’an, China,
Access, vol. 10, pp. 134606–134622, 2022. Apr. 2023, pp. 1496–1501.
[9] Y. Cheng, H. Wu, Z. Yang, and H. Wang, ‘‘Underwater target detection [30] S. Fayaz, S. A. Parah, and G. J. Qureshi, ‘‘Underwater object detection:
by measuring water-surface vibration with millimeter-wave radar,’’ Architectures and algorithms—A comprehensive review,’’ Multimedia
IEEE Antennas Wireless Propag. Lett., vol. 22, no. 9, pp. 2260–2264, Tools Appl., vol. 81, no. 15, pp. 20871–20916, Mar. 2022.
Sep. 2023. [31] R. A. Dakhil and A. R. H. Khayeat, ‘‘Review on deep learning techniques
[10] M.-F.-R. Lee and Y.-C. Chen, ‘‘Artificial intelligence based object for underwater object detection,’’ in Proc. 3rd Int. Conf. Data Sci. Mach.
detection and tracking for a small underwater robot,’’ Processes, vol. 11, Learn., Copenhagen, Denmark, Sep. 2022, pp. 49–63.
no. 2, p. 312, Jan. 2023. [32] S. Mittal, S. Srivastava, and J. P. Jayanth, ‘‘A survey of deep learning
[11] Y. Liu, Z. Liu, H. Yang, Z. Liu, and J. Liu, ‘‘Design and realization of techniques for underwater image classification,’’ IEEE Trans. Neural
a novel robotic manta ray for sea cucumber recognition, location, and Netw. Learn. Syst., vol. 34, no. 10, pp. 6968–6982, Oct. 2023.
approach,’’ Biomimetics, vol. 8, no. 4, p. 345, Aug. 2023. [33] P. Jiang, D. Ergu, F. Liu, Y. Cai, and B. Ma, ‘‘A review of YOLO
[12] Y. Zeng, S. Shen, and Z. Xu, ‘‘Water surface acoustic wave detection by a algorithm developments,’’ in Proc. Elseier 8th Int. Conf. Inf. Technol.
millimeter wave radar,’’ Remote Sens., vol. 15, no. 16, p. 4022, Aug. 2023. Quant. Manag., Chengdu, China, Feb. 2022, pp. 1066–1073.

12632 VOLUME 12, 2024

A. Khan et al.: Underwater Target Detection Using Deep Learning

[34] N. R. Nandyala and R. K. Sanodiya, ‘‘Underwater object detection using [56] S. Fu, F. Xu, J. Liu, Y. Pang, and J. Yang, ‘‘Underwater small object
synthetic data,’’ in Proc. 11th Int. Symp. Electron. Syst. Devices Comput. detection in side-scan sonar images based on improved YOLOv5,’’ in
(ESDC), Sri City, India, May 2023, pp. 1–6. Proc. 3rd Int. Conf. Geol., Mapping Remote Sens. (ICGMRS), Zhoushan,
[35] J. Li, L. Shi, and S. Guo, ‘‘YOLOv7-based land and underwater target China, Apr. 2022, pp. 446–453.
detection and recognition,’’ in Proc. IEEE Int. Conf. Mechatronics Autom. [57] I. Shogo and S. Masahiko, ‘‘Similar object identification in sonar images
(ICMA), Harbin, China, Aug. 2023, pp. 1437–1442. by machine learning,’’ in Proc. IEEE Underwater Technol. (UT), Tokyo,
[36] S. Wang, W. Wu, X. Wang, Y. Han, and Y. Ma, ‘‘Underwater optical Japan, Mar. 2023, pp. 1–6.
image object detection based on YOLOv7 algorithm,’’ in Proc. OCEANS, [58] Z. Hu and C. Xu, ‘‘Detection of underwater plastic waste based on
Limerick, Ireland, Jun. 2023, pp. 1–5. improved YOLOv5n,’’ in Proc. 4th Int. Conf. Frontiers Technol. Inf.
[37] A. Liu, X. Li, and K. Xu, ‘‘Machine learning-based image detection of Comput. (ICFTIC), Qingdao, China, Dec. 2022, pp. 404–408.
deep-sea seamounts creatures,’’ in Proc. IEEE Int. Geosci. Remote Sens. [59] M. Aubard, A. Madureira, L. Madureira, and J. Pinto, ‘‘Real-time
Symp. (IGARSS), Pasadena, CA, USA, Jul. 2023, pp. 5735–5737. automatic wall detection and localization based on side scan sonar
[38] Y. Wang, Z. Dai, J. Liu, and P. Wu, ‘‘Underwater target detection images,’’ in Proc. IEEE/OES Auton. Underwater Vehicles Symp. (AUV),
technology based on YOLOv7,’’ in Proc. IEEE Int. Conf. Sensors, Singapore, Sep. 2022, pp. 1–6.
Electron. Comput. Eng., Jinzhou, China, Aug. 2023, pp. 265–270. [60] H. Zhang, M. Tian, G. Shao, J. Cheng, and J. Liu, ‘‘Target detection
[39] X. Li, Y. Kou, and Y. Liang, ‘‘Binocular matching method for detecting of forward-looking sonar image based on improved YOLOv5,’’ IEEE
and locating marine resources using binocular cameras and deep Access, vol. 10, pp. 18023–18034, 2022.
learning,’’ in Proc. 4th Int. Conf. Comput. Vis., Image Deep Learn. [61] W. Hao and N. Xiao, ‘‘Research on underwater object detection based on
(CVIDL), Zhuhai, China, May 2023, pp. 381–384. improved YOLOv4,’’ in Proc. 8th Int. Conf. Inf., Cybern., Comput. Social
[40] A. Balaji, S. Yogesh, C. K. Kalyaan, R. Narayanamoorthi, G. Dooly, and Syst. (ICCSS), Beijng, China, Dec. 2021, pp. 166–171.
S. Dhanalakshmi, ‘‘Deep WaveNet-based YOLOv5 for underwater object [62] M. Zhang, S. Xu, W. Song, Q. He, and Q. Wei, ‘‘Lightweight underwater
detection,’’ in Proc. OCEANS, Limerick, Ireland, Jun. 2023, pp. 1–5. object detection based on YOLO v4 and multi-scale attentional feature
[41] P. Li, A. Zhao, Y. Fan, and Z. Pei, ‘‘Research on underwater robust object fusion,’’ Remote Sens., vol. 13, no. 22, p. 4706, Nov. 2021.
detection method based on improved YOLOv5s,’’ in Proc. IEEE Int. Conf. [63] W. Wang and Y. Wang, ‘‘Underwater target detection system based on
Mechatronics Autom. (ICMA), Harbin, China, Aug. 2023, pp. 1185–1189. YOLOv4,’’ in Proc. 2nd Int. Conf. Artif. Intell. Inf. Syst., Chongqing,
[42] D. A. R. Ellen, P. Kristalina, M. Z. S. Hadi, and A. Patriarso, ‘‘Effective China, May 2021, pp. 1–5.
searching of drowning victims in the river using deep learning method [64] L. Zhang, G. Zhai, B. Hu, Z. Qiao, and P. Zhang, ‘‘Fish target detection
and underwater drone,’’ in Proc. Int. Electron. Symp. (IES), Denpasar, and speed estimation method based on computer vision,’’ in Proc. IEEE
Indonesia, Aug. 2023, pp. 569–574. 6th Int. Conf. Electron. Inf. Commun. Technol. (ICEICT), Qingdao, China,
[43] C. Escobar-Amado, M. Badiey, and L. Wan, ‘‘Computer vision for Jul. 2023, pp. 1330–1336.
bioacoustics: Detection of bearded seal vocalizations in the Chukchi shelf [65] P. Yang, Y. Yu, and P. Yao, ‘‘An automatic detection system for marine
using YOLOV5,’’ IEEE J. Ocean. Eng., early access, pp. 1–12, Sep. 2023. ranch biological targets based on deep learning,’’ in Proc. 3rd Int.
[44] Y. Wei, Y. Fang, F. Cheng, M. Zhang, M. Cao, and H. Zhang, Conf. Mach. Learn., Big Data Bus. Intell. (MLBDBI), Taiyuan, China,
‘‘A lightweight underwater target detection network for seafood,’’ in Dec. 2021, pp. 696–699.
Proc. 42nd Chin. Control Conf. (CCC), Tianjin, China, Jul. 2023, [66] T. Guo, Y. Wei, H. Shao, and B. Ma, ‘‘Research on underwater
pp. 8381–8387. target detection method based on improved MSRCP and YOLOv3,’’ in
[45] A. Zhang and X. Zhu, ‘‘Research on ship target detection based on Proc. IEEE Int. Conf. Mechatronics Autom. (ICMA), Takamatsu, Japan,
improved YOLOv5 algorithm,’’ in Proc. 5th Int. Conf. Commun., Inf. Syst. Aug. 2021, pp. 1158–1163.
Comput. Eng. (CISCE), Guangzhou, China, Apr. 2023, pp. 459–463. [67] X. Cao, L. Ren, and C. Sun, ‘‘Dynamic target tracking control of
[46] E. Ranolo, K. Gorro, A. Ilano, H. Pineda, C. Sintos, and A. J. Gorro, autonomous underwater vehicle based on trajectory prediction,’’ IEEE
‘‘Underwater and coastal seaweeds detection for fluorescence seaweed Trans. Cybern., vol. 53, no. 3, pp. 1968–1981, Mar. 2023.
photos and videos using YOLOV3 and YOLOV5,’’ in Proc. 2nd [68] Z. Wang, X. Ye, Y. Han, S. Guo, X. Yan, and S. Wang, ‘‘Improved real-
Int. Conf. Innov. Technol. (INOCON), Bangalore, India, Mar. 2023, time target detection algorithm for similar multiple targets in complex
pp. 1–5. underwater environment based on YOLOv3,’’ in Proc. Global Oceans,
[47] Y. Zhang, T. Liu, and X. Yu, ‘‘Contextual and lightweight network for Biloxi, MS, USA, Singapore: U.S. Gulf Coast, Oct. 2020, pp. 1–6.
underwater object detection with self-attention mechanism,’’ in Proc. [69] J. W. Li and X. Cao, ‘‘Target recognition and detection in side-scan sonar
IEEE Int. Conf. Mechatronics Autom. (ICMA), Harbin, China, Aug. 2023, images based on YOLOv3 model,’’ in Proc. 41st Chin. Control Conf.,
pp. 1644–1649. Hefei, China, Jul. 2022, pp. 7191–7196.
[48] Z. Yin, S. Zhang, R. Sun, Y. Ding, and Y. Guo, ‘‘Sonar image target [70] X. Li, M. Tian, S. Kong, L. Wu, and J. Yu, ‘‘A modified YOLOv3
detection based on deep learning,’’ in Proc. Int. Conf. Distrib. Comput. detection method for vision-based water surface garbage capture robot,’’
Electr. Circuits Electron. (ICDCECE), Ballar, India, Apr. 2023, pp. 1–9. Int. J. Adv. Robot. Syst., vol. 17, no. 3, pp. 1–11, May 2020.
[49] M. F. Ercan, N. I. Muhammad, and M. R. N. Bin Sirhan, ‘‘Underwater [71] F. Jie and W. Pingbo, ‘‘Application of improved YOLO V3 algorithm for
target detection using deep learning,’’ in Proc. IEEE Region Conf. target detection in echo image of sonar underwater reverb,’’ J. Phys., Conf.
(TENCON), Hong Kong, Nov. 2022, pp. 1–5. Ser., vol. 1748, pp. 1–9, Nov. 2020.
[50] C. Xing, B. Sun, and W. Zhang, ‘‘Image-enhanced YOLOv5 and deep [72] B. Xie, S. He, and X. Cao, ‘‘Target detection for forward looking sonar
sort underwater multi-moving target tracking method,’’ in Proc. 5th Int. image based on deep learning,’’ in Proc. 41st Chin. Control Conf. (CCC),
Symp. Auton. Syst., Hangzhou, China, Apr. 2022, pp. 1–6. Hefei, China, Jul. 2022, pp. 7191–7196.
[51] R. Chen, S. Zhan, and Y. Chen, ‘‘Underwater target detection algorithm [73] W. Ouyang, Y. Wei, and D. Wang, ‘‘YOLOX-DC: A small target detection
based on YOLO and Swin transformer for sonar images,’’ in Proc. network up to underwater scenes,’’ in Proc. OCEANS, Hampton Roads,
OCEANS, VA, USA, Oct. 2022, pp. 1–7. VA, USA, Oct. 2022, pp. 1–6.
[52] G. Wen, F. Liu, X. Luo, and S. Li, ‘‘Underwater target detection based on [74] L. Beaudoin and L. Avanthey, ‘‘Underwater field equipment of a
modified YOLOv5,’’ in Proc. 5th Int. Conf. Intell. Auton. Syst. (ICoIAS), network of landmarks optimized for automatic detection by AI,’’ in Proc.
Dalian, China, Sep. 2022, pp. 7–12. IEEE Int. Geosci. Remote Sens. Symp. (IGARSS), Waikoloa, HI, USA,
[53] Y. Li, W. Liu, L. Li, W. Zhang, J. Xu, and H. Jiao, ‘‘Vision-based Sep./Oct. 2020, pp. 1572–1575.
target detection and positioning approach for underwater robots,’’ IEEE [75] L. Qingzhong, L. Yibing, and N. Jiong, ‘‘Real-time detection of
Photon. J., vol. 15, no. 1, pp. 1–12, Feb. 2023. underwater fish based on improved YOLO and transfer learning,’’ Pattern
[54] B. Le, H. Ying, and C. H. Kun, ‘‘A marine benthic detection algoritm Recognit. Artif. Intell., vol. 32, no. 3, pp. 193–203, 2019.
based on improved lightweight network,’’ in Proc. 34th Chin. Control [76] W. Xu and S. Matzner, ‘‘Underwater fish detection using deep learning
Decis. Conf. (CCDC), Hefei, China, Aug. 2022, pp. 5795–5800. for water power applications,’’ in Proc. Int. Conf. Comput. Sci. Comput.
[55] Z. Zhang, Q. Tong, C. Yi, X. Fu, J. Ai, and Z. Wang, ‘‘The appropriate Intell. (CSCI), Las Vegas, NV, USA, Dec. 2018, pp. 313–318.
image enhancement method for underwater object detection,’’ in Proc. [77] L. Xu, Y. Wei, X. Wang, A. Wang, and L. Guan, ‘‘Binocular vision of fish
IEEE 22nd Int. Conf. Commun. Technol. (ICCT), Nanjing, China, swarm detection in real-time based on deep learning,’’ in Proc. OCEANS
Nov. 2022, pp. 1627–1632. MTS/IEEE Charleston, Charleston, SC, USA, Oct. 2018, pp. 1–6.

VOLUME 12, 2024 12633

 A. Khan et al.: Underwater Target Detection Using Deep Learning

[78] T. Liu, N. Wang, L. Zhang, S. Ai, and H. Du, ‘‘A novel visual [99] A. Testolin and R. Diamant, ‘‘Underwater acoustic detection and
measurement method for three-dimensional trajectory of underwater localization with a convolutional denoising autoencoder,’’ in Proc.
moving objects based on deep learning,’’ IEEE Access, vol. 8, IEEE 8th Int. Workshop Comput. Adv. Multi-Sensor Adapt. Process.
pp. 186376–186392, 2020. (CAMSAP), Le Gosier, Guadeloupe, Dec. 2019, pp. 281–285.
[79] Y. Liu, W. Meng, and H. Zong, ‘‘Jellyfish recognition and den- [100] J. Liu, S. Liu, S. Xu, and C. Zhou, ‘‘Two-stage underwater object
sity calculation based on image processing and deep learning,’’ in detection network using Swin transformer,’’ IEEE Access, vol. 10,
Proc. Chin. Control Decis. Conf. (CCDC), Hefei, China, Aug. 2020, pp. 117235–117247, 2022.
pp. 922–927. [101] S. Raavi, P. B. Chandu, and T. SudalaiMuthu, ‘‘Automated recognition
[80] S. Wang, J. Guo, S. Guo, Q. Fu, and J. Xu, ‘‘Study on real-time of underwater objects using deep learning,’’ in Proc. 7th Int. Conf.
recognition of underwater live shrimp by the spherical amphibious robot Trends Electron. Informat. (ICOEI), Tirunelveli, India, Apr. 2023,
based on deep learning,’’ in Proc. IEEE Int. Conf. Mechatronics Autom. pp. 1055–1059.
(ICMA), Guilin, China, Aug. 2022, pp. 917–922. [102] Z. Shao-Kang, T. De-Yan, W. Chao, and Z. Xiao-Chuan, ‘‘Intelligent
[81] X. Zhang, Z. Liu, W. Wang, and J. Xu, ‘‘Automated detection of marine recognition of underwater acoustic target noise on underwater glider
mammal species based on short-time fractional Fourier transform,’’ in platform,’’ in Proc. Chin. Autom. Congr. (CAC), Xi’An, China, Nov. 2018,
Proc. OCEANS, San Diego, CA, USA, Sep. 2021, pp. 1–5. pp. 4189–4193.
[82] H.-Y. Lin, S.-L. Tseng, and J.-Y. Li, ‘‘SUR-Net: A deep network for fish [103] T. Hoang, K. S. Dalton, I. D. Gerg, T. E. Blanford, D. C. Brown, and
detection and segmentation with limited training data,’’ IEEE Sensors J., V. Monga, ‘‘Domain enriched deep networks for munition detection in
vol. 22, no. 18, pp. 18035–18044, Sep. 2022. underwater 3D sonar imagery,’’ in Proc. IEEE Int. Geosci. Remote Sens.
[83] A. Testolin, D. Kipnis, and R. Diamant, ‘‘Detecting submerged objects Symp. (IGARSS), Kuala Lumpur, Malaysia, Jul. 2022, pp. 815–818.
using active acoustics and deep neural networks: A test case for pelagic [104] T. Zhou, J. Si, L. Wang, C. Xu, and X. Yu, ‘‘Automatic detection of
fish,’’ IEEE Trans. Mobile Comput., vol. 21, no. 8, pp. 2776–2788, underwater small targets using forward-looking sonar images,’’ IEEE
Aug. 2022. Trans. Geosci. Remote Sens., vol. 60, Art. no. 4207912, 2022.
[84] D. Zhao, B. Yang, Y. Dou, and X. Guo, ‘‘Underwater fish detection [105] Y. Choo, K. Lee, W. Hong, S.-H. Byun, and H. Yang, ‘‘Active underwater
in sonar image based on an improved faster RCNN,’’ in Proc. 9th target detection using a shallow neural network with spectrogram-based
Int. Forum Electr. Eng. Autom. (IFEEA), Zhuhai, China, Nov. 2022, temporal variation features,’’ IEEE J. Ocean. Eng., early access, pp. 1–15,
pp. 358–363. Jul. 2022.
[85] B. Weihong, J. Yun, L. Jiaxin, S. Lingling, F. Guangwei, and J. Wa, ‘‘In- [106] H. Chen, J. Lin, L. Zhuge, and X. Xia, ‘‘Underwater image restoration and
situ detection method of jellyfish based on improved faster R-CNN and target detection based on monocular depth estimation,’’ in Proc. China
FP16,’’ IEEE Access, vol. 11, pp. 81803–81814, 2023. Autom. Congr. (CAC), Beijng, China, Oct. 2021, pp. 5597–5601.
[86] S. Li, X. Jin, S. Yao, and S. Yang, ‘‘Underwater small target recognition [107] M. S. Sangari, K. Thangaraj, U. Vanitha, N. Srikanth, J. Sathyamoorthy,
based on convolutional neural network,’’ in Proc. Global Oceans, Biloxi, and K. Renu, ‘‘Deep learning-based object detection in underwater
MS, USA, Singapore: U.S. Golf Coast, Oct. 2020, pp. 1–7. communications systems,’’ in Proc. 2nd Int. Conf. Elect., Electron., Inf.
Commun. Technol., Trichirappalli, India, Apr. 2023, pp. 1–6.
[87] M. Ogawa, H. Tanaka, J. Muramatsu, M. Nakano, K. Yoshida, and
T. Asakura, ‘‘Application of deep learning to underwater invasion [108] Y. Chai, H. Yu, L. Xu, D. Li, and Y. Chen, ‘‘Deep learning algorithms
warning system,’’ in Proc. MTS/IEEE Kobe Techno-Oceans (OTO), Kobe, for sonar imagery analysis and its application in aquaculture: A review,’’
Kobe, Japan, May 2018, pp. 1–5. IEEE Sensors J., vol. 23, no. 23, pp. 28549–28563, Dec. 2023.
[109] Z. Cao, ‘‘Deep learning methods for objective detection,’’ in Proc.
[88] Y. Wang, Z. Guo, and J. Xu, ‘‘Underwater search and rescue robot
IEEE 2nd Int. Conf. Data Sci. Comput. Appl. (ICDSCA), Dalian, China,
based on convolutional neural network,’’ in Proc. IEEE 4th Int. Conf.
Oct. 2022, pp. 1353–1357.
Civil Aviation Saf. Inf. Technol. (ICCASIT), Dali, China, Oct. 2022,
pp. 786–790. [110] A. B. Amjoud and M. Amrouch, ‘‘Object detection using deep learning,
CNNs and vision transformers: A review,’’ IEEE Access, vol. 11,
[89] K. Denos, M. Ravaut, A. Fagette, and H.-S. Lim, ‘‘Deep learning
pp. 35479–35516, 2023.
applied to underwater mine warfare,’’ in Proc. OCEANS, Aberdeen, U.K.,
Jun. 2017, pp. 1–7.
[90] Y. Liu, R. Wang, K. Cao, J. Wang, Z. Shi, Y. Wang, and Y. Zhou,
‘‘Sonar target detection based on a dual channel attention convolutional
network,’’ in Proc. 12th Int. Conf. Inf. Sci. Technol. (ICIST), Kaifeng,
China, Oct. 2022, pp. 255–264.
[91] D. Einsidler, M. Dhanak, and P.-P. Beaujean, ‘‘A deep learning approach
to target recognition in side-scan sonar imagery,’’ in Proc. OCEANS
MTS/IEEE Charleston, Charleston, SC, USA, Oct. 2018, pp. 1–4.
[92] L. Jin, H. Liang, and C. Yang, ‘‘Accurate underwater ATR in forward-
looking sonar imagery using deep convolutional neural networks,’’ IEEE
Access, vol. 7, pp. 125522–125531, 2019.
[93] X. Wang, C. Zhang, and R. Wang, ‘‘Underwater acoustic target
recognition technology based on MFA-conformer,’’ in Proc. 2nd Int. ANWAR KHAN received the M.Sc. degree in
Conf. Electron. Inf. Eng. Comput. Technol. (EIECT), Yanan, China, electronics from the Department of Electronics,
Oct. 2022, pp. 214–217. University of Peshawar, in 2006, the M.Phil.
[94] M. Thomas, F. Lionel, and D.-P. Laurent, ‘‘Propeller noise detection with degree in electronics from Quaid-i-Azam Uni-
deep learning,’’ in Proc. ICASSP - IEEE Int. Conf. Acoust., Speech Signal versity, Islamabad, Pakistan, in 2009, and the
Process. (ICASSP), Barcelona, Spain, May 2020, pp. 306–310.
Ph.D. degree in electronics from the Department
[95] X. Zhou, K. Yang, and R. Duan, ‘‘Deep learning based on striation images
of Electronics, Quaid-i-Azam University, in 2018,
for underwater and surface target classification,’’ IEEE Signal Process.
in underwater wireless sensor networks. He joined
Lett., vol. 26, no. 9, pp. 1378–1382, Sep. 2019.
the Department of Electronics, University of
[96] Q. Li, J. Li, T. Li, Z. Li, and P. Zhang, ‘‘Spectral–spatial depth-based
framework for hyperspectral underwater target detection,’’ IEEE Trans. Peshawar, in 2009, as a Faculty Member. He has
Geosci. Remote Sens., vol. 61, May 2023, Art. no. 4204615. published articles with IEEE, Springer, Elsevier, Taylor & Francis, MDPI,
[97] Y. Li, B. Wang, G. Shao, S. Shao, and X. Pei, ‘‘Blind detection of and KSII Transactions on Internet and Information Systems. He is
underwater acoustic communication signals based on deep learning,’’ also a reviewer with these publishers and one of the members of the
IEEE Access, vol. 8, pp. 204114–204131, 2020. technical program committee of several conferences. He has supervised and
[98] Q. Li, M. J. Er, L. Li, J. Chen, and J. Wu, ‘‘Underwater object detection co-supervised more than ten graduate and master’s students. His current
based on improved SSD with convolutional block attention,’’ in Proc. research interests include underwater wireless sensor networks, energy
5th Int. Conf. Intell. Auton. Syst. (ICoIAS), Dalian, China, Sep. 2022, harvesting, deep learning, machine learning, object detection, game-theoretic
pp. 37–42. optimization, and intelligent transportation.

12634 VOLUME 12, 2024

A. Khan et al.: Underwater Target Detection Using Deep Learning

MOSTAFA M. FOUDA (Senior Member, IEEE) ABDULLAH M. ALQAHTANI received the B.Sc.
received the B.S. degree (as the valedictorian) degree in communication and network engineering
and the M.S. degree in electrical engineering from King Khalid University, Abha, Saudi Arabia,
from Benha University, Egypt, in 2002 and 2007, in 2012, the M.Sc. degree in digital commu-
respectively, and the Ph.D. degree in information nications networks engineering from Southern
sciences from Tohoku University, Japan, in 2011. Methodist University, Dallas, TX, USA, in 2016,
He is currently an Associate Professor with the and the Ph.D. degree from the School of Electronic
Department of Electrical and Computer Engi- and Electrical Engineering, University of Leeds,
neering, Idaho State University, ID, USA. He is Leeds, U.K., in 2023. He is currently an Associate
currently a Full Professor with Benha University. Professor in computer and network engineering
He was an Assistant Professor with Tohoku University and a Postdoctoral with the School of Computer Science and Information Technology, Jazan
Research Associate with Tennessee Technological University, TN, USA. University. His research interests include the optimization techniques of edge
He has (co)authored more than 160 technical publications. His current computing, the Internet of Things (IoT), artificial intelligence, and optical
research interests include cybersecurity, communication networks, signal access networks.
processing, wireless mobile communications, smart healthcare, smart
grids, AI, and the IoT. He has guest-edited a number of special issues
covering various emerging topics in communications, networking, and health
analytics. He is also serving on the Editorial Board of IEEE TRANSACTIONS ON
VEHICULAR TECHNOLOGY, IEEE INTERNET OF THINGS JOURNAL, and IEEE ACCESS.
He has received several research grants, including NSF Japan–U.S. Network
Opportunity 3 (JUNO3).
DINH-THUAN DO (Senior Member, IEEE)
received the M.Sc. and Ph.D. degrees in electrical
engineering from Vietnam National University Ho
Chi Minh City (VNU-HCM), in 2007 and 2012,
respectively. Prior to joining academia, he was
a Senior Engineer in the telecommunications
industry with VinaPhone Mobile Network (the
biggest cellular network provider in Vietnam),
from 2003 to 2009. Before joining the University
of Mount Union, he was a Research Scientist with
the University of Colorado Denver, in 2022, The University of Texas at
Austin, in 2021, and an Assistant Professor with Asia University, Taiwan,
from 2020 to 2021. He is currently serving as an Associate Editor for IEEE
TRANSACTION ON VEHICULAR TECHNOLOGY and Computer Communications
(Elsevier). He has also served as a Lead Guest Editor/Guest Editor in
more than 20 special issues of journals, such as Physical Communications
(Elsevier) and Annals of Telecommunications (Elsevier). His publications
include over 120 SCIE/SCI-indexed journal articles, five edited books (IET,
Springer), and over 50 international conference papers. He was a recipient
of the 2015 Golden Globe Award by Vietnamese Ministry of Science and
Technology (top 10 outstanding scientists nationwide). He also received the
Medal of Creative Young Talents, in 2015. He was named in top 14 highly ATIQ UR RAHMAN received the Bachelor of
cited scientists at Asia University, Taiwan, in 2021 (Stanford’s list of top 2% Science (B.S.) degree in computer science from
scientists in the world, in October 2020, October 2021, and October 2022). the University of Peshawar, in 2004, the Master
ABDULAZIZ ALMALEH received the master’s of Science (M.S.) degree in computer system
and Ph.D. degrees in information science from engineering from the GIK Institute of Engineering
the prestigious University of Pittsburgh. He is cur- Sciences and Technology, Pakistan, in 2008,
rently a Renowned Academician and a Researcher and the Ph.D. degree in information technology
in the field of information science, specializing in from Universiti Teknologi PETRONAS, Malaysia,
machine learning, smart cities, and smart infras- in 2013. He was a System Engineer with Pakistan’s
tructure. With the master’s and Ph.D. degrees, National Database and Registration Authority,
he has established himself as an Expert in his from 2004 to 2006. Later, he was an Assistant Professor, the Head of
domain, with a strong understanding of emerg- the Computer Science Department, and the Head of the Quality and
ing technologies and research methodologies. Academic Accreditation Unit, Northern Border University, Saudi Arabia,
Throughout his academic journey, he has demonstrated a deep commitment from 2013 to 2019. Currently, he is an Assistant Professor with the Higher
to advancing the field of information science, focusing on the practical Colleges of Technology, United Arab Emirates. His research interests include
applications of machine learning and its implications for the development of energy-efficient design, routing, deployment, coverage, and the quality of
smart cities and infrastructure. His research interests include a wide variety service in wireless sensor networks and underwater sensor networks. He is
of subjects, including data analytics, artificial intelligence, the Internet of a Distinguished Senior Member of esteemed professional organizations,
Things (IoT), and the integration of these technologies into urban planning such as the Universal Association of Computer and Electronics Engineers
and development processes. He has an impressive publication record that (UACEE), the Institute of Research Engineers and Doctors (IRED), and the
attests to his dedication and expertise in his chosen field. His numerous works American Society for Engineering Education (ASEE). With a significant
have been published in well-respected journals and conference proceedings, number of research publications in renowned journals and conferences,
contributing to the collective knowledge and understanding of machine he actively contributes as a reviewer to several journals. He is a technical
learning applications in the creation of intelligent urban environments. program committee member at international conferences. Additionally,
His research findings have greatly influenced the way in which advanced he also participated in the ‘‘Fundamental and Advanced Assessment
technologies are applied to urban planning, offering practical solutions to Workshop’’ organized by ABET in the USA.
the modern challenges faced by cities around the world.

VOLUME 12, 2024 12635