Received 16 May 2023, accepted 1 June 2023, date of publication 7 June 2023, date of current version 23 June 2023.

Digital Object Identifier 10.1109/ACCESS.2023.3283299

Optimal Fuzzy Wavelet Neural Network Based

Road Damage Detection
MOHAMMAD ALAMGEER1 , HEND KHALID ALKAHTANI 2 , MASHAEL MAASHI3 ,
MAHMOUD OTHMAN 4 , ANWER MUSTAFA HILAL 5 , MOHAMED IBRAHIM ALSAID5 ,
AZZA ELNEIL OSMAN5 , AND AMANI A. ALNEIL5
1 Department of Information Systems, College of Science and Art at Mahayil, King Khalid University, Muhayil Assir 61421, Saudi Arabia
2 Department of Information Systems, College of Computer and Information Sciences, Princess Nourah bint Abdulrahman University, Riyadh 11671, Saudi Arabia
3 Department of Software Engineering, College of Computer and Information Sciences, King Saud University, Riyadh 11543, Saudi Arabia
4 Department of Computer Science, Faculty of Computers and Information Technology, Future University in Egypt, New Cairo 11835, Egypt
5 Department of Computer and Self Development, Preparatory Year Deanship, Prince Sattam bin Abdulaziz University, Alkharj 16273, Saudi Arabia

Corresponding author: Anwer Mustafa Hilal (a.hilal@psau.edu.sa)

The authors extend their appreciation to the Deanship of Scientific Research at King Khalid University for funding this work through
Large Groups Project under grant number (RGP2/134/44). Princess Nourah bint Abdulrahman University Researchers Supporting Project
number (PNURSP2023R384), Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia. Research Supporting Project
number(RSPD2023R787), King Saud University, Riyadh, Saudi Arabia. This study is supported via funding from Prince Sattam bin
Abdulaziz University project number (PSAU/2023/R/1444).

ABSTRACT Floods are one of the most severe and most frequent natural calamities. It causes enormous
economic damage and even leads to higher mortality rates. Studies on damage detection of roads using
artificial intelligence (AI) methods gained more attention currently, especially in the development of smart
cities. Therefore, this study designs an optimal Fuzzy Wavelet Neural Network based Road Damage
Detection (OFWNN-RDD) technique for Flooding Management. The OFWNN-RDD technique aims to
exploit the remote sensing images to classify different kinds of roads. For noise removal process, the
OFWNN-RDD technique utilizes Gabor filtering (GF) technique. In addition, the OFWNN-RDD technique
uses the DenseNet121 model for feature vector generation with modified barnacles mating optimization
(MBMO) based hyperparameter optimizer. Finally, FWNN image classification approach is used for road
damage detection. The simulation values exhibit the supremacy of the OFWNN-RDD technique over other
models with improved road damage detection accuracy of 98.56%.

INDEX TERMS Flooding, road damage, machine learning, parameter tuning, computer vision.

I. INTRODUCTION technique suits for a reliable and rapid post-disaster damages

Natural catastrophes, namely wildfires, floods, and earth- assessment since rapid accessibility and acquisition of images
quakes, cause huge losses to flatten buildings, and substruc- enable recognition of damaged road areas. Detecting injured
tures and block roads, causing heavy economic and social roads over high-resolution aerial imagers can accelerate and
losses [1]. Roads are regarded as lifelines. As soon as a enrich decision-making during a disaster [4].
disaster happened, road damage detection and valuation were Studies on artificial intelligence (AI) based road damage
the basis for emergency rescue operations. To assess, identify detection model has gained significant attention in the recent
and detect road damage, several kinds of remote sensing times [5]. Since manual road damage detection is a labo-
(RS) data like Lidar and SAR, aerial or satellite images were rious and time consuming process, the automatic analysis
widely utilized [2]. Mainly, high-resolution aerial images and follow-up of road damages can be used [6]. Government
can be gained in a controlled way, concerning both flight agency cannot keep up an accurate database of the road con-
planning and time and at radiometric, geometric and high ditions. Another problem is the short fall of professionals who
spectral resolution to permit an emergency response [3]. This could evaluate the spread and state of numerous damages,
as the assessment will often be highly subjective [7].
The associate editor coordinating the review of this manuscript and Conventional techniques to collect samples in the domain
approving it for publication was Kathiravan Srinivasan . are cost-intensive and time-consuming [8]. Thus, numerous

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.

61986 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ VOLUME 11, 2023
M. Alamgeer et al.: Optimal Fuzzy Wavelet Neural Network Based Road Damage Detection

research efforts were held to help government agency for II. RELATED WORKS
automating the sample collection process and road inspec- Al Duhayyim et al. [11] introduced a novel Road Damage
tion, using technology with various degrees of complex- Detection modality utilizing Hunger Games Search and
ity. Current research works have implemented several Deep Elman NN (RDD–HGSENN) on High-Resolution RSI. The
Learning (DL)-related and Machine Learning (ML) methods RDD–HGSENN approach mainly intends to fix road destruc-
for automated damage detection or road surface survey [9]. tions through RSIs. In the approach mentioned above, the
DL, which depends on neural networks, will be an advanced RetinaNet method has been enforced to detect road damage.
conception of ML and grants solutions in application fields Besides, the RDD–HGSENN algorithm can categorise road
that are tough to model with classical statistical methods. damage through the ENN method. The HGS system has been
Convolutional Neural Networks (CNNs) are the main cat- used in this study for tuning ENN variables mechanically.
egory of deep neural networks (DNNs) that can be generally Reference [12] presents a new method relevant to the Track-
implemented for image classification and recognition [10]. ing, Learning, and Detector (TLD) structures for detecting the
The main benefit of CNN is that it mechanically finds critical damaged road area from post-disaster high-resolution RSI.
features after the training phase by not involving any human Initially, spoke wheel operators can be used to describe the
supervision. On the other hand, hyperparameters are settings initial template of the road. Afterwards, the TLD structure
that are not learned during training, but must be set prior to was exploited to find suspected road damaged regions. Lastly,
training. They can have a significant impact on the perfor- damaged road parts were extracted by pruning false damaged
mance of the model, and selecting the optimal values can roads.
lead to better accuracy. Most of the existing works have not Ochoa-Ruiz et al. [13] aim at bridging this gap through
addressed the hyperparameter tuning process. Since the trial 2 methods. First, the author presented a novel large asphalt
and error parameter tuning is a tedious process, metaheuristic dataset, which incorporated damages not existing in earlier
optimization algorithms are preferable. research works, making it very powerful and representative of
This study designs an optimal Fuzzy Wavelet Neural some reimbursements like potholes. The images are captured
Network based Road Damage Detection (OFWNN-RDD) in various weather and illumination circumstances, and a
technique for Flooding Management. The OFWNN-RDD quality-aware data augmenting technique has been used for
method uses Gabor filtering (GF) for noise removal. The sorting out examples of bad quality, which will be helpful in
OFWNN-RDD technique uses the DenseNet121 model enriching the performance metrics on the baseline. Then, the
for feature vector generation with hyperparameter tuning author well-trained various object detection methods agree-
using the modified barnacles mating optimization (MBMO) able for mobile applications with a suitable performance for
algorithm. This study employs the FWNN image classi- several applications. Hacıefendioğlu and Başağa [14] con-
fication approach for road damage detection. A complete centrated on identifying cracks in concrete roads for several
set of series was performed to highlight the proficient illumination levels, shooting and weather conditions utiliz-
classification performance of the OFWNN-RDD algo- ing a DL-related object detection technique. In this context,
rithm. In short, the key contributions are summarized as previous cracks will be determined unusually inexpensively
follows. and quickly. A descriptive method is taken into account for
detecting cracks of images on concrete road surfaces utilizing
• An intelligent OFWNN-RDD technique comprising of pre trained Faster R-CNN.
pre-processing, DenseNet121 based feature extraction, Shim et al. [15] advanced a new sensor technology that
MBMO based hyperparameter tuning, and FWNN based finds road damage through a DL-relevant image processing
classification is presented for road damage detection. technique. The presented technology is a semi-supervised
To the best of our knowledge, the OFWNN-RDD model learning and super-resolution method related to GAN. In the
has never presented in the literature. previous days, the quality of road image so that damages
• A novel MBMO algorithm is introduced by incorpo- can be observed clearly. Such 2 techniques are allplied to
rating the concepts of self-adaptive population into the 4 lightweight segmentation neural networks. Li et al. [16]
BMO algorithm. modelled an automated common subsurface distress detec-
• Employ FWNN model for the identification and classi- tion technique unitizing conventional deep learning and sig-
fication of road damage detection process. nal processing. Wavelet transform was enforced to find the
• Hyperparameter optimization of the DenseNet121 road layer for function segmentations. A particular pseudo-
model using MBMO algorithm using cross-validation color map technique was modlle4d for converting reflected
helps to boost the predictive outcome of the signals for DL model training.
OFWNN-RDD model for unseen data. A novel large-scale manhole cover detection datasets are
The rest of the paper is organized as follows. Section II introduced utilizing smartphones for collecting road image
provides the related works and section III offers the pro- datasets, and a hierarchical classification approach related
posed model. Then, section IV gives the result analysis and to the CNN is devised in [17]. Wang et al. [18] introduced
section V concludes the paper. an improved approach depending on YoloV3 that considers

VOLUME 11, 2023 61987

 M. Alamgeer et al.: Optimal Fuzzy Wavelet Neural Network Based Road Damage Detection

the FWNN model is exploited for road damage classification

process.

A. DATA PREPROCESSING
Primarily, the OFWNN-RDD technique used the GF tech-
nique for the noise removal process. In the presented sys-
tem, the Gabor filter (GF) was exploited to relax the val-
leys and enhance the ridges by implementing a short-term
Fourier transformation involving a Gaussian window in the
spatial domain [22]. This statistical feature produced an
image feature emphasized by applying the orientations and
frequency information in fingerprint images by finetuning
GFs. A sequence of GFs is used on the image I (x, y) in
different frequencies, which have distinct orientations with
g(x, y) Gabor function as follows.
′ ′
!
x 2 + γ 2y 2 X′
 
g (x, y) = exp − cos 2π + φ (1)
2σ 2 l

where, χ ′ = xcoxθ + ysṁθ and y′ = ycoxθ − xsṁθ. This

Gabor transform is employed on the Gaussian envelope σ
along with χ and y directions.
FIGURE 1. The overall workflow of OFWNN-RDD system.
B. FEATURE EXTRACTION
The OFWNN-RDD technique applied the DenseNet121
the tininess nature and slenderness of the road damage that model for feature vector generation. CNN achieves superior
needed detailed and low-level description. Yuan et al. [19] performance in the image classification domain [23]. But
introduced EcRD: an edge-cloud-related road damage detec- training the CNN from scratch is difficult as the classifi-
tion and warning structure that uses the fast-responding cation accuracy based on hyperparameter tunings such as
advantages of edge and extensive computation resources initial weight, learning rate, optimizers, number of epochs,
and storage advantages of cloud. Samma et al. [20] intend and dropout requires higher computation power and many
to introduce a pre-trained VGG-19 leveraging a potential labeled training datasets. These problems are leveraged by
two-layer optimizer. The presented optimizer performed filter using TL technique. During the TL algorithm, training dura-
selection of the last layers of VGG-19 guided by the precision tion is minimalized by the weight accomplished from the
of the linear SVM classifier. Gan et al. [21] develops one pre-trained method that can be exploited as an initial weight
approach related to M2det that can derive the deep and shal- to train the innovative difficulties. This technique of re-using
low features. For its multi-level and multi-scale, it belongs pretrained weights outcomes in low generalization error. The
to one-stage. The author used M2det for detecting road dam- DenseNet model is widespread because the DenseNet frame-
age, training on a more image snapped by vehicle-mounted work attenuates the gradient vanishing problems, stimulates
smartphones, and then comparing it with the other one-stage feature reuse, enhances feature propagation, and decreases
approaches. the parameter count. In DCNN, every layer was linked with
the other layers as feed forward patterns. All the layers in
III. THE PROPOSED MODEL DenseNet accept the feature mas of every preceding layer as
In this study, we have designed a novel OFWNN-RDD another input and passed on the feature map to all the every
approach for road damage detection in Flooding Manage- layer. Hence, n − th layers have n inputs of every layer.
ment. The OFWNN-RDD technique focused on the effectual Generally, CNN changes its feature map size through the
classification of roads into different types. To accomplish downsampling layer. But DenseNet facilitates feature con-
this, the OFWNN-RDD technique comprises GF preprocess- catenation and downsampling by splitting the network into
ing, DenseNet121 feature extractor, MBMO hyperparame- densely connected dense blocks. The size of feature maps
ter tuning, and FWNN-based classification. Fig. 1 shows in the block remains the same, and inside dense blocks,
the overall workflow of OFWNN-RDD system. Initially, the they assist to perform concatenation, whereas, outside dense
proposed model receives the road images as input and then blocks, convolution and pooling functions are implemented
eliminate the noise exist in it using the GF based prepro- for down-sampling. A transition layer or block was added
cessing technique. Next, the DenseNet121 model extracts a at the end of the dense layers. The transition layer includes
useful set of feature vectors and the MBMO algorithm is 2 × 2 average pooling layers, a batch normalization layer,
applied to optimally choose the DenseNet121 model. Finally, and a 1 × 1 convolutional. The transition layer will change

61988 VOLUME 11, 2023

M. Alamgeer et al.: Optimal Fuzzy Wavelet Neural Network Based Road Damage Detection

the feature map size. Hence, the DenseNet includes 3 transi- where rand describes the random number within [0, 1],
tions, 1-classification, and 117 Conv, which makes the size of the recently produced offspring for exploration is produced
layer 121. using Mum’s candidates. The offspring would be evaluated
In this work, the hyperparameter tuning process uses the and added to parents for extending the solution matrix from
MBMO algorithm. In this work, the MBMO technique as candidate size. Therefore, a method has been exploited to
a hyperparameter optimizer has been employed. Barnacle arrange a single dimension, eliminating incorrect solutions.
is a specific type of arthropod that constitutes an infraclass The MBMO technique incorporates self-population-related
Cirripedia depends on lobsters and crabs [24]. They are initialization with the typical BMO technique. Like other
especially marine animals which live in shallow and tidal metaheuristics, BMO has a population-related optimized
waters. They will be available everywhere one seawater and algorithm initiated by random initialization. This shows that
raised on hard surfaces in seawater. Afterwards, they hatch it demands control parameters for determining population
eggs; barnacle larvae were dispersed in water to discover sizes. Yet, it is notable that the selection of population sizes
and stick towards hard surfaces. The hard surface covers the to overcome case problem become challenging. Here, the
body of the barnacle and improves the shell plate. They need primary population size in the first iteration was attained
to search for a balance between achieving additional mates using a self-adaptive population:
in a turbulent flow and dealing with ever-longer penises.
Depending on this behaviour, a new optimization approach, PopSize = 10 × d (9)
called the BMO algorithm, was proposed. The balance behav-
iors are developed by using Hardy Weinberg equilibrium in In Eq. (9), d denotes the problem dimension, and the
the following. The primary population of barnacles for the following expression determines it:
solution is defined as follows:
PopSizenew = max (d, round (PopSize + r×PopSize))
 χ1 χ1 
 1 N
 (10)
.. ..

X= . . (2)

 1 N In Eq. (10), r determines a random value within [−0.5,
χn χn

0.5].
In Eq. (2), n determines the candidate numbers, and N
indicates the amount of decision variables based on the upper Algorithm 1 Pseudocode of BMO
and lower limiations: Initializes population of barnacle Xi
h i
lb = lb1 , . . . , lbi (3) Evaluate the barnacle fitness valu
h i Sort for locating an optimal outcome at the population
ub = u1b , . . . , uib (4) (T=the optimal solutions)
while (I < Maximal iteration)
Now, ub and lb denote the upper and lower boundaries of Fix values of pl
parameter i. By evaluating the objective function for every if select of Dad and Mum =pl
candidate, best to worst upshots are arranged and stored for every variable
during primary iteration. The proposed technique includes Offspring generations:
exploitation and exploration. The offspring generation is end for
enforced using sperm cast as an exploration term: else if selective of Dad and Mum >pl
bD = rand (n) (5) for every variable
Offspring generation
bM = rand (n) (6)
End for
From the expression, bD and bM indicates mated parents. end if
Based on Hardy Weinberg’s concept, the BMO technique Apply existing barnacle backs when it moves
considers the inheritance feature or parents’ genotype fre- outside the boundary
quency during the generation of offspring for modelling the Evaluate barnacle fitness valu
reproduction technique: Arrange and upgrade T when there was an
optimal solution
XiNnew = pXbND + qXbNM (7) l=l+1
In Eq. (7), XbNM and XbND correspondingly characterize the end whil
variable of Mum and Dad candidates, and p determines a Return T
pseudorandom number that lies within [0,1], = (1−p).
Once the candidate selection to mate excels pl amount was The MBMO approach derived a fitness function to have
initially considered, then exploration term is performed: an enriched classifier outcome. It determined positive values
for signifying superior outcomes of candidate solutions. This
XiNnew = randd×XbnM (8) article uses the reduced classifier error rate as the fitness

VOLUME 11, 2023 61989

 M. Alamgeer et al.: Optimal Fuzzy Wavelet Neural Network Based Road Damage Detection

function, as offered in Eq. (11). TABLE 1. Details of dataset.

fitness (xi ) = ClassifierErrorRate (xi )

number o fmisclassified samples
= ∗100 (11)
Total number of samples

C. CLASSIFICATION USING FWNN MODEL

For road damage detection, the FWNN image classification
approach is employed in this work. For classification pur-
poses, the proposed method integrates a WNN and FNN [25].
The network comprises eight levels based on the proposed
method.
(1) First Layer. The input features are the initial layer of
the projected FWNNet. They are Time series, Independent
variables, and picture classification features.
X = Xj | j= 1, . . . ,n

(12)
(2) Second Layer. This layer encompasses WNN and FNN
for the approximation. The wavelet is calculated through the
wavelet portion:
xj − bkij
ψijk = ψ( ),i= 1, . . . ,N , j= 1, . . . ,n (13)
akij
N indicates the number of wavelets, whereas n indicates
the number of input features.

FIGURE 3. Confusion matrices of OFWNN-RDD approach (a-b) TRS/TSS of

80:20 and (c-d) TRS/TSS of 70:30.

be designed. Moreover, the fuzzy membership function is

calculated in the fuzzy region of second layer as follows:
−((x −ckj )/σkj 2
µkj = e j ) (16)
In Eq. (16), ckj indicates the center, and σkj shows standard
deviation for the rule k membership function. Fig. 2 illustrates
FIGURE 2. The architecture of FWNN.
the framework of FWNN.
(3) Third Layer. The third layer is the aggregation layer,
Wavelet transformation, at the same time, display function
where the output of layer 3 should be collectively multi-
and discloses its local features in the time-frequency domain.
plied. Multiple WNNs with Nk wavelet activation function
This feature makes it easy to train NN for accurately mod-
are applied in the wavelet portion:
elling tremendously non-linear data:
n
Y
9ik = ψ k ij, k, 1, . . . ,M
 
− 21 x−b (17)
ψa,b = |a| ψ( ),a, b∈R, a̸ = 0, (14)
a j=1

In Eq. (14), ψ(x) ∈L 2 (R) denotes the wavelet function as Moreover, each node in layer 3 demonstrates one fuzzy
follows: rule:
n
|ψ̂(ω)|
Z +∞ Y
Cψ = dω< +∞. (15) Ok = µkj , k, 1, . . . ,M (18)
0 ω j=1

In Eq. (15), ψ̂(ω) denotes the FTs of ψ(x). To stimulate (4) Fourth Layer. The output of the wavelet component was
multi-variable processes, a multidimensional wavelet should calculated:

61990 VOLUME 11, 2023

M. Alamgeer et al.: Optimal Fuzzy Wavelet Neural Network Based Road Damage Detection

FIGURE 4. TACC and VACC outcome of OFWNN-RDD approach.

FIGURE 5. TLS and VLS outcome of OFWNN-RDD approach.

Rk : IF x1 is Ak1 . . . AND xn is Akn , layer output neurons produced the quotient that demonstrates
THEN every WNNs output is proportional to the FWNNet results.
Nk (5) (3) (4)
X Ok = Ok ·Ok = Ok · Yk ,
Yk = wki 9ik + yk (19)
M
i=1 (6)
X (5)
O1 = Ok
where x1 , x2 , . . . ,xn signify the input feature, Y1 , Y2 , . . . ,YM k=1
characterizes the 4th layer resultant layer, and Akj signifies the M
k th fuzzy set with standard membership. The weight matrix (6) (3)
X
O2 = Ok (20)
and bias were saved in the hidden layer as wki and yk . k=1
(5) Fifth Layer. The results of FNN and WNN in the third
(7) Seventh Layer. The result of the output is gathered at
and fourth layers, Ok and Yk , are integrated. The defuzzifi-
the 7th layer.
cation inference was s in this layer, and it multiplies the 3rd
resultant layers dataset by the 4th resultant dataset. O1
(6) PM
(7) k=1 Ok Yk
(6) Sixth Layer. Two neurons correspondingly act as a y = O = (6) = P M
(21)
summing operator for 5th and 3rd layer resultant signals. 7th O2 k=1 Ok

VOLUME 11, 2023 61991

 M. Alamgeer et al.: Optimal Fuzzy Wavelet Neural Network Based Road Damage Detection

FIGURE 6. The precision-recall outcome of OFWNN-RDD approach.

TABLE 2. Road classifier outcome of OFWNN-RDD system under 70:30 of TABLE 3. Road classifier outcome of OFWNN-RDD method under 80:20
TRS/TSS. of TRS/TSS.

8) Eighth Layer. This was the network layer for feature

classification. It can be an activation function which trans- deliberated that the OFWNN-RDD process has recognized
forms data into output layer value. four types of road damage accurately.
Table 2 reports the overall road classifier outcomes of the
IV. EXPERIMENTAL VALIDATION OFWNN-RDD method with 70:30 of TRS/TSS. On 70% of
The proposed model is simulated using Python 3.6.5 tool. The TRS, the results indicated that the OFWNN-RDD technique
proposed model is experimented on PC i5-8600k, GeForce has proficiently categorized four types of road damage. It is
1050Ti 4GB, 16GB RAM, 250GB SSD, and 1TB HDD. The observable that the OFWNN-RDD technique reaches average
parameter settings are given as follows: learning rate: 0.01, accuy of 98.20%, precn of 96.37%, recal of 96.37%, Fscore of
dropout: 0.5, batch size: 5, epoch count: 50, and activation: 96.36%, AUCscore of 97.58%, and MCC of 95.17%.
ReLU.In this study, road damage classifier outcomes of the Table 3 offers the overall road classification outcomes of
OFWNN-RDD technique can be tested using the dataset [11] the OFWNN-RDD system with 80:20 of TRS/TSS. On 80%
comprising 4000 samples with 4 class labels, as represented of TRS, the results exhibited the road classifier outcome of
in Table 1. the OFWNN-RDD method with 80% of TRS. The results
The road damage identification outcomes of the indicated that the OFWNN-RDD system has proficiently cat-
OFWNN-RDD method are investigated under distinct egorized four types of road damage. It is observable that the
aspects in the form of a confusion matrix in Fig. 3. The results OFWNN-RDD technique reaches average accuy of 97.91%,

61992 VOLUME 11, 2023

M. Alamgeer et al.: Optimal Fuzzy Wavelet Neural Network Based Road Damage Detection

FIGURE 7. ROC curve outcome of OFWNN-RDD approach.

TABLE 4. Comparative analysis of OFWNN-RDD methodology with other values inferred the betterment of the OFWNN-RDD tech-
techniques.
nique over other models. It is noticed that the RetinaNet
and MobileNet models have obtained lower classifier out-
comes. Next, the AlexNet and GoogleNet models have
reached slightly enhanced performance. In contrast, the
RDD-HGSENN technique has ensured reasonable perfor-
mance with anaccuy of 98.03%. However, the OFWNN-RDD
technique surpassed recent DL models with a maximum
accuy of 98.56%. These results guaranteed the maximum
performance of the OFWNN-RDD technique on the road
classification process.

precn of 95.80%, recal of 95.81%, Fscore of 95.80%, AUCscore V. CONCLUSION

of 97.21%, and MCC of 94.41%. In this study, we have designed a new OFWNN-RDD tech-
The TACC and VACC of the OFWNN-RDD method are nique for road damage detection in Flooding Management.
inspected on road classification performance in Fig. 4. The The OFWNN-RDD technique focused on the productive clas-
results pointed out that the OFWNN-RDD algorithm has sification of roads into different types. The OFWNN-RDD
exhibited enhanced outcomes with increased values of TACC technique used the GF technique for noise removal to
and VACC. It is observable that the OFWNN-RDD system accomplish this. The OFWNN-RDD technique applied the
has reached superior TACC outcomes. DenseNet121 model for feature vector generation with hyper-
The TLS and VLS of the OFWNN-RDD system are tested parameter tuning using the MBMO algorithm. This study
on road classification performance in Fig. 5. The figure employs the FWNN image classification approach for road
inferred that the OFWNN-RDD method had enhanced perfor- damage detection. A complete set of experiments were
mance with the least values of TLS and VLS. It is perceptible performed to highlight the proficient classification perfor-
that the OFWNN-RDD technique has resulted in lower VLS mance of the OFWNN-RDD method. The simulation val-
outcomes. ues demonstrate the supremacy of the OFWNN-RDD algo-
An evident precision-recall study of the OFWNN-RDD rithm over other models with maximum accuracy of 98.56%.
methodology in the test database is shown in Fig. 6. The In future, the performance of the OFWNN-RDD method will
results implied that the OFWNN-RDD system has led to be enhanced by the ensemble learning process.
superior values of precision-recall values in four class labels.
A comprehensive ROC study of the OFWNN-RDD ACKNOWLEDGMENT
method in the test database is exposed in Fig. 7. The outcomes The authors extend their appreciation to the Deanship of
indicated that the OFWNN-RDD algorithm had revealed its Scientific Research at King Khalid University for fund-
ability to classify four class labels. ing this work through Large Groups Project under grant
A widespread comparison study of the OFWNN-RDD number (RGP2/134/44). Princess Nourah bint Abdulrah-
technique is briefed in Table 4 [11], [13]. The resultant man University Researchers Supporting Project number

VOLUME 11, 2023 61993

 M. Alamgeer et al.: Optimal Fuzzy Wavelet Neural Network Based Road Damage Detection

(PNURSP2023R384), Princess Nourah bint Abdulrahman [12] K. Zhao, J. Liu, Q. Wang, X. Wu, and J. Tu, ‘‘Road damage detection
University, Riyadh, Saudi Arabia. Research Supporting from post-disaster high-resolution remote sensing images based on TLD
framework,’’ IEEE Access, vol. 10, pp. 43552–43561, 2022.
Project number(RSPD2023R787), King Saud University, [13] G. Ochoa-Ruiz, A. A. Angulo-Murillo, A. Ochoa-Zezzatti,
Riyadh, Saudi Arabia. This study is supported via funding L. M. Aguilar-Lobo, J. A. Vega-Fernández, and S. Natraj, ‘‘An asphalt
from Prince Sattam bin Abdulaziz University project number damage dataset and detection system based on RetinaNet for road
conditions assessment,’’ Appl. Sci., vol. 10, no. 11, p. 3974, Jun. 2020.
(PSAU/2023/R/1444). [14] K. Hacıefendioğlu and H. B. Başağa, ‘‘Concrete road crack detection
using deep learning-based faster R-CNN method,’’ Iranian J. Sci. Technol.,
Trans. Civil Eng., vol. 46, no. 2, pp. 1621–1633, Apr. 2022.
REFERENCES [15] S. Shim, J. Kim, S.-W. Lee, and G.-C. Cho, ‘‘Road damage detection using
[1] S. Shim, J. Kim, S.-W. Lee, and G.-C. Cho, ‘‘Road surface dam- super-resolution and semi-supervised learning with generative adversarial
age detection based on hierarchical architecture using lightweight network,’’ Autom. Construct., vol. 135, Mar. 2022, Art. no. 104139.
auto-encoder network,’’ Autom. Construction, vol. 130, Oct. 2021, [16] Y. Li, C. Liu, G. Yue, Q. Gao, and Y. Du, ‘‘Deep learning-based pavement
Art. no. 103833. subsurface distress detection via ground penetrating radar data,’’ Autom.
[2] M.-T. Cao, Q.-V. Tran, N.-M. Nguyen, and K.-T. Chang, ‘‘Survey on per- Construct., vol. 142, Oct. 2022, Art. no. 104516.
formance of deep learning models for detecting road damages using mul- [17] B. Zhou, W. Zhao, W. Guo, L. Li, D. Zhang, Q. Mao, and Q. Li,
tiple dashcam image resources,’’ Adv. Eng. Informat., vol. 46, Oct. 2020, ‘‘Smartphone-based road manhole cover detection and classification,’’
Art. no. 101182. Autom. Construct., vol. 140, Aug. 2022, Art. no. 104344.
[3] P. Hruza, T. Mikita, N. Tyagur, Z. Krejza, M. Cibulka, A. Procházková, [18] Q. Wang, J. Mao, X. Zhai, J. Gui, W. Shen, and Y. Liu, ‘‘Improvements of
and Z. Patočka, ‘‘Detecting forest road wearing course damage using YOLOv3 for road damage detection,’’ J. Phys., Conf. Ser., vol. 1903, no. 1,
different methods of remote sensing,’’ Remote Sens., vol. 10, no. 4, p. 492, Apr. 2021, Art. no. 012008.
Mar. 2018. [19] Y. Yuan, M. S. Islam, Y. Yuan, S. Wang, T. Baker, and L. M. Kolbe, ‘‘EcRD:
[4] M. Azimi, A. Eslamlou, and G. Pekcan, ‘‘Data-driven structural health Edge-cloud computing framework for smart road damage detection and
monitoring and damage detection through deep learning: State-of-the-art warning,’’ IEEE Internet Things J., vol. 8, no. 16, pp. 12734–12747,
review,’’ Sensors, vol. 20, no. 10, p. 2778, May 2020. Aug. 2021.
[5] D. Arya, H. Maeda, S. K. Ghosh, D. Toshniwal, A. Mraz, T. Kashiyama, [20] H. Samma, S. A. Suandi, N. A. Ismail, S. Sulaiman, and L. L. Ping,
and Y. Sekimoto, ‘‘Deep learning-based road damage detection and clas- ‘‘Evolving pre-trained CNN using two-layers optimizer for road damage
sification for multiple countries,’’ Autom. Construct., vol. 132, Dec. 2021, detection from drone images,’’ IEEE Access, vol. 9, pp. 158215–158226,
Art. no. 103935. 2021.
[6] H. Maeda, Y. Sekimoto, T. Seto, T. Kashiyama, and H. Omata, ‘‘Road dam- [21] X. Gan, J. Qu, J. Yin, W. Huang, Q. Chen, and W. Gan, ‘‘Road dam-
age detection using deep neural networks with images captured through a age detection and classification based on M2det,’’ in Advances in Arti-
smartphone,’’ 2018, arXiv:1801.09454. ficial Intelligence and Security. Dublin, Ireland: Springer, Jul. 2021,
[7] M. Guerrieri and G. Parla, ‘‘Flexible and stone pavements distress detec- pp. 429–440.
tion and measurement by deep learning and low-cost detection devices,’’ [22] A. A. Albraikan, M. Aljebreen, J. S. Alzahrani, M. Othman,
Eng. Failure Anal., vol. 141, Nov. 2022, Art. no. 106714. G. P. Mohammed, and M. I. Alsaid, ‘‘Modified barnacles mating
[8] C. Chu, L. Wang, and H. Xiong, ‘‘A review on pavement distress and optimization with deep learning based weed detection model for smart
structural defects detection and quantification technologies using imag- agriculture,’’ Appl. Sci., vol. 12, no. 24, p. 12828, Dec. 2022.
ing approaches,’’ J. Traffic Transp. Eng., vol. 9, no. 2, pp. 135–150, [23] T. Chauhan, H. Palivela, and S. Tiwari, ‘‘Optimization and fine-tuning
Apr. 2022. of DenseNet model for classification of COVID-19 cases in medical
[9] C. Dewi, R.-C. Chen, and H. Yu, ‘‘Weight analysis for various prohibitory imaging,’’ Int. J. Inf. Manage. Data Insights, vol. 1, no. 2, Nov. 2021,
sign detection and recognition using deep learning,’’ Multimedia Tools Art. no. 100020.
Appl., vol. 79, nos. 43–44, pp. 32897–32915, Nov. 2020. [24] A. M. Agwa, S. K. Elsayed, and E. E. Elattar, ‘‘Extracting the parameters
[10] B. Abu-Salih, P. Wongthongtham, K. Coutinho, R. Qaddoura, of three-diode model of photovoltaics using barnacles mating optimizer,’’
O. Alshaweesh, and M. Wedyan, ‘‘The development of a road network Symmetry, vol. 14, no. 8, p. 1569, Jul. 2022.
flood risk detection model using optimised ensemble learning,’’ Eng. [25] M. Ahmadi, F. D. Ahangar, N. Astaraki, M. Abbasi, and B. Babaei,
Appl. Artif. Intell., vol. 122, Jun. 2023, Art. no. 106081. ‘‘FWNNet: Presentation of a new classifier of brain tumor diagnosis based
[11] M. Al Duhayyim, A. A. Malibari, A. Alharbi, K. Afef, A. Yafoz, R. Alsini, on fuzzy logic and the wavelet-based neural network using machine-
O. Alghushairy, and H. Mohsen, ‘‘Road damage detection using the hunger learning methods,’’ Comput. Intell. Neurosci., vol. 2021, Nov. 2021,
games search with Elman neural network on high-resolution remote sens- Art. no. 8542637.
ing images,’’ Remote Sens., vol. 14, no. 24, p. 6222, Dec. 2022.

61994 VOLUME 11, 2023