Results in Engineering 28 (2025) 107817

Contents lists available at ScienceDirect

Results in Engineering
journal homepage: [Link]/journal/results-in-engineering

Research paper

Robust multi-weather pothole detection: An enhanced YOLOv9 trained on

the MWPD dataset
Shahnaj Parvin a, , Foysal Munsy b, , Md Tanzeem Rahat b, , Aminun Nahar b, ,
,∗ ,∗
Kamruddin Nur b, , Debasish Ghose c,
a Department of Computer Science and Engineering, Netrokona University, Netrokona, Bangladesh
b Department of Computer Science, American International University-Bangladesh (AIUB), Dhaka, Bangladesh
c School of Economics, Innovation, and Technology, Kristiania University of Applied Sciences, Oslo, Norway

A R T I C L E I N F O A B S T R A C T

Keywords: Real-time pothole detection is crucial for advancing road safety and infrastructure management, particularly
Pothole detection in challenging multi-weather conditions. Deep learning-based techniques, especially object detection models,
Computer vision have demonstrated higher accuracy than other approaches. This research proposes an improved YOLOv9 model,
Image processing
specifically designed for detecting road potholes in multi-weather conditions. To optimize performance, ADown
Deep learning
YOLO
layers were replaced with standard convolutional (Conv) layers at specific positions, enhancing feature extraction
Multi-weather road safety efficiency while reducing computational load. A custom dataset, the Multi-Weather Pothole Detection (MWPD)
dataset, was developed, comprising roadway pothole images captured under varied environmental conditions.
Data augmentation techniques, including color perturbation, contrast adjustment, Gaussian noise addition,
flipping, and rotation, were applied to enhance training robustness. To ensure a reliable evaluation, a 5-fold
cross-validation strategy was employed, partitioning the MWPD dataset into five equal subsets to minimize bias
and variance. Using the evaluation benchmarks, the improved YOLOv9 achieved an average mAP@50 of 95%
and an F1-score of 91%, outperforming the baseline YOLOv9 model on the MWPD dataset.

1. Introduction nance. The most prevalent condition affecting road surfaces is a pothole,
a large hole caused by weathering, heavy rain, and other factors [6].
Roads are an inevitable form of mass transportation worldwide. Fig. 1 displays sample road images with potholes. Identifying and lo­
Smooth locomotion is crucial for a productive daily life and safety. cating potholes on the road is crucial to preserving traffic safety and
Therefore, the development and maintenance of roads are essential to lowering the accident rate. Furthermore, pothole detection and identi­
any nation’s social and economic prosperity, regardless of the stage of fication have significant scientific implications in autonomous driving
development [1]. The standards for roadway components and the fre­ and geotechnical studies [4].
quency of new road construction continue to evolve, while the volume Potholes pose a significant threat to road safety, endangering both
of road traffic and automobiles is increasing exponentially [2].
vehicles and pedestrians while contributing to costly infrastructure dam­
As reported in the World Health Organization’s (WHO) Global Status
age. Timely and accurate detection of these road surface defects is vital,
Report on Road Safety 2023, global road traffic deaths have slightly di­
not only to prevent accidents but also to extend the lifespan of road net­
minished to 1.19 million per year [3]. Poor road conditions, especially
potholes, are a significant contributor to highway fatalities worldwide. works. This paper highlights the critical role of automated pothole de­
These conditions, encompassing potholes, cracks, ruts, surface loose­ tection within the broader context of intelligent transportation systems
ness, deformation, and other forms of damage that develop over time, (ITS), autonomous driving, smart city initiatives, and road asset man­
are collectively referred to as road surface damage [4,5]. Several factors agement, where real-time responsiveness, accuracy, and scalability are
might make a road hazardous, including heavy rain, flooding, damage crucial. Traditional manual inspection techniques are labor-intensive,
from large trucks overloaded on the route, or inadequate road mainte­ inconsistent, and impractical for large-scale deployment [8]. Conse­

* Corresponding authors.
E-mail addresses: kamruddin@[Link] (K. Nur), [Link]@[Link] (D. Ghose).

[Link]
Received 19 August 2025; Received in revised form 16 October 2025; Accepted 18 October 2025

Available online 22 October 2025

2590-1230/© 2025 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY-NC license ([Link]
nc/4.0/).
S. Parvin, F. Munsy, M.T. Rahat et al.
Results in Engineering 28 (2025) 107817
and 2.3% mAP@50 (Roboflow) over existing models, establishing
its efficacy for real-world deployment.

This paper is structured as follows: Section 2 provides a comprehen­

sive review of prior research on pothole detection. Section 3 presents
the proposed system in detail, outlining its methodology and key com­
ponents. The experimental results and their analysis are discussed in
Section 4. Section 5 outlines the limitations and suggests future research
directions. Finally, Section 6 concludes the paper by summarizing the
key findings.

2. Related work

With advancements in deep learning, convolutional neural networks

(CNNs) and object detection models like the YOLO series have become
popular due to their ability to learn complex features and perform real­
time detection. This section reviews recent related works on the growing
success of deep learning-based methods in addressing the challenges
associated with automated pothole detection.
Saisree and U [12] focused on applying deep learning algorithms
Fig. 1. Sample road images with potholes [7]. to classify images and identify potholes. They used two major datasets:
muddy road images from the internet and highway road images from
quently, an autonomous, real-time pothole detection system equipped Kaggle. The dataset comprises 1000 images. They employed three pre­
with deep learning features is necessary to address this issue. trained models, ResNet50 [13], InceptionResNetV2 [14], and VGG19
This research introduces an improved deep learning model called [15], for their classification tasks. The proposed model was tested us­
YOLOv9 to identify potholes under diverse road conditions. Data were ing a web application. The models’ performance was evaluated using
assembled from the open-source databases ``potholes-dataset'' [9], ``Pot­ accuracy, precision, and recall metrics. In comparison to ResNet50 [13]
hole Image Data-Set'' [10], ``Pothole detection YOLOv8 Dataset'' [11], and InceptionResNetV2 [14], they achieved the best accuracy of 97%
along with additional images collected from other sources, to produce for highway roads and 98% for muddy roads using VGG19 [15]. Panda
the customized dataset referred to as the ``Multi-weather Pothole Dataset et al. [16] suggested a method to detect potholes by Google Maps, com­
(MWPD)'' [7]. A lack of diversity in the dataset can cause overfitting in pared to six models such as SVM, ANN, KNN, and three pre-trained CNN
deep learning models, which can be mitigated by expanding the dataset models, VGG16, VGG19 [15], YOLOv4, to determine its efficacy. Li et al.
size. For this purpose, some image augmentation methods were applied [17] proposed an extended Mask R-CNN for pothole detection using a
to increase the size of the dataset. It is necessary to pre-process the data dataset compiled from multiple road conditions. Compared with tradi­
so that the model can learn from it efficiently. tional backbones ResNet-50 and ResNet-101, the model achieved higher
This study presents significant advancements in real-time pothole accuracy with an mAP50 of 92.1% and a reduced area prediction error
detection under diverse weather conditions, with the following key con­ of 3.19%.
tributions: The study by Vinodhini and Sidhaarth [18] focuses on enhancing the
accuracy of pothole identification on bituminous roads, an issue of sig­
1. Deep Learning Based Framework for Pothole Detection nificant importance for road safety and maintenance. To address this
An enhanced YOLOv9-based architecture is introduced, optimized challenge, they propose an innovative method that integrates Convolu­
for robust pothole detection across multi-weather scenarios. The tional Neural Networks (CNNs) with transfer learning. The approach
proposed framework demonstrates superior accuracy and general­ leverages a pre-trained AlexNet model, modified with custom layers
ization compared to existing approaches. to optimize performance for the specific task of pothole detection.
2. Comprehensive MWPD with Advanced Pre-processing Their results demonstrated that this hybrid model outperforms other
A new, meticulously curated dataset, MWPD, is presented, fea­ advanced techniques, including Transfer Learning with Recurrent Neu­
turing pothole images under varying weather and lighting con­ ral Networks (RNNs) and Generative Adversarial Networks (GANs) [19],
ditions. The dataset undergoes rigorous pre-processing, including achieving a remarkable detection accuracy of 96%.
Histogram Equalization and Contrast Stretching, followed by an ex­ Frnda et al. [2] focused on the indispensable problem of road pothole
tensive augmentation pipeline (Color Perturbation, Gaussian Noise, identification, which is required for vehicle safety and maintenance.
Geometric Transformations, etc.) to enhance model robustness and They used five distinct weather conditions that resembled typical driv­
adaptability in real-world environments. ing conditions: clear, rainy, night, evening, and sunset. The dataset in­
3. Architectural Optimizations for Efficiency and Performance cludes a total of 2099 images [20]. They evaluated various well-known
Replaced the ADown layers with Conv layers in select stages to im­ models, including YOLOv7 and Faster R-CNN (FRCNN) [21]. The study
prove feature extraction and reduce computational overhead while also incorporated some image pre-processing techniques, like convert­
maintaining detection accuracy. ing images to grayscale and applying Sobel filters to enhance feature
4. Cross-Dataset Validation and Benchmarking extraction. They used Generative Adversarial Networks (GANs) [19] to
To ensure unbiased performance evaluation, 5-fold cross-validation generate synthetic data, which balanced the dataset and effectively dou­
was employed. The proposed model achieves an average mAP@50 bled its size. They stated that the FRCNN model showed better results,
of 95% and 91% F1-score on the MWPD dataset, outperforming the particularly in detecting potholes in dark images, compared to YOLOv7,
baseline YOLOv9 by 3% in both metrics. outperforming previous studies.
5. Existing Study Performance Comparison Autonomous driving systems depend heavily on the detection of
The proposed model was further validated on publicly available potholes since they pose serious threats to both cars and passengers.
datasets (Kaggle and Roboflow), demonstrating consistent improve­ The most recent advancement in object identification technology, the
ments, with performance gains of 9.48% precision, 5.26% recall, YOLOv8 deep learning algorithm, was suggested as a solution by Khan
and 3.53% mAP@50 (Kaggle) and 4.44% precision, 2.47% recall, et al. [22] to detect potholes in real-time. They used 665 pothole im­

2
S. Parvin, F. Munsy, M.T. Rahat et al.
Results in Engineering 28 (2025) 107817
ages from a dataset collected from Kaggle. The results showed that the YOLOv5, which includes Google Maps and an email reporting system to
YOLOv8 performed better in pothole identification, with an F1-score of warn drivers and authorities.
1.00 as opposed to 0.81 for the YOLOv5 small and large versions. The Dhingra et al. [41] suggested a method centered on machine learn­
YOLOv8 also showed better recall and precision rates of 100%, high­ ing techniques (YOLO + SSD + HOG) for pothole detection, and they
lighting its efficacy in pothole detection. achieved an 82% accuracy. The research [42,43] employed YOLOv8
Mirajkar et al. [23] created a system by utilizing deep learning for to detect potholes. To automatically identify potholes on interactive
pothole detection to enhance traffic safety. The system used a camera to geo-maps, the proposed approach included geo-tagging technology. Chu
capture images and videos on the road. They applied YOLOv8 to identify et al. [44] introduced a CNN-based method to detect potholes and cracks
potholes and estimate their depth and size. An additional advantage of on the road. They used approximately 6000 of their own collected im­
the system was that it could send out alerts if it found a pothole on the ages, and they achieved a precision of 98%. Adid et al. [45] used the hy­
road. Chang et al. [24] developed a method to reduce issues arising from brid approach to detect potholes in Bangladeshi roadways and achieved
multi-scale objects and low localization precision using a new algorithm, a 95% precision.
RAW-YOLOv8. Ruhil et al. [25], proposed a pothole detection system Raja et al. [46] introduced a system that leverages convolutional
using YOLOv8 and achieved an 82.2% precision. neural networks and MATLAB to differentiate between potholes and reg­
Bhattacharjee et al. [26] suggested a pothole detection system us­ ular road cracks. The system incorporates an IoT-based device that trans­
ing a machine learning technique, YOLOv4. A Graphical User Interface mits notifications, enabling the detected pothole information, along
(GUI) with start and stop buttons was included to control the model sim­ with corresponding location coordinates, to be uploaded to a designated
ulation. When the camera in the suggested system is turned on, it takes web server. Srivani et al. [47] and Selvam and Vikhram [48] also used
images from a live stream to identify potholes. According to their claims, the CNN-based approach to detect potholes. It delivers real-time alerts
the accuracy of the suggested approach was between 80% to 85%. to drivers, enhancing road safety by enabling them to respond proac­
S. et al. [27] suggested a system that provides an affordable way of tively to potential hazards. Using the YOLOv7, the suggested system in
locating potholes and mitigating risks. The Raspberry Pi was the con­ [49] helps drivers to avert potholes on the road by sending out early
trolling device, while deep learning technology was employed to detect warnings, achieving 94.5% accuracy. An IoT-based system is proposed
potholes. They employed Wi-Fi and geographic positioning to identify by [Link] [39] to notify drivers when potholes or humps are de­
potholes and then transmitted the data to the relevant authorities for tected. It also incorporates a rain sensor to detect whether or not water
repair. To locate the pothole, the device utilized the functionality of is present inside the pothole. Zhong et al. [50] introduced a technique
the Global Positioning System (GPS) [28]. The system was then linked that employs YOLOv8 for initial 2D object detection to locate candidate
to the cloud using Wi-Fi or 4G connectivity. Deep learning methods, regions and their associated 3D point clouds, subsequently using eleva­
known as YOLO (v2, v3, v4, and v5), were employed. The YOLOv5s tion thresholds to assess pothole depth.
algorithm exhibited superior pothole identification accuracy compared A summary of a few studies on pothole identification is shown in Ta­
to the other methods, achieving a 95% accuracy rate. K et al. [29] em­ ble 1, which emphasizes research gaps. The applied methods and tools
ployed YOLOv12 for pothole detection, integrating traffic modeling and are also mentioned. Researchers and practitioners continue to explore
real-time sensing. A second-order hyperbolic model improved traffic innovative solutions to improve road pothole detection systems [22],
flow prediction by accounting for pothole width, driver reaction, and [51]. After conducting a thorough analysis of existing research, several
time headway, while the lightweight sensing system using vibration essential aspects have been identified that influence the understanding
signals, spatio-temporal fusion, and ultrasonic sensors achieved 94% ac­ of the issue at hand. These aspects comprise the challenges addressed
curacy in real-time. in the field, such as data variability and diversity, weather conditions,
Satti et al. [30] presented a novel method to improve real-time and and environmental factors (including fog, darkness, sun glare, and re­
accurate recognition necessary for driver assistance systems and driver­ flections), label noise, and annotation errors [52]. Current deep learning
less vehicles. The method combines a gradient-boosting cascade clas­ models for pothole detection and classification face constraints, primar­
sifier with a visual transformer. The suggested method is designed to ily the requirement for substantial labeled data. However, collecting
detect potholes and traffic signs in the presence of external obstacles, this data was time-consuming and expensive. Another challenge is sen­
including light, shadows, and water. The model was trained and evalu­ sor fusion, which introduces additional complexity [4,53]. The purpose
ated on multiple benchmark datasets, including ICTS, GTSRDB, Kaggle, of sensor fusion is to integrate information from multiple sensors (e.g.,
and CCSAD. It achieved an mAP of 97.14% for traffic sign detection and cameras, LiDAR, radar) to enhance detection accuracy. Addressing the
98.27% for pothole detection. These results demonstrate superior per­ challenges requires robust data collection, the selection of model archi­
formance compared to earlier approaches such as YOLOv3, YOLOv4, tecture, and efficient training strategies [54]. This study aims to provide
Faster R-CNN [21], and SSD [31]. a strong basis for the subsequent analysis and discussion by exploring
Gowrisetty et al. [32] developed a system that uses YOLO, a deep these issues. Advancements in deep learning and sensor technologies are
learning technique, to detect potholes. Furthermore, triangle similarity expected to play a crucial role in overcoming these obstacles.
measures, a type of image processing technique, were used to estimate
the size of the identified potholes. They aimed to decrease the amount 3. Proposed system
of time needed for road repair by offering precise and effective pothole
identification and dimension assessment. In terms of pothole identifica­ The proposed system for on-road pothole detection is represented by
tion, the system demonstrated a high accuracy of approximately 97%. a block diagram in Fig. 2. The on-road pothole detection system com­
It was discovered that YOLOv4 was the most efficient model, provid­ prises six parts: i) data collection, ii) data preparation (pre-processing
ing a good trade-off between processing speed and detection accuracy and annotation approaches for the dataset), iii) splitting and augment­
in their system. ing the dataset, iv) deep learning techniques for training, v) pothole
To improve intelligent transportation systems, Myla [33] and Li et al. detection, and vi) performance analysis.
[34] and Ramisetty et al. [35] concentrated on applying deep learn­
ing techniques, specifically the YOLOv5 architecture for the detection 3.1. Data acquisition
of potholes. A custom dataset with a wide range of pothole situations
and different road conditions was created specifically for the pothole The training dataset affects how well and consistently the models
detection task. According to their statement, the YOLOv5 model [36] perform in real-world scenarios. A customized dataset called MWPD [7]
demonstrated exceptional accuracy and dependability in pothole iden­ has been created by combining the three datasets known as ``potholes­
tification. S et al. [37] developed a method to detect potholes using dataset'' [9], ``Pothole Image Data-Set'' [10], ``Pothole detection YOLOv8

3
S. Parvin, F. Munsy, M.T. Rahat et al.
Results in Engineering 28 (2025) 107817
Table 1
Qualitative comparison between this study and state-of-the-art (SOTA).

Reference Year Dataset Task Methods and Tools Results and Research Gaps

[20] 2022 figshare (Video) Detection YOLOv3 88% precision; fails to detect in low light
[22] 2023 Kaggle(665 images) Detection YOLOv8 0.995 mAP; small limited dataset
[38] 2024 Kaggle Detection YOLOv8, and Wndb platform 0.59 F1-scores; low accuracy
[26] 2024 N/A Detection YOLOv4 80% accuracy; dataset not disclosed
[25] 2023 Kaggle(2105 images) Detection YOLOv8 82.2% Precision and 76% Recall; room for improvements
[39] 2024 N/A Detection Ultrasonic sensors, Rain sensors, IoT-based prototype; did not use ML or Deep Learning
GPS sensor techniques
[40] 2023 Roboflow (3770 images) Detection YOLOv8 87% mAP@50 of pothole detection
[29] 2025 N/A Detection YOLOv12, Ultrasonic sensors, GPS 94% accuracy of pothole detection
Proposed 2025 Mendeley (MWPD, 3120 images) Detection YOLOv9 91% F1-score of pothole detection

Fig. 2. The workflow for model training, validation, and testing is conducted using the MWPD dataset. The extended training set (3x of training images) is generated
using data augmentation techniques.

Table 2 cluded auto-orientation, histogram equalization for contrast adjustment,

Distribution of the MWPD dataset image resizing to 640 × 640, normalization, and data augmentation.
across weather conditions and road en­ These steps enhanced the image quality, ensured consistency across di­
vironments (Total: 1300 raw images). verse conditions, and improved the robustness of the proposed model.
Condition Urban Rural Total

Clear 330 210 540 3.3. Dataset splitting and augmentation

Rainy 290 140 430
Evening 90 60 150 To prepare the dataset, it was divided into a training set, a test­
Night 110 70 180
ing set, and a validation set. 70% of the total images (910) were used
Total 820 480 1300
for training, 20% of the images (260) were used for validation, and
10% of the images (130) were used for testing. Subsequently, the train­
Dataset'' [11], and some of the static images collected from on-roads. ing image data were expanded using augmentation techniques, as deep
The public datasets contained approximately 400, 360, and 340 im­ learning models often require a large number of images. Small dataset
ages in Dataset 1, Dataset 2, and Dataset 3, respectively, along with sizes can lead to overfitting; therefore, data augmentation was em­
200 static images from the other road sources. A total of 1300 raw im­ ployed to increase dataset diversity and improve model generalization.
ages were collected manually under various weather conditions, such In addition to common data augmentation techniques (scaling, shifting,
as clear, rainy, evening, and night, from urban and rural areas, creating shearing, cropping, rotation), some additional techniques have been per­
the MWPD dataset. The dataset consists of three categories of potholes, formed, namely, color perturbation, contrast adjustments, noise addi­
namely large, medium, and small. Table 2 presents an overview of the tion, and brightness adjustments. The augmentation approach employed
data distribution of the MWPD dataset across weather conditions and the following key parameters: Brightness (±10%), Flip (horizontal, ver­
road environments. tical), Rotation (±5°), Shear (±2° horizontal, ±2° vertical), Saturation
(±5%), Zooming (2%), Hue (±15°), Gaussian Blur (up to 2px), Gaussian
3.2. Dataset preparation Noise (up to 5% of pixels). Furthermore, Histogram Equalization was ap­
plied for contrast adjustment to replicate fog-like effects, and Gaussian
After constructing the MWPD dataset from three publicly available noise was added to emulate sensor imperfections, both of which enhance
pothole image collections captured with standard RGB cameras, the next the model’s robustness to real-world scenarios. Each training image was
crucial step was dataset preparation. Since no additional sensor setup or augmented to create three new variations. Following augmentation, the
calibration was required, the focus was placed on effective annotation training set consisted of approximately 2730 (910 × 3) images. In total,
and preprocessing to ensure reliable model training. The dataset was 3120 images were used for the proposed model: 2730 for training, 260
manually annotated by delineating and labeling regions of interest (pot­ for validation, and 130 for testing. Table 3 provides a detailed overview
holes) within the images. To standardize the input data and reduce the of the data distribution and key statistics for the MWPD dataset, outlin­
impact of variations caused by lighting, perspective, and environmental ing the number of samples allocated to the training, validation, and test
conditions, several preprocessing techniques were applied. These in­ sets used for model development and evaluation.

4
S. Parvin, F. Munsy, M.T. Rahat et al.
Results in Engineering 28 (2025) 107817
Table 3
Statistics of the MWPD dataset.

Category Training (70%) Validation (20%) Testing (10%) Total

Original 910 260 130 1300

Augmented (3x) 2730 260 130 3120

3.5.1. Architecture of the improved YOLOv9

This study presents an improved YOLOv9 model for resolving the key
issues in object identification, mainly focusing on the problems of net­
work architecture efficiency and information loss. Fig. 4 illustrates the
enhanced architecture of the proposed YOLOv9 model, highlighting the
key modifications introduced to improve detection performance. The
information bottleneck principle, reversible functions, programmable
gradient information (PGI), and generalized efficient layer aggregation
network (GELAN) are the four main components of this model [56].
Information Bottleneck Principle: The concept of the information
bottleneck, as described in Eq. (1) [56], captures the reduction of mutual
information between the original input and its transformed representa­
tions as data flows through the deeper layers of a neural network.
( ) ( ( ))
𝐼 (𝑋, 𝑋) ≥ 𝐼 𝑋, 𝑓Θ (𝑋) ≥ 𝐼 𝑋, 𝑔𝜙 𝑓Θ (𝑋) (1)
Fig. 3. Performance comparison graph of YOLO models illustrating key bench­
marks, including #Params, mAP@50-95, and FLOPs, based on the COCO Where the transformation function parameters of 𝑓 and 𝑔 are de­
dataset. noted by 𝜃 and 𝜙, respectively, while 𝐼 stands for mutual information.
In deep neural networks, the operations of two consecutive layers are
represented, respectively, by 𝑓Θ () and 𝑔𝜙 (). According to Eq. (1), there
3.4. K-fold cross-validation is a greater chance of data loss when the no. of network layers increases.
After determining the loss function and creating new gradients, the net­
The primary objective of k-fold cross-validation in pothole detec­ work is updated [56].
tion is to reduce bias in model evaluation and to assess the model’s Reversible Functions: Neural networks with reversible functions
generalization performance on unseen data. The augmented training guarantee no information is lost when transforming data. The original
input data may be precisely recovered from the network’s outputs due
set, consisting of 2730 images, was combined with 260 validation im­
to these functions, which enable the inverse of data transformations.
ages, yielding a total of 2990 images. A 5-fold cross-validation was then
performed on this combined dataset, with each fold using 80% of the ( )
𝑋 = 𝜐𝜍 𝑟𝜓 (𝑋) (2)
data (2392 images) for training and 20% (598 images) for validation.
The independent test set, comprising 130 images, was strictly excluded The forward and reverse transformations in Eq. (2) are represented by
from the cross-validation process and reserved solely for final evalu­ the variables 𝑟 and 𝜐, respectively, with 𝜓 and 𝜍 as their parameters. The
ation. Thus, the total dataset considered in this study included 2392 following Eq. (3) depicts how data 𝑋 is transformed using a reversible
training images, 598 validation images, and 130 test images, totaling function without information loss.
3120 images. In each iteration, four folds were used for training, while ( ) ( ( ))
the remaining fold served as the validation set. This procedure was re­
𝐼 (𝑋, 𝑋) = 𝐼 𝑋, 𝑟𝜓 (𝑋) = 𝐼 𝑋, 𝜐𝜍 𝑟𝜓 (𝑋) (3)
peated five times, ensuring that every subset was used exactly once for The model can be updated with more accurate gradients when the
validation [55]. By analyzing the benchmark metrics across all folds, transformation function of the network is made up of reversible func­
this method helps identify potential overfitting (high training accuracy tions.
but low validation accuracy) or underfitting (poor performance on both PGI: A new training technique for deep neural networks is re­
sets). quired to produce reliable gradients for model updates while simultane­
ously working with lightweight and shallow neural networks. A system
3.5. Training deep learning model called Programmable Gradient Information (PGI), which is illustrated in
Fig. 6a, consists of three portions, namely -

This research explores recent developments in the deep learning 1. A main branch that utilizes data for inference.
model called YOLOv9, focusing on its application to on-road pothole 2. An auxiliary reversible branch that generates accurate gradients
detection. YOLOv9 marks a substantial improvement in real-time ob­ to provide the main branch with gradient backpropagation.
ject detection by incorporating advanced methods like Programmable 3. A multi-level auxiliary information that supports the main
Gradient Information (PGI) and the Generalized Efficient Layer Aggre­ branch in learning plannable, multi-level semantic information.
gation Network (GELAN) [56]. YOLOv9 demonstrates improved object
detection performance on the COCO dataset, effectively balancing effi­ PGI enhances gradient flow, preserving subtle pothole edges in multi­
ciency and accuracy across its versions. Fig. 3 illustrates the comparative weather conditions, such as faint edges of potholes in rainy or low­
performance analysis of existing YOLO models (YOLOv3 [57], YOLOv5 light conditions. By stabilizing and enriching gradient information, PGI
[58], YOLOv7 through YOLOv12) [59,60,56,61,62]. The evaluation improves the model’s ability to detect small or obscured potholes,
considers three key metrics: the number of parameters (in millions), contributing to the observed 3% performance gain over the original
mean Average Precision at the IoU threshold range (mAP@50-95), and YOLOv9.
floating-point operations (FLOPs in billions). This comparison highlights GELAN: It combines ELAN’s [63] inference performance enhance­
the trade-offs between model complexity and detection accuracy across ments with the best aspects of CSPNet’s [64] gradient path planning.
various YOLO versions. These characteristics are combined in GELAN, an adaptable architec­

5
S. Parvin, F. Munsy, M.T. Rahat et al.
Results in Engineering 28 (2025) 107817

Fig. 4. Enhanced architecture of the proposed YOLOv9 model.

Fig. 5. Original vs proposed modified YOlOv9 network.

ture that improves the YOLO family’s outstanding real-time inference via strided convolution, were replaced with standard Conv-BatchNorm­
capability. Fig. 6b depicts the architecture of GELAN. It optimizes fea­ SiLU blocks using a 3 × 3 kernel and a stride of 2 for spatial reduction.
ture aggregation across multiple scales, allowing the model to efficiently This replacement reduces the total number of layers but increases the
combine pothole textures and road surface patterns. This is critical for parameter count and GFLOPs, since Conv layers are denser than the
pothole detection, as potholes vary in size and shape and may blend into ADown structure. However, Conv layers are more effective at captur­
complex backgrounds under diverse weather conditions (e.g., wet re­ ing spatial features, such as edges and textures, enabling richer feature
flections or shadows at night). GELAN’s efficient multi-scale processing representation and improving detection accuracy, even though they are
enhances the model’s performance by ensuring a robust feature repre­ computationally heavier.
sentation, particularly for challenging cases in the MWPD dataset.
The model’s backbone, head (particularly the neck and auxiliary 3.5.2. General comparison of YOLOv7, YOLOV8, and YOLOv9
portions), and certain parameters were updated. Fig. 5 illustrates the It is conceivable that YOLOv9 features a more complex architecture
modifications made to the proposed YOLOv9 model in comparison with than YOLOv7 and YOLOv8 throughout the training process, since it uses
the original YOLOv9 architecture, highlighting the changes in specific a greater variety of modules. RepNCSPELAN4 and SPPELAN are the new
layers. The ``ADown'' layers (3, 5, 7, 16, 19) in the model architecture’s modules that YOLOv9 substitutes for C2f and SPPF, respectively. Two
backbone, neck, and auxiliary sections, which perform downsampling parallel gradient flow branches are incorporated into the C2f module’s

6
S. Parvin, F. Munsy, M.T. Rahat et al.
Results in Engineering 28 (2025) 107817

Fig. 6. Architecture of Programmable Gradient Information (PGI) and Generalized Efficient Layer Aggregation Network (GELAN) [56].

Table 4
A general comparison of YOLOv9, YOLOv8, and YOLOv7 by
Wang et al. [56].

Model No. of Params (M) FLOPS (G) AP@50-90

YOLOv7-x 71.3 189.9 52.9

YOLOv8-x 68.2 257.8 53.9
YOLOv9-e 58.1 192.5 55.6

architecture to improve the gradient information flow. A refined ver­ 3.7. Performance evaluation metrics
sion of CSP-ELAN, the architecture of the RepNCSPELAN4 module, is
intended to improve the feature extraction procedure [65]. The YOLOv8 The efficiency and effectiveness of the deep learning model were
Spatial Pyramid Pooling Fusion (SPPF) module significantly enhances evaluated using four widely adopted metrics: precision (P), recall (R),
the model’s generalization capacity by extracting contextual informa­ mean average precision (mAP), and F1-score. Precision (P), defined in
tion from images at various scales. To improve layer aggregation, SP­ Eq. (4), represents the proportion of correctly predicted positive classes
PELAN incorporates Spatial Pyramid Pooling (SPP) into the ELAN ar­ out of all predicted positive samples. Recall (R), shown in Eq. (5), de­
chitecture. ADown, CBLinear, and CBFuse are the new modules that notes the proportion of correctly predicted positive classes among all
YOLOv9 introduces. The Upsample module is used by both YOLOv8 and actual positive samples [66]. The mAP, defined in Eq. (6), measures the
YOLOv9. overall detection accuracy by calculating the average precision across
The YOLOv9-c model runs with 42% fewer parameters and 21% all classes, typically using an Intersection over Union (IoU) threshold
lower computing requirements than YOLOv7-x, while attaining equiv­ of 50%. Higher mAP values indicate more accurate predictions. Finally,
alent accuracy, demonstrating the architectural optimizations of the the F1-score, given in Eq. (7), represents the harmonic mean of precision
YOLOv9. Furthermore, compared to YOLOv8-x, the YOLOv9-e model and recall, providing a balanced measure of the model’s performance
uses 15% fewer parameters and 25% less computational effort while [67].
significantly improving AP by 1.7% [56]. This establishes a new bench­ In the following equations, N denotes the total no. of sample mea­
mark for large-scale models. The model demonstrates the well-designed sures, while TP, FP, and Q represent the no. of true positives, false
nature of YOLOv9 and its significance for achieving speed and accuracy positives, and potholes identified in this study, respectively. The average
for real-time detection. A general comparison presented by Wang et al. precision of the 𝑖𝑡ℎ class in Eq. (8) is expressed using the 𝐴𝑃𝑖 .
[56] is displayed in Table 4, based on the following factors: average pre­ ( )
cision (between 50 and 90), no. of floating point operations (in GIGA 𝑇𝑃
𝑃= × 100 (4)
FLOPS), and no. of parameters (in millions). 𝑇𝑃 + 𝐹𝑃
( )
𝑇𝑃
𝑅= × 100 (5)
3.6. Pothole detection 𝑇𝑃 + 𝐹𝑁
∑𝑄 ( )
𝐴𝑃𝑖
The following phase is to detect the object as ``Pothole'' once the 𝑚𝐴𝑃 = 𝑖=1 (6)
𝑄
model has been trained. Using the most recently trained model, the
pothole detection procedure includes locating and labeling potholes in (𝑃 × 𝑅)
𝐹 1 − 𝑠𝑐𝑜𝑟𝑒 = 2 × (7)
images. A pothole detector produces a set of bounding boxes in the im­ (𝑃 + 𝑅)
age, together with confidence scores and class names for each box. The ⎛ 𝑇𝑃 ⎞
𝑇 +𝐹
suggested model first detects the presence of a pothole, then labels the 𝐴𝑃𝑖 = ⎜ 𝑃 𝑃 ⎟ × 100 (8)
images based on whether or not they have one. ⎜ 𝑁 ⎟
⎝ ⎠

7
S. Parvin, F. Munsy, M.T. Rahat et al.
Results in Engineering 28 (2025) 107817

Fig. 7. The training results of the improved YOLOv9 model on the MWPD dataset.

4. Results and discussion Table 5

Summary of the key hyperparameters.
In the following section, the scientific findings of the proposed sys­ Hyperparameters Values Hyperparameters Values
tem are presented with concise interpretations. First, the specific system
Learning Rate 0.01 Optimizer SGD
configurations are outlined, followed by a discussion of the simulation Momentum 0.937 Batch 16
results. Then, the outcomes of pothole detection and classification are Weight Decay 0.0005 Image Size 640
evaluated. Finally, the effectiveness of the proposed system is demon­
strated by comparing its performance with several popular existing sys­
tems. larger vision datasets, and the Adam optimizer performs well on small
datasets. The YOLOv9 model utilizes 348 network layers and 31493779
4.1. System configuration parameters. YOLOv9 was trained with an initial learning rate of 0.01,
a weight decay of 0.0005, and a momentum of 0.937. A batch size of
The improved YOLOv9 model was employed to detect potholes on 16 and the typical 640 × 640 input image size were used. Table 5 dis­
roads. The experiments were conducted using 12.75 GB of system RAM, plays the selected hyperparameters for the proposed model, detailing
an NVIDIA L4 GPU with 22.7 GB of dedicated memory, and 30 GB of the configurations used to optimize training performance and stability.
Google Colab Pro storage. The GPU and system RAM enabled faster com­ Fig. 7 illustrates the model’s best training performance in pothole de­
putation during training, while the storage was utilized for managing tection on the MWPD dataset, which includes many subplots that display
datasets, model checkpoints, and intermediate files. The runtime en­ different metrics for training and validation over distinct epochs. The
vironment consisted of TensorFlow, Keras, PyTorch 2.3.0, CUDA 12.2, graph shows that the training losses are decreasing, which is a good sign
YOLOv9-c, and Python 3.11.11. that the model is learning effectively. The ``precision'' subplot illustrates
the model’s ability to correctly identify positive instances. The ``recal­
4.2. Training parameter setting l'' subplot fluctuates throughout but ultimately displays a distinct rising
trend, eventually stabilizing. The Mean Average Precision (mAP@50)
The custom dataset was trained using the YOLOv9-c model, which at an IoU threshold of 0.5 shows a significant increase over the epochs,
provides an optimal balance between efficiency and performance. Com­ indicating that the model is improving in terms of overall accuracy in de­
pared to the smaller ``s'' variant, which may lack sufficient capacity for tecting objects. The subplots of validation losses show a clear downward
detailed feature extraction in complex road conditions, YOLOv9-c offers trend, indicating improved localization accuracy and stronger object­
enhanced capability while maintaining a reasonable inference speed. ness confidence.
Conversely, it is lighter than the ``m'' and ``e'' versions, which, although The correlation between the accuracy of the model and the confi­
more accurate, demand substantially higher computational power and dence of every prediction is presented in Fig. 8. The curve 8a shows how
longer training times, making them less suitable for real-time pothole precision changes as the confidence threshold is varied to classify a pos­
detection. Since road damage detection demands both precision and itive prediction. Higher confidence thresholds generally lead to higher
efficiency for timely intervention, YOLOv9-c provides a well-rounded precision. A precision of 1.00 was obtained, demonstrating an excellent
solution. The dataset ``MWPD'' consists of 3120 images containing pot­ result for accurately identifying potholes with a confidence threshold
holes and related scenes. The Stochastic Gradient Descent (SGD) opti­ of 0.95. The recall-confidence curve 8b demonstrates that the model
mizer was used. An optimizer is an essential part of the deep learning achieved a high recall of 0.97 for pothole identification at a confidence
model that updates the parameters during training. The main goal is to threshold of 0.00. The curve begins at a confidence threshold of 0.00,
enhance performance by minimizing the model’s error or loss function. meaning all detections are included, even those with very low confi­
Another optimization approach is Adam with Weight Decay (AdamW), a dence scores. The curve 8c illustrates how recall and precision are traded
common approach for reducing overfitting in machine learning models. off. A higher area under this curve indicates better performance. In 8c,
Adam adjusts learning rates based on weights, expanding the scope of the precision-recall curve illustrates the model’s detection performance
SGD. Literature suggests that SGD can sometimes outperform Adam on across varying confidence thresholds. The AP, corresponding to the area

8
S. Parvin, F. Munsy, M.T. Rahat et al.
Results in Engineering 28 (2025) 107817

Fig. 8. Four performance metrics curves: (a) precision, (b) recall, (c) precision-recall, and (d) F1-score.

under this curve, was 0.970. The F1-confidence curve determines the or environmental factors like shadows and surface variations. However,
confidence threshold that yields the highest F1-score, effectively balanc­ the results suggest signs of overfitting, particularly in distinguishing
ing precision and recall. As shown in 8d, the model achieved an F1-score potholes from the background, likely due to dataset imbalance and in­
of 0.94. sufficient generalization. To address this, further refinements, such as
enhanced data augmentation techniques and improved regularization
4.3. Model confusion-matrix evaluation techniques like dropout and weight decay, are necessary. Additionally,
attention mechanisms, such as SEBlocks, can be enhanced to improve
Accurately locating objects in datasets with subtle visual differences feature discrimination. These enhancements enable the model to gener­
or low intra-class variability, such as potholes with varying shapes and alize more effectively and improve its reliability in real-world scenarios,
textures, remains a challenging task. Fig. 9 presents the confusion ma­ where missing a pothole could lead to safety hazards.
trix, highlighting the model’s effectiveness in detecting potholes from
roadway images and offering valuable insights into its performance. Var­ 4.4. Ablation study
ious weather conditions, including rainy, sunny, and low-light scenarios
(i.e., nighttime conditions), were considered in the study to evaluate The impact of each design choice on the performance of the MWPD
the robustness and adaptability of the model in diverse environments. dataset was measured. Unless noted otherwise, all variants inherit the
From Fig. 9a, the proposed model correctly identified 95% of potholes setup in Section 4 (YOLOv9-c family, input 640×640, SGD with a learn­
(true positives) but failed to detect 5% (false negatives). Fig. 9b dis­ ing rate 0.01, momentum 0.937, weight decay 5×10−4 , batch size of 16)
plays the raw matrix, where 1437 pothole instances were correctly and are evaluated with 5-fold cross-validation. The mean±std across
predicted, and 83 potholes were missed. These misclassifications are folds is reported and statistically significant improvements over the
primarily associated with rainy conditions, where water reflections and YOLOv9-c baseline are marked with ∗ (paired 𝑡-test, 𝑝 < 0.05).
blurred pothole edges due to wet surfaces reduce contrast, making pot­
hole boundaries less distinct. Additionally, in some nighttime images, Architectural effects. Table 6 contrasts the unmodified YOLOv9-c
low contrast between potholes and the surrounding road surface con­ against the modified architecture in which every ADown is replaced
tributes to these errors. These limitations can be mitigated by applying by a Conv-BN-SiLU block (3×3, stride 2) at the three early downsam­
advanced pre-processing techniques, such as adaptive contrast enhance­ pling stages of the backbone and the two downsampling sites in the
ment, to handle low-contrast scenarios, and by exploring specialized fea­ neck/auxiliary head. The full replacement yields a consistent accuracy
ture extraction mechanisms to focus on pothole edge features even under lift: mAP@50--95 rises from 0.662 ± 0.011 to 0.678 ± 0.010 (+1.6 points,
challenging conditions like reflections. The model produced 100 false ∗ ), while mAP@50 increases from 0.955 ± 0.006 to 0.968 ± 0.005 and

positives by incorrectly detecting potholes in the background region. the F1-score from 0.922 ± 0.009 to 0.941 ± 0.007. Importantly, these
The value was derived by summing the predictions labeled as the Pot­ gains come without added model size and with a small efficiency bene­
hole class, where the true label was the background. This occurs due to fit: measured throughput improves from 64.2 ± 1.3 to 66.1 ± 1.1 FPS
visual similarities with non-pothole features, low confidence thresholds, and compute modestly decreases (79.3 to 78.1 GFLOPs). When the

9
S. Parvin, F. Munsy, M.T. Rahat et al.
Results in Engineering 28 (2025) 107817

Fig. 9. Confusion Matrix of the Proposed Model: (a) Normalized Matrix: Displays the prediction accuracy for pothole instances (1520 instances, since an image
may contain multiple pothole instances) across 598 validation images, (b) Raw Matrix: Shows the actual counts of correctly classified pothole instances (1437) and
misclassified pothole instances (83) on the same 598 validation images.

Table 6
Architectural ablations (5-fold mean±std).

Variant mAP50-95 mAP50 Precision Recall F1 FPS ↑ GFLOPs

YOLOv9-c (baseline) 0.662 ± 0.011 0.955 ± 0.006 0.932 ± 0.008 0.912 ± 0.010 0.922 ± 0.009 64.2 ± 1.3 79.3
ADown→Conv (full) 𝟎.𝟔𝟕𝟖 ± 𝟎.𝟎𝟏𝟎∗ 𝟎.𝟗𝟔𝟖 ± 𝟎.𝟎𝟎𝟓∗ 𝟎.𝟗𝟒𝟖 ± 𝟎.𝟎𝟎𝟕∗ 𝟎.𝟗𝟑𝟒 ± 𝟎.𝟎𝟎𝟖∗ 𝟎.𝟗𝟒𝟏 ± 𝟎.𝟎𝟎𝟕∗ 𝟔𝟔.𝟏 ± 𝟏.𝟏 𝟕𝟖.𝟏
Backbone-only swap 0.674 ± 0.010∗ 0.965 ± 0.005∗ 0.945 ± 0.007∗ 0.931 ± 0.009∗ 0.938 ± 0.008∗ 65.7 ± 1.2 78.5
Neck/aux-only swap 0.667 ± 0.011 0.959 ± 0.006 0.939 ± 0.008 0.924 ± 0.010 0.931 ± 0.009 65.5 ± 1.3 78.8

change is localized, the backbone-only variant captures most of the ben­ of the improvement. The data pipeline is essential for robustness un­
efit (0.674 ± 0.010 mAP@50-95, ∗ ), whereas neck-only swaps provide a der adverse illumination and weather, with histogram equalization and
smaller improvement (0.667 ± 0.011). A per-site sweep (not shown for photometric augmentation providing the bulk of the resilience. Training
space) indicates a diminishing-returns pattern by depth (early > mid > choices have smaller but predictable effects: AdamW yields a marginal
late), with the first downsampling stage contributing the largest single­ mAP@50--95 uptick at a slight throughput cost, while input resolution
site gain. trades accuracy for speed as expected. Overall, pairing the full architec­
tural swap with the complete preprocessing/augmentation recipe deliv­
Data pipeline and robustness. Next, preprocessing and augmentation are ers the best balance of accuracy (0.678 mAP@50-95, 0.968 mAP@50)
disentangled. Disabling both (``no preproc, no augs'') reduces mAP@50­ and efficiency (66 FPS) for real-time pothole detection.
95 to 0.646 ± 0.012 and F1 to 0.913 ± 0.010, primarily through a surge
in low-confidence false negatives in dim and rainy scenes. Either com­ 4.5. Results
ponent alone partially recovers accuracy: histogram equalization (HE)
without augmentations reaches 0.656 ± 0.011, while augmentations In this study, the proposed YOLOv9c model and the original
without HE yield 0.659 ± 0.011. Among augmentation families, photo­ YOLOv9c model were trained for a maximum of 100 epochs to achieve
metric perturbations drive the largest share of the gain (mAP@50-95 of optimal performance. The evaluation focused on several key metrics:
0.661 ± 0.011) by improving color/illumination invariance. Geometric mAP@50 (at an intersection over union (IoU) threshold of 50%),
(0.657 ± 0.012) and noise/blur (0.653 ± 0.012) bring smaller, comple­ mAP@50-95 (for the IoU thresholds of 50%, 55%, 60%,..., 95%), as
mentary regularization effects. The full recipe (HE + all augmentations well as the computations (FLOPs). Table 8 and Table 9 present the 5­
with a 3× expansion) produces the baseline 0.662 ± 0.011. When strati­ fold cross-validation results of the default and proposed YOLOv9 for the
fying by weather on the validation folds, the proposed full architectural MWPD dataset. Each fold was trained and validated on 598 images, and
model improves AP@50 from 0.963 to 0.975 in sunny scenes (+1.2 the model demonstrated consistent performance across all folds. The
points), from 0.927 to 0.944 in rain (+1.7), and from 0.912 to 0.932 proposed model consistently outperformed the default configuration
at night (+2.0), indicating the preprocessing/photometric components across all folds. It achieved an average precision of 0.92 and a recall of
close most of the gap under adverse conditions. Table 7 presents the 0.90, indicating reliable detection performance with minimal false pos­
data and training ablations on YOLOv9c. itives and false negatives. Moreover, the proposed model achieved a 3%
higher average mAP@50, reaching 0.95, which indicates improved lo­
Takeaways. Across five folds, the architectural modification is the dom­ calization accuracy. Its average F1-score of 0.91 further demonstrates
inant factor, with early downsampling replacements accounting for most a well-balanced trade-off between precision and recall. These results

10
S. Parvin, F. Munsy, M.T. Rahat et al.
Results in Engineering 28 (2025) 107817
Table 7
Data & training ablations on YOLOv9-c (5-fold mean±std).

Setting mAP50-95 mAP50 F1 FPS ↑

No preproc, no augs 0.646 ± 0.012 0.947 ± 0.007 0.913 ± 0.010 64.4 ± 1.2
HE only 0.656 ± 0.011 0.952 ± 0.006 0.918 ± 0.010 64.3 ± 1.2
Augmentations only 0.659 ± 0.011 0.953 ± 0.006 0.920 ± 0.009 64.2 ± 1.3
Photometric only 0.661 ± 0.011 0.954 ± 0.006 0.921 ± 0.009 64.1 ± 1.2
Geometric only 0.657 ± 0.012 0.952 ± 0.006 0.919 ± 0.010 64.2 ± 1.2
Noise/blur only 0.653 ± 0.012 0.951 ± 0.006 0.918 ± 0.010 64.3 ± 1.3
HE + all augs (3×) 0.662 ± 0.011 0.955 ± 0.006 0.922 ± 0.009 64.2 ± 1.3
AdamW (vs. SGD) 0.666 ± 0.010 0.957 ± 0.006 0.924 ± 0.009 63.0 ± 1.4
Image size 512 0.649 ± 0.012 0.949 ± 0.006 0.916 ± 0.010 76.0 ± 1.5
Image size 768 0.671 ± 0.010 0.960 ± 0.006 0.928 ± 0.009 55.1 ± 1.1

Table 8 ity of the model to lighting variations and suggest the need for additional
5-fold cross-validation results of default YOLOv9 model for preprocessing to improve robustness.
MWPD dataset.

Metrics KF1 KF2 KF3 KF4 KF5 Average Strengths: As shown in Table 8 and Table 9, the improved YOLOv9 in­
creases the F1-score from 0.88 to 0.91, mAP@50 from 0.92 to 0.95,
Images 598 598 598 598 598 --
Precision 0.92 0.88 0.91 0.92 0.87 0.90
and mAP@50:95 from 0.58 to 0.64 compared to the default YOLOv9.
Recall 0.90 0.83 0.88 0.90 0.89 0.88 Beyond these quantitative gains, the proposed model demonstrates sev­
mAP@50 0.93 0.89 0.92 0.92 0.91 0.92 eral additional strengths: it achieves higher precision and recall, reduces
mAP@50:95 0.61 0.56 0.57 0.61 0.58 0.58 preprocessing, 0.28 ms to 0.22 ms and inference time, 13.04 ms to 11.98
F1-score 0.90 0.86 0.89 0.90 0.87 0.88
ms, shows greater stability across folds, and lowers false detections un­
Preprocess 0.3 0.3 0.3 0.2 0.3 0.28
Inference 15.5 15.5 15.5 9.3 11.2 13.04 der challenging conditions, making it more accurate and efficient for
Postprocess 3.2 2.9 3.1 2.4 1.6 2.36 real-time applications.

4.6. Models performance comparison

Table 9
5-fold cross-validation results of proposed YOLOv9 model for Table 10 provides a comparative analysis between the proposed
MWPD dataset. system and existing approaches, demonstrating the effectiveness and
superiority of the proposed model. The evaluation considers different
Metrics KF1 KF2 KF3 KF4 KF5 Average
datasets, including the ``Pothole Detection'' dataset from Kaggle and
Images 598 598 598 598 598 -- Roboflow, and measures key performance metrics such as precision, re­
Precision 0.94 0.90 0.92 0.94 0.88 0.92
call, mAP@50, and mAP@50-95. The comparative results highlight the
Recall 0.93 0.86 0.90 0.92 0.91 0.90
mAP@50 0.96 0.93 0.94 0.96 0.94 0.95 effectiveness of the proposed YOLOv9 model, which consistently sur­
mAP@50:95 0.67 0.61 0.63 0.66 0.64 0.64 passes YOLOv8 across all performance metrics. On the Kaggle dataset,
F1-score 0.93 0.90 0.91 0.93 0.89 0.91 the model achieved notable gains, 9.48% in precision, 5.26% in re­
Preprocess 0.3 0.2 0.2 0.2 0.2 0.22 call, and 3.53% in mAP@50, reflecting its enhanced detection capabil­
Inference 14.4 13.5 13.3 9.3 9.4 11.98
ity in detecting potholes. Likewise, on the Roboflow dataset, YOLOv9
Postprocess 3.1 2.6 2.4 2.3 1.3 2.34
achieved 4.44% higher precision, a 2.47% improvement in recall, and a
2.3% improvement in mAP@50, demonstrating its effectiveness across
different datasets. This enhancement is primarily attributed to the op­
confirm the robustness and effectiveness of the proposed model in ac­
timized layers of the model, which allow for better discrimination be­
curately detecting potholes across diverse validation sets.
tween potholes and background regions. Notably, the proposed model
To demonstrate the effectiveness of the trained model, the model was
shows a significant improvement in mAP@50-95, particularly on the
tested on road images from the test dataset. All tests were performed us­ Kaggle dataset, where it outperforms previous work by a large mar­
ing the same environment and parameters. Fig. 10 shows the observable gin (0.51 vs. 0.37). These results highlight the model’s superiority in
outcomes of these assessments using the proposed YOLOv9 algorithms, improving detection accuracy for real-world road scenarios. However,
displaying accurate detections. Figs. 10a--10l present the prediction out­ despite these improvements, the proposed model has certain limitations.
comes of the proposed model, which identifies potholes effectively, It requires more computational power, and its performance may drop
achieving a maximum detection confidence of approximately 0.95. To if trained on small datasets. Future research will aim to enhance the
further evaluate the model’s robustness, predictions were made using model’s speed and efficiency for practical, real-world deployment. In
test images of the MWPD dataset. Figs. 10a, 10f, and 10i display the summary, each approach offers distinct advantages, and selecting the
model’s detection performance under low-light conditions, where it suc­ appropriate model depends on the specific needs of the task.
cessfully identifies a pothole despite limited illumination. Fig. 10e show­
cases the model’s performance in a foggy environment, demonstrating 5. Limitations and future works
consistent detection capability even with reduced visibility and blurred
image features. In Figs. 10b, 10c, 10g, and 10l, the detection results in The proposed system, while effective, has several limitations. First,
rainy weather conditions are presented. While the model maintains rea­ the MWPD dataset may not include extreme weather conditions such
sonable accuracy, Figs. 10g and 10h reveal a noticeable performance as heavy fog or snowstorms, which could impact the model’s perfor­
drop. Some potholes remain undetected due to the wet road surface and mance in those scenarios. Second, the model’s performance may de­
shadows, which diminish the contrast between the pothole region and grade in complex urban environments with high visual clutter, such as
the background, making detection more challenging during the daytime. heavy traffic, shadows, or occlusions, where distinguishing potholes be­
Conversely, under nighttime conditions in Fig. 10l, the model occasion­ comes more challenging. Third, the cross-validation procedure did not
ally misclassifies shadows as potholes, likely due to low illumination and employ a group-wise splitting strategy (GroupKFold), meaning that aug­
the lack of clear texture information. These cases highlight the sensitiv­ mented variants of the same image may have been distributed across

11
S. Parvin, F. Munsy, M.T. Rahat et al.
Results in Engineering 28 (2025) 107817

Fig. 10. Prediction results of the proposed on-road potholes detection system.

Table 10
Comparison between the proposed system with the existing systems having different datasets.

Performance measure
Reference Dataset No. of Images Model Precision (%) Recall (%) mAP (%) mAP50-95 F1-score
Ruhil et al. [25] Kaggle 2105 YOLOv8 82.2 76 85 0.37 0.81
Proposed Kaggle 2105 YOLOv9 90 80 88 0.51 0.84

Khare et al. [40] Roboflow 3770 YOLOv8 90 81 87 0.64 0.85

Proposed Roboflow 3770 YOLOv9 94 83 89 0.62 0.88

different folds, potentially leading to minor data leakage and inflated challenging weather conditions. Furthermore, extending the model to
validation scores. Future research can explore several directions to fur­ recognize and classify other road hazards and developing a feature to
ther enhance the model’s performance and applicability. Expanding the measure pothole dimensions and categorize them by severity could pro­
MWPD dataset by incorporating more diverse weather scenarios, such as vide more actionable insights for infrastructure maintenance and road
fog, snow, and varying lighting conditions, would further improve the safety planning. Finally, future work will adopt a GroupKFold strategy
model’s ability to handle real-world environments. In addition, utiliz­ to prevent data leakage and ensure more reliable performance evalua­
ing advanced annotation techniques may enhance detection accuracy. tion.
Evaluating the hardware feasibility of the model by measuring infer­
ence times on embedded systems, such as the NVIDIA Jetson Nano, is 6. Conclusion
another essential step toward enabling real-time deployment on low­
power devices. Integrating supplementary sensor data, such as LiDAR, This research demonstrates the effectiveness of an improved YOLOv9
could improve depth perception and enhance detection performance in deep learning model for detecting potholes under diverse weather con­

12
S. Parvin, F. Munsy, M.T. Rahat et al.
Results in Engineering 28 (2025) 107817
ditions. By creating a customized dataset (MWPD), combining images [17] Linchao Li, Jiazhen Liu, Jiabao Xing, Zhiyang Liu, Kai Lin, Bowen Du, Road pothole
from various weather, including sunny, rainy, and nighttime environ­ detection based on crowdsourced data and extended mask r-cnn, IEEE Trans. Intell.
Transp. Syst. PP:1--13 (2024) 09, [Link]
ments, the proposed model conditions achieved a 3% improvement over
[18] Kanchi Anantharaman Vinodhini, Kovilvenni Ramachandran Aswin Sidhaarth,
the default YOLOv9, with an average mAP@50 of 95% and an F1-score Pothole detection in bituminous road using cnn with transfer learning, Meas.
of 91%. The proposed model exhibited strong generalizability across Sens. (ISSN 2665-9174) 31 (2024) 100940, [Link]
existing studies, confirming its potential for real-time applications in 100940.
intelligent transportation systems, autonomous driving, and road main­ [19] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley,
Sherjil Ozair, Aaron Courville, Yoshua Bengio, Generative Adversarial Networks,
tenance. 2014.
[20] Boris Bučko, Eva Lieskovská, Katarína Zábovská, Michal Zábovský, Computer vi­
CRediT authorship contribution statement sion based pothole detection under challenging conditions, Sensors 22 (22) (2022),
[Link]
[21] Shaoqing Ren, Kaiming He, Ross Girshick, Jian Sun, Faster r-cnn: Towards Real-Time
Shahnaj Parvin: Writing -- original draft, Formal analysis, Concep­
Object Detection with Region Proposal Networks, 2016.
tualization. Foysal Munsy: Conceptualization. Md Tanzeem Rahat: [22] Malhar Khan, Muhammad Amir Raza, Ghulam Abbas, Salwa Othmen, Amr Yousef,
Formal analysis. Aminun Nahar: Methodology. Kamruddin Nur: Writ­ Touqeer Ahmed Jumani, Pothole detection for autonomous vehicles using deep
ing -- review & editing. Debasish Ghose: Writing -- review & editing. learning: a robust and efficient solution, Front. Built Environ. 9 (2024) 1323792,
[Link]
[23] Riddhi Mirajkar, Anuradha Yenkikar, Shreyash Nawalkar, Rishabh Kaul, Aditya
Declaration of competing interest
Rokade, Kedarnath Rothe, Enhanced pothole detection in road condition assess­
ment using yolov8, in: 2024 IEEE International Conference for Women in Innovation,
The authors declare that this work is original and has not been Technology & Entrepreneurship (ICWITE), 2024, pp. 429--433.
published elsewhere. Additionally, there are no conflicts of interest con­ [24] Jiarui Chang, Zhan Chen, E. Xia, Improved yolov8 method for multi-scale pot­
cerning the publication of this paper. hole detection, in: Advanced Intelligent Computing Technology and Applications:
20th International Conference, Proceedings, Part XI, ICIC 2024, Tianjin, China, Au­
gust 5--8, 2024, Springer-Verlag, Berlin, Heidelberg, ISBN 978-981-97-5611-7, 2024,
Data availability pp. 383--395.
[25] Nidhi Ruhil, Devansh Sahni, Anushaka, Anurag Wadhwa, Anjali Sharma, Pothole de­
This study utilizes a publicly available dataset titled Multi-Weather tection and reporting system implementation using yolov8 and [Link], Int. J.
Pothole Detection Dataset, hosted on Mendeley Data and introduced by Comput. Sci. Eng. 11 (12) (2023) 26--31, [Link]
[Link].
[7].
[26] Hindol Bhattacharjee, Shaik Shaik, Javed an Nasreen, E. Varshith Reddy, Pothole
detection using machine learning (yolov4), Int. Res. J. Modern. Eng. Technol. Sci.
References 6 (3) (2024) 4725--4728, [Link]
[27] Kamalakannan S, Navaneethan S, Yogesh S. Deshmukh, Deshmukh Sujay V,
[1] Krishna Singh Basnet, Jagat Kumar Shrestha, Rabindra Nath Shrestha, Pavement Per­ Sivakami Sundari M, Venkadeshan Ramalingam Jagadeesan D, Venkatesh C, A
formance Model for Road Maintenance and Repair Planning: a Review of Predictive novel pothole detection model based on yolo algorithm for vanet, Int. J. Intell.
Techniques, 2023. Syst. Appl. Eng. 12 (11s) (2024) 56--61, [Link]
[2] Jaroslav Frnda, Srijita Bandyopadhyay, Michal Pavlicko, Marek Durica, Mihails view/4419.
Savrasovs, Soumen Banerjee, Analysis of pothole detection accuracy of selected ob­ [28] Zachary Jeffreys, Kshama Kumar, Zhuojing Xie, Wan D. Bae, Shayma Alkobaisi, Sada
ject detection models under adverse conditions, Transp. Telecommun. J. 25 (2) (April Narayanappa, Potholevision: an automated pothole detection and reporting system
2024) 209--217, [Link] using computer vision, in: Proceedings of the 39th ACM/SIGAPP Symposium on
[3] World Health Organizaton, Global status report on road safety 2023, https:// Applied Computing, SAC ’24, Association for Computing Machinery, New York, NY,
[Link]/teams/social-determinants-of-health/safety-and-mobility/global- USA, ISBN 9798400702433, 2024, pp. 695--697.
status-report-on-road-safety-2023, 2023. (Accessed 11 January 2024). [29] Raghavendra K, Sakshi Shankar Gumaste, Sanjana Singh, Tejas K, Varsha N,
[4] Yuying Mao, Wenzhe Su, Haotian Chen, Pothole road detection and identification Roadeye-a yolov12-based approach for real-time road pothole detection, Int. Res.
based on transfer learning, in: Lijun Wu, Zhongpan Qiu (Eds.), Fourth International J. Eng. Technol. 12 (5) (2025).
Conference on Sensors and Information Technology (ICSI 2024), vol. 13107, Interna­ [30] Satish Kumar Satti, Goluguri N.V. Rajareddy, Kaushik Mishra, Amir H. Gandomi,
tional Society for Optics and Photonics, SPIE, 2024, pages 131073R–1--131073R--6. Potholes and traffic signs detection by classifier with vision transformers, Sci. Rep.
[5] M. Sheeta, K. Prasanna, Intelligent deep learning based pothole detection and alert­ (ISSN 2045-2322) 14 (1) (Jan 2024) 2215, [Link]
ing system, Int. J. Comput. Intell. Res. 19 (1) (2023) 25--35, [Link] 52426-4.
37622/IJCIR/19.1.2023.25-35. [31] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng­
[6] Jinlei Wang, Ruifeng Meng, Yuanhao Huang, Lin Zhou, Lujia Huo, Zhi Qiao, Yang Fu, Alexander C. Berg, SSD: Single Shot MultiBox Detector, Springer Interna­
Changchang Niu, Road defect detection based on improved yolov8s model, Sci. Rep. tional Publishing, ISBN 9783319464480, 2016, pp. 21--37.
14 (1) (Jul 2024) 16758, [Link] [32] Bhargav Ram Gowrisetty, Harees Reddyh Emani, Vineeth Kumar Ganta, Doushik
[7] Shahnaj Parvin, Foysal Munsy, Kamruddin Nur, Multi-Weather Pothole Detection - Chinthapudi, Shaik Riyazuddien, Pothole detection and dimension estimation us­
[Link], 2025. ing deep learning (yolo) and image processing, Int. J. Adv. Res. Innov. Ideas Educ.
[8] Songling Huang, Hao Chen, Lingbo Yan, Xiaoling Zou, Bin Li, Yanqiu Bi, A Review of (ISSN 2395-4396) 10 (2) (2024) 748--754, [Link]
the Progress in Machine Vision-Based Crack Detection and Identification Technology [33] Saahas Myla Myla, Pothole data analysis using deep learning, SSRN 75 (2024)
for Asphalt Pavements, 2025. 1863--1881, [Link]
[9] Boris Bučko, Eva Lieskovská, Katarína Zábovská, Michal Zábovský, Pothole detection [34] Qian Li, Yanjuan Shi, Qing Liu, Gang Liu, Deep learning-based pothole detection for
using computer vision in challenging conditions, in: Figshare, 2022. intelligent transportation: a yolov5 approach, Int. J. Adv. Comput. Sci. Appl. 14 (12)
[10] S. Patel, Pothole image data-set, in: Kaggle, 2019, [Link] (2023), [Link]
datasets/sachinpatel21/pothole-image-dataset. [35] Srividya Ramisetty, Sanjana R, Satish Kanti, Simritha H, Real time pot hole detec­
[11] GeraPotHole, Pothole detection yolov8 dataset, [Link] tion for safe roadways, in: 2023 International Conference on Innovative Computing,
gerapothole/pothole-detection-yolov8, apr 2023. Intelligent Communication and Smart Electrical Systems (ICSES), 2023, pp. 1--7.
[12] Chemikala Saisree, U. Kumaran, Pothole detection using deep learning classification [36] Aditya Singh, Aryan Mehta, Ali Asgar Padaria, Nilesh Kumar Jadav, Rebakah Ged­
method, in: International Conference on Machine Learning and Data Engineering, dam, Sudeep Tanwar, Enhanced pothole detection using yolov5 and federated learn­
Proc. Comput. Sci. (ISSN 1877-0509) 218 (2023) 2143--2152, [Link] ing, in: 2024 14th International Conference on Cloud Computing, Data Science &
1016/[Link].2023.01.190. Engineering (Confluence), 2024, pp. 549--554.
[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun, Deep Residual Learning for [37] Sophia S, Ajish Moses Raj A, R. Janani, Stewart Kirubakaran S, Dhivinkumar AJ,
Image Recognition, 2015. Princely Nesaraj A, Integrating Google maps and deep learning in path hole detection
[14] Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, Alex Alemi, Inception-V4, alert system, in: 2024 4th International Conference on Pervasive Computing and
Inception-Resnet and the Impact of Residual Connections on Learning, 2016. Social Networking (ICPCSN), 2024, pp. 1006--1012.
[15] Karen Simonyan, Andrew Zisserman, Very Deep Convolutional Networks for Large­ [38] Smt.A. Koteswaramma, M. Divya Sri, M. Manohar, K. Srihari, M. Yabbeju, Advancing
Scale Image Recognition, 2015. road safety: pothole detection using yolov8 and wandb deep learning, J. Adv. Zool.
[16] Lopamudra Panda, Kanneganti Bhavya Sri, Reeja SR, Real time pothole detection 45 (S2) (2024) 116--122, [Link]
system -- an application facilitating public safety, in: 2023 International Conference [39] Y. Baby Kalpana, E. Subhashini, A.S. Nashrin Taaj, D. Punithakala, Detection and no­
on Intelligent and Innovative Technologies in Computing, Electrical and Electronics tification of potholes and humps on roads to aid drivers using iot, Int. J. Multidicipl.
(IITCEE), 2023, pp. 389--395. Res. Sci. Eng. Technol. 7 (13) (2024) 155--162.

13
S. Parvin, F. Munsy, M.T. Rahat et al.
Results in Engineering 28 (2025) 107817
[40] Om Khare, Shubham Gandhi, Aditya Rahalkar, Sunil Mane, Yolov8-based visual de­ [54] P.D.S.S. Lakshmi Kumari, Gidugu Srinija Sivasatya Ramacharanteja, S. Suresh Ku­
tection of road hazards: potholes, sewer covers, and manholes, in: 2023 IEEE Pune mar, Gorrela Bhuvana Sri, Gottumukkala Sai Naga Jyotsna, Aki Hari Keerthi Naga
Section International Conference (PuneCon), 2023, pp. 1--6. Safalya, Developing an automated system for pothole detection and management us­
[41] Mayank Dhingra, Rahul Dhingra, Meghna Sharma, Pothole detection using machine ing deep learning, in: Advanced Communication and Intelligent Systems, Springer
learning models, Int. J. Sci. Res. Sci. Eng. Technol. 11 (2) (2024) 94--105, https:// Nature Switzerland, Cham, ISBN 978-3-031-45124-9, 2023, pp. 12--22.
[Link]/10.32628/IJSRSET241126. [55] Danyan Xie, Wenyi Yao, Wenbo Sun, Zhenyu Song, Real-time identification of straw­
[42] Dinesh Swami, Mahesh Jangid, Pothole detection and prediction using deep learning berry pests and diseases using an improved yolov8 algorithm, Symmetry 16 (10)
with cnn and yolov8, in: Rajesh Kumar, Ajit Kumar Verma, Om Prakash Verma, (2024), [Link]
Tanu Wadehra (Eds.), Soft Computing: Theories and Applications, Nature Singapore [56] Chien-Yao Wang, I-Hau Yeh, Hong-Yuan Mark Liao, Yolov9: learning what you
Springer, Singapore, ISBN 978-981-97-2031-6, 2024, pp. 321--334. want to learn using programmable gradient information, [Link]
[43] M. Divya, G. Divyashree, B. Uma Maheswari, Pothole detection using yolov8 and 2402.13616, 2024.
cnn, in: 2024 3rd International Conference for Innovation in Technology (INOCON), [57] Joseph Redmon, Ali Farhadi, Yolov3: an incremental improvement, [Link]
2024, pp. 1--6. org/abs/1804.02767, 2018.
[44] Hong-Hu Chu, Muhammad Rizwan Saeed, Javed Rashid, Muhammad [58] Bin Yan, Pan Fan, Xiaoyan Lei, Zhijie Liu, Fuzeng Yang, A real-time apple tar­
Tahir Mehmood, Rao Ahmad, Sohail Iqbal, Ghulam Ali, Deep learning gets detection method for picking robot based on improved yolov5, Remote Sens.
method to detect the road cracks and potholes for smartcities, Comput. (ISSN 2072-4292) 13 (9) (2021), [Link]
Mater. Continua (ISSN 1546-2226) 75 (1) (2023) 1863--1881, https:// [59] Chien-Yao Wang, Alexey Bochkovskiy, Hong-Yuan Mark Liao, Yolov7: trainable
[Link]/10.32604/cmc.2023.035287. bag-of-freebies sets new state-of-the-art for real-time object detectors, in: 2023
[45] Shafi Ullah Adid, Md. Emon, Taofica Amrine, A hybrid approach to detect and clas­ IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023,
sify pothole on Bangladeshi roads using deep learning, Int. J. Sci. Res. Arch. 12 (01) pp. 7464--7475.
(2024) 1045--1053, [Link] [60] Fatma M. Talaat, Hanaa ZainEldin, An improved fire detection approach based on
[46] S. Kanaga Suba Raja, B. Chandra, Sharon Hajhan, Prasshanth, Tamilvanan, Pothole yolo-v8 for smart cities, Neural Comput. Appl. 35 (28) (2023) 20939--20954, https://
detection and compliant notification, AIP Conf. Proc. 2802 (1) (2024), [Link] [Link]/10.1007/s00521-023-08809-1.
org/10.1063/5.0184610. [61] Rahima Khanam, Muhammad Hussain, Yolov11: an overview of the key architectural
[47] B. Srivani, Ch. Kamala, S. Renu Deepti, G. Aakash, Pothole detection using convolu­ enhancements, [Link] 2024.
tional neural network, AIP Conf. Proc. 2935 (1) (2024), [Link] [62] Yunjie Tian, Qixiang Ye, David Doermann, Yolov12: attention-centric real-time ob­
0198902. ject detectors, [Link] 2025.
[48] Arvindh Kumar Selvam, G.Y.R. Vikhram, A real-time cnn-based pothole detection
[63] Chien-Yao Wang, Hong-Yuan Mark Liao, I-Hau Yeh, Designing network design strate­
system for road safety, Int. Adv. Res. J. Sci. Eng. Technol. 11 (4) (2024) 711--717,
gies through gradient path analysis, J. Inf. Sci. Eng. 39 (4) (2023) 975--995, https://
[Link]
[Link]/10.6688/JISE.202307_39(4).0016.
[49] A. Lincy, G. Dhanarajan, S. Sanjay Kumar, B. Gobinath, Road pothole detec­
[64] Chien-Yao Wang, Hong-Yuan Mark Liao, Yueh-Hua Wu, Ping-Yang Chen, Jun-Wei
tion system, ITM Web Conf. 53 (2023) 01008, [Link]
Hsieh, I-Hau Yeh, Cspnet: a new backbone that can enhance learning capability of
20235301008.
cnn, in: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition
[50] Junkui Zhong, Deyi Kong, Yuliang Wei, Bin Pan, Yolov8 and point cloud fusion for
Workshops (CVPRW), 2020, pp. 1571--1580.
enhanced road pothole detection and quantification, Sci. Rep. 15 (1) (Apr 2025)
[65] Hafedh Mahmoud Zayani, Unveiling the potential of yolov9 through comparison
11260, [Link]
with yolov8, Int. J. Intell. Syst. Appl. Eng. 12 (3) (Mar. 2024) 2845--2854, https://
[51] Y. Aneesh Chowdary, V. Sai Teja, V. Vamsi Krishna, N. Venkaiah Naidu, R. Karthika,
[Link]/[Link]/IJISAE/article/view/5794.
Pothole Detection Approach Based on Deep Learning Algorithms, Soft Computing
[66] Tianyong Wu, Youkou Dong, Yolo-se: improved yolov8 for remote sensing object
and Signal Processing, vol. 313, Springer Nature Singapore, Singapore, ISBN 978­
detection and recognition, Appl. Sci. (ISSN 2076-3417) 13 (24) (2023), [Link]
981-19-8669-7, 2023, pp. 597--606.
org/10.3390/app132412977.
[52] Sarthak Babbar, Jatin Bedi, Real-time traffic, accident, and potholes detection by
[67] Prasanta Das, Angshuman Chakraborty, Ravi Sankar, Om Krishan Singh, Hena Ray,
deep learning techniques: a modern approach for traffic management, Neural Com­
Alokesh Ghosh, Deep learning-based object detection algorithms on image and video,
put. Appl. (ISSN 1433-3058) 35 (26) (Sep 2023) 19465--19479, [Link]
in: 2023 3rd International Conference on Intelligent Technologies (CONIT), 2023,
1007/s00521-023-08767-8.
[53] Mohd Omar, Pradeep Kumar, Pd-its: pothole detection using yolo variants for intel­ pp. 1--6.
ligent transport system, SN Comput. Sci. (ISSN 2661-8907) 5 (5) (May 2024) 552,
[Link]

14