Abstract
This data article introduces a structured pavement surface image dataset developed to advance research in automated pavement condition assessment and data-driven road infrastructure monitoring. The dataset comprises three distinct pavement condition categories: (i) alligator cracking, (ii) edge-breaking distress, and (iii) undamaged (intact) pavement surfaces, each representing a prevalent form of pavement deterioration or intact condition typically observed in flexible pavements. The dataset consists of 12,000 raw images (4000 per class) collected under real-world conditions. These images represent the primary scientific contribution of the dataset. All images were standardized through resizing and normalization, and the dataset was partitioned into training, validation, and testing subsets to ensure reproducibility and consistency in data-driven experiments. Pavement images were collected from selected segments of National Highway N6 in Pabna District, Bangladesh, under natural daylight conditions using a smartphone camera during field surveys. Image acquisition was conducted following standard safety practices without disrupting traffic flow. All images were manually reviewed and labelled to ensure annotation accuracy. This dataset is intended to support research on automated pavement crack detection and classification, benchmarking of computer vision and deep learning models, and the development of lightweight and edge-deployable inspection systems.
Keywords: Pavement failure; Crack detection; Edge-breaks, Alligator crack; Deep learning: image processing
Specifications Table
| Subject | Engineering & Materials science |
| Specific subject area | Pavement Engineering, Computer Vision, Pattern Recognition, deep learning |
| Type of Data | Raw, Processed and well-structured Image data. |
| Data collection | Pavement images were acquired during field surveys using a smartphone camera under natural daylight conditions. Images of cracked and undamaged pavement surfaces were captured at close range and manually reviewed and labeled before preprocessing. |
| Data source Location | The data were collected along selected segments of National Highway N6 (Dhaka–Pabna Highway) in Pabna District, Bangladesh, within the approximate geographic range of 24.02°–24.03° N latitude and 89.22°–89.30° E longitude. |
| Data accessibility | Repository name: Zenodo Data identification number: Part 1: 10.5281/zenodo.18072573 & Part 2: 10.5281/zenodo.18114226 Direct URL to data: Part 1: https://doi.org/10.5281/zenodo.18072573 & Part 2: https://doi.org/10.5281/zenodo.18114226 |
| Related research article | None |
1. Value of the Data
-
•
This dataset is a valuable resource for the development and refining of automated road condition monitoring systems capable of accurately identifying and classifying pavement defects.
-
•
This dataset consists of class-balanced pavement imagery captured under natural lighting. It includes three distinct pavement condition categories: edge cracks, alligator cracks, and intact pavement surfaces, all captured in real-world conditions. The dataset facilitates research on automated pavement distress detection, classification, and performance evaluation, enabling the creation of reliable methods for infrastructure monitoring.
-
•
The dataset provides high-resolution, class-balanced imagery under natural lighting, supporting the development of robust infrastructure monitoring systems. The dataset comprises three clearly labelled pavement condition classes: edge crack, alligator crack, and undamaged road surface, captured under real-world conditions. It supports research on automated pavement crack detection, classification, and performance evaluation, enabling the development of robust methods for infrastructure monitoring.
-
•
This dataset supports automated pavement condition assessment and contributes to data-driven infrastructure management, enabling optimized maintenance planning and more efficient allocation of road maintenance resources.
-
•
The open-access nature of this dataset ensures its broad usability, fostering collaboration between academia and industry, and providing a foundation for further research into automated pavement inspection.
-
•
Future researchers may extend this dataset by incorporating additional pavement distress types, severity levels, and diverse environmental conditions to support broader and more comprehensive pavement condition assessment studies.
2. Background
The increasing demands placed on urban transportation infrastructure, together with the limitations of conventional pavement inspection methods, have highlighted the use of automated and cost-effective approaches for road condition monitoring. Alligator cracks and edge breaks are commonly observed forms of pavement distress and are widely recognized as surface-level indicators of roadway deterioration. Alligator cracking, and edge breaking a common distress in flexible pavements, has significant negative effects on pavement performance, service life, safety. Pavement distresses have a tremendous influence on the operational performance of traffic especially speed and capacity[1]. Fatigue, heavy vehicles, poor drainage, inadequate thickness of layers and inadequate mix design lead to the occurrence of flexible pavement failures on rural highways [2]. Higher surface roughness on rural roads increases expected crash frequency [3]. Flexible pavements are also climate sensitive and need changes and mitigation plans to minimize damages [4]. It is clear that cracks significantly compromise the structural integrity of pavement, triggering an accelerated performance, increasing maintenance costs, and causing sit-in safety risks [5]. Cracking appears in the first stages of worsening, so its detection will allow one to perform proper maintenance, saving huge amounts of money on a later restoration [6]. Machine learning and computer vision allow to detect cracks automatically and with a very high degree of accuracy, thus addressing the problem of low contrast and shadows in pavement pictures [7]. In parallel, advances in computer vision and machine learning have enabled the application of image-based data in the analysis of pavement surface conditions.
3. Data Description
The dataset consists of a labelled pavement surface image collection acquired from an urban segment of National Highway N6 in Pabna District, Bangladesh. It contains images representing three pavement condition categories: alligator cracking, edge breaking, and normal pavement captured under real-world highway conditions with natural lighting. The dataset is structured to facilitate direct implementation in machine learning, deep learning, and computer vision–based workflows, supporting tasks such as pavement distress detection, condition classification, and performance benchmarking for intelligent infrastructure monitoring systems (Fig. 1).
Fig. 1.
Structure of the main dataset.
The dataset is organized in a hierarchical folder structure separating raw and augmented images, with subfolders for alligator cracking, edge cracking, and normal pavement. The data are further partitioned into training, validation, and testing subsets with balanced class distributions to support reproducible machine learning workflows. Augmented images provided separately for model training purposes. We have also clarified in manuscript that the augmented images are not part of the primary dataset and are optional for model training (Table 1).
Table 1.
Pavement failure types and characteristics.
| Pavement Failure Type | Failure Description | Sample Captured Images |
|---|---|---|
| Alligator Cracks | A network of interconnected cracks forming a grid-like pattern, often due to repeated stress from traffic. Typically caused by repeated traffic loads. Common in flexible pavements under heavy traffic conditions. |  |
| Edge Breaking | Cracks and fragmentation occurring along the edges of the pavement. Often exacerbated by weather conditions, traffic wear, and insufficient shoulder support. Typically found along the road edges, especially on high-speed roads or areas with heavy truck traffic. |  |
| Normal Pavement | No visible signs of cracking or damage, representing a well-maintained pavement. Properly maintained, undamaged pavement surface. Can be found on newly constructed or recently repaired pavement. |  |
The Table 2 summarizes the class-wise image distribution following preprocessing, resulting in a balanced dataset contains 12,000 images, with 4000 images per class across three pavement condition categories. All images are captured in high-resolution using 50MP and 64MP smartphone cameras. The original raw images are provided in their full resolution (e.g., 3468×4624 pixels). For model compatibility, a subset of the images was resized to 512×512 pixels, and augmented versions were created by applying transformations such as flipping and brightness adjustment. These augmented images are not raw but are derived from the original images to enhance model training.
Table 2.
Class-wise crack distribution images after pre-processing.
| Serial No | Classes | Pre-Processed / raw Images | Augmented Images |
|---|---|---|---|
| 1 | Alligator Cracks | 4000 | 4000 |
| 2 | Edge Breaking | 4000 | 4000 |
| 3 | Normal Pavement | 4000 | 4000 |
| Total | 12,000 | 12,000 | |
All images in the dataset are in RGB format. Though Grayscale conversion was applied during preprocessing solely to standardize image intensity and reduce lighting variability. The downloadable dataset contains the original RGB images for use in tasks that require colour channels for feature extraction, ensuring the dataset remains suitable for modern deep learning models that utilize colour information.
The Fig. 2 presents representative pavement images after grayscale conversion applied during preprocessing. This step was performed to standardize image intensity representation and reduce colour-related variability, thereby supporting consistent and reproducible data preparation.
Fig. 2.
Grayscale preprocessing of pavement images: (a) edge breaking and (b) alligator crack.
3.1. Technical validation
A technical validation was conducted to assess the internal consistency, labeling accuracy, and learning suitability of the dataset for supervised pavement image analysis. To verify that the pavement condition categories are distinguishable and correctly annotated, a small set of representative baseline convolutional neural network architectures was employed. The objective of this evaluation was not to benchmark or compare model performance, but to confirm that the dataset supports reliable data-driven learning.
Baseline experiments were performed using ResNet-50, DenseNet-121, MobileNetV2, and EfficientNet-B0 architectures initialized with ImageNet pretrained weights. While the original high-resolution images (50MP/64MP) were used for dataset preparation, all models were trained on images resized to a uniform resolution of 512×512 pixels. Classification accuracies of 96.56 %, 98.28 %, 95.33 %, and 95.67 %, respectively, were obtained on the held-out test set. The consistently high performance across different architectures indicates clear class separability and the absence of major labeling inconsistencies, confirming that the dataset is well-structured and suitable for use in machine learning, deep learning, and computer vision–based pavement condition analysis. These results are reported solely to demonstrate dataset usability and integrity and do not represent a comparative performance study.
3.2. Model trainingsetupand computingplatform
For technical validation, model training was carried out using the PyTorch deep learning framework (version 2.9.1). All baseline models were trained with an input resolution of 512 × 512 pixels using the Adam optimizer with a learning rate of 0.0001 and a batch size of 32. Cross-entropy loss was employed as the objective function, and training was performed for 30 epochs. The dataset was divided into training, validation, and testing subsets following a 70 %, 15 %, and 15 % split. Standard data augmentation techniques, including random horizontal flipping and minor brightness adjustments, were applied during training to improve generalization.
All experiments were conducted on a Windows-based workstation equipped with an NVIDIA RTX 4090 GPU, 64 GB of system memory, and a multi-core Intel processor. CUDA and cuDNN libraries compatible with PyTorch 2.9.1 were used to enable GPU-accelerated training and inference.
4. Experimental Design, Materials and Methods
The survey was carried out on an 8.98 km long dual-lane section of the National Highway N6 that passes through Pabna Sadar in the Pabna District of Bangladesh. The section of the highway is an urban flexible pavement road that is exposed to mixed traffic conditions comprising of passenger and heavy commercial vehicles. These operational conditions contribute to the development of common pavement distresses, particularly alligator cracking and edge breaking (Fig. 3).
Fig. 3.
Location of the study area along National Highway N6 in Pabna District, Bangladesh.
Fig. 4 summarizes the sequential steps followed to develop the pavement image dataset, including road section selection, field image acquisition, manual labelling, preprocessing, data augmentation, and final dataset composition.
Fig. 4.
Overview of pavement image data collection, preprocessing, augmentation, and final dataset preparation.
The study was initiated to document and analyze common pavement surface distresses observed on urban highway sections in Pabna, particularly in areas near the university campus, as these distresses adversely affect ride quality, road safety, and maintenance planning. The absence of localized, high-quality image datasets motivated the need for structured field data collection.
A reconnaissance survey was conducted prior to formal data acquisition to evaluate pavement condition variability, traffic characteristics, and site accessibility. This preliminary inspection enabled identification of representative distress locations, assessment of safety constraints, and determination of suitable image capture points along the roadway.
Based on the reconnaissance findings, an approximately 8.98 km dual-lane section of National Highway N6 passing through Pabna Sadar, Pabna District, Bangladesh, was selected as the study area. The selected segment experiences mixed traffic conditions and displays diverse pavement surface conditions, making it appropriate for systematic data collection.
The field survey was carefully planned and executed to ensure systematic data collection, consistency across survey locations, and adherence to road safety requirements. Prior to field deployment, the survey route along the selected segment of National Highway N6 was reviewed, and pavement sections exhibiting visible surface distress as well as intact pavement conditions were identified based on the reconnaissance survey. Data collection activities were scheduled during daylight hours to ensure adequate illumination and minimize image quality variability caused by low-light conditions.
Multiple images were taken at each location to capture variations in surface texture, distress severity, and lighting conditions. Image acquisition was conducted without interrupting traffic flow. The smartphone models used for image collection included the Redmi Note 11 (50 MP), Realme C55 (64 MP), Realme Note 9 Pro Plus (50 MP), and Samsung S22 (50 MP). These devices were chosen for their ability to capture high-resolution images, ensuring the dataset is suitable for accurate and detailed pavement distress analysis.
During the survey, pavement surface images were collected under normal traffic conditions without traffic interruption. Image acquisition was conducted from safe roadside positions or designated stopping points to avoid interference with vehicular movement. A handheld smartphone camera was used to capture images at close range, maintaining a near-vertical viewing angle to minimize perspective distortion and ensure clear representation of pavement surface features. The camera was positioned consistently across locations to maintain uniformity in image scale and framing. Immediately following image acquisition, a preliminary quality check was performed in the field to identify and discard images affected by excessive blur, occlusion, shadowing, or irrelevant objects. This initial screening reduced the inclusion of low-quality images and improved overall dataset reliability.
To ensure field safety, the survey team consisted of three members. One team member was positioned at the centre to capture pavement images, while the other two members were assigned to monitor incoming vehicles from both directions. This arrangement allowed continuous awareness of traffic movement and ensured the safety of the surveyor engaged in image capture. All activities were performed from safe roadside positions, and no personnel entered active traffic lanes.
As the dataset was collected on the public National Highway N6 using a smartphone, there was potential for capturing Personally Identifiable Information (PII), such as license plates or pedestrian faces. To address privacy concerns, we carefully reviewed and screened all images during and after the data collection process. Any images containing visible PII were identified and cropped to ensure no personally identifiable information remained. This process was carried out consistently across all images before they were included in the dataset. As a result, the dataset is free from any identifiable personal information and complies with privacy standards.
The following section describes the image normalization and data augmentation procedures applied during dataset preparation:
4.1. Image normalizationand dataaugmentation
To standardize image inputs and increase dataset variability under realistic field conditions, a systematic image normalization and augmentation pipeline was applied prior to deep learning model training. All transformations were applied probabilistically to preserve the natural distribution of pavement distress patterns while introducing controlled variability.
4.2. Normalization
Image normalization was performed using ImageNet-standard mean and standard deviation values (mean: 0.485, 0.456, 0.406; standard deviation: 0.229, 0.224, 0.225). This process scales pixel intensities to a consistent distribution, reduces internal covariate shift, and supports stable optimization during convolutional neural network (CNN) training. Normalization ensures uniform illumination representation across images captured under varying environmental conditions.
4.3. Rotational augmentation
Rotational transformations were applied to simulate variations in camera orientation and vehicle movement during field acquisition. Images were randomly rotated within ±25° to represent minor viewpoint deviations, ±90° to account for portrait–landscape orientation changes, and 180° for full inversion. These rotations encourage rotation-invariant feature representation for pavement distress patterns.
4.4. Brightness and contrastadjustment
Brightness and contrast were randomly varied within ±30 % to replicate illumination changes caused by sunlight intensity, shadows, cloud cover, and time-of-day effects. This augmentation reflects real-world lighting variability commonly encountered in outdoor pavement inspection surveys.
4.5. Blur and noiseaugmentation
Gaussian blur with kernel sizes ranging from 3 to 7 pixels was introduced to simulate mild defocus, vibration-induced blur, and dust interference during image capture. Additionally, Gaussian noise with a variance limit of up to 2 % was applied to emulate sensor noise associated with field imaging devices. These operations model common image degradation effects without altering pavement distress geometry.
4.6. Motion blur
A motion blur operation using a 2-pixel kernel was applied with random probability to reproduce linear smearing effects resulting from vehicle motion, handheld camera usage, or vibration during data collection. This transformation reflects dynamic acquisition conditions encountered during roadside surveys.
4.7. Weather-based augmentation
Weather-based data augmentation was incorporated into the image preprocessing pipeline. Fog and rain simulations were applied stochastically during training, with their application probabilities randomly sampled within predefined ranges to reflect realistic variations in weather conditions. Fog augmentation was implemented using the RandomFog transformation, which simulates reduced-visibility scenarios such as early-morning fog, winter haze, and airborne dust by introducing atmospheric light scattering and attenuation effects. This process reduces global contrast, suppresses high-frequency texture information, and softens crack boundary edges, thereby increasing the difficulty of crack detection under low-visibility conditions.
Rain augmentation was implemented using the RandomRain transformation, which generates synthetic rain streaks with random orientation, length, and density, along with wet-surface reflectance effects. These effects alter local intensity distributions and specular reflections on the pavement surface, potentially leading to partial occlusion of crack patterns. Both augmentations were applied independently to each training image and combined with standard image normalization operations. Collectively, these weather-based transformations expand the diversity of environmental conditions represented in the training dataset and enhance the robustness and generalization capability of the proposed model across different seasonal and weather scenarios.
4.8. Image resizing
All images were resized to a uniform resolution of 512 × 512 pixels to ensure consistent input dimensions, preserve pavement distress geometry, and maintain computational efficiency for CNN-based processing. No cropping or zooming operations were applied, thereby retaining the spatial integrity of crack and edge-breaking features.
Limitations
The dataset currently includes only two types of pavement failures: alligator cracks and edge-breaks. While these are common issues in pavement deterioration, it does not cover other types of damage, such as potholes, tearing, or surface wear. Additionally, the data was collected exclusively along the National Highway (N6) in Pabna District, Bangladesh, which may not fully represent pavement deterioration patterns in regions with different traffic volumes, environmental conditions, or road maintenance practices. Despite efforts to capture images under varying lighting conditions, the reliance on natural light resulted in inconsistencies in brightness and contrast across the images. Lastly, the dataset does not include precise measurements of crack length or crack depth. Future datasets should include additional pavement failure types (e.g., potholes, tearing), cover broader geographic areas, incorporate controlled lighting and consistent weather conditions, and include measurements of crack length and depth for more detailed analysis.
Ethics Statement
This article does not involve any research involving human or animal subjects by any of the authors. The datasets utilized in this article are publicly accessible. The dataset was collected on a public national highway (N6) using a smartphone-based imaging device, and all images were manually screened prior to release to ensure that no Personally Identifiable Information (PII), including faces, vehicle license plates, or pedestrians, is visible. Images containing identifiable elements were excluded, and only pavement surface regions relevant to road condition assessment are retained. When utilizing these datasets, it is essential to adhere to appropriate citation guidelines. The author has read and followed the ethical requirements for Data in Brief. The methods and data associated with this paper did not involve human subjects, animal experiments, or any data collected from social media platforms. When utilizing these datasets, it is essential to adhere to appropriate citation guidelines.
Credit Author Statement
Md. Nayem Hossain: Conceptualization, Supervision, Methodology, Writing (Original draft preparation, Reviewing and Editing), Data review, Investigation; Nakib Aman: Supervision, Methodology, Data review, Investigation Writing (Original draft preparation, Reviewing and Editing); Nymur Rahaman Antor: Data Collection, Data review, Investigation; Nafisa Tasnim: Data Collection, Data review, Investigation; Md Zubair Azam: Data Collection, Data review, Investigation; Sabab Asfaq: Data Collection, Data review, Investigation.
Acknowledgements
The authors would like to express their deepest gratitude towards Department of Civil Engineering at Pabna University of Science and Technology for their contributions to conducting the field survey using proper procedures. We would like to thank all the members who conduct the survey despite of being high speed vehicles and heavy traffic at the survey locations. Additionally, we would like to acknowledge the Department of Computer Science and Engineering (CSE) for their expertise in data augmentation, image cleaning, and the development of deep learning models. The data sets mentioned in this article are publicly available; however, it is still mandatory for researchers to follow the rules of citation is invaluable. This work received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.
Declaration of Competing Interest
The authors report no financial or personal conflicts of interest between themselves and others.
Data Availability
References
- 1.Hashim I.H., Badawy R.M., Heneash U. Impact of pavement defects on traffic operational performance. Sustainability. May 2023;15(10):8293. doi: 10.3390/SU15108293. 2023, Vol. 15, Page 8293. [DOI] [Google Scholar]
- 2.Ahmed S., Al-Arkawazi F. Flexible pavement evaluation: a case study. Kurdistan J. Appl. Res. Aug. 2017;2(3):292–301. doi: 10.24017/SCIENCE.2017.3.33. [DOI] [Google Scholar]
- 3.Chen S., Saeed T.U., Labi S. Impact of road-surface condition on rural highway safety: a multivariate random parameters negative binomial approach. Anal. Methods Accid. Res. Dec. 2017;16:75–89. doi: 10.1016/J.AMAR.2017.09.001. [DOI] [Google Scholar]
- 4.Qiao Y., Dawson A.R., Parry T., Flintsch G., Wang W. Flexible pavements and climate change: a comprehensive review and implications. Sustainability. Feb. 2020;12(3):1057. doi: 10.3390/SU12031057. 2020, Vol. 12, Page 1057. [DOI] [Google Scholar]
- 5.Mehdinazar M., Kordani A.A., Arabani M. Effects of changes in runoff characteristics on asphalt mixtures’ Performance against low-temperature cracking. Results Eng. Dec. 2024;24 doi: 10.1016/J.RINENG.2024.103246. [DOI] [Google Scholar]
- 6.Gavilán M., et al. Adaptive road crack detection system by pavement classification. Sensors. Oct. 2011;11(10):9628–9657. doi: 10.3390/S111009628. 2011, Vol. 11, Pages 9628-9657. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Zou Q., Cao Y., Li Q., Mao Q., Wang S. CrackTree: automatic crack detection from pavement images. Pattern Recognit. Lett. Feb. 2012;33(3):227–238. doi: 10.1016/J.PATREC.2011.11.004. [DOI] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.