BD-PaveSurface: A multi-weather pavement surface image dataset for crack, pothole, and good pavement condition assessment in Bangladesh

Abstract

This data article describes BD-PaveSurface, a pavement surface image dataset developed for image-level assessment of three pavement surface conditions: crack, pothole, and good pavement surface. The images were collected through systematic field surveys from selected segments of National Highways N5 and N6 in Pabna District, which form part of the Pabna-Dhaka highway route in Bangladesh. Smartphone cameras were used to capture pavement surface images under natural outdoor lighting and real traffic-environment conditions, without interrupting normal traffic movement. Data collection was conducted over an extended period across rainy-season, winter, and summer conditions to capture realistic variation in pavement appearance, illumination, surface moisture, shadows, dust, and texture. The collected images underwent manual screening and image-level labeling. Labeling was conducted by a five-member annotation team, with each image independently reviewed by a second team member. Disagreements were resolved through consensus using predefined class criteria and the dominant-condition rule for mixed crack–pothole cases, followed by independent validation by a Civil Engineering faculty member. After removal of unsuitable samples and label verification, the cleaned raw dataset contained 9000 field images, with 3000 images in each pavement condition class. To support machine-learning-based pavement condition assessment, a processed/augmented subset was prepared from the cleaned images. The preprocessing workflow converted images to RGB format and resized them to 224 × 224 pixels. The cleaned images were split into training, validation, and testing subsets before augmentation, and augmentation operations were then applied only to the training subset. The validation and testing subsets contained only standardized original images and were not augmented. The cleaned raw dataset represents 9000 unique field-acquired pavement images. The processed/augmented subset contains 12,000 derivative image files generated from these field-acquired images, comprising standardized 224 × 224 versions of all 9000 cleaned originals and 3000 newly generated training-only augmented images (1000 per class). Accordingly, the dataset contains 21,000 stored image files in total; however, these files do not represent 21,000 distinct field acquisitions or physical pavement locations. Publicly available Bangladesh-focused pavement image datasets containing both crack and pothole samples remain limited. BD-PaveSurface contributes a field-collected, multi-weather pavement image resource from a South Asian road environment and can be reused for automated pavement condition assessment, crack and pothole recognition, transfer learning, data augmentation studies, computer vision benchmarking, and low-cost camera-based road monitoring.