Correction: Eco-friendly pavement raveling detection based on in-situ data and transfer learning

Engineering notes

Investments in infrastructure significantly promote economic growth by enhancing productivity, with notable effects observed in both urban and rural areas. Pavement functionality and durability significantly influence driving efficiency, comfort, and economic factors such as maintenance and vehicle operating costs, ultimately affecting daily life and regional economies [5,6] .

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

The road infrastructure forms the backbone of modern civilization, playing a vital role in economic development, social connectivity, and quality of life [1] . Investments in infrastructure significantly promote economic growth by enhancing productivity, with notable effects observed in both urban and rural areas. Interestingly, the economic impact of rural infrastructure investments is slightly higher and lasts for approximately six years in China [2] . Among these infrastructures, the road surface, as the critical layer in direct contact with vehicles, endures continuous traffic loads while ensuring safe and efficient transportation [3,4] . Pavement functionality and durability significantly influence driving efficiency, comfort, and economic factors such as maintenance and vehicle operating costs, ultimately affecting daily life and regional economies [5,6] . As societies expand their transportation networks, the demand for sustainable, resilient, and high-performance pavement systems has become increasingly prominent [7] . However, the occurrence of pavement diseases can severely compromise road functionality and safety, making their timely treatment critically important.The accurate classification of pavement diseases and identification of their causes are prerequisites for scientific maintenance decision-making. Extensive research has been conducted by scholars worldwide on the causes of pavement diseases. Generally, pavement diseases can be categorized into three types based on size and shape: texture-related, deformation-related, and crack-related [8,9] . The texture-related diseases in asphalt pavement include raveling, potholes, and bleeding. Among these, raveling is one of the most common and challenging disease mechanisms. The raveling is commonly observed in open-graded friction courses and other asphalt mixtures, particularly in airport pavements, where age-related distress results from severe erosion of the bituminous mastic [10] . The Raveling typically begins with the loss of fine aggregates, progressing to the displacement of larger particles, leading to increasingly rough and uneven textures. Various factors contribute to the formation of raveling, including material aging, insufficient binder content, improper compaction during construction, freeze-thaw cycles, oxidation, and excessive traffic loads [11] . The Raveling, as a prevalent asphalt pavement disease, significantly impacts both passenger comfort and pavement durability [12] . Shamami et al. [13] concluded that traffic loads and asphalt mixture properties significantly influence pavement distress, revealing cracking patterns and disease mechanisms. Xu [14] identified typical characteristics of various pavement diseases and developed machine learning methods for intelligent identification of disease causes. Abbassi et al. [15] found that untreated raveling reduces skid resistance, accelerates moisture infiltration, increases road noise, and poses safety hazards due to loose aggregate particles. Yajin Han et al. [16] suggested that increasing emulsion and SBR dosages can delay the onset of raveling. Pourhassan et al. [17] observed that most raveling occurs within the first 50,000 vehicle passes and that partial use of crumb rubber does not significantly impact raveling resistance.Additionally, the raveling often serves as an early indicator of more severe diseases, triggering a chain reaction of pavement degradation that, if left unaddressed, may lead to structural failures. The traditional methods for evaluating pavement diseases primarily rely on manual visual inspections and conventional pavement management systems (PMS) [18,19] . However, these traditional techniques have been the cornerstone of pavement maintenance practices for decades; they exhibit significant limitations in the context of modern infrastructure management. Specifically, effective PMS frameworks must collect data reliably and consistently, and manual surveys were utilized for data collection, resulting in these approaches being time-consuming and inefficient [20] . Due to the high operational costs of dedicated inspection vehicles, conventional PMS methodsrelying on trained inspectors for on-site identification and assessment-often suffer from limited data coverage, subjectivity, and resource intensiveness, leading to infrequent inspections that may overlook early-stage diseases [21] . The labor-intensive nature of these inspections also exposes personnel to potential traffic safety risks.Although traditional PMS frameworks provide a systematic approach for condition assessment, they often lack the granularity and timeliness needed to optimize maintenance decisions, particularly for rapidly evolving diseases like raveling. These limitations highlight the need for intelligent, datadriven approaches such as artificial intelligence to enhance detection accuracy and optimize maintenance strategies. The advent of artificial intelligence (AI), particularly deep learning, has revolutionized the potential for automated pavement disease detection and classification. Deep learning has been widely applied to road damage detection in countries such as India, Japan, and the Czech Republic [22] . With advances in computer vision and computing power, AI systems now analyze pavement images with high accuracy and efficiency [23] . Nasertork et al. [12] introduced an image-based AI system for detecting asphalt raveling. Progress in this field is driven by large labeled datasets and developments in autonomous driving [24] . Convolutional Neural Networks (CNNs), in particular, excel at identifying visual patterns of pavement diseases, as shown in studies like Li et al. [25] and other CNN-based models for rapid damage detection [26] . Nevertheless, raveling has received comparatively less attention in pavement diseases detection research than more visually distinct issues like cracks and potholes. It also faces many challenges in recognition. Raveling detection poses challenges due to its subtle texture changes, which contrast with the clear patterns of cracks or potholes. Early-stage raveling often resembles normal pavement texture, making automated detection difficult. Its progressive nature further requires methods to assess severity levels for effective maintenance planning. Addressing these challenges is vital to prevent rapid pavement deterioration, extend service life, and optimize maintenance resources. For raveling detection, deep learning techniques can analyze the intricate texture variations and aggregate displacement patterns that define this disease, potentially identifying early-stage raveling before it progresses to more severe conditions. Kulambayev et al. [27] proposed an optimized model tailored to detect fissures in diverse pavement types under varying lighting and environmental conditions, addressing challenges like shadows and debris that often hinder traditional methods. Guerrieri et al. [28] developed a dataset containing 13,135 flexible pavement disease images and 30,989 bounding boxes, achieving a mean average precision of up to 80% during testing, depending on the disease type. To achieve an optimal balance between detection accuracy and computational efficiency in automated pavement defect detection, Lin, J et al. [29] Propose a lightweight convolutional neural network architecture and demonstrate its effectiveness through rigorous experimental validation. These studies highlight the growing potential of deep learning in overcoming challenges associated with raveling.This study aims to develop a robust and efficient deep learning-based approach for the automatic identification and classification of asphalt pavement raveling. To address the challenges posed by limited training data and diverse pavement conditions, this study leverages TL to adapt advanced convolutional neural network structures-originally trained on large-scale datasets-for the task of texture-based raveling detection. The proposed method enhances detection accuracy while maintaining computational efficiency by integrating multi-scale texture feature extraction with specialized classification algorithms. The VGG16, Eff-B0, InceptionV3, and RegNet network structures are enhanced by optimizer and preprocessing techniques to determine the most appropriate network. Furthermore, data augmentation strategies are combined with fine-tuned TL to improve generalization across varying lighting conditions, pavement textures, and severity levels, thereby establishing a reliable framework for intelligent pavement raveling monitoring. The overall research framework is illustrated in Figure 1. The pavement surface data were collected using the LS-40 portable three-dimensional (3D) surface analyzer, as shown in Figure 2. The collection site was located in the areas surrounding the campus where water readily pooled on the asphalt surfaces. All the data were collected under clear weather conditions. A total of 745 images were obtained, comprising 530 images of raveling distress and 215 images of normal pavement surfaces. These images provided high-resolution surface texture data, capturing critical features necessary for raveling identification.Data preprocessing involved several steps to enhance image quality and prepare the dataset for model training. Histogram equalization was applied to reduce noise, improve contrast, and enhance the visibility of pavement texture features. To mitigate the limitation of a small original dataset and enhance model generalization, data augmentation techniquesspecifically mirroring and rotation-were applied. From an initial pool of 745 original images, a comprehensive dataset of 1600 enhanced images was produced. The augmented dataset was balanced, comprising 800 raveling-distress images and 800 normal pavement-surface images. All augmented images were subsequently resized to 224×224 pixels to meet the input requirements of the CNN models. The dataset was then randomly split into a training set (80%) and a validation set (20%). This 8:2 split ensures a balanced and representative distribution for effective model training and evaluation. The TL approach can leverage the power of pre-trained models and address the challenge posed by the limited size of the training dataset [30] . The pre-trained CNN (convolutional neural network) models, originally trained on the ImageNet dataset, were adapted for raveling detection. Although the ImageNet dataset and the pavement raveling database differ significantly in high-level semantics, the generic low-level features (e.g., edges, colors, and simple textures) learned from ImageNet are highly transferable and effective for asphalt pavement disease detection [31] . The final fully connected layer of each pre-trained model was replaced with a new layer specifically designed for binary classification, between raveling and normal This with the dataset and binary this pre-trained network from the were with model trained on the as or for the To enhance the to image classification the was This approach the initial of the convolutional for the of the features and the classification layer This method the need for labeled data and the learning Furthermore, to mitigate several techniques were including the early a of in the classification and using a learning the optimal optimizer was from three commonly learning and for model training. The are shown in Figure and the and methods of each learning model are shown in 1. and and the training for the proposed CNN models was optimized using several strategies to improve prevent and robust To develop and a robust method for raveling detection, CNN structures were The model The VGG16, developed by is a deep CNN of convolutional and fully connected The network a simple effective using convolutional by to reduce while critical features. The of feature increases from to across convolutional feature extraction for texture and recognition. The is the model of the proposed by in which to optimize and The is of designed to enhance efficiency while maintaining high of the is in 2. The model The is a deep convolutional neural network proposed by which model efficiency through the use of and The network convolutional with varying multi-scale feature it and to reduce computational while maintaining high These for detecting intricate texture such as in raveling diseases. Figure The RegNet RegNet a systematic approach to network to balance and The is to the of the of the network . The three and The of of each containing convolutional and This efficient feature extraction while critical for identifying pavement texture To achieve optimal during three learning were is shown in which the at the the at the is the learning and is the of the loss at the The model based on the of the loss with to the A learning was combined with to and prevent in the for the are as to the the first of the and the of the and are the of learning for each by the of and It of the learning to adapt during for the are as and is the average of the is the typically set to and is a small set to to prevent by The the learning by a average of ensuring This optimizer was found to the in the raveling detection task and was as the final optimizer for model this the of three and of loss were across network VGG16, Eff-B0, InceptionV3, and The most optimizer was then for The validation loss for each optimizer are shown in Figure The loss of the models are shown in Figure Among the and the final loss by with both a final loss of approximately contrast, the optimizer and its final loss on the the optimizer in of both and loss it is as the optimizer for raveling distress training across techniques were to mitigate and enhance model commonly to as was into the loss to large thereby model and generalization The is shown in Furthermore, with a were into the network to randomly during the of by the of an early was utilized to training the validation loss to improve a of ensuring the model optimal to the training the the loss of the model on the training the in the is the total of and is the that the of the in this study were conducted using a high-performance computing with the and in The the deep learning which provided robust for model and training. The for model training were based on to optimal and These are in The input image size of 224×224 pixels was to the requirements of the CNN while a size of for efficient training on the The optimizer was based on its during as in A learning of combined with a of and a of and efficient during training. The dataset was split into training data and validation data, a balanced to the models and their generalization The models were trained for with early to requires CNN models to robust feature and optimal in the task of raveling detection. The of the CNN Eff-B0, InceptionV3, and based on their in addressing the texture-based classification challenge of raveling detection in asphalt was for its feature extraction which effective for texture its large Eff-B0, with to achieve an optimal balance between accuracy and computational efficiency, making it for and robust multi-scale feature extraction through its convolutional and which both training and The RegNet a optimized that a computational making it for varying resource A structural the structural and computational of the models was provided in the for their in robust and efficient raveling detection. The models in were to a and representative framework for raveling detection, structural and varying from the of the to the intricate multi-scale of the InceptionV3, the model a of CNN both lightweight structures such as the and models including the VGG16, the framework systematic of This structural further across high-performance to enhancing the of the models for of the models was using to a comprehensive and robust the total of of a while of an of an and of a this and were to the results based on the and in to the the was to the overall classification The was by varying the classification and the the as shown in and The provided a the overall classification with a higher between These provided a comprehensive of the of the models, a of their and effectiveness in the between and such as the and were particularly for into model effectiveness accuracy comprehensive framework integrating and was to rigorous and of model A approach was to enhance with across to reduce The early validation loss not improve for a of to and a training of to efficiency and through and with a of less than observed using of learning and for early detection of or This approach a robust and of models under in accuracy and loss the training are in to which the of the deep learning models before and TL across training learning models, such as and RegNet a rapid in accuracy on both the training and validation datasets with increasing training the accuracy of models The accuracy of and significantly during the that the models are from the and RegNet both achieve accuracy levels that the fine-tuned TL models can and the of the raveling The loss of these models rapidly in the early of training and loss and RegNet the three models than the of training accuracy and deep learning models based on TL are for the raveling disease task to models trained from is a that the classification binary classification it is a of and as shown in Figure image classification the the and of the dataset, a more of the classification across classification is in in the shown in Figure The model identified images of raveling disease and of normal images, with one image as The model images as normal and identified normal images as the raveling The classification results of the original and TL models for each model are in each for the classification The Eff-B0, the InceptionV3, and the high accuracy in identifying raveling disease, between and normal images. The classification of the original and Eff-B0, InceptionV3, and on the validation The results are in which including and The the overall the original models, achieving an accuracy of and an of which precision and balanced The also with an accuracy of and an of its is than that of the The RegNet an accuracy of and an of The the with an accuracy of and an of primarily due to its of enhances and for models. RegNet the accuracy of has a precision of and across the overall with across including an accuracy of and a precision of The InceptionV3, slightly at to other models. The a significant an accuracy of and an of it the other models in overall TL significantly the of models. The of the original and TL models is illustrated in Figure The in the of each model before and TL are shown in Figure The the approaches the the the model to the the and the the original models. The an of on the validation while the approximately For InceptionV3, the increases the is which is significantly to the and the Eff-B0, and its As for the its is to the in the that its to between and is the of models with comprehensive of CNN structures their varying effectiveness in raveling detection. The and are across several including structural computational and CNN characteristics that influence its and computational The at capturing features and learning texture making it for raveling detection, it from high in and is with significant computational The efficient feature extraction by network and multi-scale feature extraction for detecting diverse raveling and with high computational The enhances precision through feature and efficient by its it is limited by training due to its and to input image The RegNet balanced through systematic architecture across varying conditions, and has resource making it for The computational resource requirements and for the structures are shown in the of the computational resource the notable the models in of their computational and as the most a of and a of which its for high-performance computing efficiency, with a a a and a rapid of This lightweight nature it particularly for where computational are limited or where is and RegNet a in of resource between and across the A critical is the and between model size and computational resource on at the of efficiency, that while it in capturing it is less for with resource contrast, a balance between accuracy and efficiency, which up for across various critical of each model and their to accuracy are in the resource of models is illustrated in Figure the between computational requirements and The of TL is in of its impact on training efficiency and model The results across models, both in of training and enhanced classification TL significantly reduces the training for models, as shown in The include the of training across models through The RegNet the of and by the with a of and and the with and The the of which can be to its more in training the efficiency of TL by pre-trained which model The a learning in the models. on detecting and in pavement images, capturing visual patterns such as and simple as shown in Figure and Figure on these features to more the between and patterns of raveling and normal Figure Deep high-level features to raveling, such as aggregate and which are for accurate classification and decision-making. Figure and Figure These the to increasingly features tailored to the task of raveling is to further the models by the of that influence the The results are in Figure Figure distinct attention patterns in the For raveling images, the models primarily on with aggregate such as areas with or loose which are critical of raveling damage and are with visual of the For normal pavement images, the models attention more across the surface texture, the of normal pavement where the TL models a more on features to the original models, that TL enhances the to critical thereby classification accuracy and the effectiveness of feature such as convolutional layer and in into and the models The results demonstrate that TL not also enhances the and accuracy of feature comprehensive of CNN structures and TL several and for models in this study limitations and for results of this study as the most model for raveling detection, a balance of high computational efficiency, and robust an accuracy of it ensures reliable detection across diverse pavement while maintaining requirements and making it for such as and the computational of the proposed method is not in and also in the as an an of image with an average power of approximately during The resulting for images is The of lightweight models more with the approximately J the as as the the computational efficiency of the its across a of Figure the results of on a The system is as the through the which to the system for image pavement images are to the where they automatic image preprocessing to by lighting variations and environmental the images are and to the for accurate pavement raveling detection. The results are using with a reliable of detection This the of the trained model in pavement inspection efficient and accurate automated pavement distress detection on computing these several limitations for The binary classification on normal raveling be to classification to detect other pavement such as potholes, and the models with such as or continuous of pavement conditions. the models to varying environmental including and is critical for enhancing in and techniques such as or data augmentation address these research is also needed to improve the of model decisions, to the This the use of such as and to provide into model and enhance in automated that the proposed framework not raveling detection also an for large-scale of intelligent pavement maintenance systems to its addressing the identified limitations and research the approach be to in infrastructure management and a comprehensive study on raveling disease in asphalt using deep learning The and of the research are as A systematic approach for raveling was by conventional CNN structures with the training through a of the for its and techniques such as early data and to prevent and enhance model The TL approach significantly enhanced the of CNN The model accuracy from to The balanced with the accuracy of the precision of the of and the of The techniques a feature extraction with early capturing textures, identifying and deep high-level features to raveling, while the on with TL models more and reliable attention to This study the potential of deep learning models in pavement combined with learning, both high accuracy and in raveling detection, thereby an efficient and for intelligent pavement automated inspection Although this study significant distress classification (e.g., potholes, with systems for continuous assessment, and model under varying environmental such as and surface study the of automated raveling detection in pavement maintenance The of TL its potential for other pavement disease high computational efficiency, and the proposed approach to the of infrastructure management

Links and sources

Comments