Multi-Objective Optimization for Sustainable Pavement Maintenance Decision...

[Full title: Multi-Objective Optimization for Sustainable Pavement Maintenance Decision Making by Integrating Pavement Image Segmentation and TOPSIS Methods]

[Find the meaning and references behind the names: Shu, Four, Doi, Human, Work, List, January, Save, Basel, Stride, Road, Low, Net, Works, Development, Time, Carbon, Edu, China, February, Multi, Ability, Chong, Size, Under, High, Few, Bridge, Cost, Open, Comfort, Parts, Study, Goal, Image, Strong, Able, Need, Liao]

Citation: Chong, D.; Liao, P.; Fu, W Multi-Objective Optimization for Sustainable Pavement Maintenance Decision Making by Integrating Pavement Image Segmentation and TOPSIS Methods Sustainability 2024 , 16 , 1257. https://doi.org/10.3390/ su 16031257 Academic Editor: El ˙zbieta Macioszek Received: 8 January 2024 Revised: 30 January 2024 Accepted: 30 January 2024 Published: 1 February 2024 Copyright: © 2024 by the authors Licensee MDPI, Basel, Switzerland This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/) sustainability Article Multi-Objective Optimization for Sustainable Pavement Maintenance Decision Making by Integrating Pavement Image Segmentation and TOPSIS Methods Dan Chong 1 , Peiyi Liao 1, * and Wurong Fu 2 1 School of Management, Shanghai University, Shanghai 200444, China; chongdan@shu.edu.cn 2 Shanghai Road & Bridge (Group) Co., Ltd., Shanghai 200433, China; fuwurong@126.com * Correspondence: lpy@shu.edu.cn Abstract: To provide a low-carbon economy maintenance strategy is the most challenging problem faced by pavement management authorities under the restricted budget and significant environmental repercussions. The development of a multi-objective optimization model for pavement maintenance decision making is essential to formulate pavements. Nevertheless, the existing automatic detection can only recognize and classify pavement distress. However, few studies are able to accurately determine the precise dimensions of specific distresses such as cracks and potholes, especially combined with the actual size of the image. This limitation hinders the ability to provide specific maintenance recommendations and make optimal maintenance decisions. Therefore, this paper develops a comprehensive and effective multi-objective decision-making framework for pavement maintenance. This framework consists of four distinct components: (1) recognizing the dimensions of pavement distresses based on the pavement image segmentation technique; (2) compiling a list of viable pavement maintenance strategies; (3) assessing the costs and carbon emissions of these strategies; and (4) optimizing decisions on pavement maintenance. We used the U-Net algorithm to accurately recognize the dimensions of pavement distresses, while an improved entropy-weighted TOPSIS model was proposed to determine the optimal pavement maintenance strategy with the lowest cost and carbon emissions. The results indicated that the pavement distress dimension recognition model achieved a high accuracy of 96.88%, and the TOPSIS model identified the optimal maintenance strategy with a score of 99.16. This maintenance strategy achieved a substantial reduction of 30.80% in carbon emissions and a cost reduction of 20.81% compared to the highest values among all maintenance strategies. This study not only provides a scientifically objective method for making pavement maintenance decisions but also offers specific, quantifiable maintenance programs, marking a stride towards more environmentally friendly and cost-effective road maintenance. It also contributes to the sustainability of pavement maintenance Keywords: sustainable pavement maintenance; pavement image segmentation; TOPSIS method; multi-objective decision making; carbon emission 1. Introduction Prolonged and extensive highway construction has brought increasing attention to pavement maintenance. To ensure the desired performance of highways, they need regular maintenance. The goal of pavement maintenance is to maintain the usability of the pavement; to restore the damaged parts in time; to ensure the safety, comfort, and smoothness of traffic; and to save transportation costs and time. In addition to maintaining the sustainability of pavement work, the environmental impacts and costs incurred in maintenance works can equally affect the sustainability of society. Efficient and accurate maintenance strategies can reduce the carbon emissions of the transportation industry, while also saving maintenance costs. The traditional approach to pavement assessment is a visual inspection that can be conducted by human experts, which is considered the easiest method [ 1 ]. Sustainability 2024 , 16 , 1257. https://doi.org/10.3390/su 16031257 https://www.mdpi.com/journal/sustainability

[Find the meaning and references behind the names: Rapid, Class, Level, Local, Bee, Life, Urban, Poor, Field, Vector, Hahn, Bio, Data, Deep, Hoang, Energy, Lies, Given, Aid, Offer, Evolution, Due, Manual, Non, Ouma, Pixel, Black, Quality, Super, Bang, Box]

Sustainability 2024 , 16 , 1257 2 of 24 Despite providing roughly exact examination outcomes, there are some downsides to the established approach such as the impact on traffic, nonobjective evaluation, hazards for site inspectors, considerable expense, and poor performance [ 2 , 3 ]. In addition to how pavement damage is detected, most of the materials currently used to maintain asphalt pavements are associated with high emissions, largely due to the nature of asphalt mixes For example, processes such as asphalt patching can lead to increased emissions due to the energy-intensive nature of the rehabilitation process [ 4 ]. While the use of recycled asphalt pavement or bio-based binders has the potential to reduce emissions, they have not yet gained widespread adoption, mainly due to concerns about their pavement performance [ 5 ]. Given these challenges, an important issue facing the maintenance of asphalt pavements is how to efficiently and objectively conduct road surface damage detection and propose maintenance solutions that are both low-carbon and economical Rapid and efficient pavement inspection facilitates the implementation of pavement maintenance tasks. Bang et al. proposed a pixel-level detection method for identifying road cracks in black-box images using a deep convolutional encoder–decoder network [ 6 ]. Li et al. employed a model-based transfer learning strategy for initializing the parameters of a Fully Convolutional Network (FCN), and their research results demonstrated a pixel accuracy of 98.61% [ 7 ]. Majidifard et al. established a U-Net-based model to quantify the severity of distress. Ultimately, by integrating the YOLO and U-net models, they developed a hybrid model for classifying distress and simultaneously quantifying its severity [ 8 ]. Their research results can be conveniently utilized for assessing the condition of pavements during their service life and aid in making effective decisions for the repair or reconstruction of roads at the appropriate time. Ouma and Hahn based their approach on multi-scale texture-based image filtering, utilized wavelet transform for text on representation, and employed the Fuzzy C-Means (FCM) algorithm for super-pixel clustering of pavement defects and non-defects. They proposed a low-cost detection method based on 2 D visual images for identifying potholes in urban asphalt pavements [ 9 ]. Hoang et al. developed a machine learning model composed of multi-class Support Vector Machines and an Artificial Bee Colony optimization algorithm for classifying pavement cracks [ 10 ]. These studies indicate that traditional manual detection methods have been phased out. The current approach to road surface damage detection primarily relies on image recognition technology for automation, achieving a high level of accuracy. However, this method can only determine the presence and type of damage. Since it cannot quantify the extent of road surface damage, it fails to provide conclusions on whether maintenance is necessary, let alone offer specific and effective maintenance strategies. Pixel-level pavement detection can provide accurate pavement parameters for pavement condition evaluation, and U-Net is a method of pixel-level segmentation [ 11 ], which is capable of segmenting the pixel information of the pavement distresses in the image. The U-Net algorithm has been used for the pavement distress dimensions recognition in this study In the field of asphalt pavement management, formulating effective pavement maintenance strategies is a significant issue. It involves creating reasonable maintenance strategies based on pavement conditions, traffic volume, and lifespan, aiming to extend the pavement’s service life while reducing maintenance costs and environmental impact [ 12 , 13 ]. Current representative research focuses on optimization algorithms, including Genetic Algorithms (GA) and Dynamic Programming (DP) methods. Genetic Algorithms, based on the concept of genetic evolution, are widely used in asphalt pavement management for their ability to handle numerous parameters and variables and search for global optima [ 14 ]. Dynamic Programming is also extensively used in asphalt pavement management. Its advantage lies in finding global optima without falling into dead loops or local optima. Integrating DP algorithms with deep learning has been explored to enhance solution accuracy [ 15 , 16 ]. However, DP has its drawbacks, such as exponential growth in computational complexity when handling high-dimensional problems, leading to prolonged solution times. DP also requires precise pavement condition data and corresponding decision variables, demanding high data accuracy and quality [ 17 , 18 ]. Despite the contributions of these

[Find the meaning and references behind the names: Top, Step, Ranking, Seg, Peer, Sis, Future, Ing, Area, Sion, Knowledge, Light, Still, Pro]

Sustainability 2024 , 16 , 1257 3 of 24 studies in developing maintenance strategies, there are still limitations to address. Decision models typically only consider the technical and economic impacts of asphalt pavement maintenance, neglecting environmental and social impacts, such as carbon emissions and societal costs [ 19 ]. Integrating these factors into decision models remains a challenge TOPSIS has the capability to generate a ranking of alternatives by effectively utilizing the attribute information of variables [ 20 ]. We have selected the TOPSIS model as our approach for multi-objective decision making. To avoid the subjectivity of indicators, we selected the entropy weighting method to determine the weights of the indicators Few studies have been able to convert the image dimensions of pavement distress to actual dimensions and on this basis suggest a specific maintenance strategy for specific pavement situations. In light of the aforementioned challenges and knowledge gaps, the objective of this study is to propose a comprehensive and effective multi-objective decision-making framework for pavement maintenance. We innovatively proposed a multiobjective decision-making method by combining the dimensions of pavement damage images, converting them to the actual dimensions of the damage, and considering both carbon emission and cost metrics. It is able to propose specific maintenance strategies for specific pavement conditions. Four distinct components comprised this framework: (1) recognizing the dimension of pavement distresses based on the pavement image segmentation technique; (2) compiling a list of viable pavement maintenance strategies; (3) assessing the costs and carbon emissions of these strategies; and (4) optimizing decisions on pavement maintenance based an improved entropy-weighted TOPSIS model. This paper is organized as follows: Section 2 details the research methodology. Section 3 describes the process and methods of data collection. Section 4 presents the experimental results and analyzes these results in depth. Section 5 summarizes this study, discusses its limitations, and provides an outlook on future research directions 2. Research Framework Our study was conducted in four steps: (1) we detected the types and recognized the dimensions of pavement distresses based on the pavement image segmentation method (Figure 1 a); (2) a list of viable pavement maintenance strategies was compiled based on the automatic pavement detection in the previous step (Figure 1 b); (3) we assessed the carbon emissions and costs of these viable strategies (Figure 1 c); and (4) a multi-objective decision-making model for pavement maintenance strategies based on the improved entropy-weighted TOPSIS method was established (Figure 1 d). Sustainability 2024 , 16 , x FOR PEER REVIEW 3 of 24 computational complexity when handling high ‐ dimensional problems, leading to pro ‐ longed solution times DP also requires precise pavement condition data and correspond ‐ ing decision variables, demanding high data accuracy and quality [17,18] Despite the con ‐ tributions of these studies in developing maintenance strategies, there are still limitations to address Decision models typically only consider the technical and economic impacts of asphalt pavement maintenance, neglecting environmental and social impacts, such as carbon emissions and societal costs [19] Integrating these factors into decision models remains a challenge TOPSIS has the capability to generate a ranking of alternatives by effectively utilizing the attribute information of variables [20] We have selected the TOP ‐ SIS model as our approach for multi ‐ objective decision making To avoid the subjectivity of indicators, we selected the entropy weighting method to determine the weights of the indicators Few studies have been able to convert the image dimensions of pavement distress to actual dimensions and on this basis suggest a specific maintenance strategy for specific pavement situations In light of the aforementioned challenges and knowledge gaps, the objective of this study is to propose a comprehensive and effective multi ‐ objective deci ‐ sion ‐ making framework for pavement maintenance We innovatively proposed a multi ‐ objective decision ‐ making method by combining the dimensions of pavement damage images, converting them to the actual dimensions of the damage, and considering both carbon emission and cost metrics It is able to propose specific maintenance strategies for specific pavement conditions Four distinct components comprised this framework: (1) recognizing the dimension of pavement distresses based on the pavement image segmen ‐ tation technique; (2) compiling a list of viable pavement maintenance strategies; (3) as ‐ sessing the costs and carbon emissions of these strategies; and (4) optimizing decisions on pavement maintenance based an improved entropy ‐ weighted TOPSIS model This paper is organized as follows: Section 2 details the research methodology Section 3 describes the process and methods of data collection Section 4 presents the experimental results and analyzes these results in depth Section 5 summarizes this study, discusses its limita ‐ tions, and provides an outlook on future research directions 2. Research Framework Our study was conducted in four steps: (1) we detected the types and recognized the dimensions of pavement distresses based on the pavement image segmentation method (Figure 1 a); (2) a list of viable pavement maintenance strategies was compiled based on the automatic pavement detection in the previous step (Figure 1 b); (3) we assessed the carbon emissions and costs of these viable strategies (Figure 1 c); and (4) a multi ‐ objective decision ‐ making model for pavement maintenance strategies based on the improved en ‐ tropy ‐ weighted TOPSIS method was established (Figure 1 d). Figure 1. Research framework 2.1. Recognizing Dimensions of Pavement Distresses The pavement condition index (PCI) depends on the actual area of pavement damage, making it crucial to accurately identify the actual dimension of pavement distress. Only seg-

[Find the meaning and references behind the names: Connect, Angle, Map, Ment, Unique, Better, Bias, Key, Path, Enough, Final, Cor, Loss, Maps, Ture, Original, Dam, Age, Contracting, Architecture, State, Jump, Rich, Fine, Fuse, Small, Scales, Shown]

Sustainability 2024 , 16 , 1257 4 of 24 mentation can achieve pixel-level recognition accuracy and output geometry information of distress that can be utilized to evaluate the state of pavement performance more practically Therefore, segmentation has become the mainstream in the field of pavement distress recognition [ 21 ]. The U-Net algorithm, which derives its name from its U-shaped structure, is a deep learning network for the task of image segmentation. U-Net is mainly used for solving tasks such as medical image segmentation, where fine boundary information is usually required. The structure of U-Net is divided into two parts: the encoder and the decoder. The encoder is responsible for gradually down-sampling the input image into a small feature map while preserving the contextual information of the image. The decoder gradually up-samples this small feature map into a segmentation prediction map of the same dimension as the original input image. This symmetrical U-shaped structure of U-Net (Figure 2 ) consists of a contracting (down-sampling) path and an expanding (up-sampling) path. This structure allows the network to capture the contextual information of an image at different scales while preserving the detailed information. It uses the jump connections to connect the feature maps of the down-sampling path to the corresponding layers of the up-sampling path. This connection helps the network to recover detailed information during the up-sampling process and reduces information loss. Jump connections allow the network to learn detailed features better even with limited training data, improving segmentation accuracy. The design of U-Net also focuses on feature learning in localized regions. This means that it can effectively learn rich local features from small samples. Due to its unique up-sampling and jump-joining structure, U-Net can effectively fuse features from different layers. This feature fusion enables the model to learn enough information to achieve accurate image segmentation. The final outcome is pixel images that exclusively depict pavement damage, as shown in Figure 3 . Sustainability 2024 , 16 , x FOR PEER REVIEW 4 of 24 Figure 1. Research framework 2.1. Recognizing Dimensions of Pavement Distresses The pavement condition index (PCI) depends on the actual area of pavement dam ‐ age, making it crucial to accurately identify the actual dimension of pavement distress Only segmentation can achieve pixel ‐ level recognition accuracy and output geometry in ‐ formation of distress that can be utilized to evaluate the state of pavement performance more practically Therefore, segmentation has become the mainstream in the field of pave ‐ ment distress recognition [21] The U ‐ Net algorithm, which derives its name from its U ‐ shaped structure, is a deep learning network for the task of image segmentation U ‐ Net is mainly used for solving tasks such as medical image segmentation, where fine boundary information is usually required The structure of U ‐ Net is divided into two parts: the en ‐ coder and the decoder The encoder is responsible for gradually down ‐ sampling the input image into a small feature map while preserving the contextual information of the image The decoder gradually up ‐ samples this small feature map into a segmentation prediction map of the same dimension as the original input image This symmetrical U ‐ shaped struc ‐ ture of U ‐ Net (Figure 2) consists of a contracting (down ‐ sampling) path and an expanding (up ‐ sampling) path This structure allows the network to capture the contextual infor ‐ mation of an image at different scales while preserving the detailed information It uses the jump connections to connect the feature maps of the down ‐ sampling path to the cor ‐ responding layers of the up ‐ sampling path This connection helps the network to recover detailed information during the up ‐ sampling process and reduces information loss Jump connections allow the network to learn detailed features better even with limited training data, improving segmentation accuracy The design of U ‐ Net also focuses on feature learning in localized regions This means that it can effectively learn rich local features from small samples Due to its unique up ‐ sampling and jump ‐ joining structure, U ‐ Net can effectively fuse features from different layers This feature fusion enables the model to learn enough information to achieve accurate image segmentation The final outcome is pixel images that exclusively depict pavement damage, as shown in Figure 3 Figure 2. U ‐ Net network architecture diagram Figure 3. The result of pavement distress image segmentation Figure 2. U-Net network architecture diagram Sustainability 2024 , 16 , x FOR PEER REVIEW 4 of 24 Figure 1. Research framework 2.1. Recognizing Dimensions of Pavement Distresses The pavement condition index (PCI) depends on the actual area of pavement dam ‐ age, making it crucial to accurately identify the actual dimension of pavement distress Only segmentation can achieve pixel ‐ level recognition accuracy and output geometry in ‐ formation of distress that can be utilized to evaluate the state of pavement performance more practically Therefore, segmentation has become the mainstream in the field of pave ‐ ment distress recognition [21] The U ‐ Net algorithm, which derives its name from its U ‐ shaped structure, is a deep learning network for the task of image segmentation U ‐ Net is mainly used for solving tasks such as medical image segmentation, where fine boundary information is usually required The structure of U ‐ Net is divided into two parts: the en ‐ coder and the decoder The encoder is responsible for gradually down ‐ sampling the input image into a small feature map while preserving the contextual information of the image The decoder gradually up ‐ samples this small feature map into a segmentation prediction map of the same dimension as the original input image This symmetrical U ‐ shaped struc ‐ ture of U ‐ Net (Figure 2) consists of a contracting (down ‐ sampling) path and an expanding (up ‐ sampling) path This structure allows the network to capture the contextual infor ‐ mation of an image at different scales while preserving the detailed information It uses the jump connections to connect the feature maps of the down ‐ sampling path to the cor ‐ responding layers of the up ‐ sampling path This connection helps the network to recover detailed information during the up ‐ sampling process and reduces information loss Jump connections allow the network to learn detailed features better even with limited training data, improving segmentation accuracy The design of U ‐ Net also focuses on feature learning in localized regions This means that it can effectively learn rich local features from small samples Due to its unique up ‐ sampling and jump ‐ joining structure, U ‐ Net can effectively fuse features from different layers This feature fusion enables the model to learn enough information to achieve accurate image segmentation The final outcome is pixel images that exclusively depict pavement damage, as shown in Figure 3 Figure 2. U ‐ Net network architecture diagram Figure 3. The result of pavement distress image segmentation Figure 3. The result of pavement distress image segmentation To make the identification of pavement distress more precise, in the pavement damage image collection phase of this study, two key measures were taken to ensure the accuracy of the data: first, the actual dimensions of the images were recorded by using box markers at the time of shooting to facilitate subsequent analysis; and second, a shooting angle perpendicular to the ground was used to minimize the bias of the image information induced

[Find the meaning and references behind the names: Stage, Own, Standard, Excellent, Minor, Excel, Large, Areas, Exceptional, Hand, Major, Point, Table, Room, Rate, Semi, Good]

Sustainability 2024 , 16 , 1257 5 of 24 by the shooting angle. In the image-processing stage, we applied the U-Net algorithm to segment the collected images and exported the segmented image pixel information to Excel format. This step enabled us to accurately count the number of pixel points representing the damaged areas of the pavement. Based on these data, we used Equations (1) and (2) to calculate the actual dimension of the pavement damage: S piexl = S actual C piexl , (1) S pavement damage = S piexl × C pavement damage , (2) where S piexl means the area of each pixel point, S actual means the actual area of the image, C piexl means the product of the pixel dimensions of the image, S pavement damage means the actual dimension of the pavement damage, and C pavement damage means the number of pixel points occupied by the pavement damage 2.2. Compiling a List of Viable Pavement Maintenance Strategies The PCI is often used as an important decision-making index when analyzing maintenance decisions and selecting maintenance countermeasures, which provides an objective and reasonable decision-making basis for pavement maintenance, and thus is also an important index widely used worldwide to determine maintenance strategies. The dimension of the PCI is determined by the Distress Ratio (DR), which is calculated according to different parameters of disease types. According to the Standard for Evaluation of Highway Technical Condition in China (JTG 5210-2018) [ 22 ], the equation for calculating the PCI is as follows: PCI = 100 − a 0 DR a 1 , (3) DR = 100 × ∑ i 0 i = 1 ω i A i A , (4) where DR is the pavement distress rate, which is the percentage of the distress area of various distresses to the total area of the pavement; A is the area of the pavement; A i is the area of the pavement with type i distress; and ω i is the weight of the distress in the i th category. The value of ω i is 0.6 or 2.0 when the unit of measure of pavement distress is length, and 1.0 if the unit of measure is area a 0 = 15, a 1 = 0.412, and i 0 is the total number of distress types In this study, the PCI is used as the primary decision-making objective for whether maintenance is needed, and the specific maintenance countermeasures are shown in Table 1 . Table 1. Pavement damage ratings and maintenance countermeasures Grading Exceptional Excellent Good Moderate Poor Very Poor PCI 100~91 90~8 l 80~71 70~51 50~31 ≤ 30 Maintenance Countermeasures - Routine maintenance Minor repair Medium repair Major repair Reconstruction In asphalt pavement maintenance, the selection of appropriate maintenance materials is crucial. Major maintenance materials include petroleum asphalt, emulsified asphalt, and modified asphalt, each of which has its own unique characteristics and application scenarios. Petroleum asphalt is solid or semi-solid at room temperature and is suitable for filling large cracks and potholes, especially in areas with low temperatures. Emulsified asphalt is an emulsion made by mixing asphalt with water and is suitable for small cracks and pavement sealing. Modified asphalt, on the other hand, improves its performance by adding specific modifiers or additives, especially in areas with heavy traffic or extreme climates. When selecting maintenance materials, it is necessary to consider the local climatic conditions, the state of pavement damage, and the feasibility of construction.

[Find the meaning and references behind the names: Mate, Tion, Ical, Dif, Matic, Lower]

Sustainability 2024 , 16 , 1257 6 of 24 In the use of asphalt materials, their environmental impact is particularly reflected in carbon emissions. Petroleum asphalt consumes large amounts of energy during production and heating, resulting in high carbon emissions. Emulsified asphalt has significant advantages in reducing carbon emissions, and due to its water-based nature, no additional heating is required during construction. Modified asphalt, on the other hand, has improved temperature resistance and durability, which helps to reduce the frequency and intensity of road maintenance, thereby reducing carbon emissions. The selection of lower carbonemitting asphalt materials and the use of efficient construction techniques are critical to reducing the carbon footprint of pavement maintenance. Each of these maintenance materials has its own advantages and disadvantages, but all of them can be used as materials for crack and pothole maintenance, and the carbon emissions and costs of using them for maintenance are not the same. Therefore, in this study, three types of asphalt, namely petroleum asphalt, emulsified asphalt, and modified asphalt, were selected as fillers for cracks and potholes By reviewing the Technical Specification for Highway Asphalt Pavement Maintenance in China (JTG 5142-2019) [ 23 ] and related information on pavement damage maintenance measures, the maintenance process for cracks and potholes is obtained, as shown in Figure 4 . Sustainability 2024 , 16 , x FOR PEER REVIEW 6 of 24 filling large cracks and potholes, especially in areas with low temperatures Emulsified asphalt is an emulsion made by mixing asphalt with water and is suitable for small cracks and pavement sealing Modified asphalt, on the other hand, improves its performance by adding specific modifiers or additives, especially in areas with heavy traffic or extreme climates When selecting maintenance materials, it is necessary to consider the local cli ‐ matic conditions, the state of pavement damage, and the feasibility of construction In the use of asphalt materials, their environmental impact is particularly reflected in carbon emissions Petroleum asphalt consumes large amounts of energy during produc ‐ tion and heating, resulting in high carbon emissions Emulsified asphalt has significant advantages in reducing carbon emissions, and due to its water ‐ based nature, no additional heating is required during construction Modified asphalt, on the other hand, has im ‐ proved temperature resistance and durability, which helps to reduce the frequency and intensity of road maintenance, thereby reducing carbon emissions The selection of lower carbon ‐ emitting asphalt materials and the use of efficient construction techniques are crit ‐ ical to reducing the carbon footprint of pavement maintenance Each of these maintenance materials has its own advantages and disadvantages, but all of them can be used as mate ‐ rials for crack and pothole maintenance, and the carbon emissions and costs of using them for maintenance are not the same Therefore, in this study, three types of asphalt, namely petroleum asphalt, emulsified asphalt, and modified asphalt, were selected as fillers for cracks and potholes By reviewing the Technical Specification for Highway Asphalt Pavement Maintenance in China (JTG 5142 ‐ 2019) [23] and related information on pavement damage maintenance measures, the maintenance process for cracks and potholes is obtained, as shown in Figure 4 Figure 4. Pavement damage maintenance process Based on the above available maintenance materials, machines, and labor, combined with the maintenance process shown in Figure 4, a maintenance option selection table can be obtained, as shown in Table 2, and then, according to the maintenance strategy selec ‐ tion table, a total of 108 feasible maintenance strategies can be combined by choosing dif ‐ ferent labor, materials, and machines Some of the 108 maintenance strategies are shown in Table 3 Figure 4. Pavement damage maintenance process Based on the above available maintenance materials, machines, and labor, combined with the maintenance process shown in Figure 4 , a maintenance option selection table can be obtained, as shown in Table 2 , and then, according to the maintenance strategy selection table, a total of 108 feasible maintenance strategies can be combined by choosing different labor, materials, and machines. Some of the 108 maintenance strategies are shown in Table 3 .

[Find the meaning and references behind the names: Breaker, Mini, Cleaning, Block, Smooth, Roller, Blower, Factor]

Sustainability 2024 , 16 , 1257 7 of 24 Table 2. Maintenance strategy selection table Pavement Damage Material Grooving Cleaning and Drying Sealing Paving Adjusting Compaction Crack Petroleum asphalt, emulsified asphalt, modified asphalt Electric concrete saw, manual grooving Handheld electric blower Asphalt crack sealer Manual adjusting Pothole Petroleum asphalt, emulsified asphalt, modified asphalt Road breaker, manual grooving Handheld electric blower Mini asphalt paver, manual paving Manual operation of an electric compactor, mini smooth wheel roller Table 3. Some of 108 maintenance strategies No. Distress Material Grooving Cleaning and Drying Sealing Paving Adjusting Compaction 1 Crack Petroleum asphalt Manual grooving Handheld electric blower Asphalt crack sealer Manual adjusting Pothole Emulsified asphalt Manual grooving Manual paving Manual operation of an electric compactor 2 Crack Petroleum asphalt Manual grooving Asphalt crack sealer Manual adjusting Pothole Emulsified asphalt Manual grooving Mini asphalt paver Manual operation of an electric compactor 3 Crack Petroleum asphalt Manual grooving Asphalt crack sealer Manual adjusting Pothole Emulsified asphalt Manual grooving Manual paving Mini Smooth wheel roller 4 Crack Petroleum asphalt Manual grooving Asphalt crack sealer Manual adjusting Pothole Emulsified asphalt Manual grooving Mini asphalt paver Mini Smooth wheel roller 5 Crack Petroleum asphalt Electric concrete saw Asphalt crack sealer Manual adjusting Pothole Emulsified asphalt Manual grooving Manual paving Manual operation of an electric compactor 6 Crack Petroleum asphalt Electric concrete saw Asphalt crack sealer Manual adjusting Pothole Emulsified asphalt Manual grooving Mini asphalt paver Manual operation of an electric compactor 103 Crack Modified asphalt Electric concrete saw Handheld electric blower Asphalt crack sealer Manual adjusting Pothole Emulsified asphalt Manual grooving Manual paving Mini Smooth wheel roller 104 Crack Modified asphalt Electric concrete saw Asphalt crack sealer Manual adjusting Pothole Emulsified asphalt Manual grooving Mini asphalt paver Mini Smooth wheel roller 105 Crack Modified asphalt Electric concrete saw Asphalt crack sealer Manual adjusting Pothole Emulsified asphalt Road breaker Manual paving Manual operation of an electric compactor 106 Crack Modified asphalt Electric concrete saw Asphalt crack sealer Manual adjusting Pothole Emulsified asphalt Road breaker Mini asphalt paver Manual operation of an electric compactor 107 Crack Modified asphalt Electric concrete saw Asphalt crack sealer Manual adjusting Pothole Emulsified asphalt Road breaker Manual paving Mini Smooth wheel roller 108 Crack Modified asphalt Electric concrete saw Asphalt crack sealer Manual adjusting Pothole Emulsified asphalt Road breaker Mini asphalt paver Mini Smooth wheel roller 2.3. Assessing the Carbon Emissions and Costs of These Strategies 2.3.1. Carbon Emissions Calculation Model In the process of measuring the carbon emission of pavement maintenance engineering construction activities, it is easier and more convenient to use the “emission factor method” in the material production, construction, and transportation stages, which only needs to obtain the amount of material or energy consumption and combine it with the corresponding emission factors to calculate the final emissions. The block diagram of carbon emission calculation for each maintenance strategy is shown in Figure 5 .

[Find the meaning and references behind the names: Raw, Chosen, Ith, Emis, Common]

Sustainability 2024 , 16 , 1257 8 of 24 Sustainability 2024 , 16 , x FOR PEER REVIEW 8 of 24 Figure 5. Framework for calculating carbon emissions from maintenance strategies The carbon emissions of the asphalt pavement maintenance phase include three sub ‐ phases: the raw material phase, the transportation phase, and the construction phase The carbon emissions generated by them can be expressed by Equation (5) {?} {?} {?} {?} , (5) where {?} means the total carbon emissions of the ith maintenance program in the asphalt pavement maintenance stage, {?} means the carbon emissions generated in the raw ma ‐ terial stage of each maintenance strategy, {?} means the carbon emissions generated in the transportation stage of each maintenance strategy, and {?} means the carbon emis ‐ sions generated in the construction stage of each maintenance strategy They are all in {?}{?} According to the Standard for Calculating Carbon Emissions from Buildings (GB/T 51366 ‐ 2019) [24], carbon emissions at the raw material stage should be calculated according to Equation (6) {?} {?} {?} , (6) where {?} means the carbon emission (kg CO 2 ) in the production phase of raw materials of the i th maintenance strategy, {?} means the consumption of the j th raw material, and {?} means the carbon emission factor of the j ‐ th raw material (kg CO 2 /unit quantity of raw material) The actual area of cracks and potholes can be obtained according to the identification of pavement damage dimensions in Section 2.1, and since the depth of cracks and potholes cannot be obtained directly by identification, the common depth of crack and pothole maintenance is chosen as the actual depth in this study, which is 2 cm and 4 cm, respec ‐ tively, and then the volume of maintenance materials can be calculated The consumption of the j th raw material {?} can be calculated by Equation (7) {?} {?} {?} (7) where {?} means the density of the j ‐ th raw material (kg/m 3 ) and {?} means the volume of the j th raw material (m 3 ) Carbon emissions from the transportation phase should be calculated according to Equation (8): {?} {?} {?} {?} , (8) Figure 5. Framework for calculating carbon emissions from maintenance strategies The carbon emissions of the asphalt pavement maintenance phase include three subphases: the raw material phase, the transportation phase, and the construction phase The carbon emissions generated by them can be expressed by Equation (5) E i = E iycl + E iys + E isg , (5) where E i means the total carbon emissions of the ith maintenance program in the asphalt pavement maintenance stage, E iycl means the carbon emissions generated in the raw material stage of each maintenance strategy, E iys means the carbon emissions generated in the transportation stage of each maintenance strategy, and E isg means the carbon emissions generated in the construction stage of each maintenance strategy. They are all in kg According to the Standard for Calculating Carbon Emissions from Buildings (GB/T 51366-2019) [ 24 ], carbon emissions at the raw material stage should be calculated according to Equation (6) E iycl = n ∑ j = 1 M j F j , (6) where E iycl means the carbon emission (kg CO 2 ) in the production phase of raw materials of the i th maintenance strategy, M j means the consumption of the j th raw material, and F j means the carbon emission factor of the j-th raw material (kg CO 2 /unit quantity of raw material) The actual area of cracks and potholes can be obtained according to the identification of pavement damage dimensions in Section 2.1 , and since the depth of cracks and potholes cannot be obtained directly by identification, the common depth of crack and pothole maintenance is chosen as the actual depth in this study, which is 2 cm and 4 cm, respectively, and then the volume of maintenance materials can be calculated. The consumption of the j th raw material M j can be calculated by Equation (7) M j = ρ j V j (7) where ρ j means the density of the j-th raw material (kg/m 3 ) and V j means the volume of the j th raw material (m 3 ) Carbon emissions from the transportation phase should be calculated according to Equation (8): E iys = n ∑ j = 1 M j D j T , (8)

[Find the meaning and references behind the names: Dis, Rial, Turn, Comes, Gram, Success, Nance]

Sustainability 2024 , 16 , 1257 9 of 24 where E iys means the carbon emission (kgCO 2 ) in the transportation phase of the i th maintenance strategy, M j means the consumption of the j-th material (t), D j means the average transportation distance of the j th material (km), and T means the carbon emission factor of transportation distance per unit weight (kg CO 2 /(t · km). It is appropriate to give priority to the actual transportation distance of materials. When the transportation distance of construction materials is unknown, the default transportation distance of 40 km given in the Standard for Calculating Carbon Emissions from Buildings (GB/T 51366-2019) [ 24 ] can be used Construction phase carbon emissions shall be calculated according to Equation (9): E isg = n ∑ j = 1 R j G j E j , (9) where: E isg means the carbon emission (kgCO 2 ) during the construction phase of the i th type of maintenance strategy, R j means the energy consumption per unit shift of the j th type of construction machine (kg/shift or kW · h/shift), G j means the consumption of the j th type of construction machine unit shift (unit shift), and E j means the carbon emission factor of the energy used by the j th type of construction machine (kgCO 2 /kg or kg CO 2 /kW · h) 2.3.2. Maintenance Costs Calculation Model Accurate costing is critical to the success of pavement maintenance projects. By comprehensively considering the cost of raw materials, labor, and machines, the maintenance cost of each maintenance program can be calculated, which in turn reflects the economic benefits of each maintenance strategy. The costing block diagram of the maintenance program in this study is shown in Figure 6 . Sustainability 2024 , 16 , x FOR PEER REVIEW 9 of 24 where {?} means the carbon emission (kgCO 2 ) in the transportation phase of the i th maintenance strategy, {?} means the consumption of the j ‐ th material (t), {?} means the av ‐ erage transportation distance of the j th material (km), and {?} means the carbon emission factor of transportation distance per unit weight (kg CO 2 /(t ∙ km) It is appropriate to give priority to the actual transportation distance of materials When the transportation dis ‐ tance of construction materials is unknown, the default transportation distance of 40 km given in the Standard for Calculating Carbon Emissions from Buildings (GB/T 51366 ‐ 2019) [24] can be used Construction phase carbon emissions shall be calculated according to Equation (9): {?} {?} {?} {?} , (9) where: {?} means the carbon emission (kgCO 2 ) during the construction phase of the i th type of maintenance strategy, {?} means the energy consumption per unit shift of the j th type of construction machine (kg/shift or kW ∙ h/shift), {?} means the consumption of the j th type of construction machine unit shift (unit shift), and {?} means the carbon emission factor of the energy used by the j th type of construction machine (kgCO 2 /kg or kg CO 2 /kW ∙ h) 2.3.2 Maintenance Costs Calculation Model Accurate costing is critical to the success of pavement maintenance projects By com ‐ prehensively considering the cost of raw materials, labor, and machines, the maintenance cost of each maintenance program can be calculated, which in turn reflects the economic benefits of each maintenance strategy The costing block diagram of the maintenance pro ‐ gram in this study is shown in Figure 6 Figure 6. Framework maintenance costing The three sub ‐ phases of the asphalt pavement maintenance phase are the raw mate ‐ rial phase, the transportation phase, and the construction phase, so the cost of the mainte ‐ nance program also comes from the labor, materials, and machines used in these three sub ‐ phases, and the cost of the whole maintenance phase can be calculated by Equation (10): {?} {?} {?} {?} , (10) where {?} means the total cost of the i th maintenance strategy in the maintenance phase of asphalt pavement, {?} means the labor cost of the i th maintenance strategy, {?} means Figure 6. Framework maintenance costing The three sub-phases of the asphalt pavement maintenance phase are the raw material phase, the transportation phase, and the construction phase, so the cost of the maintenance program also comes from the labor, materials, and machines used in these three sub-phases, and the cost of the whole maintenance phase can be calculated by Equation (10): C i = C irg + C icl + C ijx , (10) where C i means the total cost of the i th maintenance strategy in the maintenance phase of asphalt pavement, C irg means the labor cost of the i th maintenance strategy, C icl means the material cost of the i th strategy, and C ijx means the machine cost of the i th maintenance strategy. All of them are in CNY.

[Find the meaning and references behind the names: Ways, Ideal, Man, Choice, Main, Price, Days, Civil, Hours, Positive]

Sustainability 2024 , 16 , 1257 10 of 24 The labor cost of the i th maintenance strategy is calculated through Equation (11): C irg = R i D i P 1 , (11) where C irg means the cost of labor for the i th maintenance strategy (CNY), R i means the number of laborers for the i-th maintenance strategy, D i means the number of hours of labor work for the i th maintenance strategy (man-days), and P 1 means the unit price of labor for composite labor (maintenance, municipal civil works) The cost of materials for the i th maintenance strategy is calculated through Equation (12): C icl = M j P 2 j , (12) where C icl means the material cost of the i th maintenance strategy (CNY), M j means the consumption of the j th material in the i th maintenance strategy (t) and P 2 j means the price of the j th material in the i th maintenance strategy (CNY/t) The cost of machine for the i th maintenance strategy is calculated by Equation (13): C ijx = T j P 3 j , (13) where C ijx means the machine cost of the i th maintenance strategy (CNY), T j means the working hours of the j th type of machine in the i th maintenance strategy (shift), and P 3 j means the price of the j th type of machine in the i th maintenance strategy (CNY/shift) 2.4. Optimizing Decisions on Pavement Maintenance In this study, there are three objectives: the pavement condition index, the carbon emissions from maintenance strategies, and the maintenance costs. The main framework of the decision-making model is shown in Figure 7 , where the pavement condition index serves as the direct basis for deciding whether pavement damage maintenance is needed, the carbon emissions generated by the maintenance strategy and the cost of the maintenance strategy together serve as the decision-making objectives, and the optimal maintenance strategy is selected through the TOPSIS model Sustainability 2024 , 16 , x FOR PEER REVIEW 10 of 24 the material cost of the i th strategy, and {?} means the machine cost of the i th mainte ‐ nance strategy All of them are in CNY The labor cost of the i th maintenance strategy is calculated through Equation (11): {?} {?} {?} {?} , (11) where {?} means the cost of labor for the i th maintenance strategy (CNY), {?} means the number of laborers for the i ‐ th maintenance strategy, {?} means the number of hours of labor work for the i th maintenance strategy (man ‐ days), and {?} means the unit price of labor for composite labor (maintenance, municipal civil works) The cost of materials for the i th maintenance strategy is calculated through Equation (12): {?} {?} {?} , (12) where {?} means the material cost of the i th maintenance strategy (CNY), {?} means the consumption of the j th material in the i th maintenance strategy (t) and {?} means the price of the j th material in the i th maintenance strategy (CNY/t) The cost of machine for the i th maintenance strategy is calculated by Equation (13): {?} {?} {?} , (13) where {?} means the machine cost of the i th maintenance strategy (CNY), {?} means the working hours of the j th type of machine in the i th maintenance strategy (shift), and {?} means the price of the j th type of machine in the i th maintenance strategy (CNY/shift) 2.4. Optimizing Decisions on Pavement Maintenance In this study, there are three objectives: the pavement condition index, the carbon emissions from maintenance strategies, and the maintenance costs The main framework of the decision ‐ making model is shown in Figure 7, where the pavement condition index serves as the direct basis for deciding whether pavement damage maintenance is needed, the carbon emissions generated by the maintenance strategy and the cost of the mainte ‐ nance strategy together serve as the decision ‐ making objectives, and the optimal mainte ‐ nance strategy is selected through the TOPSIS model Figure 7. Framework of decision modeling In the TOPSIS methodology, a uniform conversion of all metrics to positive metrics is a prerequisite for effective comparisons Such a conversion ensures that when calculat ‐ ing the distance of each alternative from the ideal solution and the negative ideal solution, all metrics contribute to the direction of improving the comprehensive performance of the program This uniformity of metrics is the key to evaluating different scenarios and de ‐ termining the optimal choice in the TOPSIS methodology Therefore, converting metrics to positive metrics is an important step in the TOPSIS ‐ integrated evaluation method, which allows metrics of different natures to be evaluated and compared under a unified standard This conversion simplifies the decision ‐ making process and enables the TOPSIS method to be effectively applied in multi ‐ objective decision making There are two ways Figure 7. Framework of decision modeling In the TOPSIS methodology, a uniform conversion of all metrics to positive metrics is a prerequisite for effective comparisons. Such a conversion ensures that when calculating the distance of each alternative from the ideal solution and the negative ideal solution, all metrics contribute to the direction of improving the comprehensive performance of the program. This uniformity of metrics is the key to evaluating different scenarios and determining the optimal choice in the TOPSIS methodology. Therefore, converting metrics to positive metrics is an important step in the TOPSIS-integrated evaluation method, which allows metrics of different natures to be evaluated and compared under a unified standard This conversion simplifies the decision-making process and enables the TOPSIS method to be effectively applied in multi-objective decision making. There are two ways to convert a negative indicator to a positive indicator. If all the elements are positive, 1/ x can be used to convert it to a positive indicator. In addition to this, the conversion can be performed

[Find the meaning and references behind the names: Element, Less, Sum, Idea, Basic, Sample, Row, Close, Line]

Sustainability 2024 , 16 , 1257 11 of 24 exclusively by subtracting the value of the indicator from the maximum value of that indicator. This conversion of the maximum value minus the value of the indicator is intuitive but may result in the difference between the maximum value and the value of the indicator having a large impact on the result, and requires careful adjustment of the weights, or it may introduce an unjustified bias. Moreover, when the indicator value is very close to the maximum value, small changes may lead to large differences in the converted values, resulting in instability. The 1/ x approach is more stable in dealing with situations close to the maximum value and is less susceptible to the influence of extreme values. So, in this study, the 1/ x approach is used to convert negative type indicators to positive type indicators Since the units of carbon emissions and costs generated by conservation are “kg/m 3 ” and “CNY”, respectively, it is necessary to remove the excessive differences between the indicator values due to the different scales to ensure the validity of the evaluation indicators. In this paper, matrix standardization is used to eliminate the influence of different index outlines. From the 108 feasible maintenance strategies identified in Section 2.2 , there are a total of 108 objects to be evaluated in this paper, and there are 2 evaluation indicators, carbon emissions and costs, which constitute the forwarding matrix as follows: X =      x 11 x 12 · · · x 1 m x 21 x 22 · · · x 2 m . . . . . x n 1 x n 2 · · · x nm      ( n = 108, m = 2 ) Then, the matrix to which it is normalized is denoted as Z, and each element in Z: z ij = x ij / s n ∑ i = 1 x 2 ij ( i = 1, 2 · · · , 108, j = 1, 2 ) , where z ij is normalized for each element and n ∑ i = 1 x 2 ij is the sum of the squares of the elements of the column in which it is located, resulting in the final matrix: Z =      z 11 z 12 · · · z 1 m z 21 z 22 · · · z 2 m . . . . . z n 1 z n 2 · · · z nm      ( n = 108, m = 2 ) Each row of the matrix represents two indicator values for the same scenario, and each column represents the same indicator value for 108 scenarios, and the next step will be to utilize the entropy-weighting method to calculate the weight of each indicator based on this matrix The basic idea of the entropy-weighting method is to determine the weight of each indicator as the entropy value of its correlation to avoid the bias brought by subjective assignment, and at the same time, it can reflect the interrelationships among the indicators. Specifically, in the construction of a multi-objective decision-making model, for the carbon emissions and conservation costs generated by each maintenance strategy, you can calculate the maximum and minimum values in the sample data matrix, as well as the score of each sample, and then find out the normalized value and entropy value of each indicator. The advantage of this method is that it is more flexible, unconstrained, and highly operable according to the importance of the indicators to give the corresponding weight, the important indicators to give a larger weight, and the unimportant indicators to give a smaller weight, which is more in line with the nature of the weight. In this paper, a total of 108 maintenance strategies are needed to make decisions, and it is known that there are a total of two maintenance strategies produced by the carbon emissions and the cost of

[Find the meaning and references behind the names: Real, Max, Min]

Sustainability 2024 , 16 , 1257 12 of 24 the maintenance strategy of the two evaluation indicators, which can be obtained from the non-negative matrix: Z =      z 11 z 12 z 21 z 22 . . z n 1 z n 2      ( n = 108 ) The probability matrix P is then computed from the matrix Z , where each element p ij in P is computed as follows: p ij = Z ij ∑ n i = 1 Z ij For the j th indicator, its information entropy is calculated as e j = − 1 ln n ∑ n i = 1 p ij ln ( p ij )( j = 1, 2 ) The larger e j means the larger the information entropy of the j th indicator, and the larger the information entropy means the larger the amount of information that can be supplemented by its value, so the amount of information known before supplementation is smaller, so the real amount of information that can be obtained through this indicator should be d j = 1 − e j Finally, the weight of each indicator can be expressed as W j = d j / ∑ m j = 1 d j ( j = 1, 2 ) With the indicator matrix Z determined in the previous two subsections and W j for each indicator, combined with the TOPSIS evaluation method, it is possible to obtain the most optimal solution for each indicator in these 108 scenarios: Z + = Z + 1 , Z + 2 , · · · , Z + m = ( max { z 11 , z 21 , · · · , z n 1 } , max { z 12 , z 22 , · · · , z n 2 } , · · · , max { z 1 m , z 2 m , · · · , z nm } ) And the least optimal solution for each metric in these 108 scenarios: Z − = Z − 1 , Z − 2 , · · · , Z − m = ( min { z 11 , z 21 , · · · , z n 1 } , min { z 12 , z 22 , · · · , z n 2 } , · · · , min { z 1 m , z 2 m , · · · , z nm } ) After determining the weights, define the distance between the i th ( i = 1, 2, · · · , n ) maintenance strategy and the optimal solution: D + i = r ∑ m j = 1 W j Z + j − z ij 2 The distance of the i th ( i = 1, 2, · · · , n ) maintenance strategy score from the least desirable solution: D − i = r ∑ m j = 1 W j ¯ − z ij 2 Then, the score of the i th ( i = 1, 2, · · · , n ) maintenance strategy can be calculated: S i = D − i D + i + D − i . It can be seen that 0 ≤ S i ≤ 1, and the larger the S i the smaller the D + i , which indicates that the score of the maintenance strategy is closer to the optimal solution, and finally, the computed S i is sorted, and the maintenance strategy with the highest score is the optimal maintenance strategy 3. Data Collection and Processing 3.1. Data Collection The amount of data required for image segmentation is not large, but the accuracy of the image is required to be high. This study also innovatively proposes the dimension

[Find the meaning and references behind the names: Pictures, Day, Set, Present, Lens, Frame, Tress, Hold, Keep, Train, Place, Shot, Truck, Sure, Cell, Reason, Market, Constant, Camera, Ton, Match, Focus]

Sustainability 2024 , 16 , 1257 13 of 24 recognition of pavement damage, which requires the collection of pavement damage images while recording the actual dimensions of the image, so it is necessary to manually collect the dataset by using the camera of a cell phone. We set the camera’s magnification to 1 and the shooting pixels to 1920 × 1080. Then, focus the lens on the pavement distress and use the camera’s autofocus function to focus. To ensure consistency across all images, we also utilized a tripod to hold the camera in place and keep the lens at the same distance from the pavement distress. We also turned on the auto-calibration feature in the phone’s camera setting. As shown in Figure 8 , it is necessary to keep the cell phone parallel to the ground during the shooting process, to ensure that the pavement damage images obtained from the shooting and the actual pavement damage are in an equiproportional magnification relationship, which ensures the authenticity and accuracy of the actual dimension calculation of the pavement damage. To make sure the camera stayed parallel to the pavement distress, we tied a level to the camera and placed another level on the pavement. Observe the position of the bubble on the pavement level and then adjust the bubble position of the level on the camera to match that on the pavement. In addition, in the process of converting the pixel dimension to the actual dimension, to accurately obtain the actual dimension of the pavement damage, it is necessary to know the actual dimension of the photographed picture, so it is also necessary to place a constant dimension (60 × 60 cm) box on the photographed pavement damage during the shooting process. If there is a situation where a rectangular frame may not be able to encompass pavement distress, we will use the approach shown in Figure 9 to ensure the integrity of each pavement distress image. In the subsequent data processing, the pavement damage pictures within the box are intercepted as the final image segmentation dataset (Figure 10 ). A total of 700 images of different types of cracks and 80 images of potholes were collected in this study in this way, and these data will be used for the subsequent training of the pavement damage dimension recognition algorithm. The method we present in this paper is applicable to all pavement distress detection. The reason we only have pothole and crack images is that these two types of pavement distress are more common. However, this does not conflict with our proposed method. If we can collect other types of distress, we can also train the segmentation model and propose the corresponding pavement distress maintenance strategy Sustainability 2024 , 16 , x FOR PEER REVIEW 13 of 24 3. Data Collection and Processing 3.1. Data Collection The amount of data required for image segmentation is not large, but the accuracy of the image is required to be high This study also innovatively proposes the dimension recognition of pavement damage, which requires the collection of pavement damage im ‐ ages while recording the actual dimensions of the image, so it is necessary to manually collect the dataset by using the camera of a cell phone We set the camera’s magnification to 1 and the shooting pixels to 1920 × 1080 Then, focus the lens on the pavement distress and use the camera’s autofocus function to focus To ensure consistency across all images, we also utilized a tripod to hold the camera in place and keep the lens at the same distance from the pavement distress We also turned on the auto ‐ calibration feature in the phone’s camera setting As shown in Figure 8, it is necessary to keep the cell phone parallel to the ground during the shooting process, to ensure that the pavement damage images ob ‐ tained from the shooting and the actual pavement damage are in an equiproportional magnification relationship, which ensures the authenticity and accuracy of the actual di ‐ mension calculation of the pavement damage To make sure the camera stayed parallel to the pavement distress, we tied a level to the camera and placed another level on the pave ‐ ment Observe the position of the bubble on the pavement level and then adjust the bubble position of the level on the camera to match that on the pavement In addition, in the process of converting the pixel dimension to the actual dimension, to accurately obtain the actual dimension of the pavement damage, it is necessary to know the actual dimen ‐ sion of the photographed picture, so it is also necessary to place a constant dimension (60 × 60 cm) box on the photographed pavement damage during the shooting process If there is a situation where a rectangular frame may not be able to encompass pavement distress, we will use the approach shown in Figure 9 to ensure the integrity of each pavement dis ‐ tress image In the subsequent data processing, the pavement damage pictures within the box are intercepted as the final image segmentation dataset (Figure 10) A total of 700 images of different types of cracks and 80 images of potholes were collected in this study in this way, and these data will be used for the subsequent training of the pavement dam ‐ age dimension recognition algorithm The method we present in this paper is applicable to all pavement distress detection The reason we only have pothole and crack images is that these two types of pavement distress are more common However, this does not con ‐ flict with our proposed method If we can collect other types of distress, we can also train the segmentation model and propose the corresponding pavement distress maintenance strategy Figure 8. Schematic of image segmentation dataset shot Figure 8. Schematic of image segmentation dataset shot Sustainability 2024 , 16 , x FOR PEER REVIEW 14 of 24 Figure 9. Schematic diagram of how to photograph pavement distresses Figure 10. Schematic diagram of image segmentation dataset interception Through the construction market information service website, the labor information price can be queried for comprehensive labor (maintenance, municipal civil construction) per working day (8 h), which is CNY 230–256 This paper takes the average price of CNY 243 as the price of labor The unit prices of materials and machines are shown in Tables 4 and 5 Manual operation means that no additional labor cost is required if the machine is selected Without manual operation, additional labor costs are required Table 4. Materials price list No. Materials Price (CNY/t) 1 Petroleum asphalt 4950 2 Emulsified asphalt 4200 3 Modified asphalt 6050 Table 5. Machines Price List No. Machines Price (CNY/Shift) 1 Electric concrete saw (with manual operation) 303.05 2 Handheld blower (without manual operation) 2.07 3 Asphalt crack sealer (with manual operation) 209.78 4 Road breaker (with manual operation) 212.08 5 Mini asphalt paver (with manual operation) 652.84 6 Electric compactor (without manual operation) 32.08 7 Mini smooth wheel roller (with manual operation) 361.02 8 Eight ‐ ton truck (with manual operation) 605.04 The inventory of raw material production stages will be based on the input–output LCA methodology, with no further breakdown of the process and no pursuit of boundary conditions Calculations will simply be carried out using the generic “two ‐ step method”, which only requires the determination of the quantity of material and the carbon emission Figure 9. Schematic diagram of how to photograph pavement distresses.

[Find the meaning and references behind the names: Plant, Diesel]

Sustainability 2024 , 16 , 1257 14 of 24 Sustainability 2024 , 16 , x FOR PEER REVIEW 14 of 24 Figure 9. Schematic diagram of how to photograph pavement distresses Figure 10. Schematic diagram of image segmentation dataset interception Through the construction market information service website, the labor information price can be queried for comprehensive labor (maintenance, municipal civil construction) per working day (8 h), which is CNY 230–256 This paper takes the average price of CNY 243 as the price of labor The unit prices of materials and machines are shown in Tables 4 and 5 Manual operation means that no additional labor cost is required if the machine is selected Without manual operation, additional labor costs are required Table 4. Materials price list No. Materials Price (CNY/t) 1 Petroleum asphalt 4950 2 Emulsified asphalt 4200 3 Modified asphalt 6050 Table 5. Machines Price List No. Machines Price (CNY/Shift) 1 Electric concrete saw (with manual operation) 303.05 2 Handheld blower (without manual operation) 2.07 3 Asphalt crack sealer (with manual operation) 209.78 4 Road breaker (with manual operation) 212.08 5 Mini asphalt paver (with manual operation) 652.84 6 Electric compactor (without manual operation) 32.08 7 Mini smooth wheel roller (with manual operation) 361.02 8 Eight ‐ ton truck (with manual operation) 605.04 The inventory of raw material production stages will be based on the input–output LCA methodology, with no further breakdown of the process and no pursuit of boundary conditions Calculations will simply be carried out using the generic “two ‐ step method”, which only requires the determination of the quantity of material and the carbon emission Figure 10. Schematic diagram of image segmentation dataset interception Through the construction market information service website, the labor information price can be queried for comprehensive labor (maintenance, municipal civil construction) per working day (8 h), which is CNY 230–256. This paper takes the average price of CNY 243 as the price of labor. The unit prices of materials and machines are shown in Tables 4 and 5 . Manual operation means that no additional labor cost is required if the machine is selected. Without manual operation, additional labor costs are required Table 4. Materials price list No. Materials Price (CNY/t) 1 Petroleum asphalt 4950 2 Emulsified asphalt 4200 3 Modified asphalt 6050 Table 5. Machines Price List No. Machines Price (CNY/Shift) 1 Electric concrete saw (with manual operation) 303.05 2 Handheld blower (without manual operation) 2.07 3 Asphalt crack sealer (with manual operation) 209.78 4 Road breaker (with manual operation) 212.08 5 Mini asphalt paver (with manual operation) 652.84 6 Electric compactor (without manual operation) 32.08 7 Mini smooth wheel roller (with manual operation) 361.02 8 Eight-ton truck (with manual operation) 605.04 The inventory of raw material production stages will be based on the input–output LCA methodology, with no further breakdown of the process and no pursuit of boundary conditions. Calculations will simply be carried out using the generic “two-step method”, which only requires the determination of the quantity of material and the carbon emission factor to the corresponding baseline. The carbon emission factors for petroleum asphalt, emulsified asphalt, and modified bitumen according to the latest Eurobitume 2020 database are shown in Table 6 . Table 6. Carbon emission factors of materials No. Materials Carbon Emission Factors (kgCO 2 /kg) 1 Petroleum asphalt 136.8 2 Emulsified asphalt 166.3 3 Modified asphalt 259.7 The energy consumption and emission of the transportation stage are mainly considered to be generated by transporting the asphalt mixture from the mixing plant to the paving site, with the main reason for the energy consumption and emission being the diesel

[Find the meaning and references behind the names: Loop, Mode, Files, File, Adam, Cases, Duty, Speed, Tool, Put, Corre, Lot, Mask, Cross, Batch, Run, Location]

Sustainability 2024 , 16 , 1257 15 of 24 fuel consumed by the transportation vehicles. China’s asphalt pavement construction in order to match the pavement paving uses dump trucks and other transportation vehicles. In this paper, the energy consumption and emission of these vehicles are categorized into the transportation phase. According to the Construction Carbon Emission Calculation Standard (GB/T 51366-2019) [ 24 ], the carbon emission factors of various transportation modes are shown in Table 7 . Table 7. Transportation mode carbon emission factors Transportation Mode Carbon Emission Factors [kgCO 2 /(t · km)] Light-duty diesel truck transportation (2-ton capacity) 0.286 Medium-duty diesel truck transportation (8-ton capacity) 0.179 Heavy-duty diesel truck transportation (10-ton capacity) 0.162 Heavy-duty diesel truck transportation (18-ton capacity) 0.129 Heavy-duty diesel truck transportation (30-ton capacity) 0.078 Heavy-duty diesel truck transportation (46-ton capacity) 0.057 The energy consumption and emissions during the construction phase of asphalt pavement maintenance are mainly generated by the energy consumed by the construction machines, including diesel and electricity. The energy consumption of machines per shift is shown in Table 8 , according to the Cost Quota of Highway Engineering Machine Shifts (JTG/T 3833-2018) [ 25 ]. Table 8. Energy consumption of machines during the construction phase No. Machines Diesel Fuel (kg/Shift) Electricity (k · Wh/Shift) 1 Electric concrete saw - 18.95 2 Handheld blower - 0.2 3 Asphalt crack sealer 9.81 - 4 Road breaker 9.6 - 5 Mini asphalt paver 27.43 - 6 Electric compactor - 17.34 7 Mini smooth wheel roller 19.2 - 3.2. Data Processing The goal of image segmentation is to divide an image into regions, each representing a category. Such a task requires labeling each pixel in the image and, therefore, requires a lot of manual labor. Nonetheless, image segmentation is important for understanding the semantic content of an image as it provides detailed information about the shape and location of an object. Image segmentation is annotated at the pixel level, which provides more detailed information but requires more manual labor. The data annotation is performed using the sprite annotation assistant (Figure 11 ), and a total of 50 images of various types of pavement damages are collected in this study because the amount of image data required for image segmentation is small. Then, the pixel-level labeling is performed. Mask information, as shown in Figure 12 , is generated after the labeling is completed Sustainability 2024 , 16 , x FOR PEER REVIEW 16 of 24 performed using the sprite annotation assistant (Figure 11), and a total of 50 images of various types of pavement damages are collected in this study because the amount of im ‐ age data required for image segmentation is small Then, the pixel ‐ level labeling is per ‐ formed Mask information, as shown in Figure 12, is generated after the labeling is com ‐ pleted Figure 11. Data ‐ labeling tool Figure 12. Results of image segmentation labeling 4. Experiments and Results 4.1. Model Training The structure of the U ‐ Net network can make it possible to still obtain good results on a small number of datasets [26–28] This helps to reduce the volume of algorithms as well as increase the speed of algorithm training So, we have selected 40 crack images and 10 pothole images from the collected images for training Before starting the training, the images are split into a training set, validation set, and test set in the ratio of 8:1:1 Then, the corrupted images and corresponding mask files in the training set are put into the images file and labels folder under the train folder, and the corrupted images and corre ‐ sponding mask files in the validation set are put into the images file and labels folder under the valid folder, respectively In order to obtain the optimal hyperparameters, we include a loop at the entry point of the program run, and each loop randomly changes the preset hyperparameter values, as shown in Table 9 A batch and epochs were set up with reference to empirical data The optimizer Adam was chosen because it combines the ad ‐ vantages of the AdaGrad and RMSProp optimizers, it is able to adaptively adjust the learning rate, and it is both efficient and easy to configure The binary cross ‐ entropy loss was chosen for the loss function because it is particularly well suited for dealing with pixel ‐ level classification problems in networks such as U ‐ Net that are used for image seg ‐ mentation, and it is often used to differentiate the segmented object from the background In addition to this, Dropout is usually set at 0.2 to 0.3 for cases with fewer training data Then, the loop is started and, in the training loop, the training dataset is traversed, and a Figure 11. Data-labeling tool.

[Find the meaning and references behind the names: Memory, Fed]

Sustainability 2024 , 16 , 1257 16 of 24 Sustainability 2024 , 16 , x FOR PEER REVIEW 16 of 24 performed using the sprite annotation assistant (Figure 11), and a total of 50 images of various types of pavement damages are collected in this study because the amount of im ‐ age data required for image segmentation is small Then, the pixel ‐ level labeling is per ‐ formed Mask information, as shown in Figure 12, is generated after the labeling is com ‐ pleted Figure 11. Data ‐ labeling tool Figure 12. Results of image segmentation labeling 4. Experiments and Results 4.1. Model Training The structure of the U ‐ Net network can make it possible to still obtain good results on a small number of datasets [26–28] This helps to reduce the volume of algorithms as well as increase the speed of algorithm training So, we have selected 40 crack images and 10 pothole images from the collected images for training Before starting the training, the images are split into a training set, validation set, and test set in the ratio of 8:1:1 Then, the corrupted images and corresponding mask files in the training set are put into the images file and labels folder under the train folder, and the corrupted images and corre ‐ sponding mask files in the validation set are put into the images file and labels folder under the valid folder, respectively In order to obtain the optimal hyperparameters, we include a loop at the entry point of the program run, and each loop randomly changes the preset hyperparameter values, as shown in Table 9 A batch and epochs were set up with reference to empirical data The optimizer Adam was chosen because it combines the ad ‐ vantages of the AdaGrad and RMSProp optimizers, it is able to adaptively adjust the learning rate, and it is both efficient and easy to configure The binary cross ‐ entropy loss was chosen for the loss function because it is particularly well suited for dealing with pixel ‐ level classification problems in networks such as U ‐ Net that are used for image seg ‐ mentation, and it is often used to differentiate the segmented object from the background In addition to this, Dropout is usually set at 0.2 to 0.3 for cases with fewer training data Then, the loop is started and, in the training loop, the training dataset is traversed, and a Figure 12. Results of image segmentation labeling 4. Experiments and Results 4.1. Model Training The structure of the U-Net network can make it possible to still obtain good results on a small number of datasets [ 26 – 28 ]. This helps to reduce the volume of algorithms as well as increase the speed of algorithm training. So, we have selected 40 crack images and 10 pothole images from the collected images for training. Before starting the training, the images are split into a training set, validation set, and test set in the ratio of 8:1:1. Then, the corrupted images and corresponding mask files in the training set are put into the images file and labels folder under the train folder, and the corrupted images and corresponding mask files in the validation set are put into the images file and labels folder under the valid folder, respectively. In order to obtain the optimal hyperparameters, we include a loop at the entry point of the program run, and each loop randomly changes the preset hyperparameter values, as shown in Table 9 . A batch and epochs were set up with reference to empirical data. The optimizer Adam was chosen because it combines the advantages of the AdaGrad and RMSProp optimizers, it is able to adaptively adjust the learning rate, and it is both efficient and easy to configure. The binary cross-entropy loss was chosen for the loss function because it is particularly well suited for dealing with pixel-level classification problems in networks such as U-Net that are used for image segmentation, and it is often used to differentiate the segmented object from the background. In addition to this, Dropout is usually set at 0.2 to 0.3 for cases with fewer training data. Then, the loop is started and, in the training loop, the training dataset is traversed, and a batch of images and corresponding segmentation masks are taken out each time. They are fed into the model, the loss function is computed, and the weights of the model are updated using the optimizer Table 9. Model hyperparameters Parameter Value Batch 2, 4, 8 Epochs 40, 50, 60, 70 Optimizer Adam Loss function Binary Cross-Entropy Loss Dropout 0.2, 0.3 During the program loop, we found that when the batch was set to 8, there was a memory overflow. When the epochs was set to 70, the training error was much lower than the validation error, which is the phenomenon of overfitting. After all cases were trained, we found that the model had the highest segmentation accuracy when the parameters were set to the values shown in Table 10 .

[Find the meaning and references behind the names: Change, Tain, Cer, Get, Record]

Sustainability 2024 , 16 , 1257 17 of 24 Table 10. Image segmentation training parameter settings Parameter Settings Batch 4 Epochs 60 optimizer Adam loss function Binary Cross-Entropy Loss Dropout 0.3 After the training is completed, the final segmented image is obtained by processing the resulting image to contain only the background and pavement damage, and the training process is shown in Figure 13 . The change in loss value and the change in accuracy of the training process are shown in Figure 14 . Sustainability 2024 , 16 , x FOR PEER REVIEW 17 of 24 batch of images and corresponding segmentation masks are taken out each time They are fed into the model, the loss function is computed, and the weights of the model are up ‐ dated using the optimizer Table 9. Model hyperparameters Parameter Value Batch 2, 4, 8 Epochs 40, 50, 60, 70 Optimizer Adam Loss function Binary Cross ‐ Entropy Loss Dropout 0.2, 0.3 During the program loop, we found that when the batch was set to 8, there was a memory overflow When the epochs was set to 70, the training error was much lower than the validation error, which is the phenomenon of overfitting After all cases were trained, we found that the model had the highest segmentation accuracy when the parameters were set to the values shown in Table 10 Table 10. Image segmentation training parameter settings Parameter Settings Batch 4 Epochs 60 optimizer Adam loss function Binary Cross ‐ Entropy Loss Dropout 0.3 After the training is completed, the final segmented image is obtained by processing the resulting image to contain only the background and pavement damage, and the train ‐ ing process is shown in Figure 13 The change in loss value and the change in accuracy of the training process are shown in Figure 14 Figure 13. Image segmentation training process Figure 13. Image segmentation training process Sustainability 2024 , 16 , x FOR PEER REVIEW 18 of 24 Figure 14. Accuracy and loss value As can be seen from the figure, as the training proceeds, the loss value gets smaller and smaller and the accuracy gets higher and higher Although it is accompanied by cer ‐ tain fluctuations, the program will automatically record the best of the parameters as the result during the training process The lowest point of the loss value in the figure is 0.3312, which indicates that the segmentation accuracy of the model has reached 96.88% at this point 4.2. Dimension Recognition Result After the training was completed, 20 pavement distress images were chosen to be collected from one of the sections of pavement with a length of 100 m and a width of 7 m for model validation Some of the results are shown in Figure 15 The pixel values of the obtained binary images are input into Excel 2019 to get the Excel format, as shown in Figure 16, in which the cell with a pixel value of 0 represents the black background of the image, the points with pixel values that are not 0 are the pixel points of the pavement damage, and the pixel points accounted for by the damage of the pavement are counted out by counting out the number of the cells with cell values that are not 0 in the Excel file The number of pixels occupied by the damaged pavement can be obtained by counting the number of cells in the Excel file that are not 0 Figure 15. Image segmentation results Figure 14. Accuracy and loss value As can be seen from the figure, as the training proceeds, the loss value gets smaller and smaller and the accuracy gets higher and higher. Although it is accompanied by certain

[Find the meaning and references behind the names: Part, Ble]

Sustainability 2024 , 16 , 1257 18 of 24 fluctuations, the program will automatically record the best of the parameters as the result during the training process. The lowest point of the loss value in the figure is 0.3312, which indicates that the segmentation accuracy of the model has reached 96.88% at this point 4.2. Dimension Recognition Result After the training was completed, 20 pavement distress images were chosen to be collected from one of the sections of pavement with a length of 100 m and a width of 7 m for model validation. Some of the results are shown in Figure 15 . The pixel values of the obtained binary images are input into Excel 2019 to get the Excel format, as shown in Figure 16 , in which the cell with a pixel value of 0 represents the black background of the image, the points with pixel values that are not 0 are the pixel points of the pavement damage, and the pixel points accounted for by the damage of the pavement are counted out by counting out the number of the cells with cell values that are not 0 in the Excel file The number of pixels occupied by the damaged pavement can be obtained by counting the number of cells in the Excel file that are not 0 Sustainability 2024 , 16 , x FOR PEER REVIEW 18 of 24 Figure 14. Accuracy and loss value As can be seen from the figure, as the training proceeds, the loss value gets smaller and smaller and the accuracy gets higher and higher Although it is accompanied by cer ‐ tain fluctuations, the program will automatically record the best of the parameters as the result during the training process The lowest point of the loss value in the figure is 0.3312, which indicates that the segmentation accuracy of the model has reached 96.88% at this point 4.2. Dimension Recognition Result After the training was completed, 20 pavement distress images were chosen to be collected from one of the sections of pavement with a length of 100 m and a width of 7 m for model validation Some of the results are shown in Figure 15. The pixel values of the obtained binary images are input into Excel 2019 to get the Excel format, as shown in Figure 16, in which the cell with a pixel value of 0 represents the black background of the image, the points with pixel values that are not 0 are the pixel points of the pavement damage, and the pixel points accounted for by the damage of the pavement are counted out by counting out the number of the cells with cell values that are not 0 in the Excel file The number of pixels occupied by the damaged pavement can be obtained by counting the number of cells in the Excel file that are not 0 Figure 15. Image segmentation results Figure 15. Image segmentation results Sustainability 2024 , 16 , x FOR PEER REVIEW 19 of 24 Figure 16. Pixel value diagram In this study, when collecting pavement damage images for image segmentation training, the dimension of the box used is 60 × 60 cm, the pixels of the image collected by cell phone shooting is 1920 × 1080, and the pixel dimension of the image obtained by in ‐ tercepting the part of the box in it is 900 × 900, so the actual area represented by each pixel point is 3600/810,000 0.0044 cm According to the results of image segmentation, the pixel information of the picture containing pixel values is converted into Excel format, the points in which the pixel is not 0 are counted, and the number of pixels occupied by the damage to the pavement of each picture as well as the actual area is obtained, as shown in Table 11 Table 11. Number of pixels and actual area No. Pavement Damage Number of Pixels Actual Area (cm 2 ) No. Pavement Damage Number of Pixels Actual Area (cm 2 ) 1 Pothole 310,242 1365.065 11 Crack 15,092 66.4048 2 Pothole 240,391 1057.72 12 Crack 15,043 66.1892 3 Pothole 289,201 1272.484 13 Crack 12,016 52.8704 4 Crack 10,923 48.0612 14 Crack 14,230 62.612 5 Crack 13,492 59.3648 15 Crack 15,023 66.1012 6 Crack 23,410 103.004 16 Crack 14,830 65.252 7 Crack 10,231 45.0164 17 Crack 11,042 48.5848 8 Crack 16,923 74.4612 18 Crack 10,321 45.4124 9 Crack 13,921 61.2524 19 Crack 9431 41.4964 10 Crack 9102 40.0488 20 Crack 13,021 57.2924 4.3. Optimal Maintenance Strategy The total area of cracks contained in this 100 m asphalt pavement was counted to be 1003.424 cm 2 and the total area of potholes was 3695.27 cm 2 , which was brought into Equa ‐ tions (3) and (4) to obtain a pavement condition index of 78.087, which is between 71 and 80 From Table 1, it can be seen that the pavement needs minor maintenance Then, through the carbon emission calculation method and cost calculation method of the maintenance program given in Section 2.3, the values of the two indexes for the 108 feasi ‐ ble maintenance strategies for this pavement are shown in Table 12 The scatter plot and linear relationship graph between them are shown in Figures 17 and 18 Figure 16. Pixel value diagram In this study, when collecting pavement damage images for image segmentation training, the dimension of the box used is 60 × 60 cm, the pixels of the image collected by cell phone shooting is 1920 × 1080, and the pixel dimension of the image obtained by intercepting the part of the box in it is 900 × 900, so the actual area represented by each pixel point is 3600/810, 000 = 0.0044 cm 2

Sustainability 2024 , 16 , 1257 19 of 24 According to the results of image segmentation, the pixel information of the picture containing pixel values is converted into Excel format, the points in which the pixel is not 0 are counted, and the number of pixels occupied by the damage to the pavement of each picture as well as the actual area is obtained, as shown in Table 11 . Table 11. Number of pixels and actual area No. Pavement Damage Number of Pixels Actual Area (cm 2 ) No. Pavement Damage Number of Pixels Actual Area (cm 2 ) 1 Pothole 310,242 1365.065 11 Crack 15,092 66.4048 2 Pothole 240,391 1057.72 12 Crack 15,043 66.1892 3 Pothole 289,201 1272.484 13 Crack 12,016 52.8704 4 Crack 10,923 48.0612 14 Crack 14,230 62.612 5 Crack 13,492 59.3648 15 Crack 15,023 66.1012 6 Crack 23,410 103.004 16 Crack 14,830 65.252 7 Crack 10,231 45.0164 17 Crack 11,042 48.5848 8 Crack 16,923 74.4612 18 Crack 10,321 45.4124 9 Crack 13,921 61.2524 19 Crack 9431 41.4964 10 Crack 9102 40.0488 20 Crack 13,021 57.2924 4.3. Optimal Maintenance Strategy The total area of cracks contained in this 100 m asphalt pavement was counted to be 1003.424 cm 2 and the total area of potholes was 3695.27 cm 2 , which was brought into Equations (3) and (4) to obtain a pavement condition index of 78.087, which is between 71 and 80. From Table 1 , it can be seen that the pavement needs minor maintenance Then, through the carbon emission calculation method and cost calculation method of the maintenance program given in Section 2.3 , the values of the two indexes for the 108 feasible maintenance strategies for this pavement are shown in Table 12 . The scatter plot and linear relationship graph between them are shown in Figures 17 and 18 . Sustainability 2024 , 16 , x FOR PEER REVIEW 20 of 24 Table 12. Carbon emissions and costs of each maintenance strategy No. Cost (CNY) Carbon Emissions (kgCO 2 ) No. Cost (CNY) Carbon Emissions (kgCO 2 ) No. Cost (CNY) Carbon Emissions (kgCO 2 ) 1 535.32 22.44 37 543.02 31.56 73 565.90 35.96 2 556.42 33.03 38 564.09 31.69 74 587.18 36.10 3 520.18 28.61 39 527.82 30.56 75 551.13 34.98 4 541.27 29.87 40 548.89 31.95 76 572.41 36.37 5 513.41 23.84 41 521.00 33.45 77 544.73 37.89 6 534.51 34.44 42 542.07 33.58 78 566.01 38.02 7 498.27 30.01 43 505.81 32.45 79 529.96 36.90 8 519.36 31.28 44 526.88 33.83 80 551.23 38.30 9 480.13 27.58 45 487.62 33.96 81 512.19 38.43 10 501.22 38.18 46 508.69 41.53 82 533.46 44.32 11 464.99 33.75 47 472.42 35.44 83 497.41 37.92 12 486.08 44.35 48 493.49 45.82 84 518.69 48.09 13 539.48 29.16 49 546.68 32.76 85 564.02 36.84 14 560.67 29.29 50 567.77 32.89 86 585.21 36.97 15 524.54 28.16 51 531.53 31.76 87 549.08 35.85 16 545.73 29.55 52 552.62 33.14 88 570.27 37.24 17 517.98 31.06 53 524.76 34.65 89 542.51 38.74 18 539.17 31.19 54 545.85 34.77 90 563.71 38.88 19 503.04 30.06 55 509.61 33.64 91 527.57 37.75 20 524.23 31.45 56 530.70 35.03 92 548.77 39.14 21 485.10 31.58 57 491.46 35.15 93 509.64 39.27 22 506.30 40.29 58 512.55 42.16 94 530.84 44.59 23 470.16 34.41 59 476.31 35.96 95 494.70 38.08 24 491.36 44.94 60 497.40 46.27 96 515.90 48.17 25 541.55 30.30 61 554.51 34.32 97 565.93 38.19 26 562.66 30.43 62 575.68 34.45 98 587.11 38.32 27 526.45 29.29 63 539.52 33.33 99 550.95 37.19 28 547.56 30.68 64 560.70 34.72 100 572.12 38.58 29 519.72 32.18 65 532.92 36.23 101 544.34 40.09 30 540.83 32.31 66 554.09 36.36 102 565.51 40.22 31 504.61 31.17 67 517.93 35.23 103 529.36 39.10 32 525.73 32.56 68 539.10 36.62 104 550.53 40.49 33 486.51 32.69 69 499.95 36.76 105 511.38 40.62 34 507.63 40.82 70 521.12 43.20 106 532.55 45.37 35 471.41 34.83 71 484.97 36.90 107 496.39 38.76 36 492.52 45.29 72 506.14 47.14 108 517.56 48.78 Figure 17. Scatter plot of carbon emissions and costs Figure 17. Scatter plot of carbon emissions and costs Sustainability 2024 , 16 , x FOR PEER REVIEW 21 of 24 Figure 18. Linear relationship diagram of carbon emissions and costs Based on the results of the calculations and the linear relationship between the two indicators, it can be seen that the correlation between them is not strong and that they are relatively independent This indicates that a high ‐ cost maintenance strategy does not im ‐ ply low carbon, and a maintenance strategy with high carbon emissions does not imply economy Due to the low correlation and independence of the evaluation indicators, a multi ‐ objective decision ‐ making process is needed to select a low ‐ carbon and economic maintenance strategy based on these two indicators Next, by using the decision objective data processing and entropy weighting method, we calculated that the weights of costs and carbon emissions were 12.40% and 87.60%, respectively Then, according to the multi ‐ objective decision ‐ making model, the final composite score of the program was calculated, as shown in Table 13 Table 13. Maintenance strategy composite score No. Score No. Score No. Score No. Score 1 28.38 28 13.30 55 64.63 82 26.93 2 7.05 29 48.56 56 31.40 83 81.03 3 48.51 30 19.33 57 87.77 84 48.41 4 19.26 31 72.59 58 58.79 85 3.48 5 60.61 32 39.01 59 97.07 86 0.18 6 26.40 33 92.22 60 79.97 87 11.51 7 81.25 34 66.65 61 8.00 88 1.75 8 49.29 35 98.44 62 0.98 89 16.99 9 96.61 36 85.40 63 20.60 90 3.50 10 76.18 37 17.23 64 4.87 91 35.56 11 99.16 38 3.87 65 28.25 92 11.53 12 91.00 39 36.21 66 8.09 93 63.75 13 21.33 40 12.06 67 50.97 94 30.39 14 5.51 41 46.28 68 20.70 95 84.08 15 41.70 42 17.89 69 78.05 96 52.87 16 15.04 43 70.62 70 44.94 97 2.81 17 51.59 44 37.08 71 92.70 98 0.12 18 21.30 45 91.21 72 68.16 99 10.06 19 75.02 46 64.94 73 2.93 100 1.34 20 41.50 47 98.18 74 0.19 101 15.18 21 93.35 48 84.36 75 10.10 102 2.87 22 68.72 49 13.74 76 1.38 103 32.85 23 98.69 50 2.63 77 14.96 104 10.17 24 86.57 51 30.72 78 2.80 105 60.85 25 18.90 52 9.28 79 32.22 106 28.06 26 4.54 53 40.19 80 9.79 107 82.10 27 38.52 54 14.23 81 59.83 108 50.16 Figure 18. Linear relationship diagram of carbon emissions and costs.

Sustainability 2024 , 16 , 1257 20 of 24 Table 12. Carbon emissions and costs of each maintenance strategy No. Cost (CNY) Carbon Emissions (kgCO 2 ) No. Cost (CNY) Carbon Emissions (kgCO 2 ) No. Cost (CNY) Carbon Emissions (kgCO 2 ) 1 535.32 22.44 37 543.02 31.56 73 565.90 35.96 2 556.42 33.03 38 564.09 31.69 74 587.18 36.10 3 520.18 28.61 39 527.82 30.56 75 551.13 34.98 4 541.27 29.87 40 548.89 31.95 76 572.41 36.37 5 513.41 23.84 41 521.00 33.45 77 544.73 37.89 6 534.51 34.44 42 542.07 33.58 78 566.01 38.02 7 498.27 30.01 43 505.81 32.45 79 529.96 36.90 8 519.36 31.28 44 526.88 33.83 80 551.23 38.30 9 480.13 27.58 45 487.62 33.96 81 512.19 38.43 10 501.22 38.18 46 508.69 41.53 82 533.46 44.32 11 464.99 33.75 47 472.42 35.44 83 497.41 37.92 12 486.08 44.35 48 493.49 45.82 84 518.69 48.09 13 539.48 29.16 49 546.68 32.76 85 564.02 36.84 14 560.67 29.29 50 567.77 32.89 86 585.21 36.97 15 524.54 28.16 51 531.53 31.76 87 549.08 35.85 16 545.73 29.55 52 552.62 33.14 88 570.27 37.24 17 517.98 31.06 53 524.76 34.65 89 542.51 38.74 18 539.17 31.19 54 545.85 34.77 90 563.71 38.88 19 503.04 30.06 55 509.61 33.64 91 527.57 37.75 20 524.23 31.45 56 530.70 35.03 92 548.77 39.14 21 485.10 31.58 57 491.46 35.15 93 509.64 39.27 22 506.30 40.29 58 512.55 42.16 94 530.84 44.59 23 470.16 34.41 59 476.31 35.96 95 494.70 38.08 24 491.36 44.94 60 497.40 46.27 96 515.90 48.17 25 541.55 30.30 61 554.51 34.32 97 565.93 38.19 26 562.66 30.43 62 575.68 34.45 98 587.11 38.32 27 526.45 29.29 63 539.52 33.33 99 550.95 37.19 28 547.56 30.68 64 560.70 34.72 100 572.12 38.58 29 519.72 32.18 65 532.92 36.23 101 544.34 40.09 30 540.83 32.31 66 554.09 36.36 102 565.51 40.22 31 504.61 31.17 67 517.93 35.23 103 529.36 39.10 32 525.73 32.56 68 539.10 36.62 104 550.53 40.49 33 486.51 32.69 69 499.95 36.76 105 511.38 40.62 34 507.63 40.82 70 521.12 43.20 106 532.55 45.37 35 471.41 34.83 71 484.97 36.90 107 496.39 38.76 36 492.52 45.29 72 506.14 47.14 108 517.56 48.78 Based on the results of the calculations and the linear relationship between the two indicators, it can be seen that the correlation between them is not strong and that they are relatively independent. This indicates that a high-cost maintenance strategy does not imply low carbon, and a maintenance strategy with high carbon emissions does not imply economy. Due to the low correlation and independence of the evaluation indicators, a multi-objective decision-making process is needed to select a low-carbon and economic maintenance strategy based on these two indicators. Next, by using the decision objective data processing and entropy weighting method, we calculated that the weights of costs and carbon emissions were 12.40% and 87.60%, respectively. Then, according to the multiobjective decision-making model, the final composite score of the program was calculated, as shown in Table 13 . From the table, it can be seen that among all the maintenance strategies, No. 11 has the highest score, which is 99.16 points, the carbon emissions generated by this maintenance strategy is 33.75 kg, which is 30.80% lower than the highest carbon emissions (48.78 kg) among all the maintenance strategies, and the cost is CNY 464.99, which is 20.81% lower than the highest cost (CNY 587.18) among all the maintenance strategies. From the 108 maintenance strategies determined in Section 2.2 , the optimal maintenance strategy can be derived, as shown in Table 14 .

[Find the meaning and references behind the names: Long]

Sustainability 2024 , 16 , 1257 21 of 24 Table 13. Maintenance strategy composite score No. Score No. Score No. Score No. Score 1 28.38 28 13.30 55 64.63 82 26.93 2 7.05 29 48.56 56 31.40 83 81.03 3 48.51 30 19.33 57 87.77 84 48.41 4 19.26 31 72.59 58 58.79 85 3.48 5 60.61 32 39.01 59 97.07 86 0.18 6 26.40 33 92.22 60 79.97 87 11.51 7 81.25 34 66.65 61 8.00 88 1.75 8 49.29 35 98.44 62 0.98 89 16.99 9 96.61 36 85.40 63 20.60 90 3.50 10 76.18 37 17.23 64 4.87 91 35.56 11 99.16 38 3.87 65 28.25 92 11.53 12 91.00 39 36.21 66 8.09 93 63.75 13 21.33 40 12.06 67 50.97 94 30.39 14 5.51 41 46.28 68 20.70 95 84.08 15 41.70 42 17.89 69 78.05 96 52.87 16 15.04 43 70.62 70 44.94 97 2.81 17 51.59 44 37.08 71 92.70 98 0.12 18 21.30 45 91.21 72 68.16 99 10.06 19 75.02 46 64.94 73 2.93 100 1.34 20 41.50 47 98.18 74 0.19 101 15.18 21 93.35 48 84.36 75 10.10 102 2.87 22 68.72 49 13.74 76 1.38 103 32.85 23 98.69 50 2.63 77 14.96 104 10.17 24 86.57 51 30.72 78 2.80 105 60.85 25 18.90 52 9.28 79 32.22 106 28.06 26 4.54 53 40.19 80 9.79 107 82.10 27 38.52 54 14.23 81 59.83 108 50.16 Table 14. Optimal maintenance strategy Pavement Damage Material Grooving Cleaning and Drying Sealing Paving Adjusting Compaction Crack Petroleum asphalt Electric concrete saw Handheld electric blower Asphalt crack sealer Manual adjusting Pothole Emulsified asphalt Road breaker Handheld electric blower Manual paving Mini smooth wheel roller 5. Conclusions Effective pavement maintenance can significantly extend the service life of a road, and can also significantly reduce long-term maintenance costs, as well as reduce resource consumption and pollutant emissions during road construction and reconstruction. In this study, the actual dimension of pavement damage is recognized by using image segmentation technology. The PCI is calculated after pavement automatic detection, which provides a direct basis for decision making on whether or not maintenance is required. Then, combining the actual dimension of the pavement damage and the maintenance process, we summarize and generalize all the feasible maintenance schemes, calculate the carbon emissions and maintenance costs generated by all the maintenance schemes, and, finally, use the multi-objective decision-making model based on the TOPSIS model to give the optimal maintenance scheme in terms of badness benefit and economic benefit. The main conclusions of this paper are as follows: (1) The U-Net algorithm is known for its symmetric U-shaped structure and effective jump connections, which enables it to achieve accurate image segmentation even on limited datasets, and likewise U-Net has a smaller dimension compared to other image segmentation algorithms. In the training process of this paper, the algorithm has achieved a loss value of 0.3312, which indicates that the segmentation is 96.88% efficient. Through accurate image segmentation, combining picture pixels with actual

[Find the meaning and references behind the names: New, Resources, Plan, Board, Cutter, Read, Company, Take, Case, Author, Peak]

Sustainability 2024 , 16 , 1257 22 of 24 dimensions, this study accurately identifies the detected pavement damage dimensions, which provides a basis for calculating the pavement condition index, the carbon emissions of the maintenance program, and the maintenance cost (2) This study accurately categorizes and identifies the dimension of the pavement damage in combination with actual cases and gives the decision of whether maintenance is needed or not by combining the PCI calculation formula. On this basis, decomposing the pavement damage maintenance process lists the labor, materials, and machines needed for maintenance, and then proposes all feasible maintenance strategies. Then, the carbon emissions and maintenance costs of these maintenance options were calculated by combining the actual dimensions of the pavement damage. Finally, through the multi-objective decision-making method based on the improved entropy-weighted TOPSIS model, the optimal maintenance strategy for cracks, in this case, was obtained by using petroleum asphalt as the material, grooving with a motorized cutter, cleaning and drying the cracks with a handheld motorized blower, grouting with an asphalt grouting machine, and, finally, repairing manually. For potholes, the optimal maintenance strategy is to use emulsified asphalt as material, use a road breaker for grooving, portable electric blower for cleaning and drying cracks, take manual paving, and, finally, use a small light wheel roller for compaction. The comprehensive score of this program is 99.16, the carbon emission generated in this case is 33.75 kg, and the cost of maintenance is CNY 464.99 (3) From Conclusion 2, it can be seen that the multi-objective decision-making model proposed in this study is able to give a specific low-carbon and economical maintenance strategy based on specific pavement conditions instead of giving a maintenance strategy that applies to all pavements. The problem that one model cannot be applied to most pavements is solved, and the generalizability of the decision-making model is greatly improved Nevertheless, due to the limitation of resources and technology, further in-depth research can be carried out in the following three aspects in future research: (1) In this study, only the actual area of the pavement damage was recognized when the pavement damage dimension recognition was performed, and the depth of the damage could not be accurately recognized, so the common depth of these damages was taken as the actual depth of the pavement damage. Future research can combine 3 D image recognition technology to obtain a more accurate depth of pavement damage (2) Due to the limitation of obtaining the carbon emission factor of materials, this study is not comprehensive enough in the selection of materials, and in the future, if the carbon emission factor of more new materials can be obtained, a lower carbon economic maintenance strategy can be proposed Author Contributions: Resources, D.C. and W.F.; writing—review and editing, D.C.; validation, D.C.; conceptualization, P.L.; methodology, P.L.; writing—original draft, P.L.; data analysis, P.L. and W.F.; visualization, P.L. All authors have read and agreed to the published version of the manuscript Funding: Research on carbon emission accounting method, indicator system and carbon emission monitoring platform framework for near-zero carbon surface roads. The funding number is: 23 DZ 1202102 This project is from: Shanghai 2023 “Science and Technology Innovation Action Plan” Science and Technology Support Carbon Peak Carbon Neutral Project Declaration Guidelines/Low Construction Institutional Review Board Statement: Not applicable Informed Consent Statement: Not applicable Data Availability Statement: The datasets generated and/or analyzed in the current study are available from the corresponding author upon reasonable request Conflicts of Interest: Author Wurong Fu was employed by the company Shanghai Road & Bridge (Group) Co., Ltd. The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

[Find the meaning and references behind the names: Da Costa, Eng, Zhang, Liu, Art, Lee, Park, Civ, Santos, Russo, Viscione, Transport, Euro, Jet, June, Sesay, Heidari, Dong, Jing, Zhao, Ferreira, Wang, Liang, Int, Sci, Vargas, Hamdi, Chem, Jin, Marcos, Correia, Tien, Pereira, Duc, Nguyen, Adu, Parra, Wei, Pham, Costa, Gong, Bui, Mater, Gyamfi, Prod, Nhat, Zakeri, Barbosa, Zhou, Junior, Kim, Leite, Bernardino, Cong, Yazdani, Meng, Rap, Guo, Llorca, Chen, Nejad, Ignatius, Edge, Canada, Tran, Sotelo, Rola]

Sustainability 2024 , 16 , 1257 23 of 24 References 1 Zakeri, H.; Nejad, F.M.; Fahimifar, A.; Torshizi, A.D.; Zarandi, M.H.F. A Multi-Stage Expert System for Classification of Pavement Cracking. In Proceedings of the 2013 Joint IFSA World Congress and NAFIPS Annual Meeting (IFSA/NAFIPS), Edmonton, AB, Canada, 24–28 June 2013; pp. 1125–1130 2 Gavil á n, M.; Balcones, D.; Marcos, O.; Llorca, D.F.; Sotelo, M.A.; Parra, I.; Ocaña, M.; Aliseda, P.; Yarza, P.; Am í rola, A. Adaptive Road Crack Detection System by Pavement Classification Sensors 2011 , 11 , 9628–9657. [ CrossRef ] [ PubMed ] 3 Nhat-Duc, H.; Nguyen, Q.-L.; Tran, V.-D. Automatic Recognition of Asphalt Pavement Cracks Using Metaheuristic Optimized Edge Detection Algorithms and Convolution Neural Network Autom. Constr 2018 , 94 , 203–213. [ CrossRef ] 4 Oreto, C.; Veropalumbo, R.; Viscione, N.; Biancardo, S.A.; Russo, F. Investigating the Environmental Impacts and Engineering Performance of Road Asphalt Pavement Mixtures Made up of Jet Grouting Waste and Reclaimed Asphalt Pavement Environ. Res 2021 , 198 , 111277. [ CrossRef ] 5 Zhang, J.; Sesay, T.; You, Q.; Jing, H. Maximizing the Application of RAP in Asphalt Concrete Pavements and Its Long-Term Performance: A Review Polymers 2022 , 14 , 4736. [ CrossRef ] [ PubMed ] 6 Bang, S.; Park, S.; Kim, H.; Kim, H. Encoder–Decoder Network for Pixel-Level Road Crack Detection in Black-Box Images Comput. -Aided Civ. Infrastruct. Eng 2019 , 34 , 713–727. [ CrossRef ] 7 Li, S.; Zhao, X.; Zhou, G. Automatic Pixel-Level Multiple Damage Detection of Concrete Structure Using Fully Convolutional Network Comput. -Aided Civ. Infrastruct. Eng 2019 , 34 , 616–634. [ CrossRef ] 8 Majidifard, H.; Adu-Gyamfi, Y.; Buttlar, W.G. Deep Machine Learning Approach to Develop a New Asphalt Pavement Condition Index Constr. Build. Mater 2020 , 247 , 118513. [ CrossRef ] 9 Ouma, Y.O.; Hahn, M. Pothole Detection on Asphalt Pavements from 2 D-Colour Pothole Images Using Fuzzy c-Means Clustering and Morphological Reconstruction Autom. Constr 2017 , 83 , 196–211. [ CrossRef ] 10 Hoang, N.-D.; Nguyen, Q.-L.; Tien Bui, D. Image Processing–Based Classification of Asphalt Pavement Cracks Using Support Vector Machine Optimized by Artificial Bee Colony J. Comput. Civ. Eng 2018 , 32 , 04018037. [ CrossRef ] 11 Yu, G.; Dong, J.; Wang, Y.; Zhou, X. RUC-Net: A Residual-Unet-Based Convolutional Neural Network for Pixel-Level Pavement Crack Segmentation Sensors 2023 , 23 , 53. [ CrossRef ] 12 Yu, B.; Gu, X.; Ni, F.; Guo, R. Multi-Objective Optimization for Asphalt Pavement Maintenance Plans at Project Level: Integrating Performance, Cost and Environment Transp. Res. Part D Transp. Environ 2015 , 41 , 64–74. [ CrossRef ] 13 Hamdi, H.; Hadiwardoyo, S.P.; Correia, A.G.; Pereira, P. Pavement Maintenance Optimization Strategies for National Road Network in Indonesia Applying Genetic Algorithm Procedia Eng 2017 , 210 , 253–260. [ CrossRef ] 14 Santos, J.; Ferreira, A.; Flintsch, G. An Adaptive Hybrid Genetic Algorithm for Pavement Management Int. J. Pavement Eng 2019 , 20 , 266–286. [ CrossRef ] 15 Lee, J.; Madanat, S. Jointly Optimal Policies for Pavement Maintenance, Resurfacing and Reconstruction EURO J. Transp. Logist 2015 , 4 , 75–95. [ CrossRef ] 16 Heidari, M.R.; Heravi, G.; Esmaeeli, A.N. Integrating Life-Cycle Assessment and Life-Cycle Cost Analysis to Select Sustainable Pavement: A Probabilistic Model Using Managerial Flexibilities J. Clean. Prod 2020 , 254 , 120046. [ CrossRef ] 17 Vargas, D.E.C.; Lemonge, A.C.C.; Barbosa, H.J.C.; Bernardino, H.S. Solving Multi-Objective Structural Optimization Problems Using GDE 3 and NSGA-II with Reference Points Eng. Struct 2021 , 239 , 112187. [ CrossRef ] 18 Leite, B.; da Costa, A.O.S.; Costa Junior, E.F. da Multi-Objective Optimization of Adiabatic Styrene Reactors Using Generalized Differential Evolution 3 (GDE 3) Chem. Eng. Sci 2023 , 265 , 118196. [ CrossRef ] 19 Chen, Z.; Wang, D. Numerical Analysis of a Multi-Objective Maintenance Decision-Making Model for Sustainable Highway Networks: Integrating the GDE 3 Method, LCA and LCCA Energy Build 2023 , 290 , 113096. [ CrossRef ] 20 Behzadian, M.; Khanmohammadi Otaghsara, S.; Yazdani, M.; Ignatius, J. A State-of the-Art Survey of TOPSIS Applications Expert Syst. Appl 2012 , 39 , 13051–13069. [ CrossRef ] 21 Gong, H.; Liu, L.; Liang, H.; Zhou, Y.; Cong, L. A State-of-the-Art Survey of Deep Learning Models for Automated Pavement Crack Segmentation Int. J. Transp. Sci. Technol 2024 , 13 , 44–57. [ CrossRef ] 22 JTG 5210—2018 ; Highway Performance Assessment Standard. Ministry of Transport of the People’s Republic of China: Beijing, China, 2019 23 JTG 5142—2019 ; Technical Specifications for Maintenance of Highway Asphalt Pavement. Ministry of Transport of the People’s Republic of China: Beijing, China, 2019 24 GB/T 51366—2019 ; Standard for building carbon emission calculation. Ministry of Housing and Urban-Rural Development of the People's Republic of China: Beijing, China, 2019 25 JTG/T 3833—2018 ; Cost Quota of Highway Engineering Machine Shifts. Ministry of Transport of the People’s Republic of China: Beijing, China, 2019 26 Jin, Q.; Meng, Z.; Pham, T.D.; Chen, Q.; Wei, L.; Su, R. DUNet: A Deformable Network for Retinal Vessel Segmentation Knowl.-Based Syst 2019 , 178 , 149–162. [ CrossRef ]

[Find the meaning and references behind the names: Wells, Eds, Frangi, Schlemper, Schaap, Heinrich, Cham, Ideas, Navab, Fischer, Kainz, Med, Oktay, Property, Springer]

Sustainability 2024 , 16 , 1257 24 of 24 27 Schlemper, J.; Oktay, O.; Schaap, M.; Heinrich, M.; Kainz, B.; Glocker, B.; Rueckert, D. Attention Gated Networks: Learning to Leverage Salient Regions in Medical Images Med. Image Anal 2019 , 53 , 197–207. [ CrossRef ] [ PubMed ] 28 Ronneberger, O.; Fischer, P.; Brox, T. U-Net: Convolutional Networks for Biomedical Image Segmentation. In Medical Image Computing and Computer-Assisted Intervention—MICCAI 2015 ; Navab, N., Hornegger, J., Wells, W.M., Frangi, A.F., Eds.; Lecture Notes in Computer Science; Springer International Publishing: Cham, Switzerland, 2015; Volume 9351, pp. 234–241. ISBN 978-3-319-24573-7 Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.