Bridge deck surface distress evaluation using S-UAS acquired high-spatial resolution aerial imagery

ABSTRACT Bridge decks need to be routinely inspected to ensure their serviceability, capacity, and safety under current traffic conditions. Traditionally, bridge deck inspection is performed on the ground by having inspectors either visually inspect surface conditions or interpret the acoustic feedback from hammer sounding or chain dragging to determine subsurface conditions. These traditional methods have many limitations, including but not limited to, expensive, labour-intensive, time-consuming, subjective, can exhibit a high degree of variability, requiring specialized staff on a regular basis, and unsafe. Recent advancements in remote sensing, especially small-uncrewed aircraft systems (S-UAS) based airborne imaging techniques and advanced image analysis techniques, have shown promise in improving current bridge deck inspection practices by providing an above-ground inspection method. This research explored the utility of S-UAS-based airborne imaging techniques and image processing techniques to develop a complete aerial data acquisition and analysis system to accurately detect and assess bridge deck wearing surface distresses in a timely and cost-effective manner. As part of the research project, a robust tool was also developed with the aim of being able to detect, extract, and map bridge deck wearing surface distresses with an adequate degree of accuracy while maximizing the ability to assist bridge inspectors with varying expertise. Research results revealed that the developed tool is able to effectively detect and map bridge deck wearing surface distresses at a high accuracy.


Introduction
Bridges provide passage over physical obstacles to substantially reduce travel time and travel cost. They are considered as critical infrastructure assets because they are extremely important for the safety of travelling public and sustainability of economic activity (La et al. 2013). Similar to other types of transportation infrastructure, bridges deteriorate over time due to various factors such as adverse weather conditions, repeated heavy traffic loads, ageing of structural elements, and exposure to chemicals (Rashidi, Samali, and Sharafi 2016). Subsequently, bridges need to be routinely inspected by transportation agencies to detect and address deficiencies in a timely manner, and ultimately, improving public transportation safety and preserving the remaining service life of a bridge.
In the United States, bridges can be inspected by transportation agencies at all levels (e.g. federal, state, local, or tribal). An effective bridge inspection programme should be able to detect signs of deterioration, identify the causes, and assist transportation agencies in making decisions on the distribution of limited resources for maintenance, repair, and construction projects to ensure the safety, serviceability, and structural capacity of a bridge under current traffic conditions (Vaghefi et al. 2012). A national bridge inspection programme has been established by the Federal Highway Act since 1968 that subsequently requires mandatory inspection of all bridges that are managed by public entities such as state department of transportation (Abu Dabous et al. 2017).
In recent years, bridge inspections have received a great amount of attention because of catastrophic failures, deteriorating conditions, and even political pressure (Vaghefi et al. 2012). According to the data obtained from the Bureau of Transportation Statistics, in 2021 there were 619,588 bridges in the U.S., and approximately 7% of them are rated as structurally deficient and 36% of them need to be repaired because the average age of them was more than 40 years (Black 2022). Five types of bridge inspection are normally conducted in the U.S., including initial inspection (i.e. the first inspection after the bridge is built), routine inspection (i.e. a serviceability inspection once every two years), in-depth inspection (i.e. a fully detailed inspection once every five years), damage inspection (i.e. an unscheduled inspection when damage occurs), and special inspection (i.e. a recurring inspection performed as part of a bridge's monitoring programme). It should be noted that high-risk bridges or bridges in poor condition may need to be inspected on a more frequent basis than the routine inspection (Vaghefi et al. 2012;Abu Dabous et al. 2017).
Typically, a bridge has three components which include deck, superstructure (e.g. girders), and substructure (e.g. piers and abutments). Although these three components are equally critical to the safety, integrity, and serviceability of a bridge, this study is focused on bridge decks because they deteriorate more rapidly than superstructures and substructures (Scott et al. 2000). A degraded bridge deck could cause the protection afforded to the superstructure and substructure diminish which in turn could cause the deterioration of them in an accelerating mode, and ultimately, leading to catastrophic structural failures (Vaghefi et al. 2012;Varnavina 2015). Subsequently, bridge decks need to be evaluated and assessed on a more frequent schedule to identify and address any potential problems in a timely manner to achieve an effective preventive maintenance programme.
To characterize the conditions of a bridge deck, four elements are typically inspected, including wearing surfaces, deck joints, guardrails, and structural decks (IDOT 2011). A wearing surface, which is typically constructed of timber, asphalt, or concrete, is a layer placed on the structural deck. A deck joint is designed to allow for traffic between segments of a bridge while facilitating the deck's transversal, longitudinal, and rotational movement. Guardrails are designed to provide passive protection to vehicles, pedestrians, and bicyclists to keep them inside the road in a secure way. Structural decks, which are typically constructed of concrete, steel, timber, or fibre-reinforced polymer (FRP), comprise the basic plate. It is worth noting that additional elements might be inspected, depending on the specific inspection requirements. Once all elements are inspected by a team of inspectors with different expertise, a report which typically includes textual descriptions of the extent and severity of the observed defects will be prepared. Photos will also be taken by inspectors and included in the report to supplement the inspection results.
Bridge deck inspection is always challenging because inspectors need to evaluate many elements and expose themselves to traffic as well as weather conditions (Vaghefi et al. 2012). Traditionally, bridge deck inspection is conducted 'boots on the ground' by having inspectors either visually inspect surface conditions or acoustically interpret the audio feedback from chain dragging or hammer sounding to determine subsurface conditions. Inspectors may also need to conduct destructive evaluations to obtain core samples from a deck to test their mechanical and chemical properties which are used to determine subsurface conditions. More recently, some non-destructive evaluation (NDE) methods such as ground penetration radar (GPR) and infrared thermography have been explored for bridge deck subsurface condition evaluation (Sun, Pashoutani, and Zhu 2018). These methods are still primarily implemented on the ground.
All the aforementioned ground-based methods have many limitations, including, but not limited to, timeconsuming, labour-intensive, expensive, subjective, can exhibit a high degree of variability, requiring specialized staff on a regular basis, and potentially unsafe to inspectors (Moore et al. 2001;Vaghefi et al. 2015). For example, when performing an in-depth inspection, inspectors may need to close the entire bridge to travellers, which can potentially cause traffic congestion and interrupt traffic flow (Vaghefi et al. 2015). To address these limitations, the Federal Highway Administration (FHWA) conducted a study and concluded that deployment of technologies in bridge inspection holds the potential to greatly reduce costs, enhance efficiency, accuracy, reliability, and improve safety (Abu Dabous et al. 2017;Moore et al. 2001).
In an effort to adopt technologies for bridge deck inspection, some researchers explored the utility of aerial photos to observe bridge deck wearing surface conditions. The method is also known as aircraft-based evaluation, which is typically accomplished with a manned aircraft and a digital camera. The collected aerial images, which typically have a spatial resolution ranges from 7.62 cm (3-inch) to 15.24 cm (6-inch), can be used to evaluate bridge deck wearing surface distresses such as large cracks (i.e. width and length larger than 7.62 cm) and potholes (Chen et al. 2011). However, these low-spatial resolutions prevent them from being used to detect and evaluate finer-scale wearing surface distresses such as cracks at the millimetre (sub-inch) level.
Recent advancements in remote sensing technologies, including small-uncrewed aircraft systems (S-UAS)-based high-spatial resolution airborne imaging technique and aerial triangulation (AT) technique, have shown promise in inspecting bridge deck wearing surface conditions above the ground. In recent years, S-UAS have emerged as an important platform for acquiring high-spatial resolution (i.e. millimetre scale) aerial imagery that can be used to provide a synoptic view of objects on the ground such as bridges. S-UAS can fly lower to the ground than traditional manned aircraft, and therefore, they can be used to acquire more detailed data about bridge deck wearing surfaces. As an advanced photogrammetric method, AT is able to create the three-dimensional (3D) models of ground objects by analysing overlapping aerial images captured from varied perspective at approximately the spatial resolution of input images (Yuan et al. 2009;Zhang, Xiong, and Hao 2011). Therefore, when coupled with high-spatial resolution aerial imagery, AT is able to estimate horizontal and vertical measurements at the millimetre level, and ultimately, detecting and evaluating bridge deck wearing surface distresses at a finer scale (Zhang et al. 2016).
S-UAS have been widely used in many fields, and civil engineering is no exception. For example, Kang and Cha explored the utility of S-UAS for structural health monitoring (Kang and Cha 2018). Coupled with deep learning and an ultrasonic beacon system, S-UAS have proven to be able to detect cracks in structural components (Kang and Cha 2018). Ali et al. proposed a real-time structural damage mapping system based on S-UAS and deep faster region-based neural networks for GPS-denied structures, and research results revealed that the proposed system can be effectively used to detect structural defects (Ali et al. 2021). In recent years, many deeplearning algorithms have also been explored to analyse S-UAS collected aerial images to detect many types of structural defects such as cracks (Choi and Cha 2020;Kang et al. 2020;Kang and Cha 2021).
Based on the review of literature, using S-UAS-based airborne imaging technique and AT technique for bridge deck wearing surface inspection is lacking and presents a significant research gap. Therefore, this study focuses on using various S-UAS-based airborne imaging techniques and image processing techniques to develop a complete data acquisition and analysis system to detect bridge deck wearing surface distresses in an accurate, rapid, and cost-effective manner. Specifically, the intent of this study is to examine how well bridge deck wearing surface distresses can be evaluated from S-UAS acquired high-spatial resolution aerial photos.

Method
This research uses an S-UAS as the airborne platform to collect high-spatial resolution aerial photos, uses AT technique to process these aerial photos to derive intermediate products, and then use pixel-based image processing and image fusion techniques to analyse the intermediate products to detect and evaluate bridge deck wearing surface conditions. Additionally, an analysis toolset, which is compatible with standard geographic information systems (GIS) software programs, was developed to automate the image analysis and information extraction process. Specifically, methodology includes the following sections: aerial data collection; aerial data processing; analysis tool development; and accuracy validation.

Aerial data collection
DJI Mavic Pro 2 S-UAS was used to collect high-spatial resolution aerial images. Ten S-UAS were tested and DJI Mavic Pro 2 outperformed all other S-UAS in terms of operational simplicity, flight time, payload, aerial imagery quality, and cost. Table 1 lists all the tested S-UAS and detailed description of them. Further details about this selection process can be found in the research report developed by Zhang et al. (Zhang, Baros, and Susan 2020). Noting that an Uncrewed Aerial Vehicle (UAV) is different from an UAS, although they are often used interchangeably. UAV indicates the actual aircraft platform that flies around to collected aerial data. In contrast, UAS indicates the entire system needed for aerial mapping, including UAV and sensors (e.g. natural colour digital cameras). Therefore, a sensor or multiple sensors can be mounted to a UAV to create a UAS. According to the Federal Aviation Administration (FAA)'s Small Unmanned Aircraft Regulations (Part 107), an S-UAS is defined as an aircraft system that weighs more than 0.25 kg or 0.55 lbs. and less than 25 kg or 55 lbs. (Kaundinya et al. 2018). That being said, when testing the best S-UAS, it is necessary to make sure the combination of UAV and sensors are less than 25 kg. DJI Mavic Pro 2 has an onboard Hasselblad L1D-20c natural colour digital camera (20-megapixel, sensor size 13.2 × 8.8 mm, visible blue, green, and red band). A bridge located at the intersection of Edith Blvd and the North Division Channel in the City of Albuquerque, New Mexico was selected for data collection. This bridge was built in 1967, and its end of functional service life is 2033. The material for the bridge is prestressed concrete, while the material for the wearing surface is asphalt. The road width is 8.5 m (2 lanes), and there is no parallel bridge to it. When selecting the bridge for data collection, its location and surrounding environment should be considered. With regard to location, there are three aspects to consider: (1) bridges are at least five nautical miles or 9.3 km away from an airport; (2) bridges are not located on arterial roads or interstate highways to avoid high traffic volumes; (3) bridges are not located in densely populated neighbourhoods or open areas that have a large number of pedestrians. Noting bridges that are located within five nautical miles can still be imaged by S-UAS but remote pilots need to submit a near real-time authorization request to FAA for operations under 400 feet (122 metres) in controlled airspace around airports (Kaundinya et al. 2018). Regarding surrounding environment, it is suggested to make sure there are no or limited amount of power lines and trees around the bridge being imaged. The considerations described above are primarily used for ensuring safety during data collection in the context of exploratory research. That said, these limitations will not impact any operational S-UAS-based bridge inspection, but S-UAS operators do need to obtain waivers or authorizations from FAA if S-UAS are operated within five nautical miles (9.3 km) of an airport or over people and moving vehicles. An additional benefit of providing these considerations is to assist researchers with S-UAS data collection for their areas of interest.
With a grid flight path (75% forward overlap and 80% side overlap), a total of 92 geotagged aerial photos in JPEG format were collected, which meets the overlapping requirements (Zhang, Xiong, and Hao 2011). The flight altitude is 23 metres above ground level (AGL). At this flight altitude, the spatial resolution of these photos is 0.005 m (5 mm). During data collection, the sky was overcast and there are no shadows in the photos. Figure 1 shows an example of the collected aerial photos. The cracks on the wearing surface and crack sealing are very noticeable. Noting that a single aerial photo can cover the entire bridge due to its high-spatial resolution. The higher the spatial resolution, the more detail it will contain, and the smaller ground coverage it will contain.
Ten ground control points (GCP) were also collected for later image processing work. GCPs are points on the ground surface with known coordinates and they are used in AT technique to orientate and calibrate the output point cloud to ensure a high degree of global positional accuracy. The coordinates of the GCPs were collected using a Real-Time Kinematics (RTK) system in a rover/based configuration. RTK is used to enhance the precision of positional data obtained from Global Navigation Satellite System (GNSS), which is the combination of existing satellite navigation systems such as GPS, GLONASS, Galileo, and Beidou. Post-processing revealed that these GCPs have a root mean square error (RMSE) of 0.003 m + 1 ppm horizontally and 0.006 m + 1 ppm vertically. Considering the spatial resolution of the aerial photos is 0.005 m, the accuracy of the GCP is able to process the aerial photos in AT effectively.

Aerial data processing
In photogrammetry and remote sensing, the technique used to process S-UAS acquired aerial photos to generate co-registered orthophotos and digital surface models (DSMs) is referred to as AT, which is also known as Structure-from-Motion and Multiple-View Stereo (SfM-MVS). Noting that although SfM is only one of the steps in SfM-MVS, for simplification purpose, this study used SfM to indicate SfM-MVS. Figure 2 shows the general steps in SfM. Detailed discussion of each step is provided below. It should also be noted that the keypoint correspondence step and the keypoint filtering step are typically conducted concurrently. However, regardless of the algorithms being used for a specific software program, the aforementioned steps are generally involved. This study used Agisoft Metashape to process the collected aerial images and GCPs, although ten software packages were evaluated. Further details about the selection process can be found from the research report developed by Zhang et al. (Zhang, Baros, and Susan 2020). The output products from SfM include a coregistered orthophoto and DSM, which are shown in Figure 3.

Image quality assessment and import
This step involves assessing the collected high-spatial resolution aerial photos to ensure they have acceptable brightness (i.e. overall lightness), contrast (i.e. difference in brightness between objects), and sharpness (i.e. clarity of detail). In addition, each aerial photo should be assessed to examine they have been properly geotagged (i.e. longitude, latitude, and altitude). Aerial photos that look blurry, saturated, overexposed or underexposed, and noisy should be excluded from the  input image collection. For this research, the assessment revealed that all 92 aerial photos have high-quality and they can be imported into Agisoft Metashape.

Keypoint identification
This step focuses on identifying common points in the overlapping aerial images. These common points are referred as keypoints, and they are the points that allow the different aerial images to be matched and the 3D geometry to be reconstructed. That said, keypoints are the features that can be clearly identified in an aerial image. These keypoints are known as tie points as they are used to link each aerial image together in later process. Many algorithms have been developed to identify keypoints, but the widely adopted is scaleinvariant feature transform (SIFT) which focuses on recognizing feature objects (a group of pixels that comprise an object in the photo) and then locating each keypoints (Carrivick, Smith, and Quincey 2016)

Keypoint correspondence
Once all keypoints have been identified and located in each aerial image, correspondence lines between the keypoints need to be determined and established. Noting that there is no assurance that any given keypoint in an image will have a matching keypoint in another image. Therefore, this step also involves discarding keypoints that do not have a matching partner. Similar to keypoint identification, many methods exist for keypoint correspondence, but the most adopted one is the approximate nearest neighbour (ANN) method (Carrivick, Smith, and Quincey 2016).

Keypoint filtering
This step involves further processing the keypoint correspondence to filter out any erroneous matches to produce a high-quality keypoint correspondence set. Many methods are available for keypoint filtering, but one of the most robust and accurate one is Random Sample Consensus (RANSAC) filter, which assumes that all keypoints can be classified into two groups that include outliers and inliers (Carrivick, Smith, and Quincey 2016). All inliers will be used in later process while all outliers will be removed from the keypoint correspondence dataset.

SfM
This step focuses on estimating camera poses and then reconstructing the 3D geometry of a scene. The inliers keypoint correspondence is used to estimate extrinsic calibration parameters (e.g. camera position and orientation) and intrinsic camera calibration parameters (e.g. sensor distortion). Bundle block adjustment, which uses camera calibration parameters and all input aerial images simultaneously, is used to reconstruct an optimal 3D geometry which is presented in the form of a sparse point cloud. For this study, a sparse point cloud with 3,479,183 points was created.

Scaling and georeferencing
This step is critically important if users want to accurate dimensions of 3D geometry. Noting that the previous step only estimates relative camera locations and scene geometry. That said, absolute distances between cameras and reconstructed 3D points cannot be resolved from aerial images, regardless the amount of cameras, aerial images, and keypoints being used (Szeliski 2011). Therefore, the 3D sparse point cloud needs to be improved with additional surveyed points that have accurate coordinates. This step involves using a set of GCPs collected by survey-grade GPS equipment to calibrate the 3D coordinates describing the scene geometry. For this study, the RTK-surveyed GCPs were used to scale and georeference the sparse point cloud.

Mvs
This step uses SfM estimated camera calibration parameters to create depth maps that are used to reconstruct dense 3D geometry. Many methods are available for MVS, but in general they can be classified into four groups, including Voxel-based methods, surface evolution-based methods, depth map merging methods, and patch-based methods (Carrivick, Smith, and Quincey 2016). At end of this step, a scaled and georeferenced 3D dense point cloud will be created. For this study, a dense point cloud with a total of 126,609,438 points was created.

Co-registered orthophotos and DSMs
In this step, the dense point cloud in the previous step is used to create a triangulated irregular network (TIN) mesh, which is then rasterized to create a DSM (a grid representation of the ground surface). The DSM will be used as a project surface during orthocorrection of the input aerial photos. At the completion of orthocorrection, aerial photos will be mosaicked to create an orthophoto, which is co-registered with the DSM. The orthophoto and DSM are generated in a single processing routine, and they can be exported as rasters (e.g. GeoTIFF). For this study, the exported orthophoto and DSM are at a spatial resolution of 0.005 m (Figure 3).

Analysis tool development
A cracking extraction tool that is compatible with a standard GIS software program (i.e. ArcGIS ArcMap) was developed to analyse the co-registered orthophoto and DSM to extract bridge deck wearing surface distresses (i.e. cracks). It was developed in Python and it involves three types of image processing techniques, including image enhancement, image fusion, and image difference. This tool has two primary outputs, including a fused image (i.e. fuse the orthophoto and DSM) to accentuate bridge deck wearing surface cracking and a height difference image to highlight bridge deck wearing surface cracking.
In this tool, eight shaded relief images are created from the input DSM at a 45-degree interval (i.e. 45, 90, 135, 180, 225, 270, 315, 360 degrees). Then, an average of these eight shaded relief images are conducted to create a mean shaded relief image. The orthophoto is used to create an albedo image based on the average pixel value of the visible red, green, and blue bands. Finally, the mean shaded relief image and the albedo image are fused based on an average operation to create the fused image to accentuate bridge deck wearing surface cracking (Figure 4a). This tool also conducts image enhancement to the DSM image through a 5 × 5 focal statistics tool with maximum values to enhance geometric information. Then, the enhanced DSM and the original DSM are differenced based on a subtract operation to create a difference DSM image. Subsequently, the difference DSM image is reclassified to assign null values to any pixels that have an elevation value that is greater than 0.016 m (0.75 inch). According to NMDOT, most wearing surface cracks are less than 0.016 m. Figure 4b shows the height difference image to highlight bridge deck wearing surface cracking.

Accuracy validation
An accuracy validation was conducted to validate the detected bridge deck wearing surface distresses in terms of crack count, crack length, and crack width. The bridge deck wearing surface data collected through ground survey were used as the ground-truth data and then compared to the wearing surface distress maps generated through the developed cracking extraction tool. For count, the fused image and the differenced image were used individually and jointly to digitize all cracks. Then the successful cracking detection rates were analysed with descriptive statistics. For length and width, measurements were first obtained from the images, and then they were compared with the groundtruth measurements with formal statistical tests. The sample size is 16, which is less 30, and therefore, nonparametric statistical tests were used (Arnold and Emerson 2011).
Measurement comparisons were conducted with both paired and unpaired statistical tests. Paired tests are more proper if two groups of measurements are dependent (i.e. repeated measurements for the same feature but at two different times). Unpaired tests are more proper if two groups of measurements are independent (i.e. measurements in one group have no impact on measurements in another group). The relationship between ground-based manual measurements and aerial imagery-based measurements can be interpreted in both a dependent way and independent way. In a dependent way, repeated measurements of a crack were performed on the ground and from aerial imageryderived products at two different times. In an independent way, the ground-based manual measurements have no impacts on the aerial imagery-based measurements because they are from two different data sources. Since the relationship can be interpreted in both ways, to err on the side of caution, this research used both paired tests and unpaired tests. The Wilcoxon signed rank test was used for paired statistical test, while Mann-Whitney U-test was used for unpaired statistical test to examine if aerial imagery-based measurements and ground-based measurements are statistically similar.

Results and discussion
For crack count validation, the ground survey results were used as the ground-truth data. The total amount of cracks identified through field survey is 16 ( Figure 5). When using only the crack-highlighting image, 10 cracks were identified visually, including crack ID 2, 3, 5-9, 12, and 14 ( Figure 4a). However, when using only the crackaccentuating image, all 16 cracks were successfully identified ( Figure 4b). The results are summarized in Table 2.
Results revealed that the crack-accentuating image could effectively identify cracks on the wearing surface of a bridge deck. In contrast, the crack-highlighting image cannot identify all cracks. However, crackhighlighting image can assist bridge inspectors with quick engineering reviews and high-level information checks. It also provides a good starting point for crack identification. When coupled with the crackaccentuating image, crack-highlighting image can also be used to identify cracks on the wearing surface of a bridge deck. For length and width validation, the ground survey results were also used as the groundtruth data. In addition, on-screen image measurement was conducted on the derived orthophoto to record the length and width of each crack. The digitized and ground-surveyed measurement results are summarized in Table 3.
The histogram distribution pattern of length and width measurements were examined and it revealed that ground-based measurements and ground-based measurements exhibit the same histogram distribution pattern. A box plot was also used to examine the distribution pattern of length and width measurements, and it did not show a substantial difference in the medians between ground-based length measurements and orthophoto-based length measurements. Additionally, no aforementioned plots revealed a substantial difference in the shape and spread of distribution between the two sets of measurements for both crack length and width.
Continuing with visual analysis, formal statistical tests were conducted. As mentioned in the previous section, the Wilcoxon signed rank test was used to compare measurement results at the paired group level. For both length and width measurements, the p-value is greater than 0.05 (0.8036 and 0.4545, respectively), and therefore, the null hypothesis should be accepted, indicating that for both length and width the median difference between the paired ground-based measurement and orthophoto-based measurement is zero at a 5% significance level (Table 4). In other words, for both length and width, ground-based measurements and   orthophoto-based measurements are not statistically different at a 5% significance level. The Mann Whitney U Test was performed to compare the measurements in an unpaired group way. Although this test does not require normally distributed data, it requires data from each population must be an independent random sample, and the population must have equal variances. For non-normally distributed data, the Levene's Test and Bartlett's Test are usually used to examine variance equability. For the Levene's Test and Bartlett's Test, the null hypothesis is that the population variances are equal. For both length and width, the p-value is greater than 0.05 (Table 5), and therefore, the null hypothesis should be accepted, and subsequently indicating that the population variances for both length and width measurements are equal at a 5% significance level. Ultimately, the Mann Whitney U Test is appropriate for both length and width. For both length and width measurements, the p-value is greater than 0.05 (0.9699 and 0.7052, respectively), and therefore, the null hypothesis should be accepted, and therefore indicating for ground-based measurements and orthophoto-based measurements there is no significant difference in the distribution pattern at a 5% significance ( Table 6).
All the aforementioned statistical tests revealed that there is no evidence showing that the ground-based length and width measurements and orthophotobased length and width measurements are statistically different at a 5% significance level. Ultimately, these results collectively prove that the high-spatial resolution aerial imagery acquired by S-UAS and the developed analysis tool can be effectively used to characterize cracks on the wearing surface of a bridge, and the accuracy is comparable to that of ground-based manual measurement.
Further investigation revealed that orthophoto-based measurements generally have lower values than ground-based manual measurement. For example, for length measurement, orthophoto-based method has nine measurements have lower values than groundbased measurement. While for width measurement, orthophoto-based method has 10 measurements have lower values than ground-based measurement. This discrepancy could be from either method because measurements made by an inspector in the field involve random errors and equipment errors that cannot be avoided. On the other hand, the orthophoto created in the SfM technique could also produce errors that cannot be avoided.
Additional investigation also revealed that the crackhighlighting image is less accurate than the crackaccentuating image in terms of cracking identification. The expectation is the crack-highlighting image will always be less accurate because it is derived only from DSMs, while the crack-accentuating image is derived from both DSMs and orthophotos, which contain additional pictorial information. However, this less accuracy expectation warrants further investigation and it could be a future research topic. In addition, it is believed that the crack-highlighting image is still worth creating even if it is less accurate than the crack-accentuating image. This is because it can be used to provide a quick way to assist bridge inspectors in visually locating many of the large and long cracks on a bridge deck. That said, the crack-highlighting image could be used for rapid, high- Although this study only considers one bridge, the results are expected to be very generalizable. This is because the wearing surface of the studied bridge's deck is paved with asphalt, which is the most prevalent material. That being said, the proposed method can be used to characterize any cracking distresses in asphalt surfacing. However, it may not be able to characterize cracks in other wearing surface materials such as concrete and wood, but these materials are much less common to be used for bridge deck's wearing surfaces. It should also be noted that the S-UAS collected aerial photos for this study can only detect and localize large-size cracks (i.e. 0.005 m in terms of length or width). However, the minimum crack size that can be detected varies with the input aerial imagery's spatial resolution. That said, the minimum detectable crack size could be improved when higherspatial resolution aerial photos are collected and used. For example, if the input aerial images have a spatial resolution of 0.002 m (2 mm), indicating the minimum detectable crack size is 2 mm (in terms of both length and width).
It is worth noting that this research was focused on using natural colour aerial photos to detect cracks on bridge deck. It is also worth noting that bridge deck has both surface and subsurface distresses. However, only surface distresses such as cracks can be observed and detected from natural colour aerial photos. That said, natural colour aerial photos cannot be used to evaluate subsurface conditions. This is be because natural colour aerial photos cannot penetrate the wearing surface to observe subsurface conditions. However, recent advancements in remote sensing such as thermalinfrared camera and light detection and ranging (LiDAR) have been becoming commercially available in miniaturized forms for operation on UAS. Thermalinfrared cameras, while still being expensive and having low-spatial resolution, have the potential to effectively evaluate bridge deck subsurface condition. Many researchers have started exploring the utility of thermal-infrared camera to evaluate delamination conditions. However, thermal-infrared cameras have not been operationally used for UASbased evaluation due to its low-spatial resolution and effective imaging range such as less than 50 metres (Escobar-Wolf et al. 2018).
One of the drawbacks of the proposed method is that it cannot be used for automated crack detection. However, this research lays the foundation for the future application of the proposed method for automated detection and assessment of detailed bridge deck wearing surface conditions. Operationally, the proposed method could be implemented internally by transportation agencies or externally by consulting firms as a service. Ultimately, the extraction of bridge deck wearing surface defects should be automated with deeplearning or machine learning algorithms to enable cost-effective scaling of S-UAS-based bridge inspection.

Conclusion
Traditional bridge deck inspection is conducted on the ground. These traditional methods are expensive, timeconsuming, labour-intensive, unsafe, and requiring specialized staff on a regular basis. They can also exhibit a high degree of variability. To overcome these aforementioned challenges, this study explored the utility of S-UAS-based airborne imaging techniques and advanced image analysis techniques in evaluating bridge deck's surface conditions. This research developed a robust tool that be used in standard GIS to automate the crack detection and extraction process. This tool was developed with the aim of being able to detect and map bridge deck surface distresses with an adequate degree of accuracy while maximizing the ability to assist inspectors with varying expertise. Research results showed that the outputs of the tool can be effectively used to count the amount of cracks on bridge deck wearing surfaces. Research results also revealed that the orthophoto-based length and width measurements are not statistically different from the ground-based length and width measurements at a 5% significance level. Collectively, these results proved that the imagery acquired by S-UAS and the developed analysis tool can be effectively used to characterize cracks on the wearing surface of a bridge, and the accuracy is comparable to that of ground-based manual measurement. An additional benefit of the proposed bridge deck wearing surface distress inspection system is that the aerial photos also provide a visual record of the study area. Aerial photos can provide a synoptic view of the bridge being inspected; permitting the documentation the bridge at the time the photos were taken. Unlike textual documents, aerial photos can document scenes and events without topical selection and human interpretation.