Underwater Bridge Inspection: Platforms, Sensors, and Data Processing for Condition Assessment—A Review

Abstract

Underwater bridge inspection combines direct observation, acoustic imaging, optical measurement, and data interpretation to examine components that are difficult to access. This review aims to organize the principal inspection methods and explain how their operating principles and technical characteristics relate to bridge condition assessment. The qualitative synthesis covers deployment platforms, optical and acoustic equipment, image and point-cloud processing, representative applications, and assessment practice. Three themes emerge. First, platform accessibility and sensor viewing geometry jointly determine inspection coverage. Second, acoustic methods provide structural and riverbed information under low visibility, while optical, laser, and tactile methods contribute local surface detail. Third, processing methods connect observations to component identification, geometric description, and comparisons across inspections. The review distinguishes the roles of two-dimensional images, three-dimensional point clouds, and visual or contact observations, and relates these information types to scour, surface damage, joint anomalies, and deformation. Its contribution is an integrated account of the technical characteristics and complementary uses of inspection methods across the acquisition-to-assessment process. Future directions include coordinated platform operation, multisensor alignment, processing adapted to underwater data, and integration of repeated inspections into bridge maintenance.

1. Introduction

Underwater components of river-crossing and sea-crossing bridges, including piers, pile foundations, pile caps, and open-caisson foundations, are exposed to hydraulic scour, sediment abrasion, chloride attack, and surface fouling. These components are important targets for bridge safety assessment and durability maintenance. Poor visibility, restricted access, and hydrodynamic disturbances complicate the observation of scour holes, exposed foundations, concrete spalling, steel corrosion, and deformation [1]. As major sea-crossing bridges age, including the Hangzhou Bay Bridge [2] and the Hong Kong–Zhuhai–Macao Bridge [3], inspection tasks increasingly connect damage identification with condition assessment and maintenance planning.

Sea-crossing deployments involve salinity- and temperature-dependent sound speed, tides and currents, waves, platform motion, and vessel traffic. These conditions influence the operating window, sensor configuration, and coverage of underwater components. Riverine, port, and sheltered-water applications provide examples of how inspection methods are deployed under different environmental and structural conditions.

Inspection of underwater structures has evolved from diver-based surveys toward remotely coordinated operations using unmanned platforms [4]. Divers provide close visual observations, tactile examination, and simple measurements. Remotely operated vehicles (ROVs), autonomous underwater vehicles (AUVs), vessel-mounted systems, and unmanned surface vehicles (USVs) carry cameras, sonar, and positioning equipment over different spatial scales. The available working space, water depth, current, visibility, and access to piers or pile groups guide the combination of manual and unmanned approaches.

Inspection sensors fall mainly into optical and acoustic categories. Cameras and laser scanners reveal cracks, spalling boundaries, and joint details [5]; their working range varies with turbidity, illumination, scattering, and refraction [6,7]. Two-dimensional sonar provides target images under low visibility. A multibeam echo sounder (MBES) acquires swath soundings for riverbed mapping and, with suitable mounting and viewing geometry, structural surfaces. Volumetric imaging sonar and mechanically scanned sonar acquire component geometry in other configurations. Optical, laser, and tactile observations complement acoustic inspection where local surface texture or fine boundaries are of interest [8].

Data analysis connects two-dimensional (2D) images and three-dimensional (3D) observations with structural interpretation. Forward-looking sonar (FLS), side-scan sonar (SSS), and synthetic aperture sonar (SAS) produce intensity images for target detection and localization [9,10,11]. MBES integrates soundings with positioning and attitude data; volumetric sonar acquires directions within an observation volume; and mechanically scanned sonar accumulates local point clouds through rotation. Applications range from scour mapping around piers to component reconstruction, local damage description, and repeated surveys [12,13,14].

Inspection data pass through enhancement, filtering, registration, target extraction, and geometric interpretation before they support assessment. Sound-speed variation, multipath reflections, platform motion, and suspended matter influence the appearance and spatial distribution of the observations [15]. Xiong et al. [16] reviewed sonar-based underwater bridge scour inspection and identification. The present review considers a broader set of structural targets and inspection methods, linking deployment platforms, optical and acoustic equipment, processing methods, and maintenance information. Table 1 summarizes the main themes of representative related reviews.

Table 1. Main themes of representative reviews related to underwater bridge inspection.

Review	Main Theme	Connection to the Present Review
Waldner and Sadhu [1]	Underwater vehicles for structural health monitoring	Deployment and inspection access
Huy et al. [4]	Underwater perception and sensing	Sensor principles and observation types
Steiniger et al. [8]	Deep learning for sonar imagery	Image interpretation and recognition
Xiong et al. [16]	Bridge scour inspection and identification	Foundation surroundings and riverbed changes
Mousavi et al. [17]	Digital twins in bridge management	Organization of inspection and maintenance information
This review	Underwater bridge inspection methods and condition assessment	Platforms, acoustic and optical methods, data interpretation, and engineering use

The objective of this review is to synthesize underwater bridge inspection methods through their operating principles, technical characteristics, and engineering uses, and to identify future research directions. The discussion brings together three contributions: a qualitative account of platform and sensor roles; an explanation of how image and point-cloud processing supports component and damage interpretation; and a connection between inspection information, condition assessment, and maintenance practice. Figure 1 organizes the review around inspection targets, defects and changes, inspection settings, and information needs. Its four columns correspond to operational approaches (Section 3), inspection equipment (Section 4), data analysis and interpretation (Section 5), and engineering applications and condition assessment (Section 6). The outputs of these stages connect coverage and data acquisition with component geometry, defect descriptions, and maintenance records. The lower band identifies the four future directions discussed in Section 7.

Figure 1. Organization of the review: inspection scope and tasks, the four connected topics addressed in Section 3, Section 4, Section 5 and Section 6, and the future directions discussed in Section 7. Abbreviations: USV, unmanned surface vehicle; AUV, autonomous underwater vehicle; ROV, remotely operated vehicle; 2D/3D, two-dimensional/three-dimensional; FLS, forward-looking sonar; SSS, side-scan sonar; SAS, synthetic aperture sonar; MBES, multibeam echo sounder; OUT, output.

2. Scope and Organization of the Review

This article is a qualitative narrative review of underwater bridge inspection methods and their use in condition assessment. It covers diver-based and unmanned operations, optical and laser inspection, acoustic imaging and ranging, image and point-cloud processing, and the interpretation of inspection findings for maintenance. The emphasis is on differences in physical principles, information types, deployment conditions, and engineering roles.

The literature is organized around four connected topics: operational approaches, inspection equipment, data analysis, and condition assessment. Representative applications illustrate how these methods examine riverbeds, piers, pile groups, caissons, cofferdams, and local defects. The technical discussion relates method characteristics to inspection tasks and uses published equipment parameters to describe representative configurations. The comparisons concern operating principles, imaging modes, deployment requirements, and inspection uses; the selected equipment specifications describe individual configurations.

3. Operational Approaches for Underwater Bridge Structural Inspection

The inspection task, sensing mechanism, and deployment platform jointly determine how underwater bridge components are observed. Sensors provide different information types, working ranges, and spatial resolutions; platforms determine access, viewing direction, stability, and spatial coverage. Divers and ROVs operate close to components, vessels and USVs survey larger areas, and fixed installations observe selected locations over time. Route-based autonomous systems provide another deployment option in accessible underwater environments [4]. Water depth, current, turbidity, working space, and positioning conditions guide platform selection.

Divers examine suspected cracks, spalling, joint anomalies, and surface conditions through close visual observation, touch, and simple measurements [1]. Their direct access supports flexible local judgments, while depth, current, visibility, and operational windows determine the practical extent of an inspection. Pier bases, pile-group gaps, and pile-cap undersides require particular attention to access and diver safety.

ROVs are controlled from the surface and can carry cameras, sonar, lasers, and tactile sensors. Slow movement and station-keeping near piles, pile caps, or caisson walls support detailed component inspection. AUVs follow autonomous routes, with applications in underwater structure inspection and view planning providing relevant operational approaches [4,18]. Tether management and operator visibility shape ROV operations; navigation, positioning, obstacle avoidance, and mission planning shape AUV operations. Position and orientation records connect observations from either platform to structural locations.

Fixed platforms place sonar, current meters, or other sensors on or near foundations to observe scour or selected structural locations over time. Stable sensor positions and viewing directions support comparisons during floods and other short-term events. Installation access, equipment maintenance, floating debris, fouling, and impact influence their operation, making fixed monitoring particularly relevant to critical locations.

Vessels and USVs carrying MBES or SSS follow planned survey lines to examine bridge-site riverbeds, scour extents, and exposed foundations. An ROV can then approach identified regions with acoustic, optical, or contact sensors. This combination connects area coverage with local component observations through common positioning and component records [19,20,21].

Coordinated inspection links area surveys, component observations, and local verification through common coordinates, timestamps, and component identifiers. Anomaly locations connect surface-platform observations with ROV images and point clouds. Repeated survey lines, control points, and viewing directions provide a consistent spatial basis for comparing inspections.

Acquisition planning specifies coverage, overlap, repeat-survey locations, and areas requiring additional viewpoints. Synchronized records of sound speed, attitude, position, sensor mounting, and operating conditions support subsequent processing and comparisons [22]. Local visual or tactile observations provide complementary surface information [23]. Survey products identify coverage gaps and differences between overlapping observations so that the available information can be traced to the inspection setting.

4. Equipment for Underwater Bridge Structural Inspection

4.1. Underwater Optical, Stereo Vision, and Laser Inspection Equipment

Underwater vision equipment includes monocular video systems, actively illuminated cameras, stereo or multi-camera systems, structured-light systems, laser scanners, and underwater lidar. These systems directly reveal surface texture, cracks, spalling boundaries, and joint details, and can recover local geometry after scale calibration [5,6]. Their main advantage is intuitive visualization of fine details, while their limitations arise from the propagation medium. Turbidity, illumination, suspended particles, refraction, and platform station-keeping all affect image quality and measurement accuracy.

In turbid water, absorption, scattering, color distortion, and artificial-light reflections reduce image contrast. Lighting optimization, descattering, color restoration, contrast enhancement, and segmentation improve the visibility of contours and surface features [6,7,24]. Scale calibration and camera geometry connect the visible features to local dimensions.

Stereo and multi-camera systems recover local three-dimensional morphology through disparity-based triangulation. During near-field ROV operations, they can measure crack widths, spalling boundaries, and bridge pier surface defects [5,25,26]. In combined sonar and stereo vision systems, sonar first locates suspected regions and cameras then capture surface details [20,21]. Reliability depends on underwater calibration, refraction correction, camera baseline, target distance, and synchronized acquisition. Weak texture, surface fouling, and suspended particles may cause matching failures.

Structured-light systems, laser scanners, and underwater light detection and ranging (LiDAR) use active illumination to recover range or surface profiles. Structured light has been applied to underwater pipeline reconstruction [27], and sonar–laser registration has connected above-water and underwater models of partially submerged structures [28]. Multi-camera calibration and refraction correction support bridge-pier surface measurements [25,26,29]. Turbidity, scattering, incidence angle, occlusion, and platform stability govern the usable optical view.

Optical and laser equipment provide local surface detail in clear water or at close range. In turbid-water inspection, acoustic observations can guide the platform toward a selected region before local optical imaging. Tactile imaging offers contact-based surface information where visual access is poor [23]. Common coordinates, scale calibration, and the identification of the same component connect these observations.

4.2. Two-Dimensional Sonar Inspection Equipment

FLS, SSS, and SAS provide different two-dimensional acoustic views [8,30]. FLS images a forward sector, SSS records successive side-looking swaths, and SAS combines coherent observations along a trajectory to form a synthetic aperture. Their characteristic views support near-field target localization, area surveys, and detailed seabed imaging, respectively.

FLS transmits fan-shaped beams and receives echoes at different azimuths and ranges to form two-dimensional acoustic images ahead of the sensor in real time. It is widely used for underwater robot localization, obstacle avoidance, and near-field target searches [30]. A single frame projects the three-dimensional scene onto an azimuth–range plane, losing elevation information and introducing ambiguity in target height and spatial correspondence. Forward-looking sonar odometry addresses this problem by constraining platform motion through frame-to-frame and submap matching. It suppresses accumulated drift and provides a relatively stable pose basis for three-dimensional reconstruction from multiple frames [9]. Method development also includes fractional-order feature propagation for object detection in forward-looking sonar imagery [31].

SSS records bilateral intensity images along a survey track and is used to examine riverbed morphology, pipelines, and component surroundings [32]. Target height can be estimated from acoustic-shadow geometry when sonar altitude and target position are known. Echo intensity describes the strength of the returned signal; it does not directly determine height. Shadow interpretation depends on the viewing geometry, bottom shape, and towfish attitude. Deconvolution improves boundary definition [33], while convolutional neural network (CNN)-based recognition, transfer learning, YOLOv8 variants, and DNTFE-Net support target detection and extraction [10,34,35]. Physics-informed geometric and frequency decoupling has also been used for weak-target perception in side-scan imagery [36].

Recent detector families include Ultralytics YOLO11 [37] and the attention-centric YOLOv12 [38].

SAS synthesizes a virtual aperture through platform motion. When trajectory and coherent-processing requirements are met, it provides high azimuth resolution that is relatively insensitive to range. This makes it suitable for detailed screening over wide riverbed swaths and underwater targets. Semantic segmentation networks enhance target boundary representation in SAS images [11], while deep supervision supports large-area seabed characterization and uncertainty estimation [39]. Object-detection algorithms have also been developed specifically for SAS images [40].

Three-dimensional information can be obtained directly from acoustic ranges and directions or reconstructed from two-dimensional image sequences. Direct acquisition uses MBES, volumetric imaging, or mechanical scanning. Indirect reconstruction combines image observations with platform motion and multi-view geometry. Section 4.3 and Section 5.1 describe the corresponding acquisition and processing approaches.

4.3. Three-Dimensional Sonar Inspection Equipment

Three-dimensional acoustic observations combine range and direction information. MBES acquires swath soundings and forms spatial surfaces by integrating position, attitude, sound speed, and survey-line observations. Volumetric imaging sonar resolves azimuth, elevation, and range within an observation volume. Mechanically scanned sonar accumulates directions as the sonar head rotates. These acquisition mechanisms provide different arrangements of spatial coverage and update rate.

(1) Multibeam echo sounders

In a conventional MBES, the transmitted beam is broad across track and narrow along track. The receiving array electronically forms multiple narrow beams across the swath. Echo travel time and receive-beam direction provide a sounding, which is combined with sound speed, position, roll, pitch, heading, and mounting information. The soundings acquired along adjacent survey lines are calibrated, filtered, and merged to form georeferenced surfaces or point clouds [12,41].

MBES development has gradually shifted from increasing coverage toward improving measurement accuracy and controlling uncertainty. Early systems such as Sea Beam extended seabed surveys from individual lines to fan-shaped swaths [42]. Studies of attitude, sound speed, sidelobes, and beam footprints subsequently clarified the sources of systematic error [43]. The CUBE method uses statistical redundancy in dense observations to handle outliers and estimate uncertainty [44]. MBES measurements are jointly affected by beam count, swath width, positioning, attitude, sound speed, and post-processing parameters [22].

MBES data are now used for seabed morphology interpretation, anomaly identification, backscatter correction, and sediment classification. Cui et al. [45] combined prior constraints with geostatistical prediction to identify anomalous points while preserving terrain features. Hiroji and Clarke [46] corrected system artifacts through radiometric beam-pattern extraction and relative calibration. For sediment classification, CNNs with feature optimization [47], decision fusion of deep and shallow learning [48], and multiple imputation [49] address feature representation, classification, and missing data, respectively. These methods broaden MBES applications, but their results remain dependent on survey-line design, angular response, and field calibration quality.

MBES supports continuous riverbed mapping, scour surveys, and inspection of exposed foundation surfaces. It can also survey vertical or steep structures, including caisson walls and pile sidewalls, when mounting orientation, steering range, and survey geometry provide a suitable view. Coverage of a particular surface depends on beam direction, incidence angle, footprint, and occlusion. Thus, downward-looking and side-looking configurations serve different inspection geometries.

(2) Three-dimensional imaging sonar

Three-dimensional imaging sonar encompasses more than one equipment type. Real-time matrix-array or phased-array sonar uses a two-dimensional transducer array to resolve azimuth, elevation, and range simultaneously within a single transmission. It directly displays volumetric images at relatively high update rates. Three-dimensional mechanically scanned sonar combines a multibeam sonar head with mechanical rotation, progressively accumulating sector or spherical scans to generate a point cloud. The former suits monitoring of dynamic operations and rapid component imaging; the latter suits localized, high-density measurements under stable deployment conditions. Both are three-dimensional sonar systems, but differ in real-time capability, scanning method, and error sources.

Three-dimensional acoustic imaging has developed around array beamforming, computational efficiency, and multi-view fusion. Early research established frameworks for three-dimensional acoustic image generation and processing [50]. Moving acoustic-lens experiments and registration of multiple range views further demonstrated the feasibility of spatial observation and scene reconstruction [51,52]. Subsequently, fast beamforming methods, including the chirp z-transform (CZT), reduced the computational demands of three-dimensional imaging [53]. Online modeling methods converted sequential range images into structural point clouds [54]. Together, these advances established the technical basis for real-time volumetric imaging and multi-view point-cloud fusion.

Matrix-array sonar is a principal form of real-time volumetric acoustic imaging. Its observations can be processed as images or point clouds for component extraction and reconstruction. Sewer point-cloud segmentation provides an example of processing underwater engineering geometry [55]. The real-time detection transformer (RT-DETR) has also been used for image-based underwater bridge-pier crack detection [56]. Mechanically scanned systems accumulate point clouds through rotation, so scan duration and platform stability are integral to their acquisition process.

Bridge engineering cases illustrate these operating roles. Yan et al. [13] used an Echoscope 4G C500 to observe caisson-foundation sidewalls and chamfers from close viewpoints. Zhang et al. [57] investigated denoising of locally acquired three-dimensional sonar point clouds, and Yan et al. [58] used sonar point clouds for steel-structure extraction, reconstruction, and corrosion-pit description. These applications connect acoustic imaging with local component morphology and damage interpretation.

Engineering deployments combine acquisition modes according to the target. Caisson-joint inspection can use MBES for initial coverage and three-dimensional scanning sonar for local observations [59]. Oscillatory FLS provides another route to spatial reconstruction through changing viewpoints [60]. Optical structured-light measurements supply detailed surface geometry [27].

4.4. Technology Selection and Acoustic-Optical Coordination

Equipment selection follows the information required from the target. MBES provides riverbed and structural soundings across a swath; volumetric sonar acquires an observation volume; and mechanically scanned sonar collects directions over a scan. Backscatter and angular response add information about the surveyed surface [61]. Table 2 compares acquisition principles, the information produced, and engineering uses. Acoustic–optical coordination begins with acoustic localization of a region and continues with close optical or laser observation of the same component. Extrinsic calibration, time alignment, refraction correction, and shared coordinates connect the two views. Acoustic geometry describes the broader structural setting, while optical texture and surface boundaries support local interpretation.

Table 2. Comparison of MBES and three-dimensional imaging sonar for underwater structural inspection.

Acquisition Type	Operating Principle	Characteristic Information	Engineering Use
MBES	Transmitted swath and electronically formed receive beams; position and attitude link successive soundings.	Georeferenced swath soundings and backscatter.	Riverbed and scour mapping; structural surfaces viewed with suitable mounting and survey geometry.
Volumetric imaging sonar	A two-dimensional array resolves range, azimuth, and elevation within an observation volume.	Rapid volumetric updates and component point samples.	Near-field structural observation and monitoring of dynamic operations.
Mechanically scanned sonar	A rotating sonar head accumulates directions over a scan.	Local point clouds assembled over the scan duration.	Detailed inspection of accessible components from a stable deployment.

Table 3 presents representative equipment configurations and selected published parameters. Frequency, range, field of view, and acquisition mode describe how the systems operate and the scale of their intended observations. The entries are organized by technology type to complement the principle-based discussion.

Table 3. Representative acoustic equipment, acquisition modes, and selected published specifications.

Technology and Example	Selected Published Parameters	Acquisition Characteristics	Typical Role
FLS: Seaview FLS800D [62]	850/1150 kHz; maximum range 120/40 m; horizontal view 130°/100°.	Forward-looking 2D intensity images; dual-frequency operation.	Target localization and ROV navigation.
SSS: EdgeTech 4205 Tri-Frequency [63]	120/410/850 or 230/540/850 kHz; range per side varies from 90 to 600 m across these frequencies.	Bilateral intensity swaths along a survey track.	Area surveys and acoustic-shadow interpretation.
SAS: HISAS 1032, HUGIN 3000 configuration [64]	Published image resolution 5 cm × 5 cm; range 240 m at 3 kn and 185 m at 4 kn.	Coherent aperture synthesis along platform motion.	Detailed seabed and underwater-target imaging.
MBES: Hydro-Tech MS400P [65]	400 kHz; 512 beams; nominal swath angle 143°.	Swath soundings combined with position, attitude, and sound speed.	Riverbed, scour, and accessible structural-surface surveys.
Volumetric sonar: Echoscope 4G C500 [66]	375/630 kHz; more than 8000 points per transmission; models up to 20 Hz.	Range and two angular coordinates within an observation volume.	Near-field component observation and dynamic updates.
Mechanical scan: BlueView BV5000 MK2 [67]	Field of view 42° × 1°; scanning range 45–360°; published range 30 m.	Rotational scanning accumulates a local point cloud.	Local inspection of piles, caissons, and cofferdams.

Notes: The parameters refer to the cited model and configuration. A sonar update rate and the time required to complete a mechanically scanned point cloud describe different acquisition processes. The EdgeTech tri-frequency and HUGIN 3000 HISAS configurations are used in the corresponding entries. Sources: [62,63,64,65,66,67].

5. Data Analysis and Interpretation

5.1. Two-Dimensional Sonar Image Processing and Indirect Three-Dimensional Reconstruction

Image processing extracts target information from acoustic intensity observations. Enhancement, detection, segmentation, and multi-frame reconstruction address different tasks, from improving visible contours to recovering spatial structure. Table 4 summarizes these methods and their engineering outputs.

(1) Image enhancement and target recognition

SSS and FLS images are affected by sidelobes, water-column reverberation, and noise, producing blurred edges and uneven contrast. Deconvolution and multiscale filtering restore contours [15,33], while simulated noise and near-surface artifact models support targeted denoising [68,69]. The preservation of target boundaries is particularly important when the enhanced image feeds subsequent segmentation or reconstruction.

FLS interpretation is affected by missing elevation information and nonuniform resolution. Conventional recognition methods use histograms of oriented gradients (HOG), local binary patterns (LBP), and acoustic shadows. Deep networks learn multiscale features from training images. FLS odometry provides pose constraints for reconstruction across frames [9], while lightweight segmentation models offer architectures for resource-constrained underwater platforms [70].

Multiscale networks address variation in target size [71]. The convolutional neural network–Receptance Weighted Key Value hybrid (CNN-RWKV) combines local texture with sequence information [72], while imaging and reflection models describe differences among seabed materials [73]. Studies combining sonar recognition with building information modeling (BIM) in bridge applications connect detected regions with structural information [74,75]. Detection and segmentation identify where a target or region occurs; geometric analysis then describes its dimensions and spatial relationship to a reference.

(2) Indirect three-dimensional reconstruction under multi-view constraints

Indirect three-dimensional reconstruction from two-dimensional sonar sequences depends on platform motion, multi-view observations, and pose fusion. Rapidly exploring random tree (RRT) viewpoint planning can optimize multi-view observation paths [18]. Adaptive region-of-interest (ROI) Radon transforms improve the stability of inter-frame rotation estimation at low signal-to-noise ratios [76]. Fusion of Doppler velocity log (DVL) and inertial measurement unit (IMU) data provides pose constraints for multiple two-dimensional observations [77]. Orthogonal dual-sonar and probabilistic reconstruction methods improve the sparsity and incompleteness of single-view point clouds [78,79,80]. Compared with direct three-dimensional sonar measurements, indirect reconstruction reuses existing two-dimensional equipment. However, its point-cloud accuracy is more sensitive to pose drift, observation overlap, and geometric assumptions, making it suitable for cost-constrained applications or localized reconstruction.

(3) Deep learning and foundation models in sonar image processing

Vision Transformers and sonar-adapted Segment Anything Model (SAM) methods have been applied to acoustic image segmentation. Fine-tuning, collaborative prompts, and embedding compensation adapt image representations to blurred boundaries and sonar-specific appearance [81,82]. Acoustic shadows, reverberation, and intensity variation also motivate specialized enhancement and segmentation. Point Transformer, PointNet, and PointNet++ provide point-set representations for subsequent analysis of reconstructed geometry [83,84,85].

Table 4. Processing tasks, representative approaches, and the interpretation of their outputs.

Task	Representative Approaches	Output and Interpretation	Evaluation Focus	References
Enhancement and denoising	Deconvolution, multiscale filtering, and artifact removal	Improved visibility and retained target boundaries	Contrast, edge preservation, and residual noise	[15,33,68,69]
Detection	Transfer learning, YOLO-family detectors, and feature fusion	Target regions in acoustic images	Precision and recall under a defined matching rule	[10,34,35,36,37,38,71,72]
Segmentation	DNTFE-Net, sonar-adapted SAM, and point-set networks	Image or point labels for target extraction	Region overlap and boundary agreement	[11,35,39,81,82,83,84,85]
Pose and registration	Odometry, frame matching, and sensor fusion	Relative viewpoints and aligned observations	Pose consistency and registration residuals	[9,76,77]
Reconstruction	View planning, dual-sonar geometry, and probabilistic fusion	Spatial surfaces from multiple observations	Coverage and correspondence to reference geometry	[18,78,79,80]
Geometric interpretation	Reference fitting, sections, and surface comparison	Dimensions and local deviations	Dimensional residuals and root-mean-square error	[12,14,25,26,58]

Table 4. Processing tasks, representative approaches, and the interpretation of their outputs.

Task	Representative Approaches	Output and Interpretation	Evaluation Focus	References
Enhancement and denoising	Deconvolution, multiscale filtering, and artifact removal	Improved visibility and retained target boundaries	Contrast, edge preservation, and residual noise	[15,33,68,69]
Detection	Transfer learning, YOLO-family detectors, and feature fusion	Target regions in acoustic images	Precision and recall under a defined matching rule	[10,34,35,36,37,38,71,72]
Segmentation	DNTFE-Net, sonar-adapted SAM, and point-set networks	Image or point labels for target extraction	Region overlap and boundary agreement	[11,35,39,81,82,83,84,85]
Pose and registration	Odometry, frame matching, and sensor fusion	Relative viewpoints and aligned observations	Pose consistency and registration residuals	[9,76,77]
Reconstruction	View planning, dual-sonar geometry, and probabilistic fusion	Spatial surfaces from multiple observations	Coverage and correspondence to reference geometry	[18,78,79,80]
Geometric interpretation	Reference fitting, sections, and surface comparison	Dimensions and local deviations	Dimensional residuals and root-mean-square error	[12,14,25,26,58]

5.2. Acquisition of Three-Dimensional Sonar Point Clouds

Three-dimensional point clouds represent riverbed and component surfaces as spatial samples. They support contour extraction, damage localization, and geometric description [12,13,14]. Calibration, sound-speed correction, attitude compensation, coordinate transformation, and overlapping views establish their spatial arrangement. Figure 2 connects acoustic scanning of a pier with the Acquire–Align–Extract–Describe sequence and the sensor-to-platform-to-world coordinate transformations.

Figure 2. Acquisition and processing of three-dimensional sonar point clouds: (a) acoustic scanning of a bridge pier; (b) acquisition of raw scans, alignment through calibration, filtering and registration, extraction through segmentation and outlier removal, and description through geometry fitting and defect measurement; and (c) chained transformations from the sensor frame (s) through the platform frame (p) to the world frame (w). The two transformations represent the calibrated sensor mounting position and orientation and the platform position and attitude, respectively. D and H illustrate component diameter and height. ENU, east–north–up; GNSS, Global Navigation Satellite System; IMU, inertial measurement unit.

(1) Direct acquisition

Direct acquisition of three-dimensional point clouds primarily uses MBES, three-dimensional imaging sonar, three-dimensional mechanically scanned sonar, and real-time three-dimensional acoustic cameras. These systems calculate range from echo travel time. Beam azimuth and elevation, platform attitude, and positioning data are combined to transform observation points from the sonar coordinate system to the engineering coordinate system. Acquisition also requires time synchronization, sound-speed profile correction, and consistent sensor extrinsic parameters. In Figure 2c, the mounting calibration defines the sensor-to-platform transformation. For a surface-based platform, Global Navigation Satellite System (GNSS) positioning and IMU attitude locate the observations in an east–north–up (ENU) world frame.

Huang et al. [12] used MBES riverbed point clouds around bridge piers to describe scour-hole morphology. Xiao [59] used three-dimensional scanning sonar and a measurement bracket for local caisson-joint observations. These examples represent area-scale riverbed inspection and close component inspection. Filtering, registration, and fusion connect observations from different views before target extraction and geometric analysis [14].

(2) Indirect acquisition

Indirect acquisition uses FLS, SSS, or sonar profile sequences to recover three-dimensional morphology. It relies on platform motion, mechanical sweep scanning, multi-view observations, dual-sonar constraints, or pose fusion. The key is to establish correspondence among pixels or range observations, beam geometry, and platform pose. This compensates for the elevation and height information missing from individual two-dimensional sonar frames. Specific methods are discussed in Section 5.1 (2).

The acquisition route follows the target scale, viewing geometry, and deployment setting. MBES provides swath observations of riverbeds and accessible structural surfaces. Volumetric sonar supplies rapid near-field updates, while mechanical scanning accumulates dense local samples over time. Reconstruction from 2D sequences uses platform motion and overlapping views to recover spatial information. Scan duration, accessible viewpoints, and the coordinate reference distinguish the practical use of these routes.

5.3. Three-Dimensional Sonar Point-Cloud Processing and Metric Extraction

Point-cloud processing extracts riverbed and component geometry from observations containing outliers, water-column returns, density variations, and missing regions. Figure 2b groups the processing tasks into four stages. Acquire supplies raw scans from different viewpoints; Align combines calibration, filtering, and multi-scan registration; Extract separates the component from riverbed and outlier points; and Describe fits reference geometry and measures component dimensions or local deviations. Figure 2c shows the coordinate transformations underlying scan alignment, while the methods below address the processing steps within this sequence.

(1) Denoising, registration, and quality control

Preprocessing brings observations into common coordinates and removes isolated or inconsistent points. Filtering, simplification, registration, and fusion retain the surfaces and boundaries relevant to the task, including pier bases, scour edges, corrosion pits, and component contours.

The Point Cloud Library (PCL) provides filtering, feature estimation, registration, and segmentation tools [86]. For bridge applications, Yan et al. [13] examined reconstruction from sonar point clouds, and Luo et al. [87] combined filtering, morphological processing, and wavelet denoising to recover pier contours. Rectangular information granulation [88] and optimized outlier detection [89] address 3D sonar data and MBES outliers. Point density, range, and target scale guide the choice of neighborhoods and filtering parameters.

Registration establishes the spatial relationship among views and across inspections. Iterative closest point (ICP) [90] and multi-view surface registration [91] provide fine alignment based on corresponding geometry. Simultaneous localization and mapping (SLAM) combines sonar, vision, inertial measurements, and pressure information [92]. Sonar–laser registration connects submerged and exposed surfaces [28], while visual–inertial–depth fusion and loop closure address accumulated drift [93]. Probabilistic registration, 3D scan matching, and FLS-based SLAM provide further approaches for sparse acoustic observations [94,95,96,97]. Overlap and stable geometric features influence the resulting alignment.

(2) Target extraction and semantic segmentation

Target extraction and semantic segmentation separate components, riverbed surfaces, fouling, and water-column noise, providing prerequisites for damage localization and geometric quantification. Rule-based methods depend on geometric priors and parameter calibration, whereas learning-based methods depend on samples, class definitions, and stability across scenes.

Geometric-prior methods use component shape to guide extraction with relatively small training requirements. Curve fitting, region growing, minimum cuts, and density-based spatial clustering of applications with noise (DBSCAN) are representative approaches. Applications include separating pile groups from riverbeds [98], locating damage through robust statistics and clustering [14], and extracting pipe cross-sections from multibeam water-column data [99,100]. Point density, range, and target scale determine suitable fitting and clustering settings.

PointNet and PointNet++ model unordered point sets and local neighborhoods, while dynamic graph convolutional neural networks (DGCNNs), RandLA-Net with random sampling and local aggregation, and Point Transformer provide graph, large-scale, and attention-based representations [83,84,85,101,102]. In acoustic point clouds, noise, variable density, and blurred boundaries affect neighborhoods and learned features. These architectures provide modules for methods adapted to underwater inspection data.

Underwater point-cloud methods address bridge foundations, pipelines, and sewer networks. Lightweight graph convolution supports bridge-foundation and terrain segmentation [103]. Cross-scale and convolutional point Transformers represent small targets and variable sampling density [104,105], while semi-supervised learning uses limited annotations and addresses class imbalance [106]. Table 5 summarizes these methodological groups and their characteristic uses.

Table 5. Comparison of target extraction and semantic segmentation methods for three-dimensional sonar point clouds.

Method Group	Examples	Characteristic Use	Data Considerations	References
Geometric priors	Curve fitting, region growing, minimum cuts, DBSCAN	Extraction of piles, riverbeds, and local anomaly regions using shape information	Neighborhood size, density, and fitting assumptions	[14,98,99]
General point-set models	PointNet, PointNet++, DGCNN, RandLA-Net, Point Transformer	Learned point classification and local or global feature representation	Training labels, sampling density, and boundary quality	[83,84,85,101,102]
Underwater-oriented models	Lightweight graph convolution, Cross-PIC, convolutional point Transformers, semi-supervised segmentation	Representation of underwater foundation, pipeline, and sewer geometry	Underwater noise, nonuniform samples, and class balance	[103,104,105,106]

(3) Geometric feature representation and damage quantification

A geometric description links an extracted component to a defined reference. Common quantities include cross-sectional dimensions, surface flatness, scour depth and extent, foundation exposure, surface-loss area, and displacement. Coordinate directions, the analysis region, and the reference surface determine how these quantities are obtained.

Normals, curvature, shape descriptors, and model fitting represent point-cloud geometry. Combined with segmentation, they locate bridge-surface defects [107] and describe pier morphology [25]. Scour is measured relative to a riverbed surface, design elevation, or historical observation [12]. Component sections can be compared with design geometry, fitted intact regions, or earlier models to describe relative deviations [14].

Local damage analysis describes defect location, dimensions, area, depth, and distribution. Multi-distance clustering and adaptive alpha shapes have been used to characterize steel-cofferdam corrosion pits [58], while refraction correction and multi-camera calibration support optical description of pier surfaces [26]. Scour prediction models describe risk trends [108], and field observations provide the corresponding riverbed geometry.

Nonplanar components and irregular defects affect reference fitting and boundary extraction [29]. Acoustic observations locate surface-loss and suspected damage regions, while close optical, laser, or tactile observations reveal fine boundaries and texture [58]. Comparisons with design geometry or historical models describe translation, rotation, and overall component deformation.

Figure 3 summarizes representative deterioration processes and associated geometric changes in water-exposed concrete bridge substructures. Local scour removes surrounding bed material and increases foundation exposure, whereas concrete deterioration modifies the structural surface and cross-section. Surface abrasion involves the progressive removal of cement paste and aggregate, while localized section loss represents a larger-scale reduction in structural geometry. Under suitable exposure conditions, reinforcement corrosion can generate expansive products that promote cover cracking and spalling. These processes may coexist, so similar surface features do not necessarily indicate the same underlying cause. Their geometric manifestations provide inspection targets, but identifying the governing mechanism requires evidence beyond surface geometry alone.

Figure 3. Representative deterioration in water-exposed concrete bridge substructures: (a) local scour and foundation exposure; (b) localized section loss; (c) surface abrasion; and (d) corrosion-induced cracking and cover spalling. Blue arrows indicate flow, particle transport, or chloride ingress; orange arrows indicate corrosion-product expansion. Dashed lines mark the original bed level in (a) and structural profile in (b). Schematics are not to scale.

Acoustic imaging describes component geometry and the surrounding riverbed, while local optical or contact observations resolve surface texture, cracks, and spalling boundaries. Recording these observations against the same component and coordinate reference connects the location of a geometric anomaly with its surface description.

5.4. Measurement Principles and Interpretation of Inspection Results

Acoustic ranging relates echo travel time to distance. In a homogeneous medium, the two-way travel relation and the ideal bandwidth-limited range resolution are given by Equation (1). For a real aperture at small angles, Equation (2) relates cross-range resolution to range and angular width. These relations explain how bandwidth, aperture, and wavelength influence the detail represented by acoustic observations.

$$r={\displaystyle \frac{\mathit{ct}}{2}},\hspace{1em}\Delta r\approx {\displaystyle \frac{c}{2B}}$$

(1)

Here, r is range (m), c is sound speed (m/s), t is two-way travel time (s), B is signal bandwidth (Hz), and Δr is ideal range resolution (m).

$$\delta x\approx \mathit{R\theta }\approx {\displaystyle \frac{\mathit{R\lambda }}{D}}$$

(2)

In Equation (2), δx is cross-range resolution (m), R is target range (m), θ is angular width (rad), λ is wavelength (m), and D is the physical aperture (m). The relation describes real-aperture geometry; SAS forms its along-track aperture through coherent platform motion.

Processing metrics describe different parts of the inspection task. Precision is the fraction of predicted detections that are correct, and recall is the fraction of reference targets detected, under the stated matching rule. Intersection over union (IoU) is the overlap of predicted and reference regions divided by their union; mean IoU (mIoU) averages this value over the evaluated classes. Boundary F1 is the harmonic mean of boundary precision and recall, calculated with a stated spatial matching tolerance. Root-mean-square error (RMSE) is the square root of the mean squared difference between an estimated quantity and its reference. Region and boundary metrics describe segmentation, while dimensional residuals describe geometric measurements.

Spatial resolution describes the separation at which features can be distinguished. Point density describes the number of samples in an area or volume. Repeatability describes the spread of repeated observations under the stated conditions, and measurement accuracy concerns agreement with a reference quantity. These concepts connect the sensing mechanism, sampling pattern, and processing result to different aspects of inspection quality.

Table 6 relates image, point-cloud, contact, and time-series information to inspection tasks. The principal outputs range from visible surface features to component geometry and changes between observations. Table 7 summarizes physical and processing factors that influence these outputs and the corresponding acquisition or processing controls.

Table 6. Representative inspection outputs and their use in structural interpretation.

Information Type	Representative Methods	Inspection Interpretation	Examples
Surface images	Cameras, stereo vision, and laser imaging	Texture, cracks, spalling, and joint boundaries	[5,6,7,25,26,27,28,29]
Acoustic intensity images	FLS, SSS, and SAS	Target location, surface response, and acoustic shadows	[8,9,10,11,30,31,32,33,34,35,36,39,40]
Spatial point samples	MBES, volumetric imaging, and mechanical scanning	Riverbed morphology and component shape	[12,13,14,58,59,60]
Local contact observations	Diver examination and tactile imaging	Close surface detail at an identified component	[1,23]
Temporal records	Repeated surveys and fixed monitoring	Scour evolution and changes in component geometry	[16,22,109,110]

Table 7. Physical and processing factors affecting underwater inspection information.

Factor	Effect on Inspection Information	Acquisition or Processing Response	References
Sound speed	Changes the range scale and propagation direction.	Sound-speed profiling and range/beam correction.	[22,41,43]
Position, attitude, and mounting	Changes the location and orientation assigned to an observation.	Positioning, synchronization, attitude compensation, and extrinsic calibration.	[22,41,92,93]
Viewing geometry and occlusion	Changes which surfaces are sampled and the projected footprint.	Additional viewpoints and planned overlap.	[13,18]
Turbidity, scattering, and refraction	Changes optical contrast and the relation between image rays and geometry.	Lighting design, image processing, and underwater calibration.	[6,7,25,26,29]
Filtering and sampling density	Changes local surface smoothness and representation of small features.	Density-aware neighborhoods and feature-preserving filtering.	[15,57,87,88,89]
Registration	Changes the alignment of different views or survey epochs.	Stable reference geometry, overlap, and coarse-to-fine alignment.	[28,90,91,92,93,94,95,96,97]
Segmentation and reference fitting	Changes the selected boundary or reference used to describe a feature.	Component-specific labels, geometric models, and local reference surfaces.	[14,25,26,58,98,99,100,101,102,103,104,105,106,107]

Scour depth is measured vertically from the reference riverbed level to the scoured bed at the piles, as represented in Figure 3a. The foundation exposure describes the portion of the foundation uncovered by sediment removal. Local section dimensions, surface-loss depth and area, crack width, and spalling extent describe the component changes represented in Figure 3b–d. Joint opening and offset are additional inspection quantities discussed in the engineering applications. Component location, viewing geometry, and the selected reference identify where and how each quantity is measured.

6. Engineering Applications and Condition Assessment

6.1. Engineering Cases of Underwater Bridge Structural Inspection

Three-dimensional sonar has been applied to caisson foundations, bridge piers, steel cofferdams, and bridge-site riverbeds. Yan et al. [13] reconstructed caisson geometry from close acoustic observations of sidewalls and chamfers. Zhang et al. [57] examined denoising of local sonar point clouds, and Yan et al. [58] characterized corrosion pits on steel cofferdams. High-resolution sonar and deep learning have also been used for local damage recognition, concrete-crack detection, and pier-surface interpretation [111,112,113]. Table 8 summarizes representative engineering targets and inspection outputs.

In scour inspection, Huang et al. [12] compared MBES-derived riverbed sections with design references to reconstruct scour morphology and its changes. Zhang et al. [14] separated pile groups from riverbed point clouds using minimum cuts. Xiao [59] combined initial MBES screening with detailed three-dimensional scanning sonar measurements to inspect caisson joints. Field research on Qiantang River bridge piers also demonstrated the applicability of multi-source sonar to scour morphology surveys [114]. These engineering results show that combining broad surveys with detailed local measurements improves anomaly detection efficiency and reduces missed detections caused by limited single-sensor coverage.

Pipeline, offshore-foundation, and harbor-caisson applications provide additional examples of underwater filtering, segmentation, target extraction, and geometric interpretation [32,59,100,115,116]. Guided-wave ultrasonic non-destructive testing (NDT) supplies a different information type, illustrated by void imaging in underwater concrete slabs [117]. These applications show how sensing mechanisms address surface geometry, structural surroundings, and internal features.

Across these applications, acoustic imaging supplies geometry under low visibility, and the acquisition viewpoint determines which surfaces are observed [12,13,14,57,58]. Filtering, registration, and component separation organize the spatial information before local feature analysis. Known dimensions, repeated observations, and other sensing modalities connect the resulting geometry with the inspected component [58,59].

Table 8. Engineering cases of three-dimensional sonar inspection of underwater bridge components.

Target or Setting	Equipment or Information	Inspection Task	Reported Output	References
Open-caisson bridge foundation	Echoscope 4G C500 volumetric sonar	Component reconstruction and foundation surroundings	Foundation geometry and scour information	[13]
Underwater steel cofferdam	High-resolution 3D sonar point clouds	Surface-loss and corrosion-pit inspection	Pit location, area, and depth	[58]
Riverbed around bridge piers	MBES point clouds	Local scour inspection	Scour extent and riverbed morphology	[12]
Harbor caisson joints	MBES and 3D scanning sonar	Area inspection followed by local joint observation	Joint location and local geometry	[59]
Pipelines and sewer pipes	Multibeam water-column data and point clouds	Cross-section extraction and defect-region identification	Cross-sections and local anomaly regions	[99,100,115]

Table 8. Engineering cases of three-dimensional sonar inspection of underwater bridge components.

Target or Setting	Equipment or Information	Inspection Task	Reported Output	References
Open-caisson bridge foundation	Echoscope 4G C500 volumetric sonar	Component reconstruction and foundation surroundings	Foundation geometry and scour information	[13]
Underwater steel cofferdam	High-resolution 3D sonar point clouds	Surface-loss and corrosion-pit inspection	Pit location, area, and depth	[58]
Riverbed around bridge piers	MBES point clouds	Local scour inspection	Scour extent and riverbed morphology	[12]
Harbor caisson joints	MBES and 3D scanning sonar	Area inspection followed by local joint observation	Joint location and local geometry	[59]
Pipelines and sewer pipes	Multibeam water-column data and point clouds	Cross-section extraction and defect-region identification	Cross-sections and local anomaly regions	[99,100,115]

6.2. From Inspection Metrics to Condition Assessment

Condition assessment connects inspection information with the structural component and the maintenance question. Coordinates, component identifiers, and timestamps locate the observation. The selected reference surface and geometric variable describe the feature, while calibration and repeat observations characterize measurement quality. Optical, laser, or tactile information adds local surface detail. The resulting record can then be interpreted against inspection guidance, structural assessment criteria, and project requirements.

Table 9 summarizes standards and guidance relevant to these assessment tasks. IHO S-44 Edition 6.2.0 (October 2024) distinguishes a priori uncertainty estimated during survey planning from a posteriori uncertainty evaluated using acquired data [118]. Its survey-quality requirements concern hydrographic observations, while bridge-evaluation and scour guidance address structural interpretation and maintenance decisions [119,120,121,122,123]. The assessment framework depends on the component, damage mechanism, and applicable inspection or maintenance rules.

Table 9. Standards and guidance relevant to underwater inspection and condition assessment.

Framework	Main Purpose	Connection to Underwater Bridge Inspection
IHO S-44, Edition 6.2.0 (2024) [118]	Hydrographic survey quality, coverage, feature detection, and uncertainty	Quality of bathymetric and other hydrographic observations.
FHWA Bridge Inspector’s Reference Manual [119]	Underwater inspection levels and inspection practice	Organization of visual, detailed, and NDT inspection.
AASHTO Manual for Bridge Evaluation [120]	Bridge evaluation and load-rating context	Interpretation of observed component condition.
HEC-18 [121]	Scour evaluation and countermeasure planning	Interpretation of scour and foundation exposure.
DNV-ST-0126 and ISO 19902 [122,123]	Support structures and offshore steel structures	Relevant marine-structure inspection and assessment concepts.
IMCA D 014 [124]	Offshore diving practice	Planning and conduct of diver-based inspection.
JTG/T H21-2011 [125]	Highway bridge technical condition evaluation	Organization of condition information within the applicable maintenance framework.

Complementary NDT methods provide information about internal defects or material condition, while fixed sonar and other installed sensors record temporal changes [109,117]. Their information can be associated with acoustic and optical surface observations through component-based inspection records.

6.3. Linking Point-Cloud Evidence to Maintenance Decisions

Table 10 links typical inspection findings with condition information and possible maintenance responses. Shared coordinates and component identifiers connect observations from different surveys [92,93]. Scour depth, foundation exposure, surface-loss extent, joint condition, and section geometry can then be considered alongside current, sediment, and foundation information when planning repeat inspections or interventions [126,127].

Table 10. Connections between inspection findings, condition information, and maintenance responses.

Inspection Finding	Condition Information	Possible Maintenance Response	References
Scour and foundation exposure	Riverbed profile, scour extent, exposure, and temporal trend	Review hydraulic and foundation conditions; plan follow-up surveys or scour protection.	[12,16,108,118,121,127]
Surface loss and corrosion pits	Location, area, depth, and affected section	Inspect the affected region and evaluate repair or section loss.	[13,58,119,120,121,122,123]
Component shape change	Cross-section, alignment, and deviation from reference geometry	Examine structural significance and compare with the inspection history.	[14,93,110,120,123]
Cracks, spalling, and joint anomalies	Local appearance, boundary, opening, and displacement	Obtain local detail and plan component-specific evaluation or repair.	[25,26,29,58,123,125]
Change between inspections	Location, direction, and development of observed changes	Adjust inspection attention and maintenance priorities using the time series.	[22,109,110,120,121,122,123,125]

Note: The table organizes possible uses of inspection information. Maintenance actions are selected using the applicable project criteria and bridge assessment process.

Scour models support interpretation of the environmental and foundation factors associated with an observed riverbed. Explainable machine learning has incorporated current, sediment, foundation geometry, and flood resilience [126], while probabilistic models describe variability in local scour [127]. Inspection data provide riverbed profiles and foundation exposure for comparison with these model outputs.

Digital twins organize component geometry, inspection findings, and maintenance records across the asset life cycle [17,128]. Common coordinates associate raw observations, calibration information, processing versions, point-cloud products, and damage descriptions with the underwater component model [129,130]. This organization supports repeated-survey comparisons and the retrieval of evidence for maintenance decisions.

6.4. Repeated Surveys and Temporal Change Detection

Repeated observations describe the development of scour, surface loss, and deformation. Scour evolution can be examined through direct differences between registered point clouds or through changes in morphological features [16]. Common reference bases, acquisition records, and processing settings support comparisons across surveys, while survey coverage records distinguish newly observed surfaces from previously inspected areas.

Repeated point clouds are aligned through coarse registration and fine ICP registration [90,91]. Multiscale model-to-model cloud comparison (M3C2) estimates changes along surface normals and provides confidence intervals [110]. Density, local roughness, and registration affect this comparison. Continuous fixed observations describe short-period scour processes [109], whereas periodic vessel or USV surveys describe changes across larger areas. Stable components and control points link both types of records.

7. Conclusions and Future Directions

This review organizes underwater bridge inspection around operational approaches, inspection equipment, data analysis, and condition assessment. Platform accessibility and sensor viewing geometry determine how components are observed. Acoustic, optical, laser, and contact methods provide complementary information about the riverbed, component shape, and local surface condition. Image and point-cloud processing connect those observations with target identification, geometric description, and temporal comparison. The resulting synthesis explains the technical roles of inspection methods across the acquisition-to-assessment process.

Coordinated operations connect vessel or USV area surveys, near-field ROV inspection, local diver examination, and fixed monitoring. Shared coordinates, timestamps, and component identifiers provide a basis for this coordination. Future development of autonomous navigation and cooperative positioning can improve route planning and repeated access to the same structural regions.

Multisensor inspection combines acoustic coverage with optical, laser, and tactile surface detail. Joint calibration and the alignment of observations from different viewpoints are central topics for further technical development. The complementary roles of these methods are especially relevant where a single viewing geometry leaves part of a component unobserved.

Data analysis is developing toward models that account for underwater imaging characteristics, variable point density, limited annotations, and component geometry. Method development can connect denoising, registration, segmentation, and geometric interpretation more closely, so that processing preserves the features relevant to the inspection task. Repeated-survey datasets and comparisons across operating settings are useful directions for this work.

Condition assessment benefits from linking component-level observations with inspection history, structural information, and maintenance records. Future work can develop consistent descriptions of defects and temporal changes, and examine how these records inform inspection intervals and intervention priorities. Integration with digital twins offers a practical route for managing this information throughout bridge operation.