On the 3D Reconstruction of Coastal Structures by Unmanned Aerial Systems with Onboard Global Navigation Satellite System and Real-Time Kinematics and Terrestrial Laser Scanning

  Citation: Gonçalves, D.; Gonçalves, G.; Pérez-Alvávez, J.A.; Andriolo, U. On the 3D Reconstruction of Coastal Structures by Unmanned Aerial Systems with Onboard Global Navigation Satellite System and Real-Time Kinematics and Terrestrial Laser Scanning. Remote Sens. 2022,14, 1485. https://doi.org/10.3390/ rs14061485 Academic Editors: JoséJuan de SanjoséBlasco, Germán Flor-Blanco and Ramón Blanco Chao Received: 15 February 2022 Accepted: 17 March 2022 Published: 19 March 2022 Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations. Copyright: © 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/). remote sensing Technical Note On the 3D Reconstruction of Coastal Structures by Unmanned Aerial Systems with Onboard Global Navigation Satellite System and Real-Time Kinematics and Terrestrial Laser Scanning Diogo Gonçalves 1,2,* , Gil Gonçalves 1,3 , Juan Antonio Pérez-Alvávez 4and Umberto Andriolo 1 1INESC Coimbra, Department of Electrical and Computer Engineering, Polo 2, 3030-290 Coimbra, Portugal; [email protected] (G.G.); [email protected] (U.A.) 2Department of Civil Engineering, University of Coimbra, 3030-788 Coimbra, Portugal 3Department of Mathematics, University of Coimbra, 3001-501 Coimbra, Portugal 4Departamento de Expresión Gráfica, Centro Universitario de Mérida, Universidad de Extremadura, 06800 Mérida (Badajoz), Spain; [email protected] *Correspondence: [email protected] Abstract: A wide variety of hard structures protect coastal activities and communities from the action of tides and waves worldwide. It is fundamental to monitor the integrity of coastal structures, as interventions and repairs may be needed in case of damages. This work compares the effectiveness of an Unmanned Aerial System (UAS) and a Terrestrial Laser Scanner (TLS) to reproduce the 3D geometry of a rocky groin. The Structure-from-Motion (SfM) photogrammetry technique applied on drone images generated a 3D point cloud and a Digital Surface Model (DSM) without data gaps. Even though the TLS returned a 3D point cloud four times denser than the drone one, the TLS returned a DSM which was not representing about 16% of the groin (data gaps). This was due to the occlusions encountered by the low-lying scans determined by the displaced rocks composing the groin. Given also that the survey by UAS was about eight time faster than the TLS, the SFM-MV applied on UAS images was the most suitable technique to reconstruct the rocky groin. The UAS remote sensing technique can be considered a valid alternative to monitor all types of coastal structures, to improve the inspection of likely damages, and to support coastal structure management. Keywords: drone; groin; breakwater; structure from motion; 3D point cloud 1. Introduction In the coastal environment, a wide variety of hard structures protect coastal activities and communities from the action of tides and waves worldwide [ 1 – 3 ]. Structure types have different configurations and constitution. Seawalls are built parallel to the shore to defend the inland areas and prevent shoreline retreatment [ 3 , 4 ]. Detached breakwaters are built shore parallel to dissipate wave energy in the nearshore [ 3 ], while jetties and shore-connected breakwaters play the function of protecting harbours and creating a secure environment for mooring, operating, and handling ships [ 3 , 5 ]. Finally, groins are built shore-perpendicular to mitigate shoreline erosion and intersect the updrift sediments [ 1 , 6 ]. Even though different materials can be used to build coastal structures, most of them are constituted by rocks and/or concrete [1,7]. It is fundamental to monitor the integrity of coastal structures and to detect likely damaged areas, to support management in the proposal of possible intervention and repairs. For instance, the displacement of stone blocks on rocky groins, rubble mound shore-connected breakwaters and jetties may endanger the stability of the structures, lowering and even compromising their protection functions [ 1 , 8 , 9 ]. The traditional monitoring methods of rocky coastal structures are based on visual or photographic inspections, however these Remote Sens. 2022,14, 1485. https://doi.org/10.3390/rs14061485 https://www.mdpi.com/journal/remotesensing Remote Sens. 2022,14, 1485 2 of 15 needs to be performed by qualified and trained personnel, and are technically, logistically and temporally limited [ 10 – 12 ]. Topographic surveys by totals stations and Global Navigation Satellite System (GNSS) have also been used, nevertheless, despite their advantages in terms of spatial accuracy, these methods require intense human effort in the field, and often do not provide a complete survey of the structure [ 12 ]. Recently, remote sensing techniques have been applied to monitor coastal structures. Aerial and terrestrial Light Detection and Ranging (LiDAR) can provide a detailed 3D reconstruction of rocky coastal protections [ 13 ], however the expensive costs and logistical constraints of LiDAR deployment remain the biggest disadvantages of this method. A valuable alternative for monitoring coastal structures is the use Unmanned Aerial System (UAS). UAS can operate autonomously, provide high spatial and temporal resolution with relatively low-cost, and are viable tools for many operational tasks [ 14 , 15 ]. Given their adaptable and multipurpose properties, UASs have improved the coastal environmental monitoring to advance knowledge on beach–dune morphodynamics [ 16 – 23 ], coastal cliffs [ 24 – 26 ], and marine pollution [ 27 – 34 ], among others. The 3D reconstruction workflow consists in applying the Structure-from-Motion (SfM) photogrammetry techniques to UAS image bulk, to obtain a 3D dense point cloud representing the targeted area [ 35 ]. The use of Ground Control Points (GCPs) is needed to georeference the 3D dense point cloud [ 36 ], however a Real-Time Kinematic assisted UAS (e.g., DJI Phantom 4 RTK; UAS-RTK) can accurately estimate the camera position, thus the use of GCPs can be potentially avoided [37]. The 3D reconstruction of coastal structures with SfM photogrammetry may improve the knowledge of structural condition of armour layer for assessing the structure integrity over a time interval or after extreme events [ 1 ]. According to the Construction Industry Research and Information Association (CIRIA), there are four levels of measures to assess the condition of armour units [ 1 ]: (i) locate units movements by measure with GNSS survey; (ii) geometric survey to describe armour layer with similar technique of level I; (iii) armour units position and areas where core and underlayer are exposed using of photographic methods (photogrammetry, comparative photography); (iv) shape and size of armour units including armour fractures. Even though different parameters can characterize the geometry of armour layer, the most important ones are the packing density (number of units per area) and the armour layer porosity (proportion of void per volume) [ 1 ]. These two parameters have a close relationship and mainly affect the performance of the coastal structure [38]. Previous works on the 3D reconstruction of coastal structures with UAS-based imagery are limited to few examples, which focused on the accuracy evaluation of SfM photogrammetry product [ 39 – 41 ]. The positional accuracy of Digital Surface Model (DSM) in respect to GNSS receiver was assessed by Henriques et al. [ 39 ], who obtained 10 cm and 8 cm of horizontal and vertical accuracies, respectively. González-Jorge et al. [ 40 ] tested instead the performance of 3D point cloud for detecting displacements of armour layer units. Overall, these pioneering and preliminary works showed that UAS surveys can improve the inspection of coastal structures. However, this technique still needs to be better evaluated. The aim of this paper is to assess the effectiveness of UAS-RTK and TLS techniques to reproduce the 3D geometry of a rocky groin. Two surveys were conducted at Leirosa beach on the North Atlantic Portuguese coast. The L-shape groin was digitally reconstructed by 3D point clouds and DSM. We evaluated the reconstruction of a L-shaped rocky groin by SfM photogrammetry and TLS scans in terms of mean surface density and data gaps. Overall, this work evaluates the most suitable technique to monitor the armour layer of coastal structures, such as groins and breakwaters, to improve the inspection of likely damages and support coastal management. Remote Sens. 2022,14, 1485 3 of 15 2. Methods 2.1. Study Site The study site was the Leirosa beach, on the wave-dominated high energetic North Atlantic Portuguese coast (Figure 1a). In this coastal sector, the mesotidal regime has an average amplitude of 2.10 m, reaching a maximum elevation of 4 m during spring tides [ 42 ]. The dominant wave regime comes from NW with average significant wave height (Hs) of 2 m and period from 7 s to 15 s [ 43 ]. Intense erosion occurred in the last decades, mainly caused by the littoral drift retention at the Mondego estuary jetty, and by the decrease in sediment deposition from the Mondego river [ 44 , 45 ]. An average shoreline retreat of 2 m/year was registered by comparing satellite images from 1958 and 2010 [ 46 ], while the dune crest retracted of about 2 m southern the groin over 2018/2019 winter season [ 16 ]. An L-shaped rocky groin was built to protect the urban coastal agglomeration of Leirosa. The shore-perpendicular sector of the groin extends about 200 m, with an average elevation of 6 m above the mean sea level (MSL), while the shore-parallel sector is shorter (150 m) and slightly higher (7 m above MSL). Both sectors are composed by stone blocks of about 2.5 m with a shape factor of 60–80% and a mass of about 15 tonnes. Since the main longshore transport on this coast is oriented N-S [ 42 , 44 , 45 ], the groin intersects sediment updrift, determining accumulation and shoreline advance northern the structure, in front of the urbanized area of Leirosa (Figure 1b). The mean slope on the shore-perpendicular sector is 21 ◦ (1:2.7 m) while on the shoreparallel wall is 30 ◦ (1:1.75 m—Figure 1j). An initial, cursory visual inspection revealed significant displacements of stone blocks at the two heads of the groin, which can determine further dislocation and the loss of the original shape and function of the structure (Figure 1c–f). Remote Sens. 2022, 14, x FOR PEER REVIEW 3 of 16 2. Methods 2.1. Study Site The study site was the Leirosa beach, on the wave-dominated high energetic North Atlantic Portuguese coast (Figure 1a). In this coastal sector, the mesotidal regime has an average amplitude of 2.10 m, reaching a maximum elevation of 4 m during spring tides [42]. The dominant wave regime comes from NW with average significant wave height (Hs) of 2 m and period from 7 s to 15 s [43]. Intense erosion occurred in the last decades, mainly caused by the littoral drift retention at the Mondego estuary jetty, and by the decrease in sediment deposition from the Mondego river [44,45]. An average shoreline retreat of 2 m/year was registered by comparing satellite images from 1958 and 2010 [46], while the dune crest retracted of about 2 m southern the groin over 2018/2019 winter season [16]. An L-shaped rocky groin was built to protect the urban coastal agglomeration of Leirosa. The shore-perpendicular sector of the groin extends about 200 m, with an average elevation of 6 m above the mean sea level (MSL), while the shore-parallel sector is shorter (150 m) and slightly higher (7 m above MSL). Both sectors are composed by stone blocks of about 2.5 m with a shape factor of 60–80% and a mass of about 15 tonnes. Since the main longshore transport on this coast is oriented N-S [42,44,45], the groin intersects sediment updrift, determining accumulation and shoreline advance northern the structure, in front of the urbanized area of Leirosa (Figure 1b). The mean slope on the shore-perpendicular sector is 21° (1:2.7 m) while on the shoreparallel wall is 30° (1:1.75 m—Figure 1j). An initial, cursory visual inspection revealed significant displacements of stone blocks at the two heads of the groin, which can determine further dislocation and the loss of the original shape and function of the structure (Figure 1c–f). Figure 1. Study site location and morphological characterization. (a) Location of the study site and the Mondego estuary in Portugal; (b) Google satellite image of the groin in the local coordinate system (ETRS89/PT-TM06). Red and blue tags indicate the two heads of the groin shown in (c,d). Green lines indicate the two profiles shown in (j). (c,d) Pictures taken by DJI Phantom 4RTK Figure 1. Study site location and morphological characterization. ( a ) Location of the study site and the Mondego estuary in Portugal; ( b ) Google satellite image of the groin in the local coordinate system (ETRS89/PT-TM06). Red and blue tags indicate the two heads of the groin shown in ( c , d ). Green lines indicate the two profiles shown in ( j ). ( c , d ) Pictures taken by DJI Phantom 4RTK showing Remote Sens. 2022,14, 1485 4 of 15 the two groin heads (namely, sub-area 1 and 3, see Section 2.4) ( e , f ) details of the two groin heads shown in ( c , d ), respectively; ( g – i ) equipment used in the fieldwork campaign, namely the Terrestrial Laser Scanning (TLS) Leica ScanStation C10 ( g ), the Unmanned Aerial System (UAS) DJI Phantom 4RTK ( h ) and the GeoMax Zenith 10 GNSS receiver ( i ); ( j ) two cross sections of the studied groin located in (b). Dark green triangles illustrate the slope of the berms. A field work campaign was conducted with an Unmanned Aerial System (UAS) and a Terrestrial Laser Scanning (TLS) on 5 of November 2021 (Figure 1g–i). A dual frequency GNSS receiver was also used in Network Real Time Kinematic (NRTK) using the Portuguese stations ReNEP (https://renep.dgterritorio.gov.pt, accessed on 6 January 2022).to collect the coordinates of: (1) 13 control points to assess the Structure-from-Motion results (Section 2.2); and (2) 12 targets to georeference and co-register the TLS scans (Section 2.3). 2.2. Unmanned Aerial System Data Acquisition and Processing The UAS survey was carried out with a DJI Phantom 4 RTK (P4RTK), equipped with a RGB camera (1 00 CMOS sensor with 20 M effective pixels, 8.8 mm focal length and image size of 5472 × 3648 pixels). The setup of drone flight and cameras parameters was planned in DJI GST RTK, choosing a linear flight path (Figure 2a). The flight altitude was set to 50 m, considering an image overlap of 80% (both lateral and frontal). A total of 103 nadiral images was collected with a mean Ground Sample Distance (GSD) of 1.36 cm/pixel which is computed considering the flight altitude, sensor dimensions and focal length. The flight lasted about 10 min. Although the P4RTK can provide accurate georeferenced data [ 16 , 47 ], we drew 13 crosses on the blocks faces, evenly distributed along the groin, to evaluate the image positional accuracy in the post-processing phase. The measurements/collection of these points with the GNSS NRTK lasted about one hour. Remote Sens. 2022, 14, x FOR PEER REVIEW 4 of 16 showing the two groin heads (namely, sub-area 1 and 3, see Section 2.4) (e,f) details of the two groin heads shown in (c) and (d), respectively; (g–i) equipment used in the fieldwork campaign, namely the Terrestrial Laser Scanning (TLS) Leica ScanStation C10 (g), the Unmanned Aerial System (UAS) DJI Phantom 4RTK (h) and the GeoMax Zenith 10 GNSS receiver (i); (j) two cross sections of the studied groin located in (b). Dark green triangles illustrate the slope of the berms. A field work campaign was conducted with an Unmanned Aerial System (UAS) and a Terrestrial Laser Scanning (TLS) on 5 of November 2021 (Figure 1g–i). A dual frequency GNSS receiver was also used in Network Real Time Kinematic (NRTK) using the Portuguese stations ReNEP (https://renep.dgterritorio.gov.pt, accessed on 6 January 2022).to collect the coordinates of: (1) 13 control points to assess the Structure-from-Motion results (Section 2.2); and (2) 12 targets to georeference and co-register the TLS scans (Section 2.3). 2.2. Unmanned Aerial System Data Acquisition and Processing The UAS survey was carried out with a DJI Phantom 4 RTK (P4RTK), equipped with a RGB camera (1″ CMOS sensor with 20M effective pixels, 8.8 mm focal length and image size of 5472 × 3648 pixels). The setup of drone flight and cameras parameters was planned in DJI GST RTK, choosing a linear flight path (Figure 2a). The flight altitude was set to 50 m, considering an image overlap of 80% (both lateral and frontal). A total of 103 nadiral images was collected with a mean Ground Sample Distance (GSD) of 1.36 cm/pixel which is computed considering the flight altitude, sensor dimensions and focal length. The flight lasted about 10 min. Although the P4RTK can provide accurate georeferenced data [16,47], we drew 13 crosses on the blocks faces, evenly distributed along the groin, to evaluate the image positional accuracy in the post-processing phase. The measurements/collection of these points with the GNSS NRTK lasted about one hour. Figure 2. Unmanned Aerial System (UAS) data acquisition and processing. (a) field data acquisition represented on UAS orthophoto. Red arrows show the UAS flight path, white diamonds represent the control points placement for georeferencing process; (b) UAS data processing workflow: through structure from motion (SfM) and multi-view stereo technique (MVS), images and control points were processed for generating a 3D point cloud and a Digital Surface Model (DSM). Figure 2. Unmanned Aerial System (UAS) data acquisition and processing. ( a ) field data acquisition represented on UAS orthophoto. Red arrows show the UAS flight path, white diamonds represent the control points placement for georeferencing process; ( b ) UAS data processing workflow: through structure from motion (SfM) and multi-view stereo technique (MVS), images and control points were processed for generating a 3D point cloud and a Digital Surface Model (DSM). Remote Sens. 2022,14, 1485 5 of 15 The UAS imagery post-processing phase was performed with Agisoft Metashape software. The application of SfM photogrammetry followed the workflow shown in Figure 2b. Firstly, images were oriented calculating the external and internal camera parameters through bundle block adjustment. A set of 3D points (tie points) was obtained representing the scale invariant features identified in overlapping images. Secondly, since the refinement of tie points improves the accuracy of reconstruction [ 26 , 48 ], we removed points with a reprojection error greater than 0.5. Afterwards, the bundle block adjustment was repeated using the refined tie points. Thirdly, the georeferencing process was performed by selecting only 3 ground control points from the set of 13 surveys points. The remaining 10 points were used as check points. All 13 control points were manually picked in the corresponding images. Fourthly, 3D point cloud was generated through dense matching by computer vision algorithms based on multi-stereoscopy (Figure 2b). Finally, after gridding the 3D point cloud with regular cells (i.e., squares 5 × 5 cm), we generated the DSM in CloudCompare. Each cell grid was calculated by the mean height of all points within the cell. The empty cells were linearly interpolated using the nearest cells. 2.3. Terrestrial Laser Scanning Data Acquisition and Processing The TLS survey was performed with a Leica ScanStation C10, scan range of 300 m. The positional accuracy of a single TLS station is 6 mm for a range of 50 m. A total of seven scans were distributed along Leirosa groin, with most of stations chosen on the two groin heads (Figure 3a). A series of 12 circular targets were distributed in the TLS scan ranges of different station, to stich the scan data in the post-processing phase. Targets were placed among the terrain in order to be visible at least from two consecutive stations. Before starting the scanning procedure, targets were visually picked in the TLS screen, and measured with the GNSS NRTK receiver. When moving to the following station, we detected the same previous targets for building the scan network. During the TLS survey, the tide rose influencing the candidate zones for the scans position. In the end, the TLS survey lasted about 8 h. Remote Sens. 2022, 14, x FOR PEER REVIEW 6 of 16 Figure 3. Terrestrial Laser Scanning (TLS) data acquisition and processing. (a) Field data acquisition on UAS orthophoto. Green triangles show the different stations in which the TLS was placed for scanning (from 1 to 7), blue-white circles represent the targets (from 1 to 12) placed for stitching the scans acquired from different stations. (b) TLS data processing workflow: different scans are coregistered and joined using targets, for finally generating a 3D point cloud and a Digital Surface Model (DSM). The mean surface density aims at characterizing the mean points concentration over the surface, identifying zones without points (data gaps). The 3D point clouds were projected along Z direction and converted into a regular grid. The mean surface density (𝜌) was estimated by counting the points inside each grid cell and then calculating the mean of the overall cells density: ρ = 1 N∑ci N i=1 (1) where 𝑁 is the number of grid cells in the region of interest (𝐶) and 𝑐𝑖 is the point counting or local density in grid cell 𝑖. The data gaps indicate the numbers of empty grid cells in the region of interest using 𝛿𝑔= #𝐷𝑔×𝐴𝑐 (2) where #𝐷𝑔 is the cardinality of set 𝐷𝑔={𝑐𝑖:𝑐𝑖=0, for 𝑖 =1,…,𝑁} formed by the empty grid cells and 𝐴𝑐 is the grid area. 2.5. Comparative Analysis and Statistical Measures To assess the accuracy and effectiveness of UAS-RTK and TLS products for mapping groins, we implemented a two-stage methodology (Figure 4): (i) 3D approach for assessing the positional accuracy of 3D point cloud, and (ii) 2.5 D approach for assessing the DSM vertical accuracy. Figure 3. Terrestrial Laser Scanning (TLS) data acquisition and processing. ( a ) Field data acquisition on UAS orthophoto. Green triangles show the different stations in which the TLS was placed for Remote Sens. 2022,14, 1485 6 of 15 scanning (from 1 to 7), blue-white circles represent the targets (from 1 to 12) placed for stitching the scans acquired from different stations. ( b ) TLS data processing workflow: different scans are co-registered and joined using targets, for finally generating a 3D point cloud and a Digital Surface Model (DSM). The resolution of scans was divided into the following categories: (i) medium resolution, corresponding to 10 cm distance between points at 100 m; (ii) high resolution, corresponding to 5 cm distance between points at 100 m. The TLS post-processing methodology was performed with Leica Cyclone software (Figure 3b). The main steps consisted in: (i) importing the scans; (ii) reading targets coordinates; (iii) co-registration of the scans. The scans were stitched using the correspondent picked targets to produce a single 3D point cloud. The georeferencing process was also included, projecting the 3D point cloud into the Portuguese reference system. Using the same grid adopted to produce the UAS DSM (Section 2.2), we also generated the DSM from TLS 3D point cloud. 2.4. Three-Dimensional Point Clouds and Digital Surface Model We evaluated and compared the 3D point clouds of both UAS-RTK and TLS using two quality parameters, namely the mean surface density and data gaps. We computed the two parameters for the entire groin structure, and for three sub-areas. Sub-areas were chosen on the groin heads and in the central groin area (Figure 4). The head on seaside is the more affected zone by waves, and already shows damages and block displacements. The south head shows also block displacements and covers the dune ridge, which is under severe erosion [ 16 ]. Finally, the central area represents a flat wall for protecting the urban area (see also Figure 1b–f). Remote Sens. 2022, 14, x FOR PEER REVIEW 7 of 16 For the 3D approach, the distance between points of TLS and UAS was calculated. For this purpose, in CloudCompare software, two algorithms were adopted: (i) Cloud-toCloud (C2C) distance and (ii) Multiscale Model-to-Model Cloud Comparison (M3C2). The C2C computes directly the distance between two 3D point clouds based a local model fitting. This local model is calculated for each point of the reference data (TLS 3D point cloud) and is composed by: (1) the k nearest points or (2) the points inside a sphere with radius r of the target data (UAS 3D point cloud). The M3C2 measure computes the distances between two 3D point clouds through the normal direction of a local surface. This local surface is calculated by searching for points in a radius D and then a cylinder is computed along the normal direction that intersects the target data [49]. Therefore, the error of UAS point cloud (Δ𝐷) was obtained by: Δ𝐷 =[Δ𝑑1⋯ Δ𝑑𝑁] (3) where Δ𝑑𝑖 is the distance between the point 𝑖 of UAS and the correspondent in TLS point cloud, and 𝑁 is number of UAS 3D point cloud. . Hereinafter, when Δ𝐷 is mentioned, it includes distances calculated by C2C and M3C2. For the 2.5D approach, a direct vertical difference was calculated using DSM of Difference (DoD). The overall errors were obtained as: Δ𝑍 =[Δ𝑧1⋯ Δ𝑧𝑀] (4) where Δ𝑧𝑗=𝑧𝑗𝑇𝐿𝑆 −𝑧𝑗𝑈𝐴𝑆 is the vertical difference of j-th DoD cell. Finally, the statistical analysis focused at characterizing the errors provided by the 3D and 2.5D approaches (Figure 4). We firstly fitted the histograms to characterize the distribution of errors. Then, a robust statistical analysis suitable for non-normal distributions was adopted [50], in order to compare the results with the traditional statistical measures. For traditional measures, the mean, the Standard Deviation (Std) and the Root Mean Square Error (RMSE) were calculated. For robust measures of non-normal error distributions, the median and Normalized Median Absolute Deviation (NMAD) were computed. In addition, the robust RMSE (RRMSE) was also computed based on mean and standard deviation [51]. Figure 4. Data processing workflow for the comparative analysis. Three sub-areas (left box, red polygons) were defined over the whole groin structure (left box, green polygon). The distance between 3D point clouds and Digital Surface Model were computed using 3D and 2.5D approach based on point-to-point distance and vertical differences (central boxes). Finally, the errors were assessed through statistical measures computation (right box). Figure 4. Data processing workflow for the comparative analysis. Three sub-areas ( left box, red polygons) were defined over the whole groin structure ( left box, green polygon). The distance between 3D point clouds and Digital Surface Model were computed using 3D and 2.5D approach based on point-to-point distance and vertical differences ( central boxes). Finally, the errors were assessed through statistical measures computation (right box). The mean surface density aims at characterizing the mean points concentration over the surface, identifying zones without points (data gaps). The 3D point clouds were projected along Z direction and converted into a regular grid. The mean surface density ( ρ ) Remote Sens. 2022,14, 1485 7 of 15 was estimated by counting the points inside each grid cell and then calculating the mean of the overall cells density: ρ=1 N N ∑ i=1 ci(1) where N is the number of grid cells in the region of interest ( C ) and ci is the point counting or local density in grid cell i. The data gaps indicate the numbers of empty grid cells in the region of interest using δg=#Dg×Ac(2) where # Dg is the cardinality of set Dg={ci:ci=0, for i=1, . . . , N} formed by the empty grid cells and Acis the grid area. 2.5. Comparative Analysis and Statistical Measures To assess the accuracy and effectiveness of UAS-RTK and TLS products for mapping groins, we implemented a two-stage methodology (Figure 4): (i) 3D approach for assessing the positional accuracy of 3D point cloud, and (ii) 2.5 D approach for assessing the DSM vertical accuracy. For the 3D approach, the distance between points of TLS and UAS was calculated. For this purpose, in CloudCompare software, two algorithms were adopted: (i) Cloud-to-Cloud (C2C) distance and (ii) Multiscale Model-to-Model Cloud Comparison (M3C2). The C2C computes directly the distance between two 3D point clouds based a local model fitting. This local model is calculated for each point of the reference data (TLS 3D point cloud) and is composed by: (1) the k nearest points or (2) the points inside a sphere with radius r of the target data (UAS 3D point cloud). The M3C2 measure computes the distances between two 3D point clouds through the normal direction of a local surface. This local surface is calculated by searching for points in a radius D and then a cylinder is computed along the normal direction that intersects the target data [ 49 ]. Therefore, the error of UAS point cloud (∆D)was obtained by: ∆D=∆d1· · · ∆dN(3) where ∆di is the distance between the point i of UAS and the correspondent in TLS point cloud, and N is number of UAS 3D point cloud. Hereinafter, when ∆D is mentioned, it includes distances calculated by C2C and M3C2. For the 2.5D approach, a direct vertical difference was calculated using DSM of Difference (DoD). The overall errors were obtained as: ∆Z=∆z1· · · ∆zM(4) where ∆zj=zTLS j−zUAS jis the vertical difference of j-th DoD cell. Finally, the statistical analysis focused at characterizing the errors provided by the 3D and 2.5D approaches (Figure 4). We firstly fitted the histograms to characterize the distribution of errors. Then, a robust statistical analysis suitable for non-normal distributions was adopted [ 50 ], in order to compare the results with the traditional statistical measures. For traditional measures, the mean, the Standard Deviation (Std) and the Root Mean Square Error (RMSE) were calculated. For robust measures of non-normal error distributions, the median and Normalized Median Absolute Deviation (NMAD) were computed. In addition, the robust RMSE (RRMSE) was also computed based on mean and standard deviation [ 51 ]. 3. Results 3.1. 3D Point Clouds and Digital Surface Model The Agisoft Metashape processing workflow lasted about 1.5 h and returned a 3D point cloud with a georeferencing RMSE on check points equal to 3.2 cm. The 3D point cloud represented the entire groin with about 11 million points (Table 1), with a mean Remote Sens. 2022,14, 1485 8 of 15 surface density of 1.7 × 10 3 points/m 2 and without data gaps. After using the 12 targets to link the TLS scans in Cyclone that lasted about 1 h, the RMSE of TLS stitching was of 1.5 cm. The final 3D point cloud of TLS covered the groin with about 41 million points (Table 1). Even though the number of points and mean surface density of TLS point cloud was about three times the UAS one, the TLS data gap was about 16% of the entire groin area (6630 m2). Table 1. Characterization of the 3D point clouds obtained by the Unmanned Aerial System and Terrestrial Laser Scanning data. Number of points, mean surface density and data gaps are indicated for the whole groin structure and the three subareas (refer to Figure 4for sub-areas identification). Parameters Unmanned Aerial System Terrestrial Laser Scanner Whole structure (6630 m2) Number of points 11.3 ×10641.3 ×106 Mean surface density (points/m2) 1.7 ×1036.2 ×103 Data gaps (%) 0 16.8 Sub-area 1 (560 m2) Number of points 1.0 ×10615.2 ×106 Mean surface density (points/m2) 1.8 ×10327.2 ×103 Data gaps (%) 0 1.4 Sub-area 2 (530 m2) Number of points 0.9 ×1060.7 ×106 Mean surface density (points/m2) 1.6 ×1031.3 ×103 Data gaps (%) 0 0 Sub-area 3 (600 m2) Number of points 1.2 ×10621.8 ×106 Mean surface density (points/m2) 2.0 ×10336.4 ×103 Data gaps (%) 0 0.8 Overall, the number of points and mean surface density among the three sub-areas were much variable on TLS data, while did not change significantly for UAS. In fact, the mean surface density of TLS was higher on sub-areas 1 and 3, due to the more numerous scans and stations in these zones. Finally, for the sub-areas 1–3, which are the most interesting areas to be monitored on the groin structure, TLS survey returned an average 1.1% of surface data gaps. On the other hand, despite the lower means surface density, no data gaps were found on UAS survey. Figure 5depicts the 3D point clouds and Digital Surface Models (DSM) obtained in the data processing step. Although the data gaps were present on TLS 3D points, the interpolation of the empty cells (5 × 5 cm) allowed to mitigate this effect on DSM. 3.2. Comparative Analysis and Statistical Measures of 3D Point Clouds and DSM For analysing the spatial distribution of errors, Figure 6shows the results of 3D and 2.5D approaches for the sub-areas. Overall, more than 50% of errors were concentrated between 0 and 10 cm. The accumulated green bars presented higher values in sub areas 1 and 3. In fact, the presence of data gaps was proportional to the increase of errors. The roughness of these sub-areas may have penalized the errors of M3C2 algorithm where it is not possible to compute the distance in several points due to wrong calculation of the normal direction of local surfaces (see Section 2.4 and Figure 4). The errors of sub-area 2 registered lower variability among the different algorithms. These results were due to the regular roughness of the groin surface. Overall, the non-normality of the errors distribution was corroborated with a course analysis of the histogram shapes. The main reason for this assumption is the right heavy tails of the histograms. Remote Sens. 2022,14, 1485 9 of 15 Remote Sens. 2022, 14, x FOR PEER REVIEW 9 of 16 Figure 5. Products obtained by Unmanned Aerial System (a) and Terrestrial Laser Scanning (b). In details, 3D point clouds (upper row) and Digital Surface Models (DSM, lower row). Red numbered polygons represent the three chosen sub-areas (1–3, please refer to Section 2.4 for explanation). The axis labels refer to the projected Portuguese reference systems (ETRS89/PT-TM06). For an in-depth assessment, the accuracy measures were tabulated highlighting the main differences between 3D and 2.5D approaches in sub-areas (Table 2). In general, the conventional measures overestimated the errors between UAS and TLS in comparison with the robust measures (in the worst case, differences were about 14 cm). The sub-area 2 presented the lower variability both in terms of Std and NMAD, while sub-area 1 showed high dispersion in relation to the mean value. Overall, the RMSE registered values in a range of 30 cm due to the variability of errors expressed in term of Std. In fact, the formulation of RMSE considers the square of errors that magnifies the largest values mainly observed in sub-areas 1 and 3. However, it can be observed that the RRMSE presented a lower variability concentrating values in a range of 10 cm. Although the Figure 5. Products obtained by Unmanned Aerial System ( a ) and Terrestrial Laser Scanning ( b ). In details, 3D point clouds (upper row) and Digital Surface Models (DSM, lower row). Red numbered polygons represent the three chosen sub-areas (1–3, please refer to Section 2.4 for explanation). The axis labels refer to the projected Portuguese reference systems (ETRS89/PT-TM06). For an in-depth assessment, the accuracy measures were tabulated highlighting the main differences between 3D and 2.5D approaches in sub-areas (Table 2). In general, the conventional measures overestimated the errors between UAS and TLS in comparison with the robust measures (in the worst case, differences were about 14 cm). The subarea 2 presented the lower variability both in terms of Std and NMAD, while sub-area 1 showed high dispersion in relation to the mean value. Overall, the RMSE registered values in a range of 30 cm due to the variability of errors expressed in term of Std. In fact, the formulation of RMSE considers the square of errors that magnifies the largest values mainly observed in sub-areas 1 and 3. However, it can be observed that the RRMSE