COMBINING UNMANNED AERIAL SYSTEMS AND STRUCTURE FROM MOTION PHOTOGRAMMETRY TO RECONSTRUCT THE GEOMETRY OF GROINS

In the context of coastal environment, the combined use of Unmanned Aerial Systems (UAS) with Structure-from-Motion (SfM) photogrammetry has been demonstrated to be a suitable tool to collect accurate geoinformation. This work aims at investigating the effectiveness of UAS survey to accurately map the geometry of rouble mound groins for monitoring the structural integrity. A set of nadiral images was processed through SfM photogrammetry to obtain the Digital Surface Model (DSM) of the groin. The DSM accuracy was evaluated through the comparison of five profiles measured by the traditional GNSS technique. Overall, vertical accuracy returned a root mean square error of 2.6 cm. The results suggest that UAS combined with SfM photogrammetry can improve the current monitoring of coastal engineering structures by providing high accurate geospatial products.


INTRODUCTION
The coastal environment is the high dynamic interface between land and ocean. In the last decades, remote sensing techniques have been used for monitoring coastal environments (Andriolo et al., 2018;Laporte-Fauret et al., 2020;Mancini et al., 2013;Medjkane et al., 2018;Turner et al., 2016;Zimmerman et al., 2020) Among these, Unmanned Aerial Systems (UAS) allow to acquire images with high resolution to 3D reconstruct the geometry and to potentially increase the survey frequency of coastal features. UAS have been used for monitoring beach and dunes Pagán et al., 2019;Rotnicka et al., 2020), morphological change in cliffs (Gómez-Gutiérrez and Ružić et al., 2014), mapping marine litter Duarte et al., 2020) and vegetation classification (Laporte-Fauret et al., 2020). Among coastal engineering structures, groins are built shore perpendicular to maintain updrift beaches, or to restrict longshore sediment transport (CIRIA et al., 2007). The armour layer is the groin section that protects the structure from wave attacks and storms. It is usually composed of stone blocks and/or concrete (Bush et al., 2001). The structural condition of armour layer represents the assessment of structural integrity (CIRIA et al., 2007), and consequently the assessment of likely movements in armour layer units inducted by waves and/or storms. Therefore, preserving the integrity of the structure is fundamental to preserve the safety of coastal communities. In this context, the monitoring of groins should be repeatable over time. There are four levels of surveys to evaluate the condition of armour units (CIRIA et al., 2007): (1) locating armour units movements measuring with Global Navigation Satellite System (GNSS) surveys; (2) geometric survey to describe armour layer with similar techniques of level 1; (3) armour units position including voids and areas (photographic  Corresponding Author methods such as photogrammetry or comparative photography); (4) shape and size of armour units including armour fractures. Nowadays, the monitoring of groins and breakwaters is mainly done by visual inspection. Trained technicians walk through the structure to identify likely armour layer movements. This technique lacks accurate geospatial information and temporal resolution. In several cases, the likely damaged areas are almost impossible to reach due to its location in the wave breaking zone which compromises the safety of technicians (Henriques et al., 2017). The 3D reconstruction of surfaces from a UAS-based imagery survey can be a potential solution for monitoring groins and consists in applying the Structure from Motion and Multi-View Stereo (SfM-MVS) workflow. From a set of overlapping images, SfM-MVS workflow produces high details 3D point clouds that can be derived in Digital Surface Models (DSM) and/or orthophotos. Few works were devoted to the 3D reconstruction of coastal structures using UAS-based imagery surveys. Based on our knowledge, the relevant studies focused on evaluating the capability of UAS to provide accurate geoinformation of coastal structures geometry (Gonçalves et al., 2022;González-Jorge et al., 2016;Henriques et al., 2014). Henriques et al., (2014) assessed the horizontal and vertical accuracy of DSM obtaining 10 cm and 8 cm respectively from a set of check points measured with a GNSS receiver. The potentiality of UAS surveys was confirmed by identifying sub-degree sensibility to armour units rotation (González-Jorge et al., 2016). Recently, Gonçalves et al., (2022) proved the advantage of UAS surveys when compared with Terrestrial Laser Scanning surveys in terms of fieldwork time and data completeness. Among different environments, the vertical accuracy of DSM can be improved combining images with nadiral and oblique viewing angles (Manfreda et al., 2019). Broadly speaking, the accuracy of UAS products presents better horizontal than vertical accuracy (Rangel et al., 2018). The UAS with onboard Global Navigation Satellite System and Real Time Kinematics can avoid the use of GCPs by accurately estimate the images position (Taddia et al., 2019). However, the use of ground control points (GCP) in SfM-MVS workflow may improve the estimation of camera distortion parameters (Štroner et al., 2021). The damage in coastal structures is expressed by indicators that characterize armour layer geometry. The most important ones are armour layer porosity and packing density (CIRIA et al., 2007). However, the damage percentage expressed by the fraction between the number of displaced armour units and the total number of armour units can also be used to identify likely damaged areas (Musumeci et al., 2018). Another approach may consider profiles transects to estimate the erosion areas over time (Melby and Kobayashi, 1998). This work aims at validating the combination of UAS-RTK and SfM photogrammetry to reconstruct the geometry of rubble mound groins. From a UAS-based imagery network, the groin geometry was obtained through SfM-MVS and then, validated against the reference data collected with GNSS survey. We also discussed the contribution that UAS-RTK combined with SfM photogrammetry can provide to the traditional monitoring methods for assessing the structural integrity of armour layer.

STUDY SITE AND METHOD
This section describes the methodology to evaluate the geospatial information provided by an UAS survey on a rubble mound groin for armour layer monitoring. Firstly, we introduce the study site and characterize the surveys (section 2.1). Secondly, it is explained briefly the photogrammetric workflow in Agisoft Metashape to generate a 3D point cloud which is then derived in a DSM and an orthophoto (section 2.2). Finally, the accuracy assessment of vertical position is calculated through the linear regression adjustment and conventional root mean square error (section 2.3).

Study site and surveys
The study area is Leirosa Beach, located in the central zone of the Portuguese coastal zone, with NE-SW orientation (Figure 1). The groin has a L-shaped with an extension of 300 m and is composed of stone blocks of about 2.5 m and a mass of about 15 tonnes. The two groin heads show stone blocks displacements caused mainly by wave attacks. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLIII-B2-2022 XXIV ISPRS Congress (2022 edition), 6-11 June 2022, Nice, France On 5 th November 2021, a fieldwork campaign was conducted during low tide. The data was collected with a low-cost UAS DJI Phantom 4RTK (P4RTK) and a dual-frequency GNSS receiver Geomax Zenith 10. P4RTK is equipped with a 1'' CMOS sensor Sony FC6310 (20 Mpixels, a focal length of 8.8 mm and an image size of 5472×3648 pixels) and with onboard Global Navigation Satellite System and Real Time Kinematics. The UAS survey was planned in DJI GST RTK software defining a regular grid path. A set of 103 nadiral images was collected at an altitude of 50 m for ensuring a Ground Sample Distance (GSD) with centimeter order (estimated mean GSD of 1.56 cm/pixel). We selected a side and front overlap of 80% to guarantee suitable image overlap for applying the photogrammetric workflow (section 2.2). In the end, the UAS survey lasted about 10 min. From GNSS survey, it was measured a set of 54 points. 12 points were used as control points in the photogrammetric workflow (section 2.2). The remaining 42 points were used for a depth analysis of the vertical accuracy of DSM derived by the photogrammetric workflow (section 2.3).

Processing work
The images acquired with the P4RTK were processed using the SfM-MVS technique to obtain a 3D dense point cloud. The photogrammetric workflow was implemented in Agisoft Metashape and it is composed of four main steps : (1) image alignment calculating the image external orientation parameters and the camera internal orientation parameter (IOP) through bundle adjustment; (2) tie points (correspondent point in zones of image overlap) refinement; (3) re-do bundle adjustment using the refined tie points; (4) dense matching applying MVS algorithms. Between steps 3 and 4, the control points were manually marked in images selecting 3 as GCP and 9 as check points (CHP) to evaluate the georeferencing of the resulting 3D dense point cloud (Figure 1c). Therefore, a Digital Surface Model (DSM) was generated from the 3D dense point cloud with a grid size of 5 cm choosing the interpolation of the empty grid cells. Finally, the image network was orthorectified to produce the orthophoto of the groin with a pixel size of 1.5 cm.

Accuracy assessment
The accuracy assessment aimed at evaluating the feasibility of UAS-based imagery to provide valid geospatial information of the groin surface. The points measured with the GNSS survey on the groin blocks were used to assess the vertical accuracy (ZGNSS).
Considering the planimetric position of profile points, the elevation of UAS (ZUAS) was interpolated on the DSM. The vertical deviation (dZ) was therefore computed as dZ = ZGNSS -ZUAS (1) Firstly, linear regression was used to adjust the function that approximated the set of points composed by ZGNSS (independent variable) and ZUAS (dependent variable). Linear regression estimates the parameters m and b of the affine function g(x)=mx+b (2) This adjustment can be assessed through the calculation of the determination coefficient (R). R can have values in a range of 0 (worst adjustment) and 1 (perfect adjustment). Secondly, the vertical accuracy was also computed using the root mean square error (RMSE) of the deviations computed by Equation 1.

RESULTS
The photogrammetric workflow generated a high 3D dense point cloud with about 40 million points. The 3D dense point cloud was manually cropped particularly in the wave breaking zone to remove mismatches of SfM-MVS. The DSM covered the entire groin and the erosion and sediment retention zones (Figure 2a).
In georeferencing process, we used 3 GCP resulting in a RMSE of 2.6 cm. The remaining 9 points were used as CHP resulting in a RMSE of 3.4 cm.
To further test, the set of ZGNSS was divided into 1) profile points (P -34 points) and 2) sample points along the west part of groin (CHPZ -8 points; Figure 2b). All profiles measured by GNSS survey were used to evaluate the vertical accuracy on the groin surface ( Figure 2c). The profiles on DSM had a resolution of 50 cm contributing more details than the GNSS survey. Figure 3 shows a quantitative comparison between the vertical accuracy of ZGNSS and ZUAS. The histogram was plotted using a bin spacing of 1 cm. 50% of vertical deviation varied from -3.2 cm and -1.2 cm (interquartile range). Overall, the linear regression model returned a high agreement between ZGNSS and ZUAS (determination coefficient equal to 0.998). Regarding the deviations computed in RMSE, the vertical accuracy was found equal to 2.6 cm in P and 2.5 cm in CHPZ. Finally, the weighted mean (based on point number) between P and CHPZ was found approximately equal to 2.6 cm proving the stability of vertical accuracy over the DSM.

DISCUSSION
The present work demonstrates the potential of combining a Real-Time Kinematic assisted Unmanned Aerial System (UAS-RTK) and Structure from Motion (SfM) photogrammetry to reconstruct the geometry of rubble mound groins. The UAS-RTK survey provided a simple procedure with less fieldwork demanding when compared with other more accurate remote sensing techniques such as Terrestrial Laser Scanning (Gonçalves et al., 2022). The simplified Metashape workflow provided an easy-to-use implementation of SfM photogrammetry processing workflow to generate an accurate 3D dense point cloud . Since the geometry of the image network was acquired only with nadiral images, the combination of nadiral and oblique images may improve mainly the vertical accuracy (Manfreda et al., 2019) and will be also beneficial to improve the density of the 3D dense point clouds (Rossi et al., 2017). Even though UAS-RTK does not require control points for the georeferencing process, we used 3 GCPs to ensure the correction of camera internal orientation parameters (Štroner et al., 2021). A limitation of this accuracy analysis was the lack of sample points (CHPZ) in the shore-parallel sector of the groin and in the likely damaged areas of the groin (mainly in heads - Figure  1). Broadly speaking, the existing techniques for collecting geospatial data on the physical state of coastal structures are based on visual inspections that require qualified technicians. The use of UAS-RTK can provide a valuable alternative to improve the traditional monitoring methods. The resulting DSM provided high vertical accuracy of the groin surface. In fact, from the interpolation of the profiles and sample points, the results of vertical accuracy analysis returned a low Root Mean Square Error (RMSE -2.6 cm). The use of UAS-RTK improved the vertical accuracy reported by Henriques et al. (2014) who obtained a RMSE of 8 cm. Overall, we showed that the DSM generated by the UAS-RTK survey can be used to implement monitoring methods for evaluating the structural condition of armour layer. For instance, it will possible to calculate the deformations of armour layer using cross-sections (Melby and Kobayashi, 1998). On the other hand, an in-depth investigation is required to calculate the indicators that characterize the armour The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLIII-B2-2022 XXIV ISPRS Congress (2022 edition), 6-11 June 2022, Nice, France layer geometry (e.g., packing density, armour layer porosity, and damage percentage). Due to the high spatial resolution of the orthophoto (1.5 cm/pixel), the automatic segmentation of block in orthophoto based on object-based image analysis (OBIA) (Cooper et al., 2021) can be a feasible option. Therefore, combining with the information provided by the DSM, it will be possible to isolate and measure horizontal and vertical movements of the armour unit. Circles correspond to P and plus sign correspond to CHPZ.

CONCLUSION
This work showed the validation of Real-Time Kinematic assisted Unmanned Aerial Systems (UAS-RTK) surveys combined with Structure from Motion (SfM) photogrammetry to reconstruct the geometry of rouble mound groins for monitoring displacements in armour units. The UAS-RTK survey allowed to map the entire groin including the areas hard-to-reach by the technicians. The flight was automatically planned and lasted about 10 min. The vertical accuracy of the Digital Surface Model (DSM), in comparison with the traditional GNSS, returned a RMSE of 2.6 cm. Therefore, the UAS-RTK DSM and orthophoto present an effective and accurate alternative to evaluate the geometry of rubble mound groins. Future work will investigate the segmentation techniques of orthophotos based on Object-Based Image Analysis, in order to automatically isolate armour layer units to track vertical and horizontal movements. As result, it will be computed the main armour layer indicators (armour layer porosity and packing density) for assessing the integrity of the structure.