Close-range airborne Structure-from-Motion Photogrammetry for high-resolution beach morphometric surveys: Examples from an embayed rotating beach

    Close-range airborne Structure-from-Motion Photogrammetry for highresolution beach morphometric surveys: Examples from an embayed rotating beach Guillaume Brunier, Jules Fleury, Edward J. Anthony, Antoine Gardel, Philippe Dussouillez PII: DOI: Reference:

S0169-555X(16)30061-7 doi: 10.1016/j.geomorph.2016.02.025 GEOMOR 5526

To appear in:

Geomorphology

Received date: Revised date: Accepted date:

15 June 2015 11 January 2016 20 February 2016

Please cite this article as: Brunier, Guillaume, Fleury, Jules, Anthony, Edward J., Gardel, Antoine, Dussouillez, Philippe, Close-range airborne Structure-from-Motion Photogrammetry for high-resolution beach morphometric surveys: Examples from an embayed rotating beach, Geomorphology (2016), doi: 10.1016/j.geomorph.2016.02.025

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

ACCEPTED MANUSCRIPT Close-range airborne Structure-from-Motion Photogrammetry for high-resolution beach morphometric surveys: Examples from an

SC R

IP

T

embayed rotating beach

a

Guillaume BRUNIER a

a*

MA

Edward J. ANTHONY

NU

Jules FLEURY

b

Antoine GARDEL

a

CE P

a

TE

D

Philippe DUSSOUILLEZ

Aix-Marseille Université, CEREGE UM 34, CNRS, IRD, , Europôle Méditerranéen de

l’Arbois - Avenue Louis PHILIBERT - BP 80 -, 13545 Aix en Provence, France *

b

AC

Institut Universitaire de France (IUF) CNRS Guyane, USR 3456 and LOG, UMR CNRS 8187, 28 avenue Foch, BP 80 62930

Wimereux, France

Corresponding author: Guillaume BRUNIER, [email protected]

1

ACCEPTED MANUSCRIPT Abstract The field of photogrammetry has seen significant new developments essentially

T

related to the emergence of new computer-based applications that have fostered the growth

IP

of the workflow technique called Structure-from-Motion (SfM). Low-cost, user-friendly SfM

SC R

photogrammetry offers interesting new perspectives in coastal and other fields of geomorphology requiring high-resolution topographic data. The technique enables the construction of topographic products such as digital surface models (DSMs) and

NU

orthophotographs, and combines the advantages of the reproducibility of GPS surveys and the high density and accuracy of airborne LiDAR, but at very advantageous cost compared to

MA

the latter. Three SfM-based photogrammetric experiments were conducted on the embayed beach of Montjoly in Cayenne, French Guiana, between October 2013 and 2014, in order to

D

map morphological changes and quantify sediment budgets. The beach is affected by a

TE

process of rotation induced by the alongshore migration of mud banks from the mouths of the

CE P

Amazon River that generate spatial and temporal changes in wave refraction and incident wave angles, thus generating the reversals in longshore drift that characterise this process. Sub-vertical aerial photographs of the beach were acquired from a microlight aircraft that flew

AC

alongshore at low elevation (275 m). The flight plan included several parallel flight axes with an overlap of 85% between pictures in the lengthwise direction and 50% between paths. Targets of 40 x 40 cm, georeferenced by RTK-DGPS, were placed on the beach, spaced 100 m apart. These targets served in optimizing the model and in producing georeferenced 3D products. RTK-GPS measurements of random points and cross-shore profiles were used to validate the photogrammetry results and assess their accuracy. We produced dense point clouds with 150 to 200 points per m², from which we generated DSMs and orthophotos with respective resolutions of 10 cm and 5 cm. Compared to the GPS control points, we obtained a mean vertical accuracy less than +/- 10 cm, with a maximum of 20 cm in marginal sectors with sparse vegetation and in the lower intertidal zone where water-saturated surfaces generated lower-resolution data as a result of a lack of coherence between photographs. The 2

ACCEPTED MANUSCRIPT overall results show that SfM photogrammetry is a robust tool for beach morphological and sediment budget surveys. Our SfM workflow enabled the discrimination of beach surface features at a scale of a few tens of centimetres despite the low textural contrasts exhibited by

IP

T

the quartz beach sand and the relatively uniform upper beach topography, as well as the calculation of beach sediment budgets. 66,000 m³ of sand were removed from the northern

SC R

sector of the beach, of which 22,000 m³ were transferred to the southern sector in the course of rotation. Finally, we briefly highlight: (1) the advantages of SfM photogrammetry compared

NU

to other high-resolution survey methods, (2) the advantages and disadvantages of, respectively, a microlight aircraft and an unmanned aerial vehicle (UAV) in undertaking SfM

MA

photogrammetry, and (3) areas of potential future improvement of the SfM workflow technique. These concern more extensive cross-shore deployment of ground control points

TE

D

to reduce possible tilt, and oblique cross-shore photography to improve parallax.

Keywords: Low-cost photogrammetry; Structure-from-Motion photogrammetry; digital

AC

Highlights

CE P

surface model; beach rotation; beach morphometry.



Low-cost photogrammetry enables high-resolution beach morphological monitoring.



Beach rotation in French Guiana was monitored by photogrammetry in 2013 and 2014.



We produced three digital surface models (DSMs) and sediment budgets for the beach.



We obtained a mean DSM accuracy <10 cm on the upper dry beach.



The DSMs show beach morphological contrasts and rapid beach rotation over a year.

3

AC

CE P

TE

D

MA

NU

SC R

IP

T

ACCEPTED MANUSCRIPT

4

ACCEPTED MANUSCRIPT 1. Introduction

T

Data products generated by accurate and high-resolution topographic surveying are

IP

becoming increasingly important in understanding beach morphological changes and

SC R

processes at timescales of days, months or years, as well as in the monitoring of the impacts of episodic events such as storms. There are various techniques to monitor beach morphology and evolution with derived digital elevation models (DEMs). These include video-

NU

imaging ARGUS systems, Airborne Light Detection and Ranging (LiDAR), ground-based Terrestrial Laser Scanning (TLS), traditional topographic monitoring using a Real Time

MA

Kinematic Differential Global Positioning System (RTK-DGPS), or a total station, and photogrammetric techniques. Each of these techniques has advantages and limitations in

D

terms of spatial and temporal coverage, accuracy, and operational expertise and software

TE

needs and costs (James et al., 2013). For example, TLS is very efficient for an object size of 10s to 100s of metres. Despite the increasing use of this tool on beaches or rocky cliffs and

CE P

its accuracy, it still requires significant expertise in data collection and processing and is a costly technique. LiDAR allows for rapid collection of data and operates over large areas (> 1 km) at high spatial density. Its vertical accuracy, close to 10 cm, is well within that of beach

AC

topographic changes that commonly occur following storms, and, consequently, it is adequate to monitor such changes (e.g., Sallenger et al., 2003). However, LiDAR surveys are commonly conducted rather infrequently due to high cost and the prior required careful organization (James et al., 2013; Ouédraogo et al., 2014a), and there are, indeed, only few examples in the literature involving multiple LiDAR surveys over short periods of a year or two (e.g., Montreuil et al., 2014). Video imaging systems such as ARGUS (Plant and Holman, 1997) allow for high frequency surveys (10 min to 1 h interval range) over areas of 100 m to several km (Harley et al., 2011). However, despite their low cost, these techniques require specialist analytical software and technical and scientific expertise, and are, therefore, not readily accessible. In contrast, RTK-DGPS and total station techniques are

5

ACCEPTED MANUSCRIPT accurate and accessible but they suffer from the low spatial density of points necessary for DEM construction. This disadvantage of areal coverage may be partially offset by the mounting of RTK-DGPS stations on all-terrain vehicles that can cover large areas (e.g.,

IP

T

Harley et al., 2011), but vehicle tracks can alter surface features being studied.

SC R

New recent developments in photogrammetry are such that this technique is now emerging as an alternative and complementary tool for coastal scientists and managers (e.g., Gonçalves and Henriques, 2015). The technique, based on stereoscopy between

NU

image pairs, is not new, as it has been used for decades to reconstruct landform topography and to produce maps. The manual alignment of stereoscopic image pairs is a time-

MA

consuming task based on input from aerial cameras. The accuracy of the DEMs generated by this technique was lower than that of LiDAR (Ouédraogo et al., 2014a). Advances in

D

computer vision and image analysis are, however, generating innovative developments in

TE

photogrammetry through the technique of Structure-from-Motion (SfM), which offers an automated method for the production of high-resolution digital surface models (DSMs) with

CE P

standard cameras (Fonstad et al., 2013; Javernick et al., 2014; Agisoft, 2015). SfM photogrammetry has been employed in recent years in the morphometric

AC

reconstruction of landforms, geological outcrops (Marzolff et al, 2009; Westoby et al., 2012), and braided river channels (Javernick et al., 2014). The technique has been recently applied by Harwin and Lucieer (2012), James et al. (2013), Mancini et al. (2013) and Casella et al. (2014) to beach morphological studies. However, none of these studies has used SfM photogrammetry to carry out beach morphodynamic assessments and sediment budget quantification, both of which require repeated high-resolution surveys. In this work, we address this research gap by using close-range photographic surveys from a microlight vessel. The implementation of this technique is accessible to non-specialist users, as demonstrated by Westoby et al. (2012), Fonstad et al. (2013), Hugenholtz et al. (2013), James et al. (2013), Javernick et al. (2014), Ouedraogo et al. (2014b), and Gonçalves and Henriques (2015). This low-cost method includes the advantages of reproducibility and

6

ACCEPTED MANUSCRIPT accuracy of measurement of RTK-DGPS and LiDAR surveys (e.g., Montreuil et al., 2014). Moreover, the field protocol is easy to organise and reproduce: it combines an aerial

T

photographic survey and deployment of targets on the beach georeferenced by RTK-DGPS.

IP

We tested SfM photography on Montjoly beach (Fig. 1), a highly dynamic embayed

SC R

beach in Cayenne, French Guiana, affected by a unique type of beach rotation influenced by the alongshore migration of mud banks formed north of the mouths of the Amazon River in Brazil. We highlight, from three experiments conducted between October 2013 and October

NU

2014, the utility of this photogrammetric technique in the generation of high-quality beach morphometric data products such DSMs and orthophotos and in the quantification of

D

MA

morphological and mass budget changes associated with beach rotation.

TE

2. Study site

CE P

Montjoly beach is a 3.5 km-long body of sand between rocky headlands in Cayenne, French Guiana (Fig. 1). The beach, composed essentially of quartz sand (> 90%), has an

AC

average width of 100 m, and bounds a lagoon characterised by an inlet that is seasonally closed but which is now kept manually open to avoid flooding of the neighbouring urban zones of Cayenne. Montjoly beach lies along the pathway of large mud banks that migrate alongshore from the mouths of the Amazon River to those of the Orinoco in Venezuela (Anthony et al., 2010, 2014), sourced by the large mud supply of the Amazon. Each mud bank can be up to 5 m thick, 10 to 60 km long and 20 to 30 km wide, and migrates at rates of 1 to 5 km/year (Gardel and Gratiot, 2005). Headland-bound beaches on this 1500 km-long muddy coast of South America are limited to the vicinity of Cayenne and Kourou in French Guiana (Fig. 1), the only sectors where notable outcrops of bedrock composed of migmatites and granulites occur. Montjoly and the rare sandy beaches on this muddy coast are important both economically and ecologically, providing outlets for recreation and routes, and

7

ACCEPTED MANUSCRIPT nesting sites for protected marine turtles (Lepidochelys olivacea, Cheloniamydas, Eretmochelys imbricata, Dermochelys coriacea). The presence of mud significantly alters the behavioural patterns of Montjoly and these sandy beaches, by modulating the influence of

IP

T

seasonal changes in trade-wind wave energy. Mud directly welds onto the beaches for months to years as a migrating mud bank approaches, leading to unique examples where

SC R

ocean-facing beaches are completely incorporated into a temporarily prograded muddy intertidal-to-shoreface system. During such phases, the mud-bound foreshore and muddy

NU

shoreface are rapidly colonized by mangroves. Subsequent mud erosion and mangrove forest destruction during inter-bank phases (corresponding to the space between two

MA

successive alongshore-migrating wave-dissipating mud banks) mark the resumption of normal beach dynamics. This involves the restitution, to the beach sand budget, of sand sequestered in the previous bank phase within the shoreface mud deposits. The main effect

TE

D

of these changes on the longer bay beaches is a form of beach rotation (Anthony et al., 2002; Anthony and Dolique, 2004), unique at the global scale. Such rotation does not result

CE P

from climate-induced variations in deepwater wave approach directions, as is generally reported for rotating beaches (e.g., Thomas et al., 2010), but from short to medium-term (order of a few years) changes in nearshore bathymetry induced by the migrating mud

AC

banks. These bathymetric changes affect wave refraction and dissipation patterns, inducing strong longshore gradients in waves. These wave-energy gradients generate, in turn, longshore movements of sand in these headland-bound beaches, resulting in alternations of erosion and accretion areas over time. Anthony and Dolique (2004) defined these beach morphological changes in terms of a simple, four-stage conceptual model, comprising bank, inter-bank and transitional phases that are characterised by dramatic beachface retreat or advance of up to 50 m a year (Fig. 1), which is much larger than the seasonal cycle of beach change (±10-20 m) characterising the storm-free tropical beach of Montjoly (Anthony et al., 2015).

8

ACCEPTED MANUSCRIPT Montjoly beach is affected by trade winds from the northeast that are mainly active from January to April. Waves impinging on the beach have periods ranging from 6 to 12 s, indicating a mix of trade wind-waves and longer swell, while offshore modal heights are up to

IP

T

1.5 m (Gratiot et al., 2007). The longer-period waves (> 10 s) are large swell waves generated by North Atlantic depressions in autumn and winter and by Central Atlantic

SC R

cyclones in summer and autumn. The most energetic trade wind-waves are observed between January and April (Fig. 1), in response to peak wind activity, whereas swell waves

NU

appear to be most frequent in autumn and winter, reinforcing the relatively high winter to early spring wave-energy regime induced by the trade winds. Tides are semi-diurnal and the

MA

spring tidal range is mesotidal (2.5 to 3.1 m). There is presently no certainty regarding the duration of a cycle of beach rotation. Cycle durations are considered as highly variable, depending on numerous large-scale controls such as Amazon mud supply, mud-bank

TE

D

formation processes at the mouth of the Amazon, mud bank size, the Atlantic wave climate and long-term tidal cycles (Anthony et al., 2014). A 60-year aerial photographic coverage of

CE P

the Cayenne coast (1950-2010) suggests rotation cycles of 12 to 20 years (Anthony et al., 2015). The current cycle duration has now attained 13 years (2002-2015), and the inter-bank phase terminated, given the proximity of the current incoming mud bank. It is within the

AC

framework of this phase and the new beach rotation cycle that the photogrammetric surveys were carried out in October 2013, March 2014 and October 2014. Rotation does not affect the net long-term (order of decades) beach sand budgets although there are signs that illegal sand extractions for the growing building industry in Cayenne will ultimately affect the beaches.

3. Materials and methods 3.1.

General workflow

9

ACCEPTED MANUSCRIPT 3.1.1.

Data collection

T

This section presents a general outline of our data collection protocol. Each survey

IP

was conducted during the low tide peak, in a one-hour time window, and covered the entire

SC R

subaerial beach. About the same area was covered in the three flights, except for that of October 2013 for which we lack images of the southern sector of the beach and of the intertidal zone. Atmospheric and tide conditions necessitated minor adaptations, especially

NU

concerning the number of flight axes and, consequently, the number of photographs taken. The aerial photographic surveys were carried out using a microlight aircraft with a low-

MA

altitude (<300 m above ground) flight permit. The photographic material consisted of a fullframe Digital Single Lens Reflex (DSLR) CANON 5D Mark II camera with a 36 x 24 mm

D

CMOS sensor with a photosite dimension (the size of the sensor's photosensitive element) of

TE

6.25 μm. A special frame was built on the aircraft to position the camera for vertical shooting. The flight plan and parameters were designed in a way as to make up for the lack of strongly

CE P

contrasting morphologies and textures on the sandy beach. The essential parameters we tried to optimize were longitudinal parallax and ground size dimension (GSD) of image pixels. We chose to work with a 50 mm CANON USM lens to produce pictures of 3.5 cm GSD and a

AC

200 x 130 m picture footprint from a flight height of 280 m. The overlap between pictures was about 85% in the lengthwise flight direction, and about 50% between paths (Fig. 2). The DSLR camera was installed in the aircraft with the longest sensor parallel to the flight axis in order to improve along-side parallax and the Base Height (B:H) ratio. The flight speed was around 100 km/h. The interval between two consecutive pictures was set to one second in order to obtain a degree of redundancy that compensates for eventual flight navigation inaccuracies. The shutter speed was set at 1/3000s to avoid blur from aircraft motion. This speed corresponds to a maximum motion of 1/3 pixel of the sensor. The sensitivity of the camera was set high enough, at 400 ISO, to mitigate variability in luminosity during acquisition. The autofocus was disabled and a manual focus was set to infinite. The

10

ACCEPTED MANUSCRIPT diaphragm aperture was maintained at around f:8 in order to increase depth of field and for optical quality. The entire flight coverage of Montjoly beach needed at least three alongshore axes and an inter-band spacing of 50 m (Fig. 2). During the flight, a differential GPS Trimble

IP

T

GeoXT GeoExplorer 2008 series, synchronized with the DSLR camera and set at a frequency of 1 Hz, enabled control of the aircraft trajectory and good adherence to the flight

SC R

plan.

The number of pictures acquired in the course of each experiment is shown in Table

NU

1. The pictures were saved in JPEG and RAW file formats. SfM photogrammetry is based on the identification of the ground texture pattern. In order to optimize the latter, we used RAW

MA

development to enhance texturing by increasing the contrast and to correct chromatic lens aberrations. The semi-diurnal tidal context and mesotidal tide range of Montjoly beach

D

required flying during the low-tide peak in order to capture as much uncovered beach surface

TE

as possible. The tropical convection processes prevailing in French Guiana during the day create atmospheric turbulences with risks of thunder showers that necessitated flights being

CE P

carried out no later than mid-day. The flight times corresponded to the sun at zenith, in order to avoid illumination and shadows variations. The nadir angle of the sun was, therefore,

AC

optimal for minimizing shadows that could perturb texture identification. There was, each time, a flight window of at least one hour corresponding to these criteria. Prior to each flight, a team on the ground deployed a minimum of thirty targets (Table 1) georeferenced using a RTK-DGPS Trimble R8 (Fig. 2). The targets were alternately deployed every 100 m between the breaker zone and the upper beach. The targets served as ground control points (GCPs) for georeferencing and optimization of aerotriangulation in the photogrammetric workflow. Each target measured 40 x 40 cm and was designed with a checkered pattern in black and white or black and yellow (Fig. 2). The intersection of the two black triangles representing the centre had to be clearly visible in the picture for postprocessing.

11

ACCEPTED MANUSCRIPT For control points, random and cross-shore topographic profile points were acquired with the RTK-DGPS in order to assess model quality. These are named ground truth points (GTPs) (Fig. 2). To avoid georeferencing problems with local coordinate systems and to

IP

T

render data comparable, our datasets were georeferenced into the same reference. We used the Universal Transverse Mercator 22 Northern zone (UTM 22N) projection associated with

SC R

the World Geodetic System 1984 (WGS 84) datum. The elevation reference was based on the global geoid model Earth Gravitational Model 1996 (EGM96). Aspects of each of the

NU

three work sessions are summarized in Table 1. Between the October 2013 session and March 2014 session, we improved the field protocol by adding ground-truth points and

MA

increasing the number of flight axes from 2 to 3.

D

SfM photogrammetry workflow

TE

3.1.2.

CE P

Current implementations of photogrammetry based on the SfM technique involve the use of various commercial and non-commercial software such as Agisoft Photoscan Professional, MicMac open-source from Institut Géographique National in France (IGN),

AC

SfM-PMVS open-source, Pix4D© and Trimble Business Center©. We selected Agisoft Photoscan Professional software because it integrates and automates all the SfM protocol stages with a user-friendly interface, and, above all, has an error control for each step. It also allows for easy transfer of results to a GIS. The main improvement in photogrammetry and its increasingly widespread use nowadays are linked to progress in computer vision algorithms. This new paradigm does away with what was hitherto a tedious task wherein the operator had to find homologous points for aerotriangulation, this task now being automatically accomplished using the Scale Invariant Feature Transform (SIFT) algorithm, an object recognition system (Lowe, 2004).

12

ACCEPTED MANUSCRIPT The photogrammetric workflow implemented in Agisoft Photoscan uses known algorithms such as those included in other recent software packages, but it is slightly modified and fully coded by the company (Agisoft, 2015). The stages of the SfM

IP

T

photogrammetry workflow have been described in a number of studies, including those of James et al. (2013), Javernick et al. (2014) and Gonçalves and Enrique (2015). Aspects of

SC R

the workflow stages and reconstruction of 3D models pertinent to our study site are briefly described below.

1. Picture setting and control. This stage consists in sorting out the pictures and

NU



synchronizing GPS data with each photo. We implemented a Matlab© script that

MA

carries out time synchronization and adds coordinates to the EXIF information on each picture. All areas with moving objects, such as water, cars or people, and the

D

microlight shadow on each photograph, had to be masked. Standing water also

CE P

2. Picture alignment. This is one of the most important procedures in the SfM protocol. At this stage Photoscan uses a tracking algorithm to identify, match and monitor the movement of the unique features of the object (Verhoeven, 2011; Javernick et al., 2014; Agisoft, 2015). This algorithm is based on an improved version

AC



TE

needed to be removed as light reflection varies through time.

of the object recognition SIFT system (Lowe, 2004). It does so by reconstructing the optimal picture positions and improves them with a bundle-adjustment algorithm (Robertson and Cipolla, 2009; Verhoeven, 2011; Javernick et al., 2014). SfM is innovative because, regarding scene reconstruction, it does not require 3D picture locations and orientations unlike traditional photogrammetry. As a result, sparse point clouds - named SIFT clouds - and a set of picture positions in a relative coordinate system are formed. Following this, the first picture alignment, together with the SIFT cloud, must be transformed into an absolute spatial reference. To this end, it is georeferenced

13

ACCEPTED MANUSCRIPT and optimized with GCPs in the working coordinate system of the area. The GCPs are checked on each picture at infra-pixel scale. The accuracy of the final estimations depends on various factors, such as overlapping of adjacent photos, and the shape of

IP

T

the forms photographed. This can lead to non-linear deformations of the final model. The non-linear deformations of the model can be removed by optimizing the

SC R

estimated point cloud and image orientation parameters based on the known reference coordinates. Completing the optimization process on the sparse cloud

NU

significantly enhances the accuracy of the sparse cloud model. Indeed, the GCPs’ initial position accuracy is < 1 m, but significantly improving to 1 cm by virtue of such



MA

optimization.

3. Point cloud and mesh building. The next stage consists in building a dense point

D

cloud, and then a mesh. Based on the estimated picture positions, the programme

TE

calculates depth information from correlation of each pair of images, and then combines all these into a single dense point cloud. The dense point cloud has almost

CE P

the same density as LiDAR point clouds with 100 to 300 points per m² for the highest reconstruction quality. A dense point cloud is manually edited prior to proceeding to

AC

3D mesh model generation with the filtering of points representing the water surface. Photoscan reconstructs a 3D polygonal mesh representing the object’s surface based on the dense point cloud. 

4. Exporting DSM and orthophoto. The final stage consists in exporting the mesh model to a georeferenced Digital Surface Model (DSM). We chose to interpolate all of our DSM exports with a 10 cm resolution. An orthophoto with a 5 cm resolution was also generated. Each of the stages summarized above involved both manual and automated

processes. The overall SfM workflow and the export of DSM and orthophoto endproducts for Monjotly beach necessitated a time investment of one week for each survey

14

ACCEPTED MANUSCRIPT and powerful memory and computation resources (Table 2). The first and second stages, which involved a manual check on the images, marking off the masks and targets, and generating the SIFT cloud points, required two days of work. The third stage, point cloud

IP

T

generation at a high density, took 72 hours with our workstation configuration (Table 2). The mesh calculation process in high quality lasted 3 hours. The end-product exports

SC R

were generated rapidly, in less than thirty minutes. Finally, it is important to note that SfM photogrammetry implementation consumes a lot of disk space with each project involving

3.1.3.

MA

NU

between 50 and 100 Giga bytes of data, including picture set and model generation.

Model quality control

TE

D

The assessment of the quality of the produced 3D-model is both qualitative and quantitative. The qualitative control concerns the extent to which the user considers that the

CE P

3D models render a coherent image of identifiable landforms, whereas the quantitative control concerns estimation of the deviation of the model from field measurements. Statistics on model accuracy may yield satisfactory values but an expert could argue that real

AC

geomorphological shapes are not preserved in the model. The qualitative coherence of the model is thus essential in order to go further into quantitative measurements. In this section, we will discuss only quantitative control. Qualitative aspects will be treated in the results section. The accuracy of the model is evaluated by the software in the first place. After optimizing the photogrammetric network from the distribution of the GCPs, a residual error and a root mean square error (RMSE) are calculated for each point in both projection and pixel units. We next used external RTK-DGPS GTPs to calculate the vertical accuracy. The strategy is to obtain as many GTPs as possible, first randomly and then on representative transects or landform shapes. The DSM accuracy is based on evaluation of vertical accuracy

15

ACCEPTED MANUSCRIPT from a comparison of the DSM and GTP elevations. The distribution of the vertical difference between the two, named ∆h, is obtained using a statistical method based on Höhle et al. (2009). The total number of GTPs is named n. The GTP samples are plotted in a histogram.

IP

T

In the case of a normal distribution, the statistical estimators for DSM accuracy are standard ones: mean error µ, standard deviation σ and root mean square error RMSE. In our case, we

SC R

have many outlier values due mainly to water-saturated surfaces and to surfaces with a homogeneous texture. The distribution is then non-normal. Consequently, we conducted

NU

statistical analysis with estimators such as the median m and normalized median absolute deviation NMAD (eq. 1). The NMAD is proportional to the median of absolute differences

MA

between errors and the median of errors. It can be considered as a more robust estimator than the standard deviation in the case of a non-normal distribution, or can replace the latter if the number of GTPs

is sufficiently large. It appears to be more adequate for DSM

TE

D

accuracy assessment from the SfM technique.

CE P

(1) NMAD = 1.4826 · medianj (|∆hi - m∆h|) Furthermore, we estimated the DSM accuracies for various beachface classes that we extracted from the orthophotos. We distinguished 5 such classes from photo-

AC

interpretation (Fig. 3): very wet sand with water run-off (exfiltration), wet sand, dry sand with texture (local concentrations of heavy minerals), dry sand with poor texture, and dry sand with sparse low vegetation. In addition to these beachface classes, we also defined areas of dense low vegetation.

3.2.

DSM production

The main parameters used in the SfM workflow for processing the pictures acquired in the course of each flight and generating the DSMs of Montjoly beach (Table 3) are described below. First, masks were created on unnecessary areas, water and moving 16

ACCEPTED MANUSCRIPT objects. The maximum number of SIFT points sought on each picture was set to 40,000, but we set no limit for the number of tie points for constructing the sparse cloud. The first alignment was made with generic pair selection that first looked for matching points in each

IP

T

pair at a sub-sampled low resolution. This was aimed at selecting pertinent pairs enabling calculation of full resolution matching points. The resulting sparse point cloud comprised

SC R

815,000 points for a density close to 1 points/m². Outliers were filtered manually and semiautomatically, based on several criteria. From experience based on reprojection criterion, we

NU

found that an expert manual filtering is more efficient and enables removal of true outliers. A classification of ground points and other classes was offered by the workflow protocol but

MA

was not used in our case study which involves gentle beach morphology, as the algorithm searches for abrupt elevation changes and breaks in slope. GCPs were next imported and picked out on each picture. The software rendered this task easy by automatically proposing

TE

D

a location for the GCP after it has been identified on a minimum of two aligned images. Optimization was then carried out using GCPs that reduced non-linear errors and

CE P

georeferenced the model. The residual and RMSE errors were calculated and allowed us to validate the optimization step. From this point, we calculated the dense cloud. We used moderate filtering at this stage in order to reconstruct the morphology of the beach. The

AC

resulting dense cloud comprises about 258,000,000 points, corresponding to a density of 200 points/m2, or to a point every 7 cm (Table 2). We chose to export from the dense cloud the DSM in UTM 22N projection, WGS84 datum, and EGM 1996 elevation reference, with a pixel size of 10 cm. Every point of the beach is covered by an average of nine overlapping images, thus yielding a good picture redundancy and parallax geometry. Overall, all the DSMs clearly portray a large cross-shore band comprising all of the five beachface classes listed above.

4. Results and discussion

17

ACCEPTED MANUSCRIPT

DSM quality assessment

T

4.1.

IP

Fig. 4 depicts the northern sector of the beach in the vicinity of the inlet of Montjoly

SC R

lagoon in 2014 as well as the intertidal zone, and highlights various features such as the lagoon's ebb tidal delta and its bedforms, the beach berm, erosion scarps, an upper beach runnel, small barkhan-like embryo dunes, the backshore with vegetation, rock revetments on

NU

the beach, houses and roads. The beach morphology, including features with very poor contrast such as a mild scarp in the October 2014 DSM, appears clearly and with a realistic

MA

shape. In the intertidal and surf zones, however, the morphology is less clear as one goes from the upper to the lower beachface, which corresponds to the boundary of the DSM (Fig.

D

4). We observed this deterioration on each DSM.

TE

The errors generated using Photoscan Professional are summarized in Table 3 for

CE P

each flight. The values are quite acceptable. The estimated easting and northing coordinates of GCPs have an accuracy of 1 mm. The estimated elevation of GCPs shows an accuracy of 1 cm. The picture average reprojection error is less than 1 pixel. These indicators provide an

AC

assessment of the internal coherence of the photogrammetric process and of the quality of alignment, and, consequently, of the depth reconstruction from which are derived the dense point clouds. These accuracy values seem very promising but may be misleading. Indeed, some non-linear errors persisted in the DSMs, and the precision of the model is more objectively evaluated from points that did not serve to process the model. We chose to calculate the vertical DSM deviation relative to the RTK-DGPS survey of October 2014. In the course of this survey, we established more than one thousand GTPs (Table 1, 3) in order to assess the real quality of the model. The three models are generated using the same protocol, and the quality assessment for the October 2014 DSM proved to be representative also of those of the October 2013 and March 2014 DSMs.

18

ACCEPTED MANUSCRIPT Fig. 5a shows the general distribution of ∆h which appears to be nearly normal. The overall quality of the DSM is centred about a mean error of 0.66 cm (Table 3). The σ and the RMSE are similar with values of 12.85 and 12.87 cm, respectively, but these indicators are

IP

T

more sensitive to outlier values of ∆h. Less sensitive to these outliers, the median and the NMAD indicators show values of -1.79 and 7.42 cm, indicating a robust reconstruction of the

SC R

beach topography. We calculated 46% of ∆h values within a ±5 cm interval and 70% within ±10 cm. We obtained the same range of accuracy on the March 2014 DSM (Table 3). Fig. 5b

NU

plots GTP elevations estimated from the DSM as a function of GTP elevations obtained by RTK-DGPS. The figure clearly shows a linear relation for the bulk of GTPs of the beach and

MA

the upper intertidal zone with a coefficient of determination (R²) of 0.98. However, the cloud point deviations increase for low elevation GTPs which correspond to the lower intertidal

D

zone.

TE

Fig. 4 depicts a focus on the spatial distribution of GTPs near the inlet of Montjoly lagoon and the ∆h values from the elevation of the DSM. It shows an overall accuracy of 0 to

CE P

10 cm for the lower beach. The incipient aeolian dunes are well reconstructed with an accuracy less than 5 cm. As expected from water-saturated surfaces, ∆h increases towards the backshore boundary of the DSM with an accuracy of 10 to 30-40 cm. We observed this

AC

pattern in each DSM of the same zone. Fig. 6 shows the variation of ∆h on a cross-shore transect. The deviation of values from the aerial beach to the lower intertidal zone clearly appears. This can be explained by the homogeneous radiometry of wet sand and water surfaces that are not coherent in consecutive pictures. Consequently, such surfaces do not allow the SIFT algorithm to extract invariant SIFT points of good quality. Afterwards, we examined the role of ground texture in the distribution of ∆h. Our results are summarized in Fig. 7, which shows the cumulative number of ∆h for each class with a more refined range of 2.5 cm, 5 cm, then 10 cm. We chose to limit the histogram to a ∆h range of ±30 cm, and to group the outliers in a class labeled ’larger than ±30 cm’. The results show that the GTPs in dry sand, including homogeneous and non-textured sands,

19

ACCEPTED MANUSCRIPT which represent 47% of the entire set of GTPs, are concentrated around a median of -0.9 cm and a NMAD of 6 cm. These results demonstrate that the method performs well on a sandy beach, and also on poorly-textured objects. Moreover, the GTPs in the vegetation class

IP

T

display good values of ∆h especially on sparse vegetation with a median of -0.9 cm and a NMAD of 7.4 cm. The good overall quality of these results is explained first by the fine

SC R

resolution of the pictures and the quality of the optical lens and sensor, but also by the RAW post-processing that increased the dynamic range. Unsatisfactory values and outliers,

NU

considered to be in the range above 15 cm, were found to be increasingly present in very wet sand near water surfaces. However, the distribution of GTPs in wet and saturated sand in the

MA

intertidal zone are not very high with a median value of ∆h at -0.8 cm and a NMAD at 7.8 cm. This deterioration in quality towards the sea must be taken into account to set the limit of

D

consistent values.

TE

Fig. 8 displays a classical phenomenon of deformation of the model inherited from the photogrammetric algorithms. This defect, called the ’bowl effect’, has been described by

CE P

James and Robson (2014) and Ouédraogo et al. (2014b). It corresponds to a non-linear parabolic distortion of the model with over-estimation of elevation near the image block border. In cases where the model is constrained with GCPs, this defect is attenuated and

AC

may not be perceptible. In our case, Fig. 8a, b, depicting the Montjoly lagoon inlet, show clearly that there is no ’bowl effect’ in the zone covered by GCPs. We plotted ∆h values as a function of their distance from GCPs and on the basis of ground texture class. It appears that none of the data deviates from a ± 10 cm accuracy, regardless of distance from GCPs, except for GTPs located in sectors of wet sand. In contrast, Fig. 8c, d highlights a relatively sparse cover of GCPs in the northwestern sector of the beach. It can be seen that this low GCP density results in a large DSM deformation tantamount to a decrease in ∆h accuracy values that exceeds 20 cm, no matter what the ground texture.

4.2.

Beach morphodynamic applications 20

ACCEPTED MANUSCRIPT

In addition to relevant information on beach morphological features, our datasets

T

enabled extraction of information on morphological changes and sediment budgets. Fig. 9

IP

depicts, for example, the morphological changes yielded by the three DSMs along a crossshore transect. The profiles show a large beach retreat of 25 m between October 2013 and

SC R

March 2014 that reflects sand removal in this sector and transport towards the southeast associated with the rotation process, under the influence of the currently approaching mud

NU

bank. The profiles also show important changes in beachface forms linked to wave forcing. The October 2013 profile displays cusp horns dissecting the berm, manifesting reflective

MA

beach conditions under relatively energetic waves of 1.5 m at the break point. This reflective beach morphology is typical of a situation of susceptibility to erosion of a well-developed

D

berm under energetic waves (Wright and Short, 1984), as rotation proceeds in this case. The

TE

March 2014 profile shows a 2 m-high beach erosion scarp that appears to mark the end of the transition phase in the 4-phase rotation model of Anthony and Dolique (2004). In

CE P

contrast, the October 2014 profile displays a regular slope that reflects trampling from beach users, leatherback turtle nesting, and falling wave energy levels as the approaching mud bank increasingly dissipates incident waves, heralding a future shift from transition to mud

AC

bank phase. The beach volumetric changes during this transitional phase are summarized in Fig. 10. We selected a DSM limit between dry (beachface) and wet (lower intertidal zone) sand in order to exclude the area affected by the ‘bowl effect’ and by picture orientation problems. We then split our area into 6 sectors that emphasize alongshore morphological differences. The volumetric differential displays a large loss of 66,000 m³ in the northwestern sector and a transfer of 22,000 m³ towards the southeastern sector between October 2013 and October 2014 that highlights the rotation process. Given the fact that the sand budget of Montjoly beach is known to be stable, barring illicit extractions that are now severely controlled, the budget difference between these two sectors reflects: (a) cross-shore transfer of sand eroded from the subaerial beach towards the lower foreshore and inner shoreface where this sand is sequestered by mud, and (b) under-estimation of volume changes 21

ACCEPTED MANUSCRIPT between October 2013 and March 2014, especially in the southeastern sector in accretion,

T

due to the lack of data on the intertidal zone in the October 2013 DSM.

SC R

IP

4.3. Advantages relative to other beach morphometric survey tools, and areas of improvement

Efficient morphometric surveys of beaches require using tools that combine accuracy,

NU

measurement density and reproducibility. Currently, three tools are commonly used to this end, and a fourth more rarely: respectively, the total station and RTK-DGPS, LiDAR,

MA

Terrestrial Laser Scanning (TLS) and the ARGUS system. SfM photogrammetry using aircraft surveys appears superior in all fields to other techniques, with the exception of

D

accuracy compared to laser technology. SfM implementation is of low cost in terms of

TE

materials, necessitating only a good full frame sensor and lens, a powerful computer, and simple field accessories such as targets. Nevertheless, it requires prior RTK-DGPS

CE P

operations for precise georeferencing of GCPs for generating the model. DSM/DEM quality is superior to that of topographic surveys with RTK-DGPS or total station in terms of spatial

AC

coverage and data density, and similar to that of laser-based techniques. Although the first two survey methods give an extremely precise location and elevation of each point surveyed, interpolation of the low density of points associated with these methods impairs the quality of DEMs they generate, compared to SfM photogrammetry. The precision of close-range SfM is close to that of laser-based techniques. However, outside of the area covered by GCPs, this technique suffers from non-linear error propagation, mainly linked to camera calibration. The error propagation on DSM/DEM products from laser technology is more homogeneous (Ouédraogo et al., 2014a; 2014b). Despite this shortcoming, photogrammetry probably offers the best compromise, as far as cost, accuracy, coverage and reproducibility are concerned, for repeated monitoring of

22

ACCEPTED MANUSCRIPT landforms potentially subject to rapid morphological and sediment budget changes such as beaches, spits and aeolian dunes.

T

One way of enhancing the possibilities offered by SfM photogrammetry consists in using

IP

a UAV (drone) rather than a microlight aircraft which requires a licensed pilot onboard. A

SC R

number of studies have used UAVs for SfM photogrammetry (Harwin and Lucieer, 2012; Mancini et al., 2013; Casela et al., 2014; Gonçalves and Enrique, 2015). UAVs offer a significant advantage in terms of reproducibility of measurements: (1) the user does not need

NU

to be, or have recourse to, a certified aircraft pilot, (2) UAVs are, all said and done, safer to use than an aircraft, (3) and yield a better image resolution, because they can fly at

MA

elevations <100 m, whereas a microlight aircraft is limited to a minimum flight ceiling of 280 m close to urban areas. However, recourse to a certified pilot with a microlight aircraft

D

dispenses the potential photogrammetry user from a range of technical and regulatory

TE

restrictions relating to flights, whereas the use of a UAV needs to take into consideration these time-consuming tasks. Moreover, reasonably-priced civil, off-the-shelf UAVs cannot

CE P

embark a DSLR sensor because of payload restrictions. Their flight autonomy is also still relatively limited, which also thus restricts their areal coverage, although the more expensive

AC

models perform better on these points. Two areas of potential future improvement of the SfM workflow on beaches concern ground control points and photograph setting. Beaches are essentially longitudinal features, thus rendering difficult correction of model tilt in the cross-shore direction. This tilt can be reduced by increasing the number of GCPs deployed across-shore and especially in the backshore area. This, however, could be a time-consuming operation and may be rendered impossible by dense vegetation. Finally, the low elevation of some of the studied beach forms, ranging from a few centimetres to metres, could be better captured by improving the parallax, and thus the geometrical setting of photographs, with oblique views in the crossshore axis. A brushless motorized gimbal could be mounted on the aircraft to this purpose.

23

ACCEPTED MANUSCRIPT

T

5. Conclusion

IP

SfM-based photogrammetry is a good-compromise technique between accuracy, data

SC R

density and measurement reproducibility for morphometric surveys of geomorphologically dynamic landforms such as beaches. Using an SfM workflow with adapted parameters, we produced three DSMs of Montjoly beach, a sandy beach in French Guiana subject to rotation

NU

under the influence of mud banks. The three DSMs captured part of the rotational sequence in the course of which the beachface retreated several metres a month between October

MA

2013 and October 2014 as a result of sand transfer from its eroding northwestern sector towards its accreting southeastern sector. Our DSMs also enabled volumetric calculations

D

and had a very high resolution of 10 cm and an accuracy of less than 10 cm. We determined

TE

that the ‘bowl effect’ and water saturation associated with beach exfiltration are the main

CE P

limitations of SfM photogrammetry as far as beach surveys are concerned. They generated a loss of accuracy but this concerned exclusively the low intertidal zone. This study also demonstrates that photogrammetry based on SfM is suitable for high-

AC

resolution morphometric surveys of features of poor surface textural contrast, which is commonly the case of beaches. Our results also highlight the potential of this technique in high-resolution monitoring of various other types of coastal features evincing rapid morphological and sediment budget changes, including aeolian dunes, spits, and estuarine and deltaic forms such as channels, bars and islets.

Acknowlegments We acknowledge support from CEREGE UM 34 for the October 2013 experiment. We thank the CNRS Mission Interdisciplinaire and the Pépinière Interdisciplinaire de Guyane GUIASANDBEACH project for funding in 2014. We acknowledge support from the Belmont 24

ACCEPTED MANUSCRIPT Forum Project ‘BF-Deltas: Catalyzing Action Towards Sustainability of Deltaic Systems with an Integrated Modeling Framework for Risk Assessment’. Finally, we appreciate the operational support from the coastal team of the CNRS French Guiana office. We

SC R

IP

T

acknowledge two anonymous reviewers for their detailed and very constructive remarks.

References

NU

Agisoft, 2015. Agisoft Photoscan User Manual: Professional Edition. Version 1.1. url:

MA

http://www.agisoft.com/pdf/photoscan-pro_1_1_en.pdf. Anthony, E.J., Gardel, A., Dolique, F., Brunier, G., Péron, C., 2015. Mud banks, sand flux and beach morphodynamics: Montjoly lagoon beach, French Guiana. In: M. Maanan, M.

TE

D

Robin (eds.), Sediment Fluxes in Coastal Areas, Coastal Research Library 10, p. 75-90. DOI 10.1007/978-94-017-9260-8_4.

CE P

Anthony, E.J., Gardel, A., Gratiot, N., 2014. Fluvial sediment supply, mud banks, cheniers and the morphodynamics of the coast of South America between the Amazon and

AC

Orinoco river mouths. In: Martini, I. P., Wanless, H. R. (eds), Sedimentary Coastal Zones from High to Low Latitudes: Similarities and Differences. Geological Society, London, Special Publications, 388, 533–560. Anthony, E.J.,Gardel, A., Gratiot, N., Proisy, C., Allison, M.A., Dolique, F., Fromard, F., 2010. The Amazon-influenced muddy coast of South America: A review of mud-bank–shoreline interactions. Earth-Science Reviews 103, 99–121. Anthony, E.J.,Dolique, F., 2004. The influence of Amazon-derived mud banks on the morphology of sandy headland-bound beaches in Cayenne, French Guiana: a short- to long-term perspective. Marine Geology 208, 249–264.

25

ACCEPTED MANUSCRIPT Anthony, E.J.,Gardel, A., Dolique, F., Guiral, D., 2002. Short-term changes in the plan shape of a sandy beach in response to sheltering by a nearshore mud bank, Cayenne, French

T

Guiana. Earth Surface Processes & Landforms, 27, 857-866.

IP

Casella, E., Rovere, A., Pedroncini, A., Mucerino, L., Casella, M., Cusati, L.A.,Vacchi, M.,

SC R

Ferrari, M., Firpo, M., 2014. Study of wave runup using numerical models and low-altitude aerial photogrammetry: A tool for coastal management. Estuarine, Coastal and Shelf Science 149, 160-167.

NU

Fonstad, M.A., Dietrich, J.T., Courville, B.C., Jensen, J.L., Carbonneau, P.E., 2013.

MA

Topographic structure from motion: a new development in photogrammetric measurement. Earth Surface Processes and Landforms 38, 421–430.

D

Gardel, A., Gratiot, N. 2005. A satellite image-based method for estimating rates of mud

TE

bank migration, French Guiana, South America. Journal of Coastal Research, 21, 720– 728.

CE P

Gonçalves, J.A., Henriques, R., 2015. UAV photogrammetry for topographic monitoring of coastal areas. ISPRS Journal of Photogrammetry, Volume 104, Pages 101-111, ISSN

AC

0924-2716.

Gratiot, N., Gardel, A., Anthony, E.J., 2007. Trade‐wind waves and mud dynamics on the French Guiana coast, South America: input from ERA‐40 wave data and field investigations. Marine Geology, 236, 15‐26. Harley, M.D., Turner, I.L., Short, A.D. and Ranasinghe, R. 2011. Assessment and integration of conventional RTK-GPS and image-derived beach survey methods for daily to decadal coastal monitoring. Coastal Engineering 58, 194-205. Harwin, S., Lucieer, A., 2012. An accuracy assessment of georeferenced point clouds produced via multi-view stereo techniques applied to imagery acquired via unmanned aerial vehicle. International Archives of the Photogrammetry, Remote Sensing and Spatial 26

ACCEPTED MANUSCRIPT Information Sciences, Volume XXXIX-B7, 2012. XXII ISPRS Congress, 25 August – 01 September 2012, Melbourne, Australia.

T

Höhle, J., 2009. Dem generation using a digital large-format frame camera.Photogrammetry

IP

Engineering Remote Sensing 75, 87–93.

SC R

Hugenholtz, C.H., Whitehead, K., Brown, O.W., Barchyn, T.E., Moorman, B.J., LeClair, A., Riddell, K., Hamilton, T., 2013. Geomorphological mapping with a small unmanned aircraft system (sUAS): feature detection and accuracy assessment of a photogrammetrically

NU

derived digital terrain model. Geomorphology 194, 16–24.

MA

James, M.R., Robson, S., 2014. Mitigating systematic error in topographic models derived from UAV and ground-based image networks. Earth Surface Processes and Landforms,

D

Volume 39, Issue 10, pages 1413–1420,

TE

James, M.R., Ilic, S., Ruzic, I., 2013. Measuring 3D coastal change with a digital

CE P

camera.Coastal Dynamics, 893-904. Javernick, L., Brasington, J., Caruso, B., 2014. Modelling the topography of shallow braided

AC

rivers using structure-from-motion photogrammetry. Geomorphology 213, 166–182. Lowe, D., 2004.Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision 60, 91–110 (440). Mancini, F., Dubbini, M., Gattelli, M., Stecchi, F., Fabbri, S., Gabbianelli, G., 2013. Using unmanned aerial vehicles (UAV) for high-resolution reconstruction of topography: the structure from motion approach on coastal environments. Remote Sensing, 5 (12), pp. 6880-6898. Marzolff, I., Poesen, J., 2009. The potential of 3D gully monitoring with GIS using high resolution aerial photography and a digital photogrammetry system. Geomorphology 111, 48–60.

27

ACCEPTED MANUSCRIPT Montreuil, A.L., Levoy, F., Bretel, P., Anthony, E.J., 2014. Complex sediment recirculation on the ebb delta of a macrotidal inlet (Normandy, France): A multiple LIDAR dataset

T

approach. Geomorphology, 219, 114-125.

IP

Ouédraogo, M.M., Degré, A., Debouche, C., 2014a. Synthèse bibliographique : le modèle

SC R

numérique de terrain de haute résolution, ses erreurs et leur propagation. Biotechnologies, Agrononomie, Société, Environnement 18(3), 407-421. (in French, with English Abstr.)

NU

Ouédraogo, M.M., Degré, A., Debouche, C., Lisein, J., 2014b. The evaluation of unmanned

MA

aerial system-based photogrammetry and terrestrial laser scanning to generate DEMs of agricultural watersheds. Geomorphology 214, 339–355.

D

Plant, N.G. and Holman, R.A., 1997.Intertidal beach profile estimation using video

TE

images.Marine.Geology. 140, 1-24.

Robertson, D.P., Cipolla, R., 2009. Structure from Motion. In: Varga, M. (Ed.), Practical

CE P

Image Processing and Computer Vision. John Wiley, Chichester. Sallenger, A.H., Krabill, W.B., Swift, R.N., Brock, J., List, J., Hansen, M., Holman, R.A.,

AC

Manizade, S., Sontag, J., Meredith, A., Morgan, K., Yunkel, J.K., Frederick, E.B., Stockdon, H., 2003. Evaluation of airborne topographic lidar for quantifying beach changes. Journal of Coastal Research 19(1), 125-133. Thomas, T., Phillips, M.R., Williams, A.T., 2010.Mesoscale evolution of a headland bay: Beach rotation processes. Geomorphology 123, 129141.doi:10.1016/j.geomorph.2010.06.018 Verhoeven, G., 2011. Taking computer vision aloft — archaeological three-dimensional reconstructions from aerial photographs with Photoscan. Archaeological Prospection18, 67–73.

28

ACCEPTED MANUSCRIPT Westoby, M.J., Brasington, J., Glasser, N.F., Hambrey, M.J., Reynolds, J.M., 2012. ‘Structure-from-Motion’ photogrammetry: a low-cost, effective tool for geoscience

T

applications. Geomorphology 179, 300–314.

IP

Wright, L.D. and Short A.D., 1984.Morphodynamic variability of surf zones and beaches: a

AC

CE P

TE

D

MA

NU

SC R

synthesis. Marine Geology, 56, 93-118.

29

ACCEPTED MANUSCRIPT Figure captions

Figure 1. General setting of the French Guiana coast and Montjoly beach. (a) Regional

IP

T

coastal context under the influence of massive mud supply from the Amazon (for more details, see Anthony et al., 2010). (b) Rotation of Montjoly beach (red arrows) hinged on

SC R

bank, inter-bank and transitional phases as mud banks migrate alongshore (from Anthony and Dolique, 2004). (c) Orthophoto of Montjoly beach from the October 2014

NU

photogrammetry survey (this study) at the end of the ’transition’ phase between the interbank and bank phases depicted in sketch (b). (d) Cross-shore profile variations (red line in

MA

(c)) in the northern part of the beach showing extreme variability during a complete cycle of rotation from severe erosion (April 2000) to massive accretion (March 2011). MHWS = mean

TE

D

high water spring tide level; MLWS = mean low water spring tide level.

CE P

Figure 2. SfM field protocol. (a) Overview of the October 2014 flight plan in 3 axes and GCPs. (b) 2014 orthophoto showing picture overlap parameters along an axis and between axes, and GTPs. (c1) Photograph of the Savannah microlight aircraft, and (c2) outside lateral

AC

support of our full-frame DSLR sensor. (d) A GCP target near the Trimble R8 RTK-DGPS mobile receiver.

Figure 3. Ground surface texture classification produced by photo-interpretation, and groundtruth points (GTPs) measured in each textural class.

Figure 4. DSMs of Montjoly beach for each survey (a, b, c). The October 2013 survey shows a lack of data in the lower intertidal zone and in the northwestern and southeastern extremities of the beach due to problems in data acquisition. Focus on Montjoly lagoon inlet

30

ACCEPTED MANUSCRIPT from the October 2014 orthophoto (d) and DSM (e). Various morphological features such as the lagoon’s ebb delta, an erosion scarp, a bulldozer excavation imprint and mildly developed barkhans, are clearly distinguished. The October 2014 DSM absolute error l ∆h l is depicted

IP

T

by proportional coloured triangles. The highest values of l ∆h l are mainly situated in the

SC R

water-saturated lower intertidal zone, and are related to beach texture.

Figure 5. October 2014 DSM accuracy based on 1090 GTPs. The histogram (a) shows the

NU

general distribution of ∆h errors at 5 cm intervals. The mean error (µ), median (m), NMAD,

MA

standard deviation (σ) and RMSE indicators are plotted on this figure. The overall accuracy of the October 2014 DSM is very satisfactory with a large proportion of GTPs centred on a mean error of 0.6 cm and a median of -1.79 cm. The presence of outliers explains the non-

D

normal distribution, and the relatively large standard deviation (12.85 cm) and RMSE (12.87

TE

cm). Finally, the NMAD (7.4 cm) shows clearly the main behaviour of the ∆h distribution. The

CE P

graph (b) shows the elevation estimate from the DSM as a function of the RTK-DGPS elevation for each GTP. It displays a very good linear match (R²=0.99), but also highlights the

AC

low-elevation outliers.

Figure 6. October 2014 DSM showing cross-shore transect with GTPs (dots) (a), and corresponding beach profile (b). MHWS = mean high water spring tide level.

Figure 7. October 2014 DSM accuracy as a function of ground texture classification, and general statistical indicators. The histogram displays the cumulative distribution of ∆h error ordering as a function of ground texture, with gradual intervals of 2.5 cm, 5 cm and 10 cm. The outliers are filtered to a limit of | 30 cm | (which represents ≈ RMSE x 3) and are assembled in two class limits. The ∆h errors show that the DSM accuracy is clearly good,

31

ACCEPTED MANUSCRIPT including on very homogeneous dry sand. However, ∆h errors increase for wet and very wet

T

sand.

IP

Figure 8. The ‘bowl effect’ on the October 2014 DSM. The map (a) displays the model

SC R

accuracy from absolute error l ∆h l distribution (in proportional coloured triangles) inside a GCP constraint zone near Montjoly lagoon inlet. The graph (b) represents the error ∆h as a function of the distance of GTPs from GCP locations and GTP texture class. These results

NU

show that the ‘bowl effect’ is eliminated where the GCP coverage is good. The presence of

MA

outliers depends in this case on ground texture, the presence of water in sand inducing uncertainty values exceeding ±15 cm. In contrast, the map (c) of the northern part of the beach shows a lack of GCPs at its northern extremity. Consequently, the absolute error l ∆h l

D

distribution displays an important deterioration of the DSM quality. The graph (d) shows, no

TE

matter what the ground texture is, that the presence of outliers is linked to distance to the

AC

accuracy.

CE P

nearest GCP. Where this distance is large, the ensuing ‘bowl effect’ generates low DSM

Figure 9. Beach profile changes during the inter-bank to bank ’transition’ phase (see Fig. 1) from DSM comparison in the northwestern part of the beach (a). The profiles highlight severe erosion concomitant with rotation to the benefit of the southeastern end of the beach. The profiles are characteristic of a steep, erosion-prone reflective beachface (b and e). The March 2014 profile marks the end of this ’transition’ phase following a beachface retreat of nearly 30 m (c and e). Slope stabilization between the March 2014 and October 2014 surveys (d and e) is indicative of the onset of the full bank phase as the migrating mud bank offshore increasingly dissipates waves throughout Montjoly beach, thus temporarily ‘turning off’ the rotation process.

32

ACCEPTED MANUSCRIPT Figure 10. Beach volume changes from comparison of DSMs. The beach is split into six alongshore sectors evincing variable morphodynamic behaviour: sector 1 is in the most protected part of Montjoly beach with rock outcrops in the lower intertidal zone. Sector 2 is

IP

T

the most exposed part of the beach when rotation, involving sand transfer to the southeast, sets in. Sector 3 covers the Montjoly lagoon inlet, a dynamic zone characterised by ebb and

SC R

flood tidal flows. Sectors 4 and 5 highlight the gradient in beach erosion from the northwest to the southeast, the latter marking the transition to sector 6, the down-drift accumulation

MA

NU

zone in the rotation process.

Table captions

D

Table 1. Characteristics of photogrammetry data acquisition on Montjoly beach. Absolute

TE

GCP and GTP accuracies are derived from GNSS RTK post-processing using the permanent

problems.

CE P

station in Kourou. GTPs were not collected for the October 2013 due to data acquisition

AC

Table 2. (a) Computer set-up and (b) Photoscan workflow times, and additional information on the number of tie points and the density of final point clouds.

Table 3. Error values of GCPs (a) measured on the dense point cloud 3D model after optimization, and of GTPs (b) measured on interpolated DSMs. These errors are derived from differencing with RTK-DGPS measurements that serve as a reference.

33

AC

CE P

TE

D

MA

NU

SC R

IP

T

ACCEPTED MANUSCRIPT

Figure 1

34

D

MA

NU

SC R

IP

T

ACCEPTED MANUSCRIPT

AC

CE P

TE

Figure 2

35

D

MA

NU

SC R

IP

T

ACCEPTED MANUSCRIPT

AC

CE P

TE

Figure 3

36

AC

CE P

TE

D

MA

NU

SC R

IP

T

ACCEPTED MANUSCRIPT

Figure 4

37

SC R

IP

T

ACCEPTED MANUSCRIPT

AC

CE P

TE

D

MA

NU

Figure 5

38

SC R

IP

T

ACCEPTED MANUSCRIPT

AC

CE P

TE

D

MA

NU

Figure 6

39

D

MA

NU

SC R

IP

T

ACCEPTED MANUSCRIPT

AC

CE P

TE

Figure 7

40

MA

NU

SC R

IP

T

ACCEPTED MANUSCRIPT

AC

CE P

TE

D

Figure 8

41

AC

CE P

TE

D

MA

NU

SC R

IP

T

ACCEPTED MANUSCRIPT

Figure 9

42

AC

Figure 10

CE P

TE

D

MA

NU

SC R

IP

T

ACCEPTED MANUSCRIPT

43

ACCEPTED MANUSCRIPT Table 1 16/10/2013

20/03/2014

08/10/2014

Flight axes

2

6

4

Number of camera

270

780

680

50

50

T

Date of survey

Camera focal length

Mean flight elevation

SC R

(mm) 310

Number of GCPs

40

Absolute GCP

±2

accuracy (cm) Number of GTPs

-

Absolute GTP

-

290

3.35

3.5

29

30

±2

±2

106

1091

±4

±4

AC

CE P

TE

D

accuracy (cm)

MA

3.8

50

280

NU

(m) Mean GSD (cm)

IP

stations

44

ACCEPTED MANUSCRIPT Table 2 (a) Configuration of workstation CPU

RAM memory

GPU

HP Z800

Intel XEON E5607, 2.27 GHz, 16 cores

48 Gb DDR3 RAM

NVIDIA GeForce GTX 780 Ti, 2880 CUDA cores and 3Gb RAM

SC R

IP

T

Workstation

(b) Characteristics of each Agisoft Photoscan project and duration (in hours) of the various stages of the SfM workflow 16/10/2013

780

680

Time of picture setting and control (1)

10

26

24

Time of picture alignment (2)

0.5

2

2

1

5

0.8

4

9

8

Time of dense point cloud generation (1)

36

76

72

Point density of

175

225

205

Time of mesh and DSM generation (1)

1

3

3

Total number of hours

52

116

109

MA

270

08/10/2014

Number of tie points (millions)

AC

CE P

Time of GCP target picking and optimization (1, 2)

TE

D

Number of images

23/03/2014

NU

Date of survey

dense cloud (points per m²)

CPU : Central processing unit GPU : Graphic processing unit RAM : Random access memory (1) : manual process ; (2) : automated process 45

ACCEPTED MANUSCRIPT Table 3 16/10/2013

Date of survey

20/03/2014

08/10/2014

0.50

Residual Y error (cm)

0.72

0.43

Residual Z error (cm)

2.12

1.44

RMSE (cm)

2.30

Mean reprojection

0.61

IP

0.56

SC R

Residual X error (cm)

T

(a) GCP error

error (pixel)

0.44 0.25 0.93

1.58

1.05

0.67

0.72

-

Standard deviation

-

(cm) -

Median (cm)

-

NMAD (cm)

-

D

RMSE (cm)

MA

Mean (cm)

NU

(b) GTP Z error

0.65

0.66

17.60

12.85

17.62

12.87

-0.09

-1.79

6.96

7.42

AC

CE P

TE

RMSE : Root Mean Square Error; NMAD : Normalized Median Average Deviation

46