Ground control point requirements for structure-from-motion … · 2017-01-09 · Ground control...

Ground control point requirements for structure-from-motion derived topography in low-slope coastal environments Evan B. Goldstein1*, Amber R. Oliver2, Elsemarie deVries1, Laura J. Moore1, Theo Jass1 1Department of Geological Sciences, University of North Carolina at Chapel Hill, 104

South Rd, Mitchell Hall, Chapel Hill, NC 27599 USA 2Division of Earth and Ocean Sciences, Nicholas School of the Environment, Duke

University, Durham, NC 27708 USA *Corresponding Author: Email: [email protected] Twitter: @ebgoldstein Abstract Vegetated coastal dunes grow as a result of feedbacks between vegetation and sand transport. Observing the coevolution of vegetation and the sand surface is therefore critical for unraveling the dynamics of coastal dune growth. Capturing synchronous topography and photography at high spatial resolution and high temporal frequency using traditional techniques (airplane-based aerial photography, LiDAR) is expensive and time-consuming. Structure-from-Motion combined with Multiview-Stereo, a photogrammetry workflow that uses low-cost, consumer-grade equipment, is an economical alternative to traditional collection methods. This workflow still requires the definition of ground control points (GCPs) — locations with known coordinates — to develop accurate digital surface models. In this contribution we address how the number of GCPs used impacts the accuracy of digital surface models. We flew a 9-foot single-line delta kite attached to a consumer-grade camera to photograph the beach and dune of Hog Island, VA, a site that contains 178 high precision GCPs over an area of ~0.025 km2 (as part of an ongoing field experiment). We then processed the 318 photographs using Agisoft Photoscan and compared the elevation accuracy of digital surface models rendered using SfM, with varying GCPs, to points surveyed by a total station. Our results suggest that there is ‘diminishing returns’ when greater than 10 GCPs are used. Results from this study can be used to inform future Structure-from-Motion studies using UAVs or kites in flat, low-sloping coastal environments. 1. Introduction Vegetated coastal dunes grow as the result of feedbacks between the growth of dune vegetation and aeolian sand transport. As a result, serial observations of topography and vegetation must be acquired to understand the morphodynamics of coastal dunes (for a recent example, see Keijsers et al., 2015). A recently developed technique for creating synchronous topography and images is a photogrammetry workflow that combines Structure-from-Motion with Multiview-Stereo algorithms (hereinafter ‘SfM’). Many pictures of a static subject, taken at different angles, are combined through the SfM workflow to create a digital surface model (DSM) and a composite image (a mosaic) with images acquired using a typical digital camera. James and Robson (2012) assess SfM

PeerJ PrePrints | https://dx.doi.org/10.7287/peerj.preprints.1444v1 | CC-BY 4.0 Open Access | rec: 22 Oct 2015, publ: 22 Oct 2015

DSMs at 3 different scales (hand sample, coastal cliff, volcano caldera) and demonstrate that Structure-from-Motion is able to theoretically produce results that are similar to traditional photogrammetry, but with much more ease and flexibility in the workflow and using inexpensive equipment. This technique holds much promise in studies of surface processes, including coupled biological-physical systems (e.g., James and Robson, 2012; Westoby et al., 2012; Fonstad et al., 2013). As such many papers have been published in recent years using SfM to study landscape form and process from both biologic and geomorphic perspectives — a quick glance at this literature includes SfM used in research pertaining to forest dynamics (e.g., Dandois and Ellis, 2013), fault zone topography (e.g., Johnson et al., 2015), coastline dynamics (e.g., Harwin and Lucieer, 2012; Mancini et al., 2013; Gibbs et al., 2015, Kinsman et al., 2015), coral reef structure (e.g., Burns et al., 2015), glacial dynamics (e.g., Ryan et al., 2015), and snow depth estimation (e.g., Nolan et al., 2015). Recent work has systematically advanced our understanding of sources of error, accuracy, precision, and validation procedures in the SfM workflow. A critical point for reducing error in the SfM workflow is georectification — after images are taken and combined to form a DSM and mosaic image, the SfM products must tied to a known coordinate system. This is accomplished by pairing the SfM products to ground control points (GCPs; positions with known coordinates) that are visible in the image set (and therefore the mosaic image). James and Robson (2012) note that a minimum of 3 GCPs are technically required, however ‘…more control provides a more robust solution which is less sensitive to error on any one point’. Furthermore, James and Robson (2012) demonstrate that DSM position error decreases with the use of an increasing number of GCPs, a trend also seen by Clapuyt et al. (2015) and Harwin and Lucieer (2012). James and Robson (2012) also assess the location of GCPs, showing that widely dispersed GCPs aid in decreasing reconstruction error. This echoes work by Vericat et al (2009) who determined that a perimeter distribution (GCPs oriented primarily along the perimeter) was optimal for reducing error in individually rectified aerial photos. We extend previous studies of the SfM workflow by exploring, in a quantitative and systematic manner, how the number of GCPs influences vertical accuracy of the SfM-derived DSMs. Images of an ongoing field experiment in a low-slope coastal setting are acquired using a kite-mounted camera — a method that has previously been used in near-earth aerial photography (Marzolff and Poesen, 2009; Smith et al., 2009; Bryson et al., 2013; Lorenz and Scheidt, 2014) and is low cost, nonintrusive, and often does not require special permits (compared to UAVs; e.g., Vincent et al., 2015). Using kite images we construct ten separate DSMs using between 5 and 30 GCPs. GCPs are selected from 178 possible sites using a deterministic selection strategy. The remaining sites that are not selected as GCPs are used to test the accuracy of the constructed DSM and illustrate the decay in error as GCPs are added. We compare the accuracy of the DSMs developed using different numbers of GCPs to empirically derived precision ratios and find that increasing GCPs beyond 10 does not increase DSM accuracy.


2. Methods The study site is located on the accreting southern end of Hog Island, VA, USA (property of The Nature Conservancy; Figure 1). In May of 2014, 180 plants were transplanted along two cross-shore swaths that extend from the foredune crest to the beach (an elevation range of ~0 to 4 m). At each plant site a 122 cm yellow fiberglass stake was inserted into the sand. As part of the experimental planting, the ground surface at each site was measured monthly (May-October) using a Nikon DTM-322 total station (angle error = 5 arcseconds) located over a benchmark installed in a concrete-filled PVC pipe on a secondary dune ridge. Coordinates of the benchmark location (and an associated backsight of the same construction) were determined by post-processing (using the National Geodetic Survey’s Online Positioning User Service; http://www.ngs.noaa.gov/OPUS/) a global positioning system survey collected using a Trimble GNSS R6 antenna. More details on the field experiment can be found in Jass (2015). We acquired aerial images on May 13, 2015 using a kite-based camera flown at an average altitude of 20 m that captured 318 images synchronously with the total station measurements. We used a 9’ Gomberg delta kite attached to a Picavet camera suspension system holding a Redstone camera rig (www.KAPtery.com) and a Canon S100 running an intervalometer script (http://chdk.wikia.com/wiki/CHDK). In the images, we were able to identify 178 plant sites (2 stakes had been dislodged during the previous year of the field experiment). To decide which plant sites to use as GCPs (for tests of DSM vertical accuracy) we employed a maximum dissimilarity algorithm (MDA) to determine the sites that were most widely distributed in space (as per the suggestion by Vericat et al., 2009). As a deterministic selection routine, the MDA selects data points that are

Figure 1: Location of study site along the US East Coast. Top panel modified from Jass (2015); background image is a 1999 NASA LANDSAT 7 scene of the Virginia portion of the Delmarva Peninsula, USA. Bottom panel shows the study site with plant sites (black dots) and survey benchmarks (red dots) relative to the foredune (top left) and shoreline (bottom right).


maximally dissimilar from all previously selected points. Camus et al. (2011) provides a detailed introduction to the MDA selection routine as well as comparisons to other common clustering and section routines. This MDA has previously been employed to select representative samples from a range of given data (e.g., Goldstein and Coco, 2014). Starting with the plant location having the largest Northing value (initialization requires the user to select a starting point), an iterative process selects the next data point, which is the site with the maximum distance from all previously selected sites (Camus et al., 2011). The MDA routine continues selecting sites until the user-defined number of sites is attained (here we halt the algorithm at 30 sites). The selected sites, and the order in which they are selected, are shown in Figure 2. The kite images were processed separately ten times using differing numbers of GCPs (between 5 and 30). All kite images were processed using Agisoft Photoscan Professional Edition (http://www.agisoft.com), a commercial SfM software suite, resulting in a composite orthophoto and a DSM. We used the bundled program Agisoft Lens to calculate lens distortion for the Canon S100 and applied the known lens parameters within the SfM workflow. The motion of the camera attached to the kite resulted in mostly off-nadir images and several images from different elevations (i.e., captured during takeoff and landing) which help to mitigate systematic error (‘doming’) during the DSM generation phase, especially in the flat, low slope coastal environment (e.g., James and Robson, 2014). Additionally the use of GCPs in the processing routine is known to reduce systemic error such as doming (e.g., Javernick et al., 2014). 3. Results Here we evaluate how the change in the number of GCPs used to generate and georeference the DSM affects the vertical accuracy by using the 148 sites not selected by the maximum dissimilarity algorithm as the testing dataset. We compare the elevation at plant sites from the testing dataset (measured with the total station) with estimates from the SfM DSM. The resulting root mean squared error (RMSE) for the DSMs using 5-30 GCPs is shown in Figure 3. Vertical error decreases with the addition of GCPs, but

Figure 2: Selected plant sites (used as GCPs) are red. Site 1 is the starting point, and all sites are then selected sequentially. For example, the DSM for 10 GCPs is developed using sites 1-‐10. The black sites are unselected, and are used to test the accuracy of the SfM DSM against the total station measurement.

1

2

3

4 5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

Nor

thin

g

Easting

0 10

Student Version of MATLAB


beyond 10 GCPs remains constant at ~0.04 m. A curve can be fit to the data using an exponential function (dotted line; Figure 3), which asymptotically approaches a RMSE of 0.039 m:

𝑅𝑀𝑆𝐸 = 0.0392+ 1.85𝑒 !!.!"!×!"# Smith and Vericat (2015) compiled 50 studies and observed that DSM accuracy is related to range (camera to subject distance) with a relationship of 1:639 (e.g., 1 m of error for pictures taken at a range of 639 m). Our range in this study is 20 m, corresponding to an expected RMSE of 0.031 m (solid line in Figure 3). This expected RMSE is effectively identical to the value that our relationship asymptotically approaches (0.039 m) and the 0.04 m RMSE we observe at 10 GCPs.

Figure 3 Left Panel: Error (RMSE) as a function of the number of ground control points. Dashed line is the fitting equation presented in this work. Right Panel: Elevations derived from Structure-‐from-‐Motion (using the DSM built with 10 GCPs) versus those derived from the Total Station (the 148 plant sites not used as GCPs); RMSE = 0.04 m.

Using the kite-based images and 10 GCPs processed through the SfM workflow yields a final composite image with a resolution of 0.005 m/pixel (Figure 4) and a DSM with a resolution of 0.02 m/pixel (Figure 5).

0 5 10 15 20 25 300

0.05

0.1

0.15

0.2

0.25

Number of GCP

RM

SE (m

)

This StudySmith and Vericat, 2015

1 1.5 2 2.5 3 3.5 41

1.5

2

2.5

3

3.5

4

Total Station (m)

Kite

10

GC

Ps (m

)

Student Version of MATLAB


Figure 4: True color mosaic image (orthophotmosaic) of the field site from May 13, 2015. The blue object in the picture is a tent and black dots are plant sites. The mosaic covers approximately 0.025 km2

Figure 5: Digital Surface Model of the field site rectified using 10 GCPs. Elevation is in meters. Black dots are plant sites.

4. Discussion and Conclusion Images captured with a kite-based consumer-grade camera flown at 20 m, capturing a scene of ~0.025 km2, and processed using the Structure-from-Motion workflow with 10 ground control points yield a DSM with a vertical RMSE of 0.04 m. Increasing the number of GCPs (in this setting, and with this scene size) did not result in a further decrease in error. This error is smaller than the 0.15 m reported by Sallenger et al. (2003)


during a comprehensive evaluation of airborne LiDAR. The results presented here suggest that kite-based SfM is a valid technique for examining ground surface change in low-lying coastal environments. Future work will focus on evaluating the relationship between vertical RMSE and GCPs when using different observation distances and scene sizes at other low-slope, sparsely vegetated coastal sites. Because SfM captures a synchronous aerial image in addition to topographic data, the technique is especially useful for studies that benefit from synchronous collection of both types of data. For example, investigations of the growth of coastal foredunes, restoration of coastal dunes, or quantification of overwash delivery may especially benefit from the use of SfM. Acknowledgements Funding was provided by NSF-GLD (EAR-1324973 to LJM) and the ACC-IAC Summer fellowship through Duke University (to ARO). We thank The Nature Conservancy for permission to perform this research on Hog Island, VA, and the staff of the UVA ABCRC for assistance with field logistics. We thank the Public Lab community for notes on kite aerial photography (www.publiclab.org). EBG thanks Tony Gilman (Terrane Aerial Mapping; http://terraneaerial.com) for many helpful discussions. References Burns, J.H.R., Delparte, D., Gates, R.D., and Takabayashi, M., (2015), Integrating structure-from-motion photogrammetry with geospatial software as a novel technique for quantifying 3D ecological characteristics of coral reefs. PeerJ 3:e1077; DOI 10.7717/peerj.1077 Bryson, M., Johnson-Roberson, M., Murphy, R.J., Bongiorno, D., (2013), Kite Aerial Photography for Low-Cost, Ultra-high Spatial Resolution Multi-Spectral Mapping of Intertidal Landscapes. PLoS ONE 8(9): e73550. doi:10.1371/journal.pone.0073550 Camus, P., Mendez, F. J., Medina, R., Cofiño, A. S., (2011), Analysis of clustering and selection algorithms for the study of multivariate wave climate, Coastal Eng., 58, 453–462. Clapuyt, F., Vanacker, V., Van Oost, K., (2015), Reproducibility of UAV-based earth topography reconstructions based on Structure-from-Motion algorithms, Geomorphology, http://dx.doi.org/10.1016/j.geomorph.2015.05.011 Dandois, J.P., Ellis, E.C., (2013), High spatial resolution three-dimensional mapping of vegetation spectral dynamics using computer vision. Remote Sensing of Environment 136, 259–276. Fonstad, M.A., Dietrich, J.T., Courville, B.C., Jensen, J.L., Carbonneau, P.E., (2013), Topographic structure from motion: a new development in photogrammetric


measurement. Earth Surface Processes and Landforms 38, 421–430. Gibbs, A., Nolan, M.,Richmond, B. M., (2015), Evaluating changes to Arctic coastal bluffs using repeat aerial photography ‘Structure-from-Motion’ elevation models. Coastal Sediments 2015:doi:10.1142/9789814689977_0080 Goldstein, E. B., Coco, G., (2014), A machine learning approach for the prediction of settling velocity, Water Resour. Res., 50, doi:10.1002/ 2013WR015116. Harwin, S., Lucieer, A., (2012), Assessing the accuracy of georeferenced point clouds produced via multi-view stereopsis from unmanned aerial vehicle (UAV) imagery. Remote Sens., 4, 1573–1599. James, M. R., Robson, S., (2012), Straightforward reconstruction of 3D surfaces and topography with a camera: Accuracy and geoscience application, J. Geophys. Res., 117, F03017, doi:10.1029/2011JF002289. James, M. R., Robson, S., (2014), Mitigating systematic error in topographic models derived from UAV and ground-based image networks Earth Surf. Process. Landforms 39, 1413–1420. Jass, T., (2015), Environmental controls on the growth of dune-building grasses and the effect of plant morphology on coastal foredune formation. Unpubl. masters’ thesis, University of North Carolina, Chapel Hill, NC. Javernick, L., Brasington, J., Caruso, B., (2014), Modelling the topography of shallow braided rivers using Structure-from-Motion photogrammetry. Geomorphology 213, 166–182. Johnson, K., Nissen, E., Saripalli, S., Arrowsmith, J.R., McGarey, P., Scharer, K., Williams, P., Blisniuk, K., (2014), Rapid mapping of ultrafine fault zone topography with structure from motion. Geosphere. http://dx.doi.org/10.1130/ GES01017.1. GES01017.1. Keijsers J. G. S., De Groot, A.V., Riksen, M. J. P. M., (2015), Vegetation and sedimentation on coastal foredunes, Geomorphology, (228)723–734 ; http://dx.doi.org/10.1016/j.geomorph.2014.10.027 Kinsman, N., Gibbs, A., Nolan, M., (2015), Evaluation of vector coastline features extracted from ‘Structure-from-Motion’-derived elevation data. Coastal Sediments 2015: doi: 10.1142/9789814689977_0251 Lorenz, R.D., Scheidt, S.P., (2014), Compact and inexpensive kite apparatus for geomorphological field aerial photography, with some remarks on operations. GeoResJ, v.3-4: 1, p2214-2428 http://dx.doi.org/10.1016/j.grj.2014.06.001


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 Sens. 5, 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. Nolan, M., Larsen, C., Sturm, M., (2015), Mapping snow depth from manned aircraft on landscape scales at centimeter resolution using structure-from-motion photogrammetry, The Cryosphere, 9, 1445-1463, doi:10.5194/tc-9-1445-2015. Ryan, J. C., Hubbard, A. L., Box, J. E., Todd, J., Christoffersen, P., Carr, J. R., Holt, T. O., Snooke, N., (2015), UAV photogrammetry and structure from motion to assess calving dynamics at Store Glacier, a large outlet draining the Greenland ice sheet, The Cryosphere, 9, 1-11, doi:10.5194/tc-9-1-2015. Sallenger, A., Krabill, W., Swift, R., Brock, J., List, J., Hansen, M., Holman, R.A., Manizade, S., Sontag, J., Meredith, A., Morgan, K., Yunkel, J.K., Frederick, E., Stockdon, H., (2003), Evaluation of airborne scanning lidar for coastal change applications. Journal of Coastal Research 19, 125–133. Smith, M. J., Chandler, J., Rose, J., (2009), High spatial resolution data acquisition for the geosciences: Kite aerial photography, Earth Surf. Processes Landforms, 34, 155–161, doi:10.1002/esp.1702. Smith, M. W., Vericat, D., (2015), From experimental plots to experimental landscapes: topography, erosion and deposition in sub-humid badlands from Structure-from-Motion photogrammetry, Earth Surf. Processes Landforms, doi:10.1002/esp.3747 Vericat D, Brasington J, Wheaton J, Cowie M., (2009), Accuracy assessment of aerial photographs acquired using lighter-than-air blimps: low cost tools for mapping river corridors. River Research and Applications 25: 985–-1000. Vincent, J.B., Werden, L.K., Ditmer, M.A., (2015), Barriers to adding UAVs to the ecologist's toolbox. Frontiers in Ecology and the Environment 13: 74–75. http://dx.doi.org/10.1890/15.WB.002 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 applications. Geomorphology 179, 300–314.


Ground control point requirements for structure-from-motion … · 2017-01-09 · Ground control...

Documents

Transcript of Ground control point requirements for structure-from-motion … · 2017-01-09 · Ground control...