1Department of Earth Science, Utah Valley University, 800...

1
High-Resolution Topographic Mapping for Geologic Hazard Studies Using Low-Altitude Aerial Photographs and Structure from Motion Software: Methods, Accuracy, and Examples Michael Bunds 1 ([email protected]), Nathan Toké 1 , Christopher DuRoss 2 , Ryan Gold 2 , Nadine Reitman 2 , Kendra Johnson 3 , Lia Lajoie 3 , Stephen Personius 2 , Richard Briggs 2 , and Andrew Fletcher 1 1 Department of Earth Science, Utah Valley University, 800 W. University Pkwy, Orem, UT, USA 84058, 2 U.S. Geological Survey Geologic Hazards Science Center, 3 Department of Geophysics, Colorado School of Mines Aerial Photography and Processing Methods For seven sites, low altitude aerial imagery and ground control points (GCPs) were processed using Structure from Motion software to produce georeferenced point clouds, which were then rasterized into digital sur- face models (DSMs). Specifics of each step in the process are listed below. Cameras: Four camera + lens combinations were used: - GoPro Hero 3 Black Edition; 12 Mpixel fisheye lens, 103 o horizontal field of view (fov) - Sony A5100 24.3 Mpixel APS-C sensor with 16 mm pancake lens, 72 o horizontal fov - Sony A5100 24.3 Mpixel APS-C sensor with 20 mm pancake lens, 61 o horizontal fov - Canon SX230 camera, 64 o horizontal fov Aerial Platforms: - DJI Phantom 2, fixed camera mount for Sony cameras, 2 or 3 axis gimbal for GoPro - Allsop Helikite, used with Canon camera Ground Control Points: - Measured with RTK GPS/GNSS SfM Processing: - Agisoft Photoscan software - GoPro photographs were pre-processed with Adobe Raw to remove distortion (method was developed prior to release of fish eye processing mode in Photoscan) Raster Processing: - Rasters were produced from point clouds using a triangulated irregular mesh (TIN) or a binning algorithm; the method used for each model is noted in the summary table to the right. -TINs were made in Photoscan on ‘high’ or custom setting. Binning was performed using GEON Points2grid or Global Mapper. Introduction We built digital surface models (DSMs) of seven study sites using low altitude aerial imagery collected using two platforms and four camera + lens combina- tions and processed with Structure-from-Motion (SfM) software to assess DSM accuracy and to address geologic problems. For five of the sites RMS error rela- tive to surveyed bare-earth checkpoints was determined. For four of the sites we built DSMs using the same photographs while varying the number of ground control points (GCPs) used to build each DSM to evaluate the effect of GCP abundance on DSM accuracy. The impact of ground sample distance (GSD; the ground distance between centers of adjacent pixels) on DSM accura- cy was also estimated. Preliminary results derived from our aerial imagery are presented for two sites to illustrate some potential applications of UAVs in geoscience. DJI Phantom 2 UAV with Sony A5100 camera and 16 mm pan- cake lens mounted. Lost River Fault City Creek Wasatch Fault at Box Elder Cyn Oquirrh Fault at Flood Cyn Dry Lake Valley Locations of Study Sites 0 200 km 0 200 400 m Lost River Fault at Doublespring Pass Rd. Location: central portion of surface rupture of 1983 M w 6.9 Borah Peak Earthquake, Idaho Goals: Detailed mapping of surface rupture, displacement measurements, stress modeling of fault trace sections Surveyed May, 2015, in collaboration with USGS (DuRoss, Gold, Briggs, Personius) and Colo rado School of Mines (Johnson, LaJoie) Ground Control Point Checkpoint Hillshade, 060 o illumination direction Wasatch Fault Near Box Elder Cyn. Location: near the northern end of the Provo segment of the Wasatch Fault, Utah (see location map above) Goals: Map subtle fault scarps, rule out additional active fault traces; error analysis. Notes: Initiated as a class project by Michael Arnoff, October 2014 Ground Control Point Checkpoint Hillshade, 090 o illumination direction San Andreas Fault Creeping Segment at Dry Lake Valley ± 0 75 150 225 300 37.5 Meters SAF The Dry Lake Valley Paleoseismic Site Figure 9 from Titus et al., 2006 Creepmeters” BSSA, Vol. 96 No. 4B, pp S250-S268. DLV Trench is located here. ± 0 6 12 18 24 3 Meters ± 0 0.5 1 1.5 2 0.25 Meters crack sets enechelon cracks 1 meter of tape A Fracture Set along SAF at the DLV Site Nine centimeter Structure from Motion (SfM) semi-transparent slopeshade / color digital elevation model (DEM) over LiDAR slopeshade with inset one centimeter SfM slopeshade One centimeter SfM slopeshade with overlays of crack set and left-stepping opening-mode fracture mapping T1 T6 T9 T8 & T5 T4 T2 T7 Sag T3 2013 Excavation 2012 Excavation Young Fan is a westward sloping alluvial fan complex cut by stepping surface traces of the SAF. These steps have developed a series of sag ponds and tall fault scarps. We trenched both old and young fan stratigraphy as well as sag pond fill. Despite the presence of tall scarps and nine exposures, no good evidence of large ground ruptures was exposed at this site. In 2014 we returned to this site to document surface fracturing due to creep that had developed in the stiff and dry soils during the recent drought. We used structure from motion (SfM) to document the fractures and compare them with deformation structures within the 2012-2013 trenches. Structure from Motion orthophoto at the DLV Site: Left-Stepping, en-echelon surface rupture due to creep is observed in the mm-scale imagery and the cm-scale elevation data. Bunds et al., 2015 Fracture Sets Fractures Fracture Sets Fractures Results Fractures were determined to be left stepping and to occur in left stepping sets that were interpreted to be R shears. The right-lateral fault movement required to produce the fractures was estimated to be 3.3 to 5.1 cm. Total offset on the fault during the ~1.8 year time of formation of the fractures was estimated to be 5.4 cm based on an average creep rate of ca. 3 cm / yr (e.g., Burford and Harsh, 1980 and Titus et al., 2006). Thus approximately 60 to 95% of fault creep may be accounted for in brit- tle fracturing and ground rupture across a narrow zone on the SAF scarp. Surveyed September, 2014 Location: Creeping segment of the San Andreas Fault (SAF) 85 km north of Parkfield, California Background: A recent paleoseismic study at the site (Toke et al., 2015) sought to distinguish between creep and possible coseismic structures in trenches. In September, 2014, surface fracturing was identified for the first time along the SAF scarp at the site. Limited time and resources were available for field work, but we were able to obtain extensive imagery using our DJI Phantom 2 - mounted GoPro and a handheld Nikon D90 camera in one day. Goals: Generate millimeter - resolution orthophotos and < 5 cm grid DSMs of creep-induced fracture sets along the SAF scarp to enable detailed mapping of fractures, modeling of associated fault displacement, and comparison to seismic fractures from other sites. Processing notes: Models were produced at three scales. Two were made from a UAV-mounted GoPro, one from a handheld Nikon D90. The UAV - derived DSMs were used for mapping fracture sets, and an orthophoto made from the Nikon photos was used for mapping individual fractures. Only five GCPs and no checkpoints were surveyed due to lack of cellular coverage. Lost River Fault, Warm Spring Section Location: northernmost surface rupture from 1983 M w 6.9 Borah Peak Earthquake, Idaho Goals: Rupture is spillover across a segment boundary; project broadly seeks to investigate processes and characteristics of a multi-segment rupture Surveyed May, 2015, in collaboration with USGS (DuRoss, Gold, Briggs, Personius) and Colorado School of Mines (Johnson, LaJoie) Processing notes: Project was divided into 15 chunks, for which cameras were aligned separately. Chunks were merged in Photoscan and a single point cloud was generated for entire project using ‘high’ setting. A cluster of three workstations was used to process the dense point cloud. The point cloud was exported from Photoscan in tiles, the tiles were individually binned into rasters in Global Mapper and then made into a mosaic in ArcMap. In addition, a point cloud and DSM was generated for each chunk individually; RMSe for that model is given. Ground Control Point Checkpoint 0 500 1000 m Hillshade, 060 o illumination direction Wasatch Fault at Box Elder Cyn City Creek Location: North end of Salt Lake Valley, in footwall of Wasatch Fault (see location map below); steep hillside Goals: Map landslides, error analysis, compare to 0.5 m LiDAR. Notes: Model was initiated as a class project by Brandon Powell, October 2014 City Creek Oquirrh Fault at Flood Cyn Ground Control Point Checkpoint Slopeshade References Bunds, M.P., Toke, N., Lawrence, A., Arrowsmith, J.R., and Salisbury, J.B., 2015, Insights into Surface Manifestation of Aseismic vs. Coseismic Strike-Slip Faulting from UAV Imagery of Creep-Induced Surface Fracturing Along the Central San Andreas Fault, 2015 AGU Fall Meeting Program with Abstracts. Burford, R. O., and P. W. Harsh, 1980, Slip on the San Andreas fault in central California from alignment array surveys, Bull. Seism. Soc. Am. 70, 1233–1261. Doser, D.I., and Smith, R.B., 1985, Source Parameter of the 28 October 1983 Borah Peak, Idaho Earthquake from Body Wave Analysis, Bulletin of the Seismological Society of America, v.75, p.1041-1051. Johnson, K., Nissen, E., Saripalli, S., Arrowsmith, J.R., McGarey, P., Scharer, K., Williams, P., and Blisniuk, K., 2014, Rapid mapping of ultra-fine fault zone topography with structure from motion, Geosphere, v.10, doi:10.1130/GES01017.1. Reitman, N.G., Bennett, S.E.K., Gold, R.D., Briggs, R.W., and DuRoss, C.B., 2015, High-Resolution Trench Photomosaics from Image-Based Modeling: Workflow and Error Analysis, Bulletin of the Seismological Society of America, v.105. Titus, S.J., C. DeMets, and B. Tikoff, 2006, Thirty-five-year creep rates for the creeping segment of the San Andreas fault and the effects of the 2004 Parkfield earthquake: constraints from alignment arrays, continuous GPS, and creepmeters: Bulletin of the Seismological Society of America, v. 96, no. 4B, p. S250-S268 doi: 10.1785/0120050811 Toke, N., Bunds, M.P., Salisbury, J.B., Lawrence, A., and Arrowsmith, J.R., 2015, Rupture Patterns Due to Aseismic Creep During the 2013-2014 Drought Match Deformation Structures Observed in the 5000 Year Dry Lake Valley Paleoseismic Record, 2015 SCEC Annual Meeting Program with Abstracts, v. 25, p.189. U.S. Geological Survey and California Geological Survey,Idaho Geological Survey, and Utah Geological Survey, 2006, Quaternary fault and fold database for the United States, accessed Jan 9, 2006, from USGS web site: http//earthquake.usgs.gov/hazards/qfaults/. Warm Spring Section Doublespring Pass Rd. Oquirrh Fault at Flood Canyon Location: near the northern end of the Oquirrh Fault, Utah (see location map above) Goals: Map fault trace; measure offset of Lake Bonneville shorelines to estimate fault slip rate, error analysis. Notes: Site was surveyed as three parts: northern, mid, and southern. The northern and mid portions were photographed using the Sony A5100 camera on October 11, 2015. The southern portion was photographed using the Hero GoPro in October 2014. Initiated as class projects by Andrew Fletcher, Jeremy Andreini, Michael Arnold, Kenneth Larson. Preliminary Results Cumulative net vertical displacement (NVD) of the fault was estimated by measuring the elevation of five offset Lake Bonneville shoreline features in the footwall and hangingwall of the fault. The measured features are the Bonneville highstand bench, Provo shoreline bench, and three sub-Provo shorelines. The highstand bench elevation was estimated by constructing profiles perpendicular to the bench, fitting least-squares best-fit lines to the bench and wave-cut face on each profile, then calculating the point of intersection of the bench and face. The Provo and three sub-Provo shoreline elevations were estimated by con- structing shoreline-parallel profiles along each shoreline and fitting lines to the shorelines in the footwall and hangingwall. The shoreline - parallel lines were projected to meet a line fit to the fault scarp, and NVD was taken to be the difference in elevation of the shoreline - parallel features at the scarp face. Preliminary Estimated net vertical displacements Highstand: 3.2 - 3.6 m NVD Sub-provo level 2: 6.7 m NVD Provo level: 3.2 m NVD Sub-provo level 3: 8.5 m NVD Sub-provo level 1: 2.8 m NVD Mid 1560 1570 1580 1590 1600 1610 1620 1630 Profile 5 Bonneville highstand footwall 1575 1580 1585 1590 1595 1600 1605 1610 1615 Profile 7 Bonneville Highstand hangingwall modified? 1471 1472 1473 1474 1475 1476 1477 1478 1479 Profile 4 Provo level 1460 1461 1462 1463 1464 1465 200 250 150 100 50 0 Profile 3 Sub-Provo level 1 1440 1442 1444 1446 1448 1450 1452 Profile 2 Sub-Provo level 2 1430 1432 1434 1436 1438 1440 Profile 1 Sub-Provo level 3 0 250 500 m 1550 1560 1570 1580 1590 1600 1610 1620 1630 1640 Profile 6 Bonneville Highstand footwall 1580 1585 1590 1595 1600 1605 1610 1615 Profile 8 Bonneville Highstand hangingwall South North p1 8.5 m NVD p2 6.7 m NVD p3 2.8 m NVD p4 3.2 m NVD p5 1591.5 m elev p6 1591.8 m elev p7 1589.0 m elev? p8 1588.3 m elev 0 25 125 225 325 meters elev (m) 0 100 200 0 120 40 0 100 200 300 80 120 40 0 80 130 50 10 90 Ground Control Point Checkpoint Hillshade, 240 o illumination direction Preliminary Lake Bench and Fault Scarp Profiles 8.5 m NVD 6.7 m NVD 2.8 m NVD 3.5 m NVD 1591.5 m elev 1591.8 m elev 1589.0 m elev 1588.3 m elev SfM - Derived DSM Accuracy Conclusions Ground Control Points (GCPs): Our results that five to seven GCPs can be sufficient match those of Reitman et al. (2015), and show the result can be extrapolated to large mapped areas. Furthermore, in some cases the addition of a GCP can increase error, suggesting that GCP accuracy is more important that the quantity of GCPs Ground Sample Distance (GSD): DSM error is correlated with GSD; higher resolution photo- graphs tend to produce more accurate DSMs. GSD should be considered in designing sur- veys. Cameras: We have produced slightly higher accuracy DSMs from the 24.3 Mpixel Sony A5100 than from the GoPro; the difference may be greater at larger GSD. Still, very good results were obtained from the GoPro, which is a lightweight, cost-effective option for collecting low alti- tude imagery. Photograph overlap: We have not systematically assessed the impact of overlap on model accuracy. However, in three cases (City Creek, Flood Canyon, Loading Dock), models were built using more photos than used in the models presented here, with no reduction in error. It is likely that overlap is greater than necessary in some or all of the models presented here. 0 1 2 3 4 5 6 7 8 9 10 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 Ground sample distance (cm) RMS error (cm) DSM Accuracy Assessment Methods Vertical accuracy of DSMs was assessed relative to bare-ground checkpoints measured with RTK GNSS/GPS. Checkpoint residuals were calculated by extracting SfM-derived DSM eleva- tions at the location of the checkpoints and subtracting the checkpoint elevations from the associated DSM elevation. RMS error of the residuals for a model were calculated as: RMSe = Σ i=1 n r i 2 n 1 r i = residual of i th checkpoint = (DSM elevation minus checkpoint elevation) n = number of checkpoints RMSe is presented in two ways: 1. RMSe was calculated and is shown for the optimal DSM made for five surveyed sites. These RMSe values are plotted against the average ground sample distance (GSD) for each model to evaluate the effect of GSD on DSM accuracy. GSD is the ground distance between adjacent photograph pixels. 2. For four surveyed sites, multiple DSMs were generated using the same photographs while varying number of GCPs (n = 3 to 19), and RMSe was calculated for each DSM to evaluate the effect of GCP abundance on RMSe. 0 0.05 0.1 0.15 0.2 0.25 0.3 2 4 6 8 10 12 14 16 18 20 Vertical RMS error (m) Number of GCPs used in model City Creek 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 2 4 6 8 10 12 14 16 Vertical RMS error (m) Number of GCPs used in model Lost River Flt at Doublespring Pass Rd. 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 2 4 6 8 10 Number of GCPs used in model Vertical RMS error (m) Wasatch Flt at Box Elder Cyn 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3 1.4 1.5 Vertical RMS error (m) Oquirrh Flt at Flood Canyon Number of GCPs used in model 2 4 6 8 10 12 14 16 5 pts truncated DSM Accuracy Assessment Model Parameters City Creek Dry Lake Valley Loading Dock (map/hillshade not shown) Lost River Fault at Doublespring Pass Rd. Lost River Fault, Warm Spring section Oquirrh Fault Wasatch Fault at Box Elder Cyn Small scale Med. Scale Large scale GoPro Sony A5100, 16mm Sony A5100, 20mm Optimal GCP vs RMSe North Mid South RMSe (cm) 7.4 n/a n/a n/a 2.9 3.23 3.1 3.8 3.8 5.8 - 6.4 5.4 6.3 9.5 4.3 Mapped area (km 2 ) 0.09 0.019 0.019 0.00018 0.010 0.23 0.12 0.90 0.90 4.6 1.23 1.1 0.85 0.10 No. of Cameras 177 184 171 377 92 211 198 1399 1399 12567 1545 1464 430 187 Camera type GoPro GoPro GoPro Nikon D90 GoPro Sony A5100 Sony A5100 Sony A5100 Sony A5100 Sony A5100 Sony A5100 Sony A5100 GoPro GoPro Mean camera altitude (m agl) 47 61 19.3 1.6 19.2 107.3 95.9 100.1 100.1 92.6 134.8 124.4 78.1 52.3 GSD (cm/pixel) 2.9 3.3 1.0 0.042 0.010 2.4 1.68 2.27 2.27 2.1 3.1 2.8 4.1 2.7 Dense cloud setting high high high high High High High high medium high medium high high high Point cloud size 39 x 10 6 51 x 10 6 48 x 10 6 254 x10 6 27 x10 6 115 x10 6 623 x 10 6 131 x 10 6 4,410 x 10 6 92 x 10 6 461 x 10 6 142 x 10 6 46 x 10 6 Mean point density (pts/m 2 ) 460 268 2515 1409560 2471 500 693 146 958.7 74.8 415.1 166.8 358 Mean point spacing (cm) 4.8 6.1 2.0 0.08 1.99 4.47 3.8 8.3 3.2 11.6 4.9 7.7 6 No. of GCPs 19 5 5* 5* 12 7 7 15 15 141 (9.4 per chunk) 9 8 16 9 Overlap (images/GCP) 46.6 16.9 22.3 12 33.4 28.3 25.1 17 17 39.8 49.6 38.6 11.9 26.3 Rasterization method binning TIN TIN TIN binning TIN TIN TIN TIN binning binning binning binning binning DSM grid (cm) 6 9 1 0.01 5 5 5 5 10 5 10 5 10 6 No. of checkpoints 52 n/a n/a n/a 53 53 53 71 71 98 36 33 43 36 118 x 10 6 515 4.4 Ground Sample Distance vs RMSe Summary of Structure from Motion / DSM Model Parameters Approx. GPS resolution Approx. GPS resolution Approx. GPS resolution Approx. GPS resolution pts truncated

Transcript of 1Department of Earth Science, Utah Valley University, 800...

Page 1: 1Department of Earth Science, Utah Valley University, 800 ...research.uvu.edu/Bunds/GSA_poster_2015_final_compressed.pdf · High-Resolution Topographic Mapping for Geologic Hazard

High-Resolution Topographic Mapping for Geologic Hazard Studies Using Low-Altitude Aerial Photographs and Structure from Motion Software: Methods, Accuracy, and ExamplesMichael Bunds1 ([email protected]), Nathan Toké1, Christopher DuRoss2, Ryan Gold2, Nadine Reitman2, Kendra Johnson3, Lia Lajoie3, Stephen Personius2, Richard Briggs2, and Andrew Fletcher1

1Department of Earth Science, Utah Valley University, 800 W. University Pkwy, Orem, UT, USA 84058, 2U.S. Geological Survey Geologic Hazards Science Center, 3Department of Geophysics, Colorado School of Mines

Aerial Photography and ProcessingMethods

For seven sites, low altitude aerial imagery and ground control points (GCPs) were processed using Structure from Motion software to produce georeferenced point clouds, which were then rasterized into digital sur-face models (DSMs). Speci�cs of each step in the process are listed below.

Cameras: Four camera + lens combinations were used: - GoPro Hero 3 Black Edition; 12 Mpixel �sheye lens, 103o horizontal �eld of view (fov) - Sony A5100 24.3 Mpixel APS-C sensor with 16 mm pancake lens, 72o horizontal fov - Sony A5100 24.3 Mpixel APS-C sensor with 20 mm pancake lens, 61o horizontal fov - Canon SX230 camera, 64o horizontal fov

Aerial Platforms: - DJI Phantom 2, �xed camera mount for Sony cameras, 2 or 3 axis gimbal for GoPro - Allsop Helikite, used with Canon camera

Ground Control Points: - Measured with RTK GPS/GNSS

SfM Processing: - Agisoft Photoscan software - GoPro photographs were pre-processed with Adobe Raw to remove distortion (method was developed prior to release of �sh eye processing mode in Photoscan)

Raster Processing: - Rasters were produced from point clouds using a triangulated irregular mesh (TIN) or a binning algorithm; the method used for each model is noted in the summary table to the right. -TINs were made in Photoscan on ‘high’ or custom setting. Binning was performed using GEON Points2grid or Global Mapper.

IntroductionWe built digital surface models (DSMs) of seven study sites using low altitude aerial imagery collected using two platforms and four camera + lens combina-tions and processed with Structure-from-Motion (SfM) software to assess DSM accuracy and to address geologic problems. For �ve of the sites RMS error rela-tive to surveyed bare-earth checkpoints was determined. For four of the sites we built DSMs using the same photographs while varying the number of ground control points (GCPs) used to build each DSM to evaluate the e�ect of GCP abundance on DSM accuracy. The impact of ground sample distance (GSD; the ground distance between centers of adjacent pixels) on DSM accura-cy was also estimated. Preliminary results derived from our aerial imagery are presented for two sites to illustrate some potential applications of UAVs in geoscience.

DJI Phantom 2 UAV with Sony A5100 camera and 16 mm pan-cake lens mounted.

Lost River Fault

City Creek

Wasatch Fault atBox Elder Cyn

Oquirrh Fault atFlood Cyn

Dry Lake Valley

Locations of Study Sites

0 200 km

0 200 400 m

Lost River Fault at Doublespring Pass Rd.Location: central portion of surface rupture of 1983 Mw6.9 Borah Peak Earthquake, Idaho

Goals: Detailed mapping of surface rupture, displacement measurements, stress modeling of fault trace sectionsSurveyed May, 2015, in collaboration with USGS (DuRoss, Gold, Briggs, Personius) and Colo rado School of Mines (Johnson, LaJoie)

Ground Control Point

Checkpoint

Hillshade, 060o illumination direction

Wasatch Fault Near Box Elder Cyn.Location: near the northern end of the Provo segment of the Wasatch Fault, Utah (see location map above)

Goals: Map subtle fault scarps, rule out additional active fault traces; error analysis.

Notes: Initiated as a class project by Michael Arno�, October 2014

Ground Control Point

Checkpoint

Hillshade, 090o illumination direction

San Andreas Fault Creeping Segment at Dry Lake Valley

± 0 75 150 225 30037.5Meters

crack sets

enechelon cracks

SAF

The Dry Lake Valley Paleoseismic Site

Figure 9 from Titus et al., 2006

Creepmeters” BSSA, Vol. 96 No. 4B, pp S250-S268.

DLV

Tre

nch

is lo

cate

d he

re.

± 0 6 12 18 243Meters

crack sets

enechelon cracks

± 0 0.5 1 1.5 20.25Meters

crack sets

enechelon cracks

1 meter of tape

A Fracture Set along SAF at the DLV Site

Nine centimeter Structure from Motion (SfM) semi-transparent slopeshade / color digital elevation model (DEM) over LiDAR slopeshade with inset one centimeter SfM slopeshade

One centimeter SfM slopeshade with overlays of crack set and left-stepping opening-mode fracture mapping

T1T6

T9

T8 & T5

T4

T2

T7

Sag

T3

2013 Excavation

2012 Excavation

Young Fan

is a westward sloping alluvial fan complex cut by stepping surface traces of the SAF. These steps have developed a series of sag ponds and tall fault scarps. We trenched both old and young fan stratigraphy as well as sag pond �ll. Despite the presence of tall scarps and nine exposures, no good evidence of large ground ruptures was exposed at this site. In2014 we returned to this site to document surface fracturing due to creepthat had developed in the sti� and dry soils during the recent drought. We used structure from motion (SfM) to document the fractures and compare them with deformation structures within the 2012-2013 trenches.

Structure from Motion orthophoto at the DLV Site: Left-Stepping, en-echelon surface rupture due to creep is observed in the mm-scale imagery and the cm-scaleelevation data.

Bunds et al., 2015

Fracture Sets

Fractures

Fracture Sets

Fractures

Results

Fractures were determined to be left stepping and to occur in left stepping sets that were interpreted to be R shears. The right-lateral fault movement required to produce the fractures was estimated to be 3.3 to 5.1 cm. Total o�set on the fault during the ~1.8 year time of formation of the fractures was estimated to be 5.4 cm based on an average creep rate of ca. 3 cm / yr (e.g., Burford and Harsh, 1980 and Titus et al., 2006). Thus approximately 60 to 95% of fault creep may be accounted for in brit-tle fracturing and ground rupture across a narrow zone on the SAF scarp.

Surveyed September, 2014

Location: Creeping segment of the San Andreas Fault (SAF) 85 km north of Park�eld, California

Background: A recent paleoseismic study at the site (Toke et al., 2015) sought to distinguish between creep and possible coseismic structures in trenches. In September, 2014, surface fracturing was identi�ed for the �rst time along the SAF scarp at the site. Limited time and resources were available for �eld work, but we were able to obtain extensive imagery using our DJI Phantom 2 - mounted GoPro and a handheld Nikon D90 camera in one day.

Goals: Generate millimeter - resolution orthophotos and < 5 cm grid DSMs of creep-induced fracture sets along the SAF scarp to enable detailed mapping of fractures, modeling of associated fault displacement, and comparison to seismic fractures from other sites.

Processing notes: Models were produced at three scales. Two were made from a UAV-mounted GoPro, one from a handheld Nikon D90. The UAV - derived DSMs were used for mapping fracture sets, and an orthophoto made from the Nikon photos was used for mapping individual fractures. Only �ve GCPs and no checkpoints were surveyed due to lack of cellular coverage.

Lost River Fault, Warm Spring SectionLocation: northernmost surface rupture from 1983 Mw6.9 Borah Peak Earthquake, Idaho

Goals: Rupture is spillover across a segment boundary; project broadly seeks to investigate processes and characteristics of a multi-segment rupture

Surveyed May, 2015, in collaboration with USGS (DuRoss, Gold, Briggs, Personius) and Colorado School of Mines (Johnson, LaJoie)

Processing notes: Project was divided into 15 chunks, for which cameras were aligned separately. Chunks were merged in Photoscan and a single point cloud was generated for entire project using ‘high’ setting. A cluster of three workstations was used to process the dense point cloud. The point cloud was exported from Photoscan in tiles, the tiles were individually binned into rasters in Global Mapper and then made into a mosaic in ArcMap. In addition, a point cloud and DSM was generated for each chunk individually; RMSe for that model is given.

Ground Control Point

Checkpoint0 500 1000 m

Hillshade, 060o illumination direction

Wasatch Fault atBox Elder Cyn

City CreekLocation: North end of Salt Lake Valley, in footwall of Wasatch Fault (see location map below); steep hillside

Goals: Map landslides, error analysis, compare to 0.5 m LiDAR.

Notes: Model was initiated as a class project by Brandon Powell, October 2014

City Creek

Oquirrh Fault atFlood Cyn

Ground Control Point

Checkpoint

Slopeshade

ReferencesBunds, M.P., Toke, N., Lawrence, A., Arrowsmith, J.R., and Salisbury, J.B., 2015, Insights into Surface Manifestation of Aseismic vs. Coseismic Strike-Slip Faulting from UAV Imagery of Creep-Induced Surface

Fracturing Along the Central San Andreas Fault, 2015 AGU Fall Meeting Program with Abstracts.Burford, R. O., and P. W. Harsh, 1980, Slip on the San Andreas fault in central California from alignment array surveys, Bull. Seism. Soc. Am. 70, 1233–1261. Doser, D.I., and Smith, R.B., 1985, Source Parameter of the 28 October 1983 Borah Peak, Idaho Earthquake from Body Wave Analysis, Bulletin of the Seismological Society of America, v.75, p.1041-1051.Johnson, K., Nissen, E., Saripalli, S., Arrowsmith, J.R., McGarey, P., Scharer, K., Williams, P., and Blisniuk, K., 2014, Rapid mapping of ultra-�ne fault zone topography with structure from motion, Geosphere, v.10,

doi:10.1130/GES01017.1.Reitman, N.G., Bennett, S.E.K., Gold, R.D., Briggs, R.W., and DuRoss, C.B., 2015, High-Resolution Trench Photomosaics from Image-Based Modeling: Work�ow and Error Analysis, Bulletin of the Seismological

Society of America, v.105.Titus, S.J., C. DeMets, and B. Tiko�, 2006, Thirty-�ve-year creep rates for the creeping segment of the San Andreas fault and the e�ects of the 2004 Park�eld earthquake: constraints from alignment arrays,

continuous GPS, and creepmeters: Bulletin of the Seismological Society of America, v. 96, no. 4B, p. S250-S268 doi: 10.1785/0120050811 Toke, N., Bunds, M.P., Salisbury, J.B., Lawrence, A., and Arrowsmith, J.R., 2015, Rupture Patterns Due to Aseismic Creep During the 2013-2014 Drought Match Deformation Structures Observed in the 5000 Year

Dry Lake Valley Paleoseismic Record, 2015 SCEC Annual Meeting Program with Abstracts, v. 25, p.189.U.S. Geological Survey and California Geological Survey,Idaho Geological Survey, and Utah Geological Survey, 2006, Quaternary fault and fold database for the United States, accessed Jan 9, 2006, from USGS

web site: http//earthquake.usgs.gov/hazards/qfaults/.

Warm Spring Section

Doublespring Pass Rd.

Oquirrh Fault at Flood CanyonLocation: near the northern end of the Oquirrh Fault, Utah (see location map above)

Goals: Map fault trace; measure o�set of Lake Bonneville shorelines to estimate fault slip rate, error analysis.

Notes: Site was surveyed as three parts: northern, mid, and southern. The northern and mid portions were photographed using the Sony A5100 camera on October 11, 2015. The southern portion was photographed using the Hero GoPro in October 2014. Initiated as class projects by Andrew Fletcher, Jeremy Andreini, Michael Arnold, Kenneth Larson.

Preliminary ResultsCumulative net vertical displacement (NVD) of the fault was estimated by measuring the elevation of �ve o�set Lake Bonneville shoreline features in the footwall and hangingwall of the fault. The measured features are the Bonneville highstand bench, Provo shoreline bench, and three sub-Provo shorelines. The highstand bench elevation was estimated by constructing pro�les perpendicular to the bench, �tting least-squares best-�t lines to the bench and wave-cut face on each pro�le, then calculating the point of intersection of the bench and face. The Provo and three sub-Provo shoreline elevations were estimated by con-structing shoreline-parallel pro�les along each shoreline and �tting lines to the shorelines in the footwall and hangingwall. The shoreline - parallel lines were projected to meet a line �t to the fault scarp, and NVD was taken to be the di�erence in elevation of the shoreline - parallel features at the scarp face.

Preliminary Estimated net vertical displacementsHighstand: 3.2 - 3.6 m NVD Sub-provo level 2: 6.7 m NVDProvo level: 3.2 m NVD Sub-provo level 3: 8.5 m NVDSub-provo level 1: 2.8 m NVD

Mid

1560

1570

1580

1590

1600

1610

1620

1630

Pro�le 5 Bonneville highstand footwall

1575

1580

1585

1590

1595

1600

1605

1610

1615Pro�le 7 Bonneville Highstand hangingwall

modi�ed?

1471

1472

1473

1474

1475

1476

1477

1478

1479

Pro�le 4 Provo level

1460

1461

1462

1463

1464

1465

200 250150100500

Pro�le 3 Sub-Provo level 1

1440

1442

1444

1446

1448

1450

1452

Pro�le 2 Sub-Provo level 2

1430

1432

1434

1436

1438

1440 Pro�le 1 Sub-Provo level 3

0 250 500 m

1550

1560

1570

1580

1590

1600

1610

1620

1630

1640Pro�le 6 Bonneville Highstand footwall

1580

1585

1590

1595

1600

1605

1610

1615Pro�le 8 Bonneville Highstand hangingwall

South

North

p1 8.5 m NVD

p2 6.7 m NVD

p3 2.8 m NVD

p4 3.2 m NVD

p5 1591.5 m elev

p6 1591.8 m elev

p7 1589.0 m elev?

p8 1588.3 m elev

0 25 125 225 325meters

elev (

m)

0 100 200

0

120400

100 200

300

80

120400 80

1305010 90

Ground Control Point

Checkpoint

Hillshade, 240o illumination direction

Preliminary Lake Bench and Fault Scarp Pro�les

8.5 m NVD6.7 m NVD

2.8 m NVD3.5 m NVD

1591.5 m elev 1591.8 m elev

1589.0 m elev1588.3 m elev

SfM - Derived DSM Accuracy ConclusionsGround Control Points (GCPs): Our results that �ve to seven GCPs can be su�cient match those of Reitman et al. (2015), and show the result can be extrapolated to large mapped areas. Furthermore, in some cases the addition of a GCP can increase error, suggesting that GCP accuracy is more important that the quantity of GCPs

Ground Sample Distance (GSD): DSM error is correlated with GSD; higher resolution photo-graphs tend to produce more accurate DSMs. GSD should be considered in designing sur-veys.

Cameras: We have produced slightly higher accuracy DSMs from the 24.3 Mpixel Sony A5100 than from the GoPro; the di�erence may be greater at larger GSD. Still, very good results were obtained from the GoPro, which is a lightweight, cost-e�ective option for collecting low alti-tude imagery.

Photograph overlap: We have not systematically assessed the impact of overlap on model accuracy. However, in three cases (City Creek, Flood Canyon, Loading Dock), models were built using more photos than used in the models presented here, with no reduction in error. It is likely that overlap is greater than necessary in some or all of the models presented here.

0

1

2

3

4

5

6

7

8

9

10

0 0.5 1 1.5 2 2.5 3 3.5 4 4.5Ground sample distance (cm)

RMS

erro

r (cm

)

DSM Accuracy Assessment MethodsVertical accuracy of DSMs was assessed relative to bare-ground checkpoints measured with RTK GNSS/GPS. Checkpoint residuals were calculated by extracting SfM-derived DSM eleva-tions at the location of the checkpoints and subtracting the checkpoint elevations from the associated DSM elevation. RMS error of the residuals for a model were calculated as:

RMSe = Σi=1

n ri2

n1 ri = residual of ith checkpoint

= (DSM elevation minus checkpoint elevation)n = number of checkpoints

RMSe is presented in two ways:

1. RMSe was calculated and is shown for the optimal DSM made for �ve surveyed sites. These RMSe values are plotted against the average ground sample distance (GSD) for each model to evaluate the e�ect of GSD on DSM accuracy. GSD is the ground distance between adjacent photograph pixels.

2. For four surveyed sites, multiple DSMs were generated using the same photographs while varying number of GCPs (n = 3 to 19), and RMSe was calculated for each DSM to evaluate the e�ect of GCP abundance on RMSe.

0

0.05

0.1

0.15

0.2

0.25

0.3

2 4 6 8 10 12 14 16 18 20

Vert

ical

RM

S er

ror (

m)

Number of GCPs used in model

City Creek

0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

2 4 6 8 10 12 14 16

Vert

ical

RM

S er

ror (

m)

Number of GCPs used in model

Lost River Flt at Doublespring Pass Rd.

0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

2 4 6 8 10

Number of GCPs used in model

Vert

ical

RM

S er

ror (

m)

Wasatch Flt at Box Elder Cyn

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

1.1

1.2

1.3

1.4

1.5

Vert

ical

RM

S er

ror (

m)

Oquirrh Flt at Flood Canyon

Number of GCPs used in model2 4 6 8 10 12 14 16

5 pts truncated

DSM Accuracy Assessment

Model Parameters

City Creek

Dry Lake Valley Loading Dock (map/hillshade not shown)

Lost River Fault at Doublespring Pass

Rd.

Lost River Fault, Warm Spring section

Oquirrh Fault Wasatch Fault at

Box Elder Cyn

Small scale

Med. Scale

Large scale

GoPro Sony A5100, 16mm

Sony A5100, 20mm

Optimal GCP vs RMSe

North Mid South

RMSe (cm) 7.4 n/a n/a n/a 2.9 3.23 3.1 3.8 3.8 5.8 - 6.4 5.4 6.3 9.5 4.3

Mapped area (km2) 0.09 0.019 0.019 0.00018 0.010 0.23 0.12 0.90 0.90 4.6 1.23 1.1 0.85 0.10

No. of Cameras 177 184 171 377 92 211 198 1399 1399 12567 1545 1464 430 187

Camera type GoPro GoPro GoPro Nikon D90 GoPro

Sony A5100

Sony A5100

Sony A5100

Sony A5100

Sony A5100

Sony A5100

Sony A5100 GoPro GoPro

Mean camera altitude (m agl) 47 61 19.3 1.6 19.2 107.3 95.9 100.1 100.1 92.6 134.8 124.4 78.1 52.3

GSD (cm/pixel) 2.9 3.3 1.0 0.042 0.010 2.4 1.68 2.27 2.27 2.1 3.1 2.8 4.1 2.7

Dense cloud setting high high high high High High High high medium high medium high high high

Point cloud size 39 x 106 51 x 106 48 x 106 254 x106 27 x106 115 x106 623 x 106 131 x 106 4,410 x 106 92 x 106 461 x 106 142 x 106 46 x 106

Mean point density (pts/m2) 460 268 2515 1409560 2471 500 693 146 958.7 74.8 415.1 166.8 358

Mean point spacing (cm) 4.8 6.1 2.0 0.08 1.99 4.47 3.8 8.3 3.2 11.6 4.9 7.7 6

No. of GCPs 19 5 5* 5* 12 7 7 15 15 141 (9.4

per chunk) 9 8 16 9

Overlap (images/GCP) 46.6 16.9 22.3 12 33.4 28.3 25.1 17 17 39.8 49.6 38.6 11.9 26.3

Rasterization method binning TIN TIN TIN binning TIN TIN TIN TIN binning binning binning binning binning

DSM grid (cm) 6 9 1 0.01 5 5 5 5 10 5 10 5 10 6

No. of checkpoints 52 n/a n/a n/a 53 53 53 71 71 98 36 33 43 36

118 x 106

515

4.4

Ground Sample Distance vs RMSe

Summary of Structure from Motion / DSM Model Parameters

Approx. GPSresolution

Approx. GPSresolution

Approx. GPSresolution

Approx. GPS resolution

pts truncated