HIGH-RESOLUTION TOPOGRAPHIC MAPPING FOR GEOLOGIC HAZARD STUDIES USING LOW-ALTITUDE AERIAL PHOTOGRAPHS AND STRUCTURE FROM MOTION SOFTWARE: METHODS, ACCURACY, AND EXAMPLES

We used low altitude aerial photographs of six study sites, taken from two platforms and processed with Structure from Motion (SfM) software, to construct digital elevation models (DEMs) and assess their accuracy. Photographs were taken with a GoPro (12 MPixel, Hero3 Black) or Sony A5100 (24 MPixel, 16 mm lens) camera mounted to a DJI Phantom II quadcopter, and at one site photographs were taken with a Canon SX230 (12 MPixel) camera mounted on a Helikite balloon. The DEMs, ~0.1 to 4.5 km² in size, were generated from point clouds of up to 1.1x10⁹ points, made with Agisoft Photoscan software running on 1 to 5 (clustered) workstations; work is underway on point clouds having >4x10⁹ points. We assessed DEM vertical accuracy relative to bare ground checkpoints (n = 36 to 71) or an airborne LiDAR-derived DEM. For four sites, we built multiple DEMs using the same photographs while varying the number (n = 3 to 19) of ground-control points (GCPs) to assess the impact of GCP abundance on DEM accuracy. GCP and checkpoint positions were measured using RTK or fast static GPS. Error in DEMs from GoPro photographs was reduced by removing lens distortion with Adobe Raw software prior to Photoscan processing. Root-mean-square error relative to checkpoints for the best DEM from each site ranged from 4.5 to 9.6 cm and 3.8 to 5.1 cm for the GoPro and Sony cameras, respectively, and was 6.5 cm for the site photographed with the Canon camera. Maximum accuracy typically was achieved with 5 to 7 GCPs, and in some cases inclusion of an additional, inaccurate GCP increased error. We have applied this methodology to a variety of geologic hazards, including using DEMs to image active landslides, as a baseline for measuring future landslide movement, and for ongoing work mapping landslide-dammed floodplains. In a highly collaborative project, we photographed > 10 km of the surface rupture from the M6.9 1983 Borah Peak earthquake in Idaho, which we processed as a contiguous point cloud using a cluster of workstations built using high-performance gaming-oriented hardware. For a study on the central segment of the San Andreas Fault, we combined DEMs with an orthophoto made by applying SfM processing to photographs from handheld cameras to map creep-induced en echelon extensional fractures that we infer accommodated 50 to 80% of creep during the 21-month study.