Open Collections

Landslide mapping from analysis of UAV-SFM point clouds

University of British Columbia

In recent years, unmanned aerial vehicles (UAVs) equipped with digital cameras have emerged as an inexpensive alternative to light detection and ranging (LiDAR) for mapping landslides. However, mapping with UAVs typically requires a ground control point (GCP) network to achieve higher mapping accuracies. Complex natural environments often limit the number as well as the proper distribution of GCPs. In the first part of this study, aerial imagery acquired with a quadrotor UAV was processed using structure from motion (SfM) technique to produce a three-dimensional point cloud of a large landslide involving multiple steep slopes and dense tree cover. The resulting point cloud was georeferenced with six different configurations of GCPs measured with a real-time kinetic GNSS receiver to test the influence of the number and the distribution of GCPs on mapping accuracies. Horizontal and vertical mapping accuracies of 0.058 m and 0.044 m, respectively, were achieved for the most accurate GCP configuration. A separate point cloud comparison was performed on the georeferenced point clouds to assess the effect of varying topography and tree cover on mapping accuracy. The 3D change in the natural terrain measured over a 1-year period from July 2016 to July 2017 showed movements ranging from ±0.4 m to over ±1 m at the toe of the landslide. Other parts of the landslide either remained inactive or moved less than 0.1 m. The second part of this thesis involved an accuracy comparison of five different opensource algorithms, originally developed for LiDAR data, for classification of the UAV-SfM point clouds. The influences of terrain slope, vegetation and point densities, and difficult-to-filter features on classification accuracy were also evaluated. CSF and MCC algorithms produced the lowest overall errors (4%) closely followed by LASground and FUSION (5%). All algorithms suffered in areas with densely vegetated steep slopes, understory vegetation, low point density, and low objects. Although any of the tested algorithms along with careful selection of input parameters can be used to accurately classify UAV-SfM point clouds, CSF is recommended as it is computationally efficient, does not require any preprocessing, and can process very large point clouds (>50 million points).

Text

2017-11-15

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/63604

Civil Engineering

University of British Columbia

UBCO

http://creativecommons.org/licenses/by-nc-nd/4.0/