Evaluation of Photogrammetry at Different Scales

Abstract

Accelerated erosion is amplified by anthropogenic effects, which can lead to changes in stream geometry. Stream topography measurements over time quantify stream bank erosion. The goals of this project are to evaluate topographic survey methods for costs, effort, and accuracy; to develop a stream geometry monitoring method based on photogrammetry; and to evaluate stream channel adjustment from post restoration. The objectives are to study the influence of vegetation on accuracy of a photogrammetric model, compare photogrammetric surveys to total station surveys and to evaluate replication over time using photogrammetric surveys.

Photogrammetry is a measurement technique that creates 3D surface models from multiple images. Structure-from-motion (SFM) is a type of photogrammetry where ground control points are required and can be simplified through automation. SFM was used throughout this study for model evaluation. Photogrammetry accuracy depends on image quality, number of images, image angle, and obscurity by object such as vegetation. Lab and field experiments were conducted to examine these parameters. Volume models created from a 23.0cm x 20.5cm x 16.5cm box were evaluated. Treatments included two cameras, a digital single lens reflex (DSLR) and cell phone Samsung Galaxy S4; change in camera angles, and vegetation density (ornamental English ivy). A volume model was considered accurate when it matched at least 95\% of the box volume. Models with no vegetation were accurate with a minimum of 21 images from both the Samsung Galaxy S4 and DSLR.

Model error increased and accuracy diminished with fewer images. Models accuracy decreased when less than 21 images were used to develop models of the box alone and the box with vegetation. Vegetation density was evaluated based on leaf area index (LAI) and a vegetation obscurity index. Model error was greater than 5\% at an LAI above 0.07 and a vegetation obscurity of 7\%.

Field studies with two truncated soil pyramids (1m x 1m and 2m x 2m) planted with German millet were evaluated for volume accuracy. Twenty-four images were captured for the small pyramid and 40 for the large pyramid to ensure proper overlap among images with the cell phone and DSLR. Image sets were taken weekly from week zero (no vegetation) to week four (100\% vegetation cover). At week zero both pyramids had 1.5\% error (or greater) with both cameras, and at week four there was 100\% error with both cameras. There is a positive relationship between increasing vegetation density and model error.

A stream restoration site on Parkerson Mill Creek was evaluated over time for post-construction channel morphology adjustment to further evaluate the impact of vegetation on photogrammetric model accuracy. Ground control points were installed to determine stream geometry change (i.e. erosion or deposition). A base line study comparing total station transects to digital elevation models from photogrammetry show that SFM photogrammetry is an accurate method to evaluate stream geometry. A second study was conducted to determine change over time or channel evolution within DEMs. The average root mean square error for the survey comparison is less than 1m. We conclude that SFM photogrammetry taken from ground control points at stream level by a Samsung Galaxy S4 and DSLR camera may be useful as a tool to monitor stream geometry change over time. Similar points within each of the models had an RMSE of less than 1.2m. Models accuracy may have been diminished by vegetation density or shading, field of view, and placement of ground control points.