PE&RS June 2016 Full - page 420

accurate orthomosaic. These keypoints can also be viewed as a
dense point cloud which can be converted into a 3D represen-
tation of the Earth’s surface, such as a digital surface model.
Previous research has utilized a variety of homemade SfM
scripts to varying degrees of success. For example, Laliberte et
al. (2008) flew imagery over the Jornada Experimental Range in
southern New Mexico and used Autopano pro, a SIFT based soft-
ware, to generate keypoints. These keypoints were included in a
custom script known as PreSync which also incorporated a 1 m
digital orthoquad and a 10 m digital elevation model to adjust
and improve existing external orientation parameters (EOs). The
original images as well as the updated EOs were then put into
Leica Photogrammetric Suite to generate the mosaic. They were
able to obtain an overall RMSE of 47.9 cm which in part was due
to a lack of differential correction on their GPS unit. Turner et al.
(2012) created an automated technique using SIFT and SfM tech-
niques to automate the mosaicking of imagery collected over
two sites of an Antarctic moss bed. They were able to achieve
mean absolute total errors ranging from .103 m to 1.247 m.
A number of commercial SfM software packages have be-
come available since 2010. Two of the most popular include
Photoscan Professional (Photoscan Pro), and Pix4D Mapper
Pro (Pix4D). Both of these software have been successfully
used in current research (Vallet et al., 2011; Kung et al.,
2011a; Kung et al., 2011b; Verhoeven, 2011; Verhoeven et al.,
2012; Woodget et al., 2014).
The goal of this research was to compare commercially avail-
able software packages for use with imagery acquired by sUAS .
Specifically, this paper examines geometric accuracy, visual
quality, and price as important factors for selecting software to
process digital aerial photographs from sUAS . Two commercially
available SfM software packages, Photoscan and Pix4D were com-
pared, as well as the freely available image stitching software Im-
age Composite Editor, from the Microsoft Research Group. Previ-
ous studies have primarily focused solely on geometric accuracy
in two or three dimensions (Sona et al., 2014; Turner et al., 2014)
with Photoscan Pro producing the best results. However, these
studies did not consider the impact of processing on image qual-
ity which is important to photogrammetric applications.
Data
A total of 969 images were collected on 26 September 2014 at
Braeburn Marsh Preserve near Ann Arbor, Michigan, (~42° 16′ N,
~84° 4′ W). Imagery was collected using a Cannon EOS 6D digital
single lens reflex camera with a 50 mm fixed focal length lens at a
flight height of 100 m above ground level. The camera was mount-
ed on a Leptron Avenger airframe (Leptron, Golden, Colorado).
After the removal of low quality images, redundant images, and
turns in the flight lines (which reduce the accuracy of the final
product), only 223 images were retained for analysis, covering ap-
proximately 6.3 hectares with a spatial resolution of 1.26 cm. Due
to the low accuracy of the camera GPS (25 m), no GPS exchange-
able image file format (EXIF) data was retained for the analysis.
Ground control points (GCP ) were established prior in the
field season using a R8 GNSS RTK rover and a TCS3 hand held
unit (Trimble Navigation Limited, Sunnyvale, California) with
a recorded horizontal accuracy of <2 cm (95 percent confidence
interval). Each point was marked with a metal pole and colored
foam for easy recognition in the imagery. A total of 70 points
were collected. These ground control points align with field
plots set up previously for other studies (unpublished data).
Methods
Overview
The 223 images and 53 GCPs were used to create an image mo-
saic; the remaining 17 GCP were used for validation The image
mosaic was created using all three software packages: Photoscan
Pro, Pix4D, and ICE. The resulting mosaics were then compared
statistically based on geometric accuracy and visual quality.
Photoscan Pro
Created by the Russian-based company Agisoft in 2010,
Photoscan Pro is a 3D model and image stitching software
package. It utilizes an adapted form of SfM technology known
as the SIFT, proposed by Lowe (2004). At its core, this process
uses feature points, which are simply geometrically similar
and distinct regions in an image, e.g., building corners or
the top of light posts. These points are then tracked across
multiple images creating a series of connections between
photographs (Verhoeven, 2011). In addition, this algorithm al-
lows Photoscan Pro to automatically and accurately estimate
a large number of internal and external camera parameters
which previously had to be known and entered manually.
Utilizing such algorithms, software packages such as Pho-
toscan Pro are capable of matching images at the subpixel
level (Woodget et al., 2014). In addition, Photoscan Pro is
capable of utilizing GCP information directly in the program to
increase geometric accuracy.
The Photoscan Pro workflow can be broken down into four
basic steps: image alignment, dense point cloud formation,
mesh creation, and texture creation. Each of these steps is
run independent of each other, and aside from the inclusion
of ground control points, they can be run with little to no
user input. It is also important to note that these stages are
all independent and can be saved separately for later use or
revision. All four steps can be run in a batch process, if the
parameters are known in advance. The use of batch process-
ing decreases processing time, as well as, allows the software
to be processed overnight without requiring user intervention.
There are a number of parameters that must be defined, of-
ten with multiple options for each parameter. These parameters
give the user control over a number of crucial factors that affect
the overall quality of the final output, such as: the maximum
number of tie-points to include in the point cloud, what type
of surface the imagery consists of, and how to handle any gaps
in the final model. Trial and error in combination with a care-
ful review of the user’s manual was used to parameterize each
step. A complete list of the parameters can be seen in Table 1.
Table 1. Parameters for Photoscan Pro
Step Parameter Setting
Align photos
Accuracy
Pair preselection
Point limit
High
Disabled
40000
Build dense
point cloud
Quality
Depth filtering
Low
Mild
Build mesh
Surface type
Source data
Polygon count
Interpolation
Height field
Dense cloud
Medium
Enabled
Build texture
Mapping mode
Blending mode
Texture size
Adaptive orthophoto
Mosaic
4096
Pix4D
Pix4D is an alternative orthomosaic software created in 2011
by a Swiss company of the same name. The Pix4D workflow
consists of three steps: initial processing, point cloud den-
sification, and DSM and orthomosaic generation. The user
defined properties which guide the quality, accuracy, and
format of the final output are all handled through a process-
ing options dialogue box which must be set up prior to any
processing steps. The options in this box are refined into five
sections: initial processing, point cloud, DSM orthomosaic,
additional outputs, and resources. A complete list of the pa-
rameters can be seen in Table 2.
420 June 2016 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING