PE&RS January 2016 - page 32

inter-dependent and their optimization is critical to the
global DEM quality and process efficiency.
3. Data handling can be challenging after merging (Baraz-
zetti et al., 2013). A fit-for-purpose analysis has to be
undertaken before data recording to ensure the chosen
measurement resolution suits the tasks ahead. This
pre-analysis is becoming more critical, as it reduces
the computing and data storage demands, whilst it also
ensures appropriate data recording. Correct resampling
might allow for more efficient data handling, without
losing critical information.
State of the Art
Specialist photogrammetric and GIS software packages, such
as IMAGINE Photogrammetry® (previously LPS/OrthoMAX)
by Leica Geosystems, ENVI®, and Esri’s ArcGIS®, comprise a
built-in DEM merging function. Techniques commonly referred
to as “feathering” are used. Weighted-averaging is generally
used to overlap data. By smoothing (“feathering”) the overlap
region, any inconsistencies at the boundaries are reduced.
However, it is important to note that this process does not
remove the systematic errors that caused the discontinuities
in the first place (Costantini et al. , 2006).
Gallant and Austin (2009) used Esri’s “mosaic” function
to merge land and bathymetric datasets obtained in South
Australia. They showed that merging could be automated, but
only by relying on the DEMs overlap. Otherwise, the process
was described as “manually intensive, requiring a great deal
of intervention, judgement, analysis and editing.” Stojic et al.
(1998), Chandler et al. (2001 and 2002) and Wackrow (2008)
used OrthoMAX/LPS to merge overlapping DEMs obtained
with close-range stereo photogrammetry. Wackrow (2008)
reportedly used an overlap of 60 percent the size of the input
DEMs, which was the same overlap distance between two im-
ages forming a stereo pair. Despite such a high data redundan-
cy, discontinuities were still clearly visible after DEM merging
(Stojic et al., 1998; Wackrow, 2008). Reasons for the failure
were not discussed. Registration errors were pointed out as
being responsible for elevation discontinuities after merg-
ing land and bathymetric raster DEMs of Tampa Bay, Florida
(Medeiros et al., 2011). Postprocessing with ArcGIS® was
necessary to reduce the significance of the seam. Marzahn et
al. (2012) used LPS for the measurement of soil surface rough-
ness, but preferred a self-programmed merging technique to
increase the surface coverage in their study. The technique
was simply described as “a postprocessing step using image
matching techniques,” requiring a 30 percent overlap between
successive DEMs to merge. Unfortunately, the merging tech-
nique, results and evaluation were not presented.
With the recent availability of free-to-use wide-area (also
called global) DEMs, such as those obtained using Advanced
Spaceborne Thermal Emission and Reflection Radiometer
(ASTER) and Shuttle Radar Topographic Mission (SRTM), data
fusion is receiving growing attention in the scientific commu-
nity. In those cases, DEM s were generally collected using differ-
ent techniques, and subsequent data fusion aims to use redun-
dant information to obtain a more accurate surface estimation
(Papasaika et al., 2011). Although the motivations for data
fusion are not the same as for data merging, some data fusion
strategies can be transferred to DEM merging. Costantini et al.
(2006) proposed a data fusion algorithm to merge large-scale
DEMs of a test site in central Italy, originating from various data
sources. Similar to other data fusion techniques (Papasaika et
al., 2011; Papasaika et al., 2008; Schindler et al., 2011; Tran
et al., 2014), their method exploits the redundant information
contained in the area of overlap between different DEMs in
order to reduce the horizontal and vertical systematic errors.
This resembles a 3D co-registration of the individual DEMs, an
essential step for creating a seamless, merged product (Gesch
and Wilson, 2002). After alignment of the individual DEMs,
redundant elevation data are processed by means of averaging.
Weighted-averaging is the favored method when various data
sources are used. Weights can be determined based on the
theoretical or measured accuracy of the input DEMs, as well as
characteristics of the surface, such as slope and roughness.
Novel Structure-from-Motion (SfM) and Multi-View Stereo
(MVS) allow for DEM reconstruction from more than two images,
and therefore would be a viable solution to the problem inves-
tigated in this study (Dowling et al., 2009; Fonstad et al., 2013;
James and Robson, 2014; Javernick et al., 2014; Ouédraogo et
al., 2014; Stumpf et al., 2015; Westoby et al., 2012). However,
SfM/MVS currently lack the capability of traditional close-range
stereo photogrammetry in recording high-quality small-scale
DEMs, as is necessary for grain-scale fluvial roughness charac-
terization. Furthermore, it has been shown that DEMs col-
lected with these techniques may suffer from large non-linear
distortions (the so-called “dome effect”) due to inadequate lens
distortion calibration (Fonstad et al., 2013; Ouédraogo et al.,
2014); a drawback that has been resolved in traditional stereo
photogrammetry (Bertin et al., 2015; Wackrow, 2008).
Motivations and Aims
The use of non-proprietary close-range digital photogramme-
try is increasingly becoming more common for studies in flu-
vial hydraulics and the Earth sciences (Bertin and Friedrich,
2014; Bertin et al., 2014; Bouratsis et al., 2013; Musumeci
et al., 2013). We previously reported on our development of
a non-proprietary stereo-photogrammetric setup capable of
recording fluvial surfaces at the grain scale, characterized by
DEMs with 0· 25 mm sampling distance and sub-millimeter ac-
curacy (see Bertin et al. (2014 and 2015) for a detailed evalu-
ation of the technique). We also showed that the same setup
can be used for through-water recordings, with water depths
up to 200 mm (Bertin et al., 2013).
The present paper unveils an efficient and effective seam-
less DEM merging method, which supports our high-resolution
stereo-photogrammetric application, and allows us to increase
the surface coverage, and therefore the range of potential ap-
plications, without reducing DEM quality. Our proposed merg-
ing method is computationally efficient and can be adapted to
any DEM , independent on the measurement recording tech-
nique. The only prerequisite is the existence of DEM s’ overlap
and regular grid arrangement, both easily satisfied in practice.
Our DEM merging solution is applied to the study of
laboratory water-worked gravel-beds. Previous studies on the
fluvial microtopography and grain-scale roughness, using
either close-range digital photogrammetry or terrestrial laser
scanning (TLS), were limited to small surfaces (0·25 m2 and
less), in order to guarantee efficient recording and sub-milli-
meter resolution. The present paper shows that DEM merging
enables the recording and subsequent analysis of large-scale
micro-topographic information. This new process we present
herewith will now allow researchers to advance our funda-
mental fluvial knowledge by fusing grain-scale and bedform
roughness in both laboratory and field studies.
Measurement Environment and Instrumentation
A hydraulic flume (19 m long, 0.45 m wide, and 0.5 m deep)
is used for our gravel-bed studies. The setup is described
previously (Bertin and Friedrich, 2014; Bertin et al., 2014;
Heays et al. , 2014). A gravel bed (0.71 mm <d <35 mm, where
d is the intermediate axis of the sediment particles) was
prepared in a 950 mm long full-width test section, located 14
m from the flume inlet. The gravel bed was water-worked at
a constant flow rate (84 L/s) until the sediment surface was
in static equilibrium. During water-work, fine sediment at the
surface was transported downstream, uncovering coarser par-
ticles, which then formed the so-called “armor layer”. After
32 January 2016 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING