PE&RS January 2016 - page 38

Discussion and Conclusions
The results presented herewith indicate that the natural trade-
off between measurement resolution and surface coverage
encountered with digital photogrammetry and other remote
sensing techniques can be overcome by merging overlap-
ping DEMs. The presented merging approach can easily be
integrated in our stereo-photogrammetric DEM collection
workflow (Bertin et al., 2015). We present how future merging
must be considered during the design stage, by identifying the
measurement resolution needed for future grain-scale analy-
sis, and then calculating the number of DEM s needed to cover
the region of interest. During the design stage, the size of the
overlap between adjacent DEMs needs to be considered. It is
clear that the choice of overlap affects the merging poten-
tial, although only limited detailed past research on overlap
dimensions is available. Our preliminary tests suggested 30
percent to be a good compromise between data redundancy
and data handling, a value which was also reported in previ-
ous studies (e.g., Marzahn et al. (2012)). We provide informa-
tion on a best-fit approach to accurately undertake 3D co-reg-
istration of individual DEM s, using the least-squares method
and standard averaging of overlapping elevations, to ensure
seamless merging.
Our merging methodology can be adapted to DEMs collect-
ed using any measurement technique, as long as a DEM over-
lap is accounted for and measured elevations are arranged
on a regular grid (unlike non-gridded point cloud data). The
former was already identified as a natural pre-requisite for
effective DEM merging (see the State of the Art section), while
the latter is custom practice in the Earth sciences (Bouratsis et
al., 2013; Chandler et al., 2002; Chandler et al., 2001; Hodge
et al., 2009; Stojic et al., 1998; Tarolli, 2014). A feature of
interest, which is not investigated herewith, is the extent to
which the co-registration process is able to align DEMs col-
lected from very different viewpoints (e.g., when the setup is
rotated substantially between the acquisitions of successive
DEMs). For the present study it is not deemed applicable due
to the vertical nature of the used measurements.
Previous studies (Bertin et al. , 2014; Hodge et al., 2009)
interpolated raw point clouds onto grids with spacing closely
matching the original point data spacing, in order to reduce
smoothing artifacts. Here, we show that the computational
A B C D
Figure 10. Residual maps, i.e., elevation difference in mm between DEM-1 and DEM-2 after co-registration, for (A, B) a DEM sampling
distance of 2 mm; and (C, D) a DEM sampling distance of 4 mm; (A, C) The horizontal alignment precision equaled the DEM sampling
distance; and (B, D) the horizontal alignment precision was 1 mm to improve DEM co-registration. Below is the final DEM after merging.
The sampling distance is 4 mm, while the horizontal alignment was performed with DEMs resampled on 1 mm grids prior to merging.
38 January 2016 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING