hul 2004
2 compares
sing single
using
sphere can
n from
gorithms
from
ie by
ystems.
DTM.
Over not
ooth out
requency.
stem
eople
ood.
| between
ssential to
what the
effects.”
rocessing
nay never
wumber of
ring of a
lin (2003)
n LIDAR
lished but
curacy is
esence of
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004
layover and shadow, loss of coherence and the size of the
footprint.
5. BARE EARTH DATA (DTM) AND EDITING
The DSM is an accurate product derived directly from the
observations, however for many purposes a terrain model
(DTM), or bare earth model, is required. Much effort has been
expended to develop algorithms for this purpose, mainly
concentrating on LiDAR data. Sithole and Vosselman, (2003)
have reported on an ISPRS test of such filters. Less work has
been done on filtering IfSAR, where the scale is generally
smaller and the problems greater because of the footprint size
and the amount of penetration, or lack of it, of the microwaves
through vegetation.
Filtering algorithms generally incorporate a thresholding
function to decide whether a point lies on the terrain or on the
observed surface. The threshold may depend on elevation of a
point or group of points, or it may depend on slope between
adjacent points and these algorithms suffer from the problem of
assigning a value to the threshold. Figure 3 gives an example of
filtering from LiDAR carried out with the recursive terrain
fragmentation filter (RTF) developed at UCL, (Sohn and
Dowman, 2002).
E
b. LIDAR DSM.
LI
c. LIDAR DTM
Figure 3. Filtering of LIDAR data using the RTF filter.
It can be seen that although the major surface features have
been removed, the terrain is still not smooth. This is in part
due to small man made features, such as vehicles, and small
natural features, such as bushes, which fall below the assigned
threshold. Sithole and Vosselman, (2003) found problems with
complex objects, attached objects, vegetation on slopes and
discontinuities. Different filters cope differently with these
problems. Smoothing filters can be used, but they can introduce
their own errors.
With LiDAR, some of these problems can be overcome if multi
return systems are used. Figure 4 shows LiDAR returns over
forest area, taken with an Optech 2033, in which the ground
surface can be confidently predicted from the last pulse return.
As point density increases, this becomes more reliable.
Figure 4. Multiple returns from a forest canopy.
© www.infoterra-global.com
Less work has been done on testing filtering of IfSAR DSMs.
An evaluation of the Nextmap UK data was carried out at UCL.
This is discussed in detail in section 6. Figure 5 shows a
comparison of the Nextmap DSM, DTM and a GPS profile over
an unvegeatated flood plane to the left and a a forest to the right.
It can be seen that the forest has not been removed by the filter
used. Zhang et al (2004) have recently published an algorithm
developed specifically for IfSAR.
We can conclude that bare earth filtering still has problems and
that there will inevitably be a need for manual editing after the
automatic processing. Filtering of LiDAR is probably more
effective that that of IfSAR.
River Profile 3 - PT Il Data,
Comparison GPS, DSM, DTM
45 - ——
wn
og 40
2 35
£30 DTM
5 25 DSM
«e
2 GPS
Lu
Q 5 10152025 30 35 40 45 50 55 60 65
Distance along profile in meters
Profiles across a flood plane and forest from the
NextmapUK data.
Figure 5.
6. ACCURACY
The accuracy of both LiDAR and IfSAR is now quite well
established in empirical terms, but there are still error sources
which are not well understood or quantified, as discussed in
section 4.
Ahokas et al (2003) have carried out an analysis of fixed wing
and helicopter LiDAR from different altitudes, over different
surface material and also looked at the effect of observation
angle. They concluded that ‘The analysis of the factors affecting
