International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
interpolation is applied in a hierarchical framework. The main
feature of the Aierarchical robust interpolation is the creation
of a data pyramid representing the data at different resolution
levels. Robust interpolation is applied to thinned-out data first,
the interpolation results being used to eliminate off-terrain
points for the next iteration that is carried out using the data of
the next finer resolution of the data pyramid. Three steps are
carried out at each level of the data pyramid:
I. Thin out the original data according to the resolution of the
current level of the data pyramid, using only points not yet
classified as being off-terrain points
2. Generate a DTM by robust interpolation, using the thinned-
out data
3. Compare the DTM thus generated with the original data.
Data points outside a certain tolerance band are classified to
be off-terrain points and, thus, no longer considered in the
subsequent iterations.
At the finest level, the DTM is computed from all original
points classified as terrain points. Using this method, the
generation of DTM from ALS data in densely built-up areas has
been shown to be feasible (Briese, et al., 2002). Using thinned-
out data, the influence of large clusters of off-terrain points (e.g.
points on buildings) can be eliminated but the resulting DTM
also has a rather coarse resolution. The influence of low
vegetation (e.g. bushes) is eliminated using the data at a finer
resolution, a process that also results in a better DTM.
4. HIERARCHICAL ROBUST INTERPOLATION FOR
DTMS FROM IMAGE MATCHING
As mentioned above, hierarchical robust filtering was primarily
created for DTM generation from ALS data. However, in this
paper the original point cloud is a grid that was generated by an
interpolation with finite elements from the initial results of
feature based matching by MATCH-T.
4.1 Characteristics of the data
Figure 3 shows the different characteristics of point clouds from
image matching and ALS. The dots represent the grid points
derived by image matching with a low degree of smoothing in
grid interpolation. The red line shows the DSM that can be
generated from these grid points. The crosses represent the ALS
points. The green dotted line represents the DSM obtained from
ALS data. One essential difference between point clouds from
ALS and image matching is that image matching does not
deliver terrain points in wooded areas because corresponding
points arc only determined on the tops of the tree canopies. On
the other hand, ALS does provide a point cloud with a good
mixtere of terrain and off-terrain points, because the laser beam
can at least partly penetrate tree canopies. If there is no ‘good
mixture’ of terrain and off-terrain-points, robust filtering will
not be able to eliminate gross errors. As a consequence, off-
terrain points from a point cloud derived by image matching
cannot be expected to be eliminated in forests. The second big
difference between point clouds from ALS and point clouds
from image matching is that, unlike ALS data, point grids from
image matching are pre- filtered in the matching process. The
effect is that the outlines of buildings and other objects are
blurred, which in densely built-up areas might result in narrow
inner courtyards without points on the terrain. Consequently,
the areas without actual terrain points might be larger for point
clouds image matching than for ALS data. This has to be
416
considered when applying hierarchical robust filtering to point
clouds derived from image matching.
Different characteristics of point clouds from image
matching (dots) and ALS (crosses) and the resulting
DSMs; DTM - blue dotted line.
Figure 3.
4.2 Adaptation of the filter strategy
The strategy applied in this work is based on a strategy that has
been shown to give good results in DTM generation from ALS
data in low-density areas (figure 4).
Thin out the original data Create a coarse DTM
at a coarse level using RLP
Classify original data Create a coarse DIM
using a tolerance band + from terrain points
that is not too restrictive using LP
Thin out the original data | Create an intermediate
at an intermediate level DTM using RLP
Classify original data Create an intermediate
using a tolerance band |«—] DTM from terrain points
that is rather restrictive using LP
Create a DTM applying Create final DTM from
RLP to pre-classified —» terrain points at original
data at original resolution
Figure 4. Work flow for our filter strategy. LP: Linear
Prediction, RLP: Robust Linear Prediction.
The terrain type, the density of vegetation and development, and
the average building dimensions are the determining factors for
an adequate filtering strategy. It turned out to be necessary to
have three iterations of the loop of thinning out, filtering, and
eliminating points off the intermediate DTM, described in
section 3.2. In each loop, the parameters were set in a way to
take into account the peculiarities of the matched points.
4.2.1 Generation of a Coarse DTM by Rigorous Thinning
and Filtering: In this first step, a DTM is created from data
that are rigorously thinned out by selecting the lowest point
within a certain neighbourhood. The degree to which the data
are thinned out, controlled by selecting the grid width of the
thinned-out data, is of crucial importance to the success of the
whole procedure. It must not be chosen too small, because
otherwise objects such as buildings and groups of trees cannot
be eliminated. On the other hand, it must not be chosen too
large, because this would result in too high a degree of
Internatio
ES
smoothing
terrain. F
thinning
matching
thinning «
object we
areas with
consequer
densely bi
gaps betw
degree of
Having th
eliminate
this first
process: I
the DTM
cut-off cat
out to be
branch of
estimate c
implics th
this stage
function t
many terr:
to be 1.5
in the last
an optima
the first it
weights w
on the ter
smaller th
(with a gr
the influe
eliminatec
4.2.2 Inte
This seco
point cloi
iteration.
DTM are
for furthe:
width of t
to include
eliminatin
selected a
accepting
eliminate
image ma
Using a n
resulted ii
The DTM
rigorous s
off-terrain
trees are
eliminatec
The origi
again, us
iteration (
grid widtl
thinned oi
settings f
influence
low veget
estimatior