004
oint
lage
ting
has
ALS
near
and
s for
y to
and
d in
ly to
ning
data
joint
data
f the
f the
'ause
nnot
| too
e of
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
smoothing of relevant terrain structures, especially in undulated
terrain. For reasons described in section 4.1, the degree of
thinning has to be higher with point clouds from image
matching than for ALS data. We chose the grid width for
thinning out to be about half the linear extent of the largest
object we wanted to eliminate, c.g. 30 m in a data set containing
areas without terrain points with an extent of 60 x 60 m^. As a
consequence, the terrain cannot be modelled very accurately in
densely built-up areas for lack of terrain points within narrow
gaps between the individual buildings, and because a higher
degree of smoothing is required to eliminate the buildings.
Having thinned out the data, robust interpolation is applied to
eliminate off-terrain points. Selecting the filter parameters in
this first iteration is crucial for the success of the overall
process: larger objects not eliminated at this stage will remain in
the DTM until the end, whereas larger terrain features that are
cut-off cannot be regained in the subsequent iterations. It turned
out to be good practice to eliminate only points on the positive
branch of the weight function. For points below the initial
estimate of the surface, the weights remained unchanged. This
implics that outliers underneath the terrain are not eliminated at
this stage. Several iterations were carried out with a weight
function that was not too restrictive in order not to eliminate too
many terrain points (h — 0.4 m). The cut-off point / was chosen
to be 1.5 m. All points being more than 1.5 m above the DTM
in the last iteration were classified as off-terrain points. To get
an optimal estimate of the DTM representing the terrain after
the first iteration of filtering, linear prediction using the original
weights was carried out, considering only points classified to be
on the terrain. The grid width of this DTM was set to a value
smaller than the thinning parameter, e.g. 5 m in our examples
(with a grid width of the original data of 1.25 m). At this stage,
the influence of off-terrain points has not yet been completely
eliminated, and the terrain is still modelled very coarsely.
4.2.2 Intermediate Filtering to Improve the Coarse DTM:
This second iteration starts with a classification of the original
point cloud with respect to the DTM generated in the first
iteration. Points within a certain tolerance band around the
DTM are classified as potential terrain points and thus accepted
for further computation. All the other points are eliminated. The
width of the tolerance band has to be selected carefully in order
to include as many actual terrain points as possible, while still
eliminating a considerable portion of the off-terrain points. We
selected a band with a bias towards points below the terrain,
accepting points as far as 3 m below the initial DTM (to
eliminate large negative outliers delivered more frequently by
image matching methods than by ALS), but only 2 m above it.
Using a more restrictive upper threshold than 2 m would have
resulted in too great a number of terrain points to be eliminated.
The DTM from iteration 1 is too coarse to perform such a
rigorous step already at this stage of processing. Consequently,
off-terrain points at building outlines and on the tops of small
trees are still included in the data. These points are to be
eliminated in the second processing stage.
The original points classified as terrain points are thinned out
again, using a smaller thinning parameter than in the first
iteration (here: by selecting the lowest point within a regular
grid width of 3 m). Robust linear prediction is applied to the
thinned out data once again, but using more rigorous parameter
Seltings for the weight functions in order to eliminate the
influence of off-terrain points at the building outlinés and on
low vegetation (/ = 0.3 m). Unlike in the first iteration, robust
estimation was also applied to points below the terrain to take
417
into consideration the more frequent occurrence of ‘negative’
errors in image matching results compared to ALS. Again,
several iterations of robust estimation were carried out, using a
cut-off point / = 0.3 m. All points having a filter value between
-0.3 m and +0.3 m in the last iteration are considered to be
terrain points. These points are used to compute the second
approximation of the DTM by linear prediction, using the
original weights. This DTM, interpolated with a grid width of
2 m in most of our examples, is supposed to be already quite a
good approximation of the terrain, though it still contains few
off-terrain points on low vegetation.
4.2.3 Final DTM Generation: The DTM created in iteration 2
is again used to classify the original points, this time using a
more restrictive tolerance band (e.g. eliminating points more
than 2 m below or more than 1 m above the intermediate DTM),
because the approximation is a much better one than in the
previous iteration. Robust linear prediction using very
restrictive values (h = s — t = 0.15 m) is applied to remove the
remaining off-terrain points. In this final stage, robust
estimation is again only applied to points above the
intermediate DTM. This means that at this stage we assume that
all large ‘negative’ outliers have already been eliminated in the
previous filtering loop. The final DTM is created from the
points classified as terrain points in the final iteration of robust
estimation by linear prediction using the original weights. The
grid width of the final DTM has to be selected in accordance
with the resolution of the original point cloud.
S. RESULTS
In order to test our filter algorithm, three data sets of quite
different characteristics with respect to land cover and image
geometry were used.
5.1 The Test Data
The first data set, captured over Eggenburg (Lower Austria)
consisted of high-resolution aerial images of a historic town and
its surroundings and was characterised by undulating terrain
with both densely-built up areas in the town centre and forested
and agricultural regions with little but dense vegetation at the
fringes. The second data set consisted of high-resolution images
of a waste disposal site in Stockerau (Lower Austria), including
few buildings and man-made "terrain" shapes. The third data set
was captured over the Schneealm mountain range in Styria,
characterized by rugged terrain and partly by dense forest. For
all test sites, a DSM was derived using MATCH-T, selecting a
high degree of smoothing. Table | gives an overview of the
flight parameters and the parameter settings for MATCH-T. The
grid points derived by MATCH-T provided the input for our
filter algorithm.
1
Area Szl: f r terrain | Point A
[mm] | [um] | type | density | [m]
Eggenburg | 4500 152 30 U M 1.25
Stockerau | 3500 208 30 U D 1.25
Schneealm | 15000 | 214 30 M D- 10.0
0
Table |l. mage scales S, focal lengths f, and scanning
resolution r of the aerial images for the three test
sites. Terrain type (U: undulating, M: mountainous),
point density (M: medium, D: dense), and A (grid
width) are the respective parameters for MATCH-T.
