See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/255666692

Adaptive filtering of aerial laser scanning data

Article · May 2012

CITATIONS READS

17 180

2 authors:

Gianfranco Forlani Carla Nardinocchi

Università di Parma Sapienza University of Rome
120 PUBLICATIONS 1,985 CITATIONS 32 PUBLICATIONS 726 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Gianfranco Forlani on 19 November 2014.

The user has requested enhancement of the downloaded file.

 ISPRS Workshop on Laser Scanning 2007 and SilviLaser 2007, Espoo, September 12-14, 2007, Finland

ADAPTIVE FILTERING OF AERIAL LASER SCANNING DATA

Gianfranco Forlani a, Carla Nardinocchi b1

a
Dept. of Civil Engineering, Parma University, Italy
b
Dept. ITS, University Roma La Sapienza, Rome, Italy

Commission III, WG 3

KEY WORDS: LIDAR, Classification, Algorithms, DEM/DTM, Automation

ABSTRACT

Filtering non-terrain points from raw laser scanning data is the most important goal to improve productivity in DTM generation.
Filtering algorithms are built on assumptions about what discriminates terrain points from points on other objects (e.g. buildings and
vegetation). In most cases, a single measure is used to accept or reject points. In this paper a three-stage raw data classification
algorithm is presented. After a preliminary interpolation to a grid, a region growing based on height differences is applied. Segments
from the region growing are classified as terrain, building or vegetation, based on their geometric and topological description.
Terrain grid cells are conditionally low-pass filtered, to remove low vegetation. A piece-wise approximation of the terrain surface is
computed, built from the grid cells classified as terrain. Finally, raw data are accepted as terrain within a given distance from the
surface. Results obtained on a ISPRS filter test data set are shown to illustrate the effectiveness of the procedure.

1. INTRODUCTION equally well on any kind of landscape, because assumptions on

terrain characteristics or the threshold values used do not
Airborne laser scanning is today the most effective data always match reality.
acquisition technology for the production of high resolution, A first group of algorithms looks for the lowest point in a
high quality DTMs (Digital Terrain Models). The only neighborhood and label it as a terrain point. This is achieved
competing technique might be aerial photogrammetry with mostly by applying morphological filters (Kilian et al., 1996;
direct camera orientation by GPS/INS (Inertial Navigation Vosselman, 2000; Sithole, 2001) where the structuring element
Systems) and DTM generation by digital image correlation; is based e.g. on height difference or slope. Wack and Wimmer
with aerial digital cameras the automation of the workflow (2002) use grid data in a hierarchical scheme where non-object
should not be far from that of the laser scanner. Nevertheless, points are detected by using a Laplacian of Gaussian.
the preference for the laser scanner is clear and unlikely to be A second group fits an interpolating surface to the data and
reversed. Because of its characteristics (first and last pulse, accept individual points measuring their distance to the surface.
penetration rate in forested areas, narrow field angles, For instance, using linear prediction Kraus and Pfeifer (1998)
independence on shadows and object texture), laser scanning is iteratively get rid of points above the interpolating surface, so
indeed better suited and more versatile than photogrammetry that it gets closer and closer to the lowest data points;
for DTM production in urban areas as well as in forested areas. Axelsson’s (2000) algorithm works the other way around, i.e. a
Penetration of pulses under the canopy provides a key minimal set of (lowest) terrain points is progressively densified
advantage over photogrammetry, since it gives the filtering in a TIN structure by slope thresholding; Brovelli (2002)
algorithms a chance to succeed in getting rid of spots on analyses the residuals from spline interpolation to detect objects
vegetation while retaining terrain hits. contours.
Due to the scanning mechanism and aircraft movement, laser A third group aim first to segment the data based on one or
spots are scattered on terrain, vegetation, buildings and on more criteria and then try to classify them: Filin (2002) clusters
whatever target that, hit by a pulse, reflects back enough energy points in feature space, based on curvature and height
to be detected. The result is a point cloud that must be filtered difference, classifying low and high vegetation, smooth and
according to the survey purpose to get rid of unwanted echos: planar surfaces; the neighbourhood used in feature evaluation is
vegetation and buildings in DTM generation, vegetation in 3D adaptively adjusted to the slope (Filin and Pfeifer, 2006);
city models, both the terrain and the buildings in tree counting Roggero (2002) clusters points based on connectivity and a
and modeling. principal component analysis using geometric descriptors;
To reduce production costs and processing time, filtering is Nardinocchi et al. (2003) segment data in regions bordered by
performed automatically; in addition to visual inspection of the discontinuities, retrieve their geometric and topological
results for quality control, manual editing is still necessary, relationships and apply a rule-based scheme to classify the
depending on the reliability of the filtering algorithms and on segments.
the complexity of the site. Recently proposed algorithms stress the need of segmentation
and of context information to improve filter robustness: Sithole
2. PREVIOUS WORK and Vosselman (2005) aim to separate objects (natural or man-
made) from the terrain by extracting regions raised above their
Many filtering algorithms have been proposed in the last surroundings and classify them using geometric and topological
decade; witnessing the difficulty of the task, none performs relationships. Tóvári and Pfeifer (2005) group points in

1
Currently visiting professor at Delft University of Technology, Dept. of Earth Observation and Space Systems

130
 IAPRS Volume XXXVI, Part 3 / W52, 2007

segments based on consistency of normal vectors, distance to their relationships and provides the contextual information
the fitting plane and distance from seed point; robust filtering of essential to increases the probability of correct classification of
the surface is then applied, where the same weights are applied single data point in the final stage. This is not to claim that the
to group of points, rather than to single points. method is error free, but rather that a segment-based approach
In this paper a strategy for the classification and filtering of raw (as in feature-based matching) is more robust that just relying
laser scanning data is presented. The main building blocks of on point-to-point comparison in a local neighborhood (as in
the strategy (namely, data segmentation by region growing and signal-based matching). Effective filtering cannot be separated
region classification) have already been presented respectively by some sort of object recognition and identifying terrain
in (Nardinocchi and Forlani, 2001) and (Nardinocchi et al, patches or trees should not be seen as different from detecting
2003). In (Forlani et al, 2006) the capability of the method in buildings.
building detection was demonstrated, in the context of building The final stage relies completely on the correctness of the
reconstruction from laser data. In the following, Section 3 classification of the cells labeled as terrain, since the overall
presents the main features of the strategy; Section 4 reviews the approximation of the terrain is obtained only from cells
segmentation and the classification, pointing to the changes classified as terrain. Some classification errors can be tolerated:
now introduced to earlier versions and showing the small patches of low vegetation labeled as terrain are filtered
improvements. Section 5 presents the raw data filtering, that out; buildings labeled as terrain, on the contrary, will not.
was just sketched in the previous papers. Finally, Section 6 On the other hand, the further a cell is from the nearest terrain
reports on the results. Examples and results refer to site 5 of the region (or the less the terrain cells), the smaller the probability
ISPRS laser scanning test dataset (Sithole and Vosselman, that the approximating surface will truly follow the terrain and
2003). so actually will help to correctly discriminate the point class.
Data interpolation, segmentation and classification have to find
3. OUTLINE OF THE METHOD the best compromise between correct labeling of the terrain
regions and the attempt to extend them as much as possible, in
The classification strategy comprises three-stages (see Figure order to penetrate into the high vegetation areas and to reduce
1). In the first one, raw data are interpolated to a grid, taking the the number of small patches of terrain that, if completely
lowest elevation in the cell as grid value. surrounded by vegetation, would be much more difficult to
In the second stage, grid data are segmented by a region classify reliably.
growing algorithm with adaptive threshold. The geometric
characteristics and the topological relationships among the 4. DATA INTERPOLATION, SEGMENTATION AND
segments are reconstructed and, based on a set of rules, the CLASSIFICATION
segments are classified as outliers, vegetation, building or
terrain. Although each cell was assigned to a class, the raw data In the following paragraphs the three stages of the strategy are
it contains must still be classified individually. reviewed, highlighting the changes introduced with respect to
In the third and last stage of the procedure, the whole set of raw earlier versions and the improvements obtained. Attention is
data is examined. For the former, consistency is measured with also paid to using First and Last pulse and how positive and
respect to the elevations of the neighbouring terrain cells. For negative outliers are dealt with.
the latter, a piecewise approximation of the terrain with a The behaviour of the procedure is exemplified on the Site 5
continuous surface is estimated using data from cells classified dataset of the ISPRS Laserscanning test which offers a great
as terrain; consistency is measured thresholding the distance variety of environments, with step edges in the terrain, slopes
from the surface. with different orientation, high vegetation on a steep hillside
and a built up area with vegetation with a relatively low density
Raw Lidar Data
of raw data.

Re-sampling to a grid 4.1 Grid Data Interpolation

Adaptive region growing Slope-based Segmentation Grid cells are assigned the elevation of the lowest raw point in
segmentation Gradient Orientation the cell (see Figure 2). The larger the grid size, the more likely
Segmentation
this prevents the noise (such as cars, low trees and so on) to
Geometrical and Knowledge base affect the aggregation process. On the other hand, increasing it
topological relationships

Grid Data Classification

High Vegetation Building Terrain

(a)
Point-based filtering of Point-based filtering of low
high vegetation vegetation and noise

Surface Interpolation Cells classified as terrain

(b)

Figure 1. Components and main relationships of the framework for Figure 2. (a) Raw data Points; (b) Grid data Points
LIDAR data filtering. Solid lines refer to processing of grid
data; dashed lines to processing of raw data too much will affect the extraction of detailed information
(slope, aspect, …) from the grid, which may hamper the
Aggregation of raw data in segments enables a richer effectiveness of further steps. The best grid size should be
description of geometric properties and the establishment of between one or two times the raw data point spacing.
topologic relationships. This makes it possible reasoning about

131
 ISPRS Workshop on Laser Scanning 2007 and SilviLaser 2007, Espoo, September 12-14, 2007, Finland

Empty grid cells are treated as no data, unless all 8-neighbours For a given roof slope, the larger the cell size or the lower the
are non-empty: in this case, the cell value is set to the median of point density, the likelier was a fragmented segmentation.
the 8-neighbours. The same may happen with very steep terrain although, as
Figure 3 shows the TIN representation of the raw data (left) and already pointed out, in such cases the aggregation may come
of the grid data (right) in a smooth forest area. It is apparent from a smoother adjacent terrain area.
that, due to the high penetration rate, in such cases the grid Indeed, the region growing threshold should be coupled to the
representation already constitutes a good, although noisy, grid cell size and should also take into account evidence of
approximation of the terrain. surface continuity in the neighborhood. Several changes have
been made to the original implementation of the method to
address this problem. The region growing algorithm is now
steered by both the gradient orientation of the grid heights and
the slope. The seed pixels of the region growing algorithm are
chosen from regions larger than 30 m2 with homogeneous
gradient orientation while the threshold value is adaptively
adjusted to the slope of the region. Morevor, the process starts
from the regions with the lowest threshold value.
Figure 3. TIN of the raw data e TIN of the grid data on a wooded area. In large regions with homogeneous gradient orientation the
computation of the threshold will not be affected by vegetation.
4.2 Grid Data Segmentation

The region growing is the first step in data segmentation.

From a seed pixel, every of the 8-connected neighbours with a
height difference from the central pixel less than a threshold is
enclosed in the region and becomes in turn a seed point for that
region. The process goes on, until no points are added (i.e. the
region border will feature a discontinuity larger than the
threshold).
Although the result may depend on the cell size and the
threshold, the region growing separates most of the high
vegetation and of the buildings from the terrain: buildings raise
above the terrain by well defined discontinuities (edges), larger
than the threshold; laser spots on high vegetation get spread
over many very small regions.
Unless some terrain patches are completely bordered by dense
vegetation or, in case of bare earth, by a slope so steep that the
threshold is exceeded, the whole terrain may end up all in a
single region. This is because, from the seed point, the Figure 5. Gradient Orientation of the heights at Site 5. The orientation
algorithm looks for a smooth path across all the 8-neighbours: space is divided in 8 partitions.
therefore, even if in some area the terrain is steeper than the The threshold value T for the segmentation based on height
threshold would allow, the region growing may include it by differences is computed, in each region obtained from the
“sneaking through” along a smoother path. gradient orientation segmentation, as:
In the original implementation neither the choice of the seed T = min(Tmax , max(Tmin , s Δ + 2 s 2σ PL
2
+ σ H2 ))
points nor the threshold for the region growing were tied to the
morphological features of the grid data. The threshold was set where: Tmax = 2Δ; Tmin = 0.5 m s = the 75th percentile of the
to 0.5 m (i.e. about two times the height error of the data), slope distribution; Δ = cell size in m; σPL and σH are
independently of cell size and terrain slope. The drawback was respectively the planimetric and height accuracy of laser data in
that, in steep roofs or in steep terrain, several narrow regions m. With this modification the primary segmentation of the grid
may be created, affecting the success rate of building and data becomes in fact a (bounded) slope based segmentation.
terrain identification. Figure 4 illustrates the problem that arose Figure 5 shows the gradient orientation (cell size = 2m). Large
with a fixed threshold on a very steep roof. areas with the same colour correspond to regions having the
same aspect (orientation intervals are 45° large). Data holes and
flat areas are rapresented in white. Figure 6 shows the color
coded threshold values T for the same dataset.
The seed points for the region growing based on height
difference are taken from segments of the gradient orientation
segmentation larger than 50 m2, from the lower threshold values
on. The 0.5 m fixed threshold value is applied to the remaining
regions: in this way, areas with vegetation, that exhibit different
gradient orientation, or very small patches with the same
(a) (b) (c) orientation, get separated in small regions.
Figure 4. Steep roof segmentation. (a) Gradient orientation image; (b) Figure 7 shows the most significant segments with different
region growing with fixed threshold: the roof is fragmented in several colors. The red spots are very small regions (less than 3 pixels)
regions; (c) region growing with slope adaptive threshold: the house is that will be labeled as outliers or vegetation if several small
included in one single region bordered by discontinuities.
regions are contiguous. Notice that segments from the region
growing may encompass several regions with different gradient
orientation or with different slopes.

132
 IAPRS Volume XXXVI, Part 3 / W52, 2007

4.3 Data Classification contrary, terrain pixels erroneously labeled as building might be
recovered in the last stage. Figure 8 shows the result of grid
Geometric characteristics of the regions and their topological
data classification.
relationships are computed and stored in a knowledge base. A
rule-based scheme is applied to classify the regions: the 4.4 Using First and Last Pulse
outcome of the process labels each region as vegetation,
Almost every laser scanner today provides first and last (F&L)
building, terrain, outlier or unclassified (the last item tipically
pulse returns; the pattern of their difference is of great help in
being 1÷3% of the area size). Actually, each class may have
identifying vegetation. This is very important to improve both
sub-classes (e.g. courtyard as part of terrain); among
data classification as well raw data filtering: the percentage of
unclassified regions, narrow regions are defined as those
grid points in a region where F&L pulse elevations differ is
slender in shape. Points on high rise chimneys, towers, power
used to help the identification of terrain; raw data filtering (in
line poles, etc may be classified as outliers or buildings,
terrain as well as non-terrain areas) can be robustified by this
depending on shape, point density and cell size. Currently, no
information (see Section 5).
rules discriminate bridges, that are therefore included in the
terrain.

Figure 8. Grid Data Classification. Yellow: terrain; white:

building; orange: narrow regions; grey: unclassified.
Colour
T (dm) <5 5-7 7-9 9-11 11-14 14-16 16-20
In the previous implementation of the strategy, cells with
Figure 6. Color plot of the threshold values for the region growing different height in the F&L pulses were classified as vegetation
before applying the region growing and were not passed to the
region growing. This led to more fragmentation of the terrain;
now a terrain region penetrates much further into areas with
high vegetation, because the (lowest) last pulse of the cell may
have an acceptable height difference to nearby terrain cells
(whether the pulse indeed hit the terrain or rather the
vegetation, is to be clarified, of course).
Grid data under high vegetation are more noisy than those on
bare Earth; together with F&L information, this can be used in
the final filtering of raw data.

4.5 Outliers

Outliers in laser data are either “negative” (i.e. points below the
surface, mostly due to multi-path) or “positive” (i.e. points
above the surface, such as hits on birds, power cables, etc). The
segmentation makes the classification insensitive to single cells
with positive or negative outliers in two ways: if the outlier is
Figure 7. The most significant regions of the grid data segmentation by the only point in the cell, it will be put in a 1-pixel region and
the adaptive threshold classified as outlier. If there are several points in the cell, some
outliers some not, the positive outliers will be recognized in the
The current set of rules has been drawn from simple models of
final filtering stage, because they are higher than the
characteristics and relationships between terrain, building and
neighbourhood, whatever the class the cell was assigned. With
vegetation. The complexity of the task means that robustness of
negative outliers, the pixel has been labeled as outlier from the
the rule set cannot be taken for granted and that more rules
grid classification; other points of the cell may be assigned to
might have to be invoked in new scenarios. Most
terrain or vegetation, depending on the distance from the
misclassification errors occur with trees labeled as buildings,
approximating surface.
buildings as terrain and terrain as buildings. The worst
Even in case several contiguous cells contain outliers, it is very
misclassification error is a building included in the terrain,
unlikely that they end grouped in a region, because this would
because it will not be corrected in the next stage; on the

133
 ISPRS Workshop on Laser Scanning 2007 and SilviLaser 2007, Espoo, September 12-14, 2007, Finland

happen only if they have similar gradient orientation or very

small height differences.

5. EXTRACTING TERRAIN POINTS FROM RAW

DATA

The output of the grid classification can be divided in two (a)

classes: terrain and non-terrain pixels (i.e. pixels classified as
building, vegetation, outliers and pixels in unclassified regions).
Each raw data in a grid cell is now examined to label it as
terrain or as non-terrain point, comparing its distance from a
reference surface with a threshold ts depending on terrain slope
and sensor error tolerance in horizontal and elevation. (b)
The reference surface is computed from the local neighborhood Figure 9. Cross-section of a forested area on an hill side; (a)
for the former class, from a global approximation of the terrain Reference data: terrain: pink; vegetation or buildings: light
for the latter. The reason for differentiating between the two blue. (b) Terrain surface approximation: spline input: red; spline
classes is to allow more flexibility and fine-tuning for the prediction: green.
terrain cells.
Some of or all the raw data points of a cell classified as terrain
may in fact be low vegetation or noise. To check against this
possibility, a reference value href is computed from the
neighborhood using a conditional averaging filter. Let mn be the
mean of the neighbouring terrain cells, hc the elevation of the
current cell and t = 0.5 m a threshold value for low vegetation:
(a)
if (hc > mn + t ) href = mn
else href = hc

Cells unclassified or classified as non-terrain may nevertheless

contain raw data points that are in fact terrain points.
Comparing the elevation of the raw data with the predicted (b)
elevation from a surface approximating the terrain, a decision
will be made on the point class. To this aim, the most reliable
information available (i.e. the raw data labeled as terrain points)
is used to compute the approximating surface. The acceptance
threshold is computed for each cell as a function of the slope of (c)
the surface. Figure 10. Cross-section of terrain with step edges; (a)
Currently, the approximating surface is computed using bilinear Reference data: terrain: pink; vegetation: light blue. (b) Terrain
splines with relatively short spacing (3÷4 times larger than the surface approximation: input: red; spline prediction: green.
cell size); this may change in the future, to cope in a better way (c) Filtering: accepted terrain points (pink), rejected points
with discontinuities (see below). (light blue).
Points in cells classified as building do not need filtering; a
consistency check of the classification is performed, though: no Figure 11 shows a cross-section in an area with buildings and
point in such regions should fall in the acceptance band. If vegetation (Sample54) with the same color coding as Figure 10.
terrain points were erroneously identified as building, they
might now be recognized as terrain, if close enough to the
interpolating surface.

6. RESULTS ON ISPRS SITE 5 (a)

Figure 9 shows the behaviour of the raw data filtering in the
forested hillside of Sample51 (ISPRS Site5). The cross-section
(a) shows the reference data: terrain (pink) and vegetation (light
blue). In (b) the red points are input to the spline, while the
predicted value of the terrain in all cells classified as vegetation (b)
or in unclassified regions is shown in green. The approximation
of the terrain is good and the ensuing raw data classification is
correct. On the other hand, if the terrain shows step edges, as in
the quarry in Sample53, the interpolation function tipically
undershoots at the bottom and overshoots at the top (see figure
(c )
10 (a)). This smoothing of sharp edges leads to rejection of true
Figure 11. Cross section of an area with buildings and
terrain points.
vegetation. (a) Reference data: terrain: pink; vegetation or
buildings: light blue. (b) Terrain surface approximation: spline
input: red; spline prediction: green; (c) Filtering: accepted
terrain points (pink), rejected points (light blue).

134
 IAPRS Volume XXXVI, Part 3 / W52, 2007

Brovelli M.A., Cannata M., Longoni U.M., 2002 Managing and

Table 1 shows the overall results for the Samples available at processing LIDAR data within GRASS. Proc: Open source GIS
Site 5; performance and correctess of the cell classification are - GRASS users conference. Trento, Italy, 11-13 Sept. 2002.
measured respectively by the percentage of true terrain points
with respect to the total number of terrain points in the Sample Filin, S. 2002, Surface Clustering from Airborne Laser
and by the percentage of misclassified points with respect to the Scanning Data. IAPRS, Vol. 34/3 A, Graz, Austria, 2002.
number of cells labelled as terrain. Filtering errors are given
according to the Laserscanning Test definitions. Filin S., Pfeifer N., 2006. Segmentation of airborne laser
scanning data using a slope adaptive neighborhood. ISPRS
Terrain grid Raw data filtering Journal of Photogrammetry and Remote Sensing, 60(2):71-80,
classification errors April 2006
TP: Terrain #TrueTP in #FalseTP in Type I Type II
Forlani G., Nardinocchi C., Scaioni M., Zingaretti P., 2006
Points; TG wrt TG wrt #TG
Complete classification of raw LIDAR data and 3D
TG: Grid cells #True TP in cells (%)
reconstruction of buildings, Pattern Analysis Applications, 8:
class. Terrain Sample (%)
357-374. DOI 10.1007/s1004-005-0018-2.
Sample 51 85.1 6.4 8.3 8.6
Sample 52 80.8 2.2 8.5 9.6 Kraus K., Pfeifer N., 1998. Determination of Terrain Models in
Sample 53 77.8 1.0 10.7 14.3 Wooded Areas with Airborne Laser Scanner Data, ISPRS
Sample 54 85.5 7.5 4.4 12.0 Journal of Photogrammetry & R. S., vol. 53, pp. 193-203.
Table 1. Correctness of grid classification and terrain filtering
for Site 5 Samples Kilian, J., Haala, N., Englich, M., 1996. Capture and evaluation
of airborne laser scanner data. IAPRS, Vol. 31/B3, pp. 383-388.
Correctness of the terrain grid classification is normally high,
taking into account that if a cell contains more than one Kobler A., N. Pfeifer, P. Ogrinc, L. Todorovski, K. Ostir, S.
TrueTP, the others were counted as errors. Classification errors Dzeroski, 2007. Repetitive interpolation: A robust algorithm for
on the grid are higher with high vegetation and buildings DTM generation from Aerial Laser Scanner Data in forested
(Sample 51 and Sample 54) but filtering improved the results terrain. Remote Sensing of Environment, 108 (2007), 1; pp. 9 -
by more than 10% in both cases. With rough terrain, both types 23.
of filtering performed less effectively, especially the spline
interpolation. Nardinocchi C., Forlani G., 2001. Detection and segmentation
As far as raw data classification is concerned, Type I errors are of building roofs from Lidar data. ISPRS Workshop 3D digital
good and better than most Test participants, Type II are among imaging and modelling applications of: Heritage, Industries,
the largest. Medicine & Commercial Land, Padova, 3-4 April 2001.

7. CONCLUSIONS AND PERSPECTIVES Nardinocchi, C., Forlani, G., Zingaretti, P. (2003).

Classification and Filtering of laser data, IAPRS, Vol. 34 Part.
A strategy for classification and filtering of raw LIDAR data 3/W13, 79-86.
has been presented. The core of the procedure, i.e. the
classification of data segments based on their geometric and Roggero, M., 2002. Object segmentation with region growing
topological relationships looks sound enough. On the ISPRS and principal component analysis. IAPRS, Vol. 34, Part 3A,
Laser Test Site 5, grid data classification led to the reliable Graz, Austria.
identification of a percentage of true terrain points varying
Sithole G., 2001. Filtering of laser altimetry data using a slope
around 80%. Based on that information, a good approximation
adaptive filter, IAPRS Vol. 34 3/W4, pp. 203–210.
of the terrain surface can be computed. Terrain raw data close
enough to the surface are also recognized as terrain, improving Sithole G., Vosselman, G. 2003. Comparision of filtering
the percentage of success by up to 10%. algorithms. IAPRS, Vol. 34 3/W13, pp.71-78.
Problems arise with step edges in the terrain, because of over-
and undershoot of the spline functions: alternative interpolation Sithole G., Vosselman, G. 2005 Filtering of airborne laser
techniques will be tested soon (in this respect, our last stage scanner data based on segmented point clouds. In: "Laser
needs a pre-filtered input as REIN (Kobler et al., 2007). scanning 2005". IAPRS, Vol. 33,3/W19, pp. 66-71.
New approaches to rule definition in the grid classification are
also being tried: an attempt is currently underway to automate Sutton C.D., Handbook of Statistics, vol. 24, cap. 11:
the search for patterns in the data, using classification trees Classification and regression trees, bagging, and boosting,
(Sutton, 2005) such as the AdaBoost algorithm. Elsevier, 2005

Acknowledgements Tovari D., Pfeifer N., 2005 Segmentation based robust

George Vosselman and George Sithole are kindly thanked for interpolation – a new approach to laser data filtering. In: "Laser
allowing the use of the ISPRS data as well as of the manually scanning 2005". IAPRS, Vol. 33,3/W19, pp. 79-84.
classified control samples. Many thanks go to Simone Atzori
for his invaluable support in using GIS software. Vosselman G., 2000. Slope based filtering of laser altimetry
data. IAPRS, Vol. 33, Part B3, pp. 935-942.
References
Axelsson P., 2000. DEM generation form Laser Scanner Data Wack, R., Wimmer, A. 2002. Digital terrain models from
using adaptive TIN Models. IAPRS, Vol. 33 B4/1 pp. 110-117. airborne laser scanner data - a grid based approach. IAPRS,
Vol. 34/3 B, Graz, Austria, 2002.

135

View publication stats